Society

F1. Planned obsolescence and consumer products
F2. Consumer society
F3. Surveys
F4. Statistics
F5. Media violence
F6. Suspicious email
F7. Suspicious web address
F8. Artificial intelligence

F1. Planned obsolescence and consumer products

Current or durable consumer products offered on the markets do not always have the expected quality. The main causes are the following :

"Construction" defect of the product
Non-optimal user manual
Non-optimal consumer service
Premature and planned obsolescence of the product
Sources

F1.1. "Construction" defect of the product

Three types of fault exist :

Design defect : the defect is often irremediable.
Correct but flanged design : the consumer is not clearly informed.
Example 1 : many oil-filled electric radiators on casters have a reduced weight reducing their heat accumulation power and consequently their action range.
Example 2 : a large number of motion detectors with PIR (Passive InfraRed) technology operate chaotically outside the robustness limits of the electronic card or the operating ranges of the detection algorithm, particularly when they are placed outdoors in the presence wind, rain or humidity.
Produce defect : the test campaigns before leaving the factory are not always sufficient and a large number of products on the market are defective on their first use.

In France nearly 10 % of electronic devices purchased on the Internet are returned due to non-compliance.
For domestic electrical devices, the faulty component is often the electronic block of the device (examples : freezer, steam iron, thermal resistance of towel dryers, lighting of bathroom cabinets with touch control, outdoor solar light garland, DVD player).

F1.2. Non-optimal user manual

The product user manuel is sometimes succinct, poorly structured, roughly translated into French or written only in English, which does not allow consumers to use the product easily and appropriately.

F1.3. Non-optimal consumer service

Access to consumer service is not always easy (examples : unclear access methods, premium rate telephone number, unspecified waiting time).
Furthermore, the processing of requests and complaints is sometimes disconcerting or even flawed, with some operators not always having sufficient training in the products and services in the catalog, in internal procedures and in IT tools.

F1.4. Premature and planned obsolescence of the product

Planned obsolescence is a market practice justified by manufacturers for economic (see Consumer society), technical and risk prevention reasons. The consumer suffers the consequences : he is frustrated when he notices that the effective lifespan of his product does not correspond to the lifespan he expects.
But the causes of obsolescence are not only linked to the producer. They may occur during sale, distribution or consumption. This is why we speak more widely of premature obsolescence.
Different types of premature obsolescence can shorten the life of consumer products [CLI][HIP] :

Product of flanged design (see paragraph no. 1 above) whose use precautions or use limits are not clearly displayed.
Product whose functioning is degraded after a certain use period (example : women's tights which generally do not last more than six uses, whereas historically nylon stockings were extremely robust and resistant) ;
Product whose operation is blocked after a certain use period (example : some printers equipped with a smart chip, forcing the ink cartridges to be replaced before they are empty) ;
Product of which one of the hardware components is weakened from the design stage (examples : computer monitor, battery of digital landline telephones when permanently charged) ;
Product whose defective parts are almost inaccessible (examples : bearings of front-loading washing machines, electric car batteries) or cannot be removed (example : some smartphones and laptops whose battery or screen is fixed with industrial glue or screws not standard) ;
Product whose spare parts or accessories are non-existent in the manufacturers' catalog (example : LED and transformer integrated in the luminaires) or taken out of the catalog (example : old coffee maker bowl) or incompatible with the product (examples : new battery, new consumable) ;
Product whose maintenance or repair is expensive compared to the residual value of the product or the purchase price of a new product (economic obsolescence) ;
Product whose software support is discontinued and does not allow updating or downloading applications (software obsolescence) ;
Operational product but become outdated in the eyes of consumers (cultural or aesthetic or psychological obsolescence).

France is the first country in the world to have criminalized planned obsolescence through article L. 441-2 of the Consumer Code dated March 14, 2016.
Law n 2021-1485 of November 15, 2021 relating to the reduction of the digital environmental footprint in France (REEN law) then modified this article by redefining the offense as follows : "The practice of planned obsolescence is prohibited which is defined by the use of techniques, including software, by which the person responsible for placing a product on the market aims to deliberately reduce its lifespan".
Warning : there may be "obsolescence" without it being planned. It is up to the consumer to provide proof of the voluntary intention of the manufacturer to reduce the life of the product : difficult or even impossible task.

One of the solutions aimed at combating planned obsolescence is the services rental. Instead of owning a good, we would pay for the service rendered. Companies would then benefit from manufacturing durable and easily repairable goods (example : Xerox in the professional photocopier market) [WIK].

F1.5. Sources relating to planned obsolescence :

[HIP] HIPPOCAMPE, Usages numériques et terminaux.
[CLI] LE CLIMATOSCOPE, L'obsolescence prématurée de nos produits de consommation : un débat à remodeler.
[WIK] WIKIPEDIA, Obsolescence programmée.

F2. Consumer society

The consumer society is a type of society within which consumers are encouraged to consume goods and services on a regular and abundant basis [JDN].

Introduction
Results
Conclusions
Sources

F2.1. Introduction :

The consumer society appeared in Western countries from the second half of the 20th century, at the same time as some trends such as the women emancipation, industrial and technological innovation, globalization of trade.
The economy will then rely on the production capacities of companies as well as on the consumption capacities of citizens. It will increase the living standard of a large number of citizens in terms of comfort and material well-being, by increasing access to goods and services in volume and diversity.
But gradually a new paradigm is emerging : "consume more, produce more, earn more". Producers and consumers then mechanically become hyper-producers and hyper-consumers. The first ones offer an abundant supply based on products with a limited lifespan (Premature obsolescence). The seconds have demands that go well beyond their primary needs.

F2.2. Results :

This situation, where producers and consumers are closely linked around the industry of "desire", has resulted in the following consequences on the individual and their environment :

1. On an information level, the media (press, edition, poster, radio, television, cinema, Internet network, telecommunications) broadcast to the consumer (reader, listener, spectator, Internet user) a mass of more or less accurate information, amplified by social networks and content sharing platforms.
Broadcast can take different forms :

Overload information : massive dissemination of disparate information which imposes itself as a deafening source of noise.
Obfuscation (or jamming) : massive dissemination of false, imprecise or flattering information, in order to drown relevant, embarrassing or confidential information in a large volumevolume of data. Examples : creating a diversion with minor events to pass unnoticed information that is embarrassing to a government ; introducing flattering information into forums to correct a bad reputation [TOU].
Disinformation : dissemination of false or biased information (quote truncated or taken out of context, fake news, propaganda ("intox"), hoax, rumor, buzz, etc.).
Misinformation : involuntary or accidental dissemination of false information.
Harmful information : dissemination of exact information for harmful purposes (disclosure of secrets, publication of confidential data, defamation, etc.).
Non-information : concealment of important information, which may be total (concealment subject to censorship) or almost total (dissemination of vague or incomplete information in such a way that it cannot be said that it is completely concealed).
Scoop : dissemination of sensational and immediate information showing current events from the focused and biased angle of sensationalism.
Spam : dissemination of unwanted information for commercial, advertising or malicious (fraud, malware) purposes.

2. On an audiovisual communication level, the media, whose concern is to preserve a market more than to ensure the civic education of the masses, become a source of sensory pollution and overstimulation of the senses to the detriment of useful information [BEN]. See Media violence.

3. On an ecological level, overconsumption encourages waste and the production of waste which pollutes, even if progress is made to recycle it as cleanly as possible.

4. On a nutritional level, overconsumption promotes the appearance of pathologies such as diabetes and obesity.

5. On a professional level, the perpetual creation of new needs can lead to over-indebtedness, overwork or insecurity.

6. On a psychological level, the search for "always more" can lead to continual frustration which generates discomfort, depressive and even aggressive behavior.

7. On a social level, the human being has also become a "product" who must "know how to sell himself", who must enter into "competition", "at war" with everyone and others.

8. On a moral level, consumer objects become ends in themselves leading to a loss of "true purpose".

9. On the human level, social relations become simple means and are therefore artificial. We must then differentiate ourselves from others, in particular by impressing others with symbols of wealth and power to the detriment of the authenticity and depth of human relationships.

10. On a spiritual level, having replaces being, depth, introspection, reflection on identity based on something other than possessions [GUI].

F2.3. Conclusions :

Today the consumer society, which has become a society of overconsumption, is beginning to move towards a society of deconsumption along four axes : individual, community, systemic and environmental :

1. Individual sufficiency which consists of consuming less and differently, favoring quality rather than quantity.
To be applicable at the level of each consumer, this sobriety must be "happy", that is to say without the idea of restriction or sacrifice. Examples : second-hand purchase, recourse to repair, return to essential needs.
This trend would be driven in France by the elderly (with fewer needs), women (attracted towards ecology) and the most qualified (attracted towards a form of daily asceticism) [GOL][CAS].

2. Collaborative economy based on the sharing or exchange between individuals of goods (car, housing, parking, drill, etc.), services (passenger transport, DIY, etc.) or knowledge (computer courses, communities of learning, etc.), with or without monetary exchange, with or without digital networking platform.
By reducing or even eliminating intermediaries, this economic system allows you to moderate your expenses and save money.

3. Circular economy focused on a sustainable strategy which consists of sharing, reusing, repairing, renovating and recycling products and materials for as long as possible so that they retain their value.
In France, we can cite the following examples :

National harmonization of the color of waste bins subject to selective sorting, in order to guide and facilitate the sorting process. For household waste : green bin for glass, yellow bin for all packaging, gray or black bin for other waste. In companies, sorting is even more selective : blue bin for paper and cardboard, brown bin for bio-waste, other freely colored bins for company-specific waste.
End of disposable tableware in catering establishments serving on site more than 20 meals simultaneously. Meals are now served in washable and reusable tableware.
End of systematic printing of receipts. The customer can always request their ticket if they wish.
Availability of spare parts by manufacturers for a minimum period after purchase (smartphones, laptop computers, household appliances, motorized gardening and DIY tools, motorized personal transport vehicles, sporting and leisure items).
Display of a repairability index by sellers of electronic and electrical equipment (smartphones, laptops computers, televisions, porthole washing machines, top washing machines, dishwashers, vacuum cleaners, lawn mowers, high-pressure cleaners).
Market for packaging-free products, sold in bulk and in reusable or reusable containers.
Second-hand market (used or reconditioned products) making it possible to avoid using new resources to produce them (platforms Vinted, Le Bon Coin, Ebay, Emmaüs, etc.).
Rental of individual services. Instead of owning a good, the consumer pays for the service provided. Companies then manufacture durable and easily repairable goods (Xerox on the professional photocopier market, Michelin on the truck tire market, etc.).
Rental of shared services (laundries, etc.).
Rental of shared goods (cars (Autolib), bicycles (Vélib), books (libraries), etc.).

4. Waste management which consists of collecting it and then transforming it, prioritizing in order: reuse, recycling, elimination [MTE].

Reusing waste consists of using it again without any special treatment after control, cleaning or repair (worn clothes, car tires reused to protect the hull of boats or trawlers, etc.). Reuse can be simple reuse (returnable bottles refilled after cleaning, etc.).
Recycling waste consists of transforming it into raw materials (processing of paper and cardboard to make paper pulp, remelting of broken bottles to make new bottles, composting of bio-waste, etc.) or into energy (production of electricity or heat by incineration of waste, methanization of biowaste, etc.).
The elimination of "ultimate" waste consists of making it disappear by incineration or storage in a dedicated landfill (refusal of sorting after selective collection treatment by sorting centers, household waste, medical waste, paint cans, rubble, bulky items, end-of-life vehicles, etc.).

In France, in 2018, 66 % of treated waste is recycled, 7 % is incinerated with or without energy recovery and 27 % is landfilled [NOT].

F2.4. Sources relating to consumer society :

Sources relating to consumer society :
[ADE1] ADEME, Economie circulaire : notions, 2013.
[ADE2] ADEME, Potentiels d'expansion de la consommation collaborative pour réduire les impacts environnementaux, 2016.
[AGE] Agence Lucie, L'économie circulaire.
[ALT] Altermaker, Economie circulaire : définition et exemple.
[BEN] Abdellatif Bensfia, "François HEINDERYCKX (2003), La malinformation. Plaidoyer pour une refondation de l'information", Communication, Vol. 23/2 | 2005, 259-265..
[CAS] Jean-Laurent Cassely, Les diplômés, bons élèves ou cancres de l'alterconsommation ?, in Constructif, juin 2021 (N 59).
[CHI] Chilowé, Comment vivre la sobriété heureuse au quotidien ?.
[DES] Cécile Désaunay, Vers la déconsommation ?, in Constructif, juin 2021 (N 59).
[GOL] Mathilde Golla, La société de déconsommation commence à faire son chemin, Novethic.
[GUI] Valérie Guillard, La société de consommation, cours PSL Paris-Dauphine.
[IMP] impots.gouv.fr, Economie collaborative et plateformes numériques.
[JDN] JDN - Journal du Net, Consommation : définition simple.
[MAR] Margaux, Société d'hyperconsommation : comment en sommes-nous arrivés là ? Comment changer demain ?, Suricates Consulting.
[ODI] ODI - Observatoire de la Déontologie et de l'Information, L'information au coeur de la démocratie - Rapport Annuel 2017.
[RUM] Yannick Rumpala, Quelle place pour une "sobriété heureuse" ou un "hédonisme de la modération" dans un monde de consommateurs ?, in Dans L'Homme & la Société 2018/3 (n 208), pages 223 à 248.
[TOU] La Toupie, Toupictionnaire.
[WIK1] Wikipedia, Société de consommation.

Sources relating to waste management :
[ADE3] ADEME, Déchets chiffres-clés - Edition 2023.
[MTE] Ministère de la transition écologique et de la cohésion des territoires, Traitement des déchets.
[NOT] notre-environnement, Le traitement des déchets.
[REC] Recygo, Poubelles de tri sélectif, comprendre les codes couleurs.
[SEN1] Sénat, Les nouvelles techniques de recyclage et de valorisation des déchets ménagers et des déchets industriels banals.

F3. Surveys

A survey is, by definition, a statistical method of analyzing a population based on a sample of this population.
It are mainly used for political purposes (to help with decision-making), electoral purposes (to exist in the political landscape), commercial purposes (to anticipate sales of a product or find out customer satisfaction), activist purposes (to defend a cause) or media purposes (to build audience).
The quality of the information collected depends largely on the rigor surrounding the design and completion of the survey, as well as the interpretation of results.
The important parameters of a survey are as follows :

Sample size
Representativeness of the sample
Clarity of the questionnaire
Effectiveness of the questionnaire
Neutrality of the questionnaire
Conduct of the survey
Analysis of results
Adjustment of results
Publication of results
Reliability of surveys
Sources

F3.1. Sample size [GUM][SER] :

Sample size is a determining factor in obtaining reliable results.
The following formula (Cochran's formula) determines the number of people (n) to question based on the margin of error (m) that can be tolerated on a proportion of responses (p).
n = σ² x p (1 - p) / m²
p is the expected proportion of responses relative to the sample size n (when p is unknown, we use p = 0.5).
m is the error margin tolerated for the proportion p, the confidence interval equal to [p - m, p + m].
NC is the confidence level (or probability) that the answers are within the error margin (m).
z is the quantile of order α/2 of the reduced centered normal distribution for a given confidence level NC = 1 - α (z = 1.96 for NC = 95 %, z = 2.58 for NC = 99 %).
With a confidence level of 95 % and an error margin of 5 %, the calculation gives : n = 384.16
With a confidence level of 99 % and an error margin of 2 %, the calculation gives : n = 4144.14
For a sufficiently reliable market study, we most often accept a confidence level of 95 % and an error margin of 5 %, which requires interviewing 400 people.

Note that the error margin only takes into account random error, that is to say the possible differences between sample and reference population due solely to random sampling [TOU]. It does not take into account all other errors (representativeness of the sample, quality of the questionnaire, conduct of the survey, analysis of the results).

Proof of Cochran's formula [WIK1][WIK2][WIK3] :

We seek to estimate the real proportion (p) of a population likely to respond positively to a questionnaire.
To do this, we carry out a survey on a restricted population by submitting the questionnaire to a number (n) of individuals drawn at random.
The survey is assumed to be based on a simple random sample from a large population.
If we repeat the operation several times, the proportion of responses obtained then follows a binomial law with mean p and standard deviation σ = (p (1 - p) / n)^1/2
For n sufficiently large, this binomial law is very close to a normal law with mean p and standard deviation σ
To obtain the margin of error (m) on the proportion p, we simply multiply the standard deviation σ by the factor z of the reduced centered normal distribution for a given confidence level (NC), which gives :
m = z σ = z (p (1 - p) / n)^1/2
Hence the formula :
n = σ² x p (1 - p) / m²
Note that the margin of error is maximum for p = 0.5

F3.2. Representativeness of the sample :

The sample must be representative of the reference population if we want to extrapolate the survey results to the entire population.
- If it is a random survey carried out from a sample drawn at random from the entire population, it must be ensured that the sampling base is complete, up to date and without repetition, and containing all categories of society (origin , sex, age, profession, region, etc.) [GUM][LAL]. This is the method used for example for major surveys carried out by INSEE [TOU].
- If it is a selection survey with judicious choice (quota methods), it must be ensured that the quotas are substantially proportional to the fraction of the population represented by each category. This option is favored by polling institutes in France and applied to much smaller samples, generally 1000 or 2000 people [TOU].
- Internet surveys allow you to interview large numbers of people at low cost. However, email lists are generally not representative and 20 % of the population, not "connected to the Internet", is omitted [LAL].

F3.3. Clarity of the questionnaire :

- Indicate the total estimated time to answer the questionnaire, which should not be too long as it risks tiring the participant.
- Limit the questionnaire to 15 questions and each question to a single subject comprising 20 words maximum. For example, asking "How do you rate the quality of our product and customer service ?" leads to confusion which leads to ignoring the question.
- Avoid sensitive questions linked in particular to money, religion, intimacy, sexuality or family conflicts.
- Adapt the vocabulary to that of the population interviewed. For example, the question "France devotes around a quarter of national income to financing social protection. Do you consider this to be excessive, normal or insufficient ?" can be advantageously simplified by the question "Do you think that enough is spent on social protection ?".
- Give a definition of acronyms, technical words and little-used words.
- Do not use restrictive words such as "always", "never", "all" or "none", which prevent people from qualifying their comments and responding objectively.
- Do not ask imprecise questions, such as : "Have you been to the cinema recently ?" or "How much gasoline do you consume on average ?".
- Do not ask questions based on statements, such as : Do you agree with the statement "The police should not intervene" ?
- Do not ask questions containing superlatives or subjective adjectives, even emotionally charged or controversial, such as : "Was our customer service excellent ?" or "How amazing was your stay at our hotel ?".
- Do not ask questions containing implying words, such as : "Are you afraid of noise in your home ?" or "Are you excited to discover our new online store ?".
- Do not ask questions containing negative terms, such as : "Do you prefer not to use electrically assisted bicycles during your bike rides ?".
- Never use a double negative such as : "In your opinion, is the use of glycophate not without danger ?".

F3.4. Effectiveness of the questionnaire :

- Prefer closed questions (single choice or multiple choices, to be selected from a list of options) to open questions (collection of opinions), the latter leading to a long and difficult statistical analysis of the responses.
- For each closed question, limit the number of options, without forgetting the "Not concerned" or "Without opinion" option.
- Avoid imprecise options, such as "usually" or "often", replace with a clear concept such as "every day" or "more than twice a week".
- Avoid the single binary choice "Yes"/"No" which leads people to respond in a less thoughtful way and favors the "Yes" answer to appear more pleasant (positivity bias or tendency to acquiesce [JAC][SOU]). Instead, offer a single choice between two or more response options.
- For single choices with more than 2 options, prefer the choice with 4 balanced options, such as : "Very satisfied", "Satisfied", "Dissatisfied", "Very dissatisfied".
- For single choices with more than 2 options consisting of ranges, ensure that they are mutually exclusive, such as : "2 years or less", "3 to 5", "6 or more".
- For single choices with more than 2 options consisting of grades (1 to 4 for example), explicitly mention that grade 1 corresponds to the lowest grade.
- For multiple choices, classify them in alphabetical order, without forgetting the "Other" option.
- Add contextual information to explain and guide the participant in their responses.

F3.5. Neutrality of the questionnaire :

You should only give credit to surveys whose questions are as neutral as possible. Some questions call for answers. For example [LAL] :
- IF we ask "Do you prefer to use the old version or this improved version of the website ?", this is a biased question since the question implies by the word "improve" that the latest version of the site is best.
- If we ask "Do you think airlines should allow the use of cell phones in flight ?" the answers will not be the same as if we ask "Do you think airlines should ban cell phone use in flight ?"
- If we ask "Do you find it important that security rules at airports are tightened ?", most people will answer in the affirmative. However, if we tell them about the conditions that will accompany this greater security (longer waits, increased police presence, increased ticket costs, etc.), their opinion may be different.
- If we ask "Are you in favor or against sanctioning judges in the event of a fault leading to a judicial error ?", who would not want to sanction someone who has committed a fault ?

F3.6. Conduct of the survey :

When the survey is carried out in the "field" (face-to-face or telephone survey), it is necessary to ensure that :
- The investigators are serious, professional, well trained, neutral in appearance, welcoming and relational.
- The place, time and circumstances are propitious. For example, only launch the survey when the media context is relatively neutral in relation to the subject of the survey.
- The participant is available to answer the questionnaire.

F3.7. Analysis of results :

The understanding and interpretation of the results are sometimes very different depending on the analyst. For example [LAL] :
- If business travelers are asked "Do you often, occasionally, rarely or never use Internet access in your hotel room ?", the results will differ greatly depending on the groupings that we will provide answers. Does the "yes" result corresponding to the "often" case also include the "occasionally" case, or even the "rarely" case ?
- If the results show that "60 % of Canadians use their credit card to pay for their vacations", then the information should not be condensed into the form "Canadians travel on credit".

F3.8. Adjustment of results [DAN] :

Adjustment is a common practice in the world of surveys.
It consists of applying weightings to individuals to increase the weight of those belonging to categories under-represented in the sample surveyed compared to the reference population, and to simultaneously reduce the weight of those who are over-represented.
Adjustment on a single variable corresponds to a simple rule of 3. Adjustment on several variables at the same time is more complex and requires lengthy computer calculations.
But be careful, excessive weightings pose a certain risk to the quality of the results. Thus, it seems absurd to give 10 times more weight to the answers of the only person we managed to interview in a given category.
It is also possible that the calculations do not allow the sample questioned to converge towards the desired distribution.
Thus, if data is false from the start, if it was poorly collected, incorrectly entered, or if it contains too much missing data, there will be no point in trying to correct it.
Hence the extreme importance of ensuring that samples consistent with the reference population are obtained in advance.

F3.9. Publication of results :

Any published survey must be accompanied by certain information.
The French law of July 19, 1977 regulates surveys made public on the national territory and linked to the electoral debate. They must mention at least the purpose of the survey, the name of the organization, the name of the sponsor, the number of people questioned, the date, the exact wording of the questions asked and the proportion of "unanswered" [DOC].

F3.10. Reliability of surveys :

Concerning the reliability of the surveys, opinions are divided :
For some people, despite their faults, surveys are good indicators of trends and opinions [LAL]. They are not wrong when they are well done and observed in a trending manner.
For others, surveys reflect an opinion formed on the spot. Responding to a pollster doesn't commit you to anything. This can even be an opportunity to play, or even to express your anger at the moment.
The surveys also reflect an opinion drowned out in the media noise surrounding the question asked.
But most often, the survey is used to build audience. Its relevance is almost never mentioned.
Only in-depth surveys, repeated from year to year, are of interest in understanding changes in values over the long term [COS].

F3.11. Sources relating to surveys :

[COS] Centre d'observation de la société, Le sondage - un mauvais outil pour comprendre la société.
[DAN] Gérard Danaguezian, Le redressement d'échantillons, Survey Magazine.
[DOC] Doctrine, Loi n 77-808 du 19 juillet 1977 relative à la publication et à la diffusion de certains sondages d'opinion.
[GUM] Hervé Gumuchian et Claude Marois, Chapitre 6 - Les méthodes d'échantillonnage et la détermination de la taille de l'échantillon, in Initiation à la recherche en géographie, Presses de l'Université de Montréal.
[JAC] Marc Jacquemain, Méthodologie de l'enquête, Institut des Sciences Humaines et Sociales, Université de Liège.
[LAL] Michèle Laliberté, L'art des sondages ou comment éviter les pièges, Réseau Veille Tourisme.
[PER] Anne Perrut, Cours de probabilités et statistiques.
[SER] Zineb Serhier, Comment calculer la taille d'un échantillon pour une étude observationnelle ?.
[SOU] Dr. Ghomari Souhila, Techniques d'Enquête, Université de Tlemcen.
[TOU] Hugo Touzet, Connaître et mesurer l'opinion publique : utilité et limites des sondages, Ressources en Sciences Economiques et Sociales.
[WIK1] Wikipedia, Intervalle de confiance.
[WIK2] Wikipedia, Marge d'erreur.
[WIK3] Wikipedia, Loi normale.

F4. Statistics

When we have a data set, for example from numerous measurements, it is often useful to characterize it by parameters of central tendency (such as the arithmetic mean or the median) and dispersion (such as the standard deviation or the interquartile range).
It is also important to know the influence of aberrant data on these parameters.
Finally, the statistical method has a certain number of traps into which even experienced users can fall.
Consider a list of n data x_i, the index i ranging from 1 to n. The most common statistical parameters are then the following :

Aberrant data
Parameters of central tendency
Dispersion parameters
Misleading statistics
Sources

F4.1. Aberrant data [WIK5]

Aberrant data is data that contrasts greatly with other data, in an abnormally low or high manner.
Aberrant date is due to :
- either a measurement error in which case it is necessary to delete the aberrant data or use robust statistical indicators in the face of aberrant data,
- or to a highly asymmetrical data distribution in which case one must be very careful in using tools or reasoning designed for a normal distribution.

Different methods exist to identify aberrant date present in a dataset.
The simplest [KHA] consists of classifying the data in ascending order then identifying the first quartile (Q1), the third quartile (Q3) and the interquartile range EI = Q3 - Q1.
Any data x_i verifying one of the following relationships will then be considered aberrant :
x_i < Q1 - 1,5 EI    or    x_i > Q3 + 1,5 EI

The Figure above shows an example dataset with aberrant data (shown in bold font).
The table on the right shows the influence of these aberrant data on the calculation of the various common statistical parameters for four particular cases :
   case a : set of 5 data without any aberrant data
   case b : set of 6 data including those of case a supplemented by an abnormally low data
   case c : set of 6 data including those of case a supplemented by an abnormally high data
   case d : set of 7 data including those from case a supplemented by aberrant data from cases b and c
The analysis of this table is given below for each statistical parameter.

F4.2. Parameters of central tendency

The main parameters of central tendency are as follows, including the median which is robust to aberrant data.

Generalized mean M_p [WIK2] :

For n strictly positive data x_i, the generalized mean of non-zero order p (or H lder mean) is the quantity : M_p = ( (1/n) ∑_i[x_i^p] )^(1/p)
This mean is used in practice with :
   p → -∞ for the data minimum Min{x_i}
   p = -1 for the harmonic mean M_H
   p → 0 for the geometric mean M_G
   p = 1 for the arithmetic mean M_A
   p = 2 for the root mean square M_Q
   p → +∞ for the data maximum Max{x_i}
Between different means, we have the following relationship : Min{x_i} ≤ M_H ≤ M_G ≤ M_A ≤ M_Q ≤ Max{x_i}
The generalized mean does not depend on the order of the data.
The generalized mean is homogeneous : for any strictly positive constant k, we have the relation : M_p(k x_i) = k M_p(x_i)
The generalized mean is cumulative : if the data list is divided into several sublists, the mean of the overall list is the weighted mean of the means of the sublists, with the coefficients of each sublist being the number of terms concerned.
The generalized mean can be expressed in the form of a norm of order p :

Proof of the generalized mean for p → 0 [WIK3] :

For p → 0, M_p takes the indeterminate form 1^∞
We then rewrite M_p in the form : M_p = exp[X] with :
X = ln[ ( (1/n) ∑_i[x_i^p] )^(1/p) ] = (1/p) ln[ (1/n) ∑_i[x_i^p] ] = f(p)/g(p)
f(p) = ln[ (1/n) ∑_i[x_i^p] ]
g(p) = p
Having furthermore f(0) = ln[1] = 0 and g(0) = 0, we can apply the Hospital rule subject to the existence of f^'(p) and g^'(p) :
f^'(p) = ∑_i[ x_i^p ln[x_i] ] / ∑_i[x_i^p] using the relations ln^'(u) = u^'/u and (x^p)^' = x^p ln[x], and provided that x is strictly positive.
g^'(p) = 1
Hence (Hospital rule) :
lim_{(p → 0)}(X) = lim_{(p → 0)}( f(p)/g(p) ) = f^'(0)/g^'(0) = (∑_i[ 1 ln[x_i] ] / ∑_i[1]) / 1 = ∑_i[ ln[x_i] ] / n = ln[ Produit_i[x_i] ] / n
The exponential function being continuous everywhere and defined at 0, we can then write :
lim_{(p → 0)}(M_p) = lim_{(p → 0)}(exp[X]) = exp[ lim_{(p → 0)}(X) ] = exp[ ln[ Produit_i[x_i] ] / n ] = (exp[ ln[ (Produit_i[x_i] ] ])^1/n = (Produit_i[x_i])^1/n
which is indeed the expression of the geometric mean.

Proof of the generalized mean for p → ±∞ [WIK3] :

For large values of p such that p → +∞, we can write :
∑_i [x_i^p] = x_max^p ∑_i [ (x_i/x_max)^p ] = x_max^p
where x_max = Max{x_i}
From where :
lim_{(p → +∞)} ( (1/n) ∑_i[x_i^p] )^(1/p) = lim_{(p → +∞)} ( (1/n) x_max^p )^(1/p) = lim_{(p → +∞)} ( (1/n)^(1/p) x_max ) = x_max

For small values of p such that p → -∞, we can write :
∑_i [x_i^p] = x_min^p ∑_i [ (x_min/x_i)^-p ] = x_min^p
where x_min = Min{x_i}
From where :
lim_{(p → -∞)} ( (1/n) ∑_i[x_i^p] )^(1/p) = lim_{(p → -∞)} ( (1/n) x_min^p )^(1/p) = lim_{(p → -∞)} ( (1/n)^(1/p) x_min ) = x_min

Harmonic mean M_H [WIK1] :

For n strictly positive data x_i, the harmonic mean is the quantity : M_H = ( (1/n) ∑_i[x_i^-1] )^(-1) which can also be written : 1/M_H = (1/n) ∑_i[1/x_i]
The harmonic mean is to be used when we seek to mean a quantity which influences according to an inverse proportionality in a physical phenomenon (example : mean speed of a vehicle on journeys of the same length).
The harmonic mean minimizes the squared deviation defined by the sum ∑_i[ (1/x - 1/x_i)² ]
Example [BIB] : you take a bike ride. You start by climbing a hill of length L at speed v1 = 20 km/h, then you go back down this same hill at speed v2 = 30 km/h. What is your mean speed v ? Please note, this is not 25 km/h found by taking the arithmetic mean.
If t1 = L/v1 is the time taken to ascend and t2 = L/v2 the time to descend, then the total time t = 2 L/v is written : t = t1 + t2 = L/v1 + L/v2 , or again : 2/v = 1/v1 + 1/v2
The mean speed v is therefore the harmonic mean of the two speeds v1 and v2, i.e.: v = 24 km/h.

The M_H harmonic mean is highly sensitive to aberrant data when it is abnormally low relative to other data (see Table above).

Geometric mean M_G [WIK1] :

For n strictly positive data x_i, the geometric mean is the quantity : M_G = ( Produit_i[x_i] )^1/n which can also be written : ln[M_G] = (1/n) ( ∑_i[ ln[x_i] ] )
The geometric mean is to be used when we seek to have a balanced representation of the influence of low data and high data thanks to their logarithmic transformation.
The geometric mean minimizes the squared deviation defined by the sum ∑_i[ (ln[x] - ln[x_i])² ]
Example [BIB] : at the end of a demonstration, the police announce x₁ = 100 demonstrators and the organizers x₂ = 900. What is the real number x of demonstrators ? Please note, this is not 500 found by taking the arithmetic mean.
If we assume that the police and the organizers "cheat" in the same way, then the police announce (x/k) demonstrators and the organizers (x k), k being a multiplier coefficient. By taking the geometric mean, we then find the exact result : x = [(x/k) (x k)]1/2 = 300.

The M_H geometric mean is highly sensitive to aberrant data when they are abnormally low relative to other data (see Table above).

Arithmetic mean M_A (or "usual mean") [WIK1] :

For any n data x_i, the arithmetic mean (or "usual mean") is the quantity : M_A = (1/n) ∑_i[x_i]
The arithmetic mean minimizes the squared deviation defined by the sum ∑_i[ (x - x_i)² ]

The arithmetic mean M_A is sensitive to aberrant data (see Table above). It should only be calculated on a normal distribution or after identifying and removing aberrant data.

Root mean square M_Q [WIK1] :

For any n data x_i, the quadratic mean is the quantity : M_Q = ( (1/n) ∑_i[x_i²] )^(1/2) which can also be written : M_Q² = (1/n) ∑_i[x_i²]
The quadratic mean is to be used when we seek to mean a quantity which has a square influence in a physical phenomenon (example : mean speed of a particle involved in kinetic energy) or which includes values oscillating around zero (example : electrical signal).
The square mean minimizes the squared deviation defined by the sum ∑_i[ (x² - x_i²)² ]

The Root mean square M_q is highly sensitive to aberrant data (see Table above).

Weighted mean [WIK1] :

When the data are individually assigned coefficients m_i (called weights), positive and not all zero, the previous means each have a weighted version as follows :
- Weighted generalized mean : M_p = ( ∑_i[m_i x_i^p] / ∑_i[m_i] )^(1/p)
- Weighted harmonic mean : M_H = ∑_i[m_i] / ∑_i[m_i/x_i]
- Weighted geometric mean : M_G = ( Produit_i[x_i^m_i] )^1/∑_i[m_i])
- Weighted arithmetic mean (or barycenter) : M_A = ∑_i[m_i x_i] / ∑_i[m_i]
- Weighted quadratic mean : M_Q = ( ∑_i[m_i x_i²] / ∑_i[m_i] )^(1/2)
If the weights are of integer value, they define the number of repetitions of each data.
When all weights are equal, the weighted mean is the same as the unweighted mean.

Median Q2 [WIK4] :

For any n data x_i, the median (or second quartile) is the data Q2 located in the middle of the data x_i when classified in ascending order [WIK6] :
- If n is odd, then Q2 = x_{(1/2)(n + 1)}
- If n is even, then Q2 = (1/2)(x_n/2 + x_{(n/2 + 1)})
The median should be used when we seek to minimize, or even ignore, the influence of aberrant data.
The median minimizes the difference defined by the sum ∑_i[ |x - x_i| ]
Example : in digital image processing, the median filter reduces noise while preserving the contours of the objects in the image. The Figure above shows the example of an aberrant pixel (value 97) replaced by the median value (equal to 4) of its neighborhood made up of the eight surrounding pixels.

The Q2 median is robust to aberrant data (see Table above) and should be preferred over other means.

Weighted median [WIK9] :

When the data is individually assigned coefficients m_i (called weights), which are positive and not all zero, the median has a weighted version which is the data x_k which divides the mass of the weights in two when the data is classified in ascending order. We can also say that x_k is the data having a cumulative weight equal to half the sum of all the weights.
k is then solution of : ∑_{i = 1, k-1} [m_i] ≤ (1/2) ∑_i[m_i] and ∑_{i = k+1, n} [m_i] ≤ (1/2) ∑_i[m_i]
When two values of k satisfy the above condition (k inf and k sup), then the weighted median is : (1/2)(x_{k inf} + x_{k sup})
If the weights are of integer value, they define the number of repetitions of each data.
When all weights are equal, the weighted median is the same as the unweighted median.

F4.3. Dispersion parameters

The main dispersion parameters are as follows, including the interquartile range which is robust to aberrant data.

Scope e [WIK4] :

For any n data x_i, the scope (or amplitude) is the quantity : e = Max{x_i} - Min{x_i}.

The range e is highly sensitive to aberrant data (see Table above).

Mean deviation EM [WIK4] :

For any n data x_i, the mean deviation is the quantity : EM = (1/n) ∑_i[ |x_i - M_A| ]
where M_A is the arithmetic mean of the data x_i.
The mean deviation is to be used when trying to calculate the arithmetic mean of the absolute value of the deviations from the arithmetic mean.
The absolute value function not being differentiable is sometimes incompatible with certain analyses. To make the differences positive, we then resort to squaring and use the standard deviation.

The mean deviation EM is highly sensitive to aberrant data (see Table above).

Standard deviation σ [WIK4] :

For any n data x_i, the standard deviation is the quantity : σ = ( (1/n) ∑_i[ (x_i - M_A)² ] )^(1/2) = ( (1/n) ∑_i[ x_i²] - M_A² )^(1/2)
where M_A is the arithmetic mean of the data x_i.
The standard deviation is to be used when trying to calculate the square mean of the deviations from the arithmetic mean.
We have the relation : EM ≤ σ

The standard deviation σ is highly sensitive to aberrant data (see Table above). It should only be calculated on a normal distribution or after identifying and removing aberrant data.

Interquartile range EI [WIK4] :

For any n data x_i, the interquartile range is the quantity EI = Q3 - Q1
Q1, called lower quartile (or first quartile), is the data below which 25 % of the data x_i is found when classified in ascending order [WIK6] :
- If the rank (1/4)(n + 3) is integer, then Q1 = x_{(1/4)(n + 3)}
- If this rank ends with 0.25 then Q1 = (1/4)(3 x_inf + x_sup)
- If this rank ends with 0.50 then Q1 = (1/2)(x_inf + x_sup)
- If this rank ends with 0.75 then Q1 = (1/4)(x_inf + 3 x_sup)
Q3, called upper quartile (or third quartile), is the data below which 75 % of the data x_i is found when classified in ascending order [WIK6] :
- If the rank (1/4)(3 n + 1) is integer, then Q3 = x_{(1/4)(3 n + 1)}
- If this rank ends with 0.25 then Q3 = (1/4)(3 x_inf + x_sup)
- If this rank ends with 0.50 then Q3 = (1/2)(x_inf + x_sup)
- If this rank ends with 0.75 then Q3 = (1/4)(x_inf + 3 x_sup)
with x_inf = x_{lower integer rank} and x_sup = x_{upper integer rank}

Interquartile range EI is robust to aberrant data (see Table above) and should be preferred over scope e, mean deviation EM and standard deviation σ.

F4.4. Misleading statistics

Statistics, like any other technique, is not always handled with care, discernment and good faith.
It is the subject of traps, misleading evidence and even scams.

Misleading scale :

Some graphs are presented on a linear vertical scale not starting at zero, or logarithmic, or missing altogether. In all three cases, if we do not pay careful attention to the scale, the interpretation of the variable represented can be erroneous.

The right definition :

Statistics calculated on a variable only have meaning relative to its definition.
Example [INS][WIK8][MON][JAI] : in March 2017, the number of unemployed in mainland France was 2.7 million according to INSEE and 3.7 million according to the Pôle Emploi. Where does this gap come from ?
For INSEE, an unemployed person is officially an "unemployed active" person (i.e. aged 15 or over, unemployed, actively looking for work and available within two weeks), which excludes so-called "inactive" people ( young people under 15 ; student ; retired ; people in training ; discouraged people ; women with young children, not available quickly ; people unable to work ; etc.
For Pôle Emploi, an unemployed person is a person registered with an agency, unemployed and actively looking for work (category A).
20 % of the unemployed identified by INSEE are not registered with Pôle Emploi (people at the end of their rights, people removed for administrative reasons, etc.). Conversely, 40 % of unemployed people registered with Pôle emploi are not recorded by INSEE, often to benefit from certain measures.
The unemployment rate is then defined as the ratio between the number of unemployed and the number of active people, the number of active people being the sum of the number of employed people and the number of unemployed people.
In 2022, according to the INSEE survey, employment is distributed as follows : 68.7 % employed active, 5.3 % unemployed active (unemployed) and 26.0 % inactive [UNE], the unemployment rate being 7.2 %

Cumulative percentages :

When we combine increases and decreases in percentages, a multiplicative perception of the percentages is necessary.
Example [DEL2] : a price increases by 15 % then decreases by 6 %. What is the percentage price change ? The answer is not +9 %
You must in fact multiply the price by 1.15 then by 0.94, which gives 1.081 and corresponds to an increase of 8.1 %.
Note that the reversed operation "decrease by 6 % then increase by 15 %" gives the same result, the multiplication being commutative.

Percentage and absolute value :

A quantity can decrease in percentage each year at the same time as it increases in absolute value.
Example [DEL2] : A member of the government assures that the increase in deficit which was 15 % last year has been reduced to 14 % this year. The opposition claims on the contrary that the deficit which was 15 billion euros last year has increased again this year by more than a billion euros. Who is right ?
The 15 billion euros of last year's deficit corresponds to 15 % of the initial deficit (from two years ago). This was therefore 100 billion euros. Last year, the deficit increased from 100 billion to 115 billion. If, as the first statement indicates, the increase in deficit was 14 %, this year the increase therefore reached 14 % of 115 billion, or 16.1 billion. This is consistent with the second statement that the deficit increased by more than a billion. The two statements are perfectly compatible.

The aberrant mean :

The arithmetic mean has little meaning when the data distribution is unbalanced.
Example [DAN] : in a company of 200 people, 80 % of the workforce receives 1,500 euros per month and the remaining 20 % receives 10,000 euros per month. The average remuneration then amounts to 3,200 euros per month, which is an unrepresentative result when we take a simple arithmetic mean.

The hidden common cause :

Confusion is frequently made between correlation of factors (often linked by a common cause) and causality of facts (meaning cause and effect relationship).
Example 1 [ANI] : a person who has diabetes will have high blood sugar levels and feel excessive hunger. The two factors are linked and come from the same cause, which is diabetes. But high sugar levels do not cause feelings of excessive hunger, and vice versa. These are related facts but not by causality.
Example 2 [TER] : people who wear shoes larger than 45 commit three times more murders than those who wear shoes between 40 and 42. Does this mean that large shoes induce murderous behavior ? The real correlation is actually linked to gender : it turns out that homicidal behavior is found primarily in male individuals.

Law of small numbers :

The law of small numbers pushes individuals to believe that a small number of observations can accurately reflect the general population.
Example 1 [ANI] : if we count 70 % women at a meeting and we use this proportion of 70 % to guess the number of women in the world, this generalization is false because it is not representative of reality nor the randomness of the sample.
Example 2 [ANI] : it is not because we were cured using a grandmother's remedy that it really works for the entire population. This is why in medicine many tests are carried out to find out whether or not a treatment has a beneficial effect.

Paradox of the average number of children :

"Taking a family at random" and "taking a child at random" are not the same thing.
Example [DEL2] : An exhaustive survey carried out in a city indicates that families with children under 18 are distributed as follows : 10 % families with 1 child, 50 % with 2 children, 30 % with 3 children , 10 % to 4 children. The average number of children per family (among those with children) is therefore (10 + 100 + 90 + 40)/100 = 2.4.
To check this statistic, the administrative authorities carry out a survey. We survey 1,000 children under the age of 18 carefully chosen at random and ask them how many children there are in their family, including themselves. By averaging the answers, we get 2.67 ! Why is this gap so significant with the 2.4 in the statistic which took into account all families with children ?
The answer lies in the fact that by interviewing children at random, you will interview 4 times more children from families with 4 children than you will from families with 1 child, which skews the average. If there are 1000 families, there will be 100 only children, 1000 children belonging to a family of 2 children, 900 children belonging to a family of 3 children, 400 children belonging to a family of 4 children. In total, the answers given by these 2,400 children will lead to the result of 2,666... children per family.
The survey carried out does not evaluate the average number of children in a family taken at random, but the average number of children found in the family of a child taken at random.

Simpson paradox :

A phenomenon observed in several groups of data can be reversed when the groups are brought together. If we want to obtain sensible conclusions, the aggregation of results must respect certain rules of homogeneity [DEL2].
Example [WIK7][SCI][DEL1] : A patient has kidney stones. His doctor offers him two alternatives: treatment A and treatment B. To help him make his choice, the doctor informs him that a study was carried out on 700 patients. Half of them (i.e. 350) received treatment A for which there were 273 cures (i.e. 78 % of cases), and the others received treatment B for which there were 289 cures (i.e. 83 % of cases).
We also know that there are two types of calculations : small and large (see small table above).
- Treatment A is successful in 81 cases out of 87 for small stones (i.e. 93 % of cases) and in 192 cases out of 263 for large stones (i.e. 73 % of cases).
- Treatment B is successful in 234 cases out of 270 for small stones (i.e. 87 % of cases) and in 55 cases out of 80 for large stones (i.e. 69 % of cases).
In both cases (small or large stones), treatment A is more effective, while for the overall result, treatment B is more effective.
What creates the paradox, and the erroneous impression that B is more effective overall, is that treatment A was given much more often for large stones, which are more difficult to treat.
To occur, the paradox requires two conditions :
- existence of an often hidden variable (called a confounding factor) which significantly influences the final result. In this example, the size of the stones affects the likelihood of treatment success.
- heterogeneous distribution of the sample studied. Treatment A is in fact more often given to large stones and B to small ones.
In science, we carry out "randomized" experiments, which ensure a homogeneous distribution : for example, if you have kidney stones and you participate in an experiment to compare treatments, you are randomly assigned treatment A or B , without the size of the calculations influencing the decision. We thus erase the heterogeneity of distribution, and the paradox disappears : treatment A will be clearly seen as being the best [MAT].
When Simpson's paradox occurs, one of the solutions to eliminate it is to make the distribution homogeneous by modifying the numbers in each data group while retaining the percentages. In the "Small calculations" line, the ratio 81/87 can be replaced by 251.38 /270 giving the same percentage 93 %. In the "Large calculations" line, the ratio 55/80 can be replaced by 180.81 /263 giving the same percentage 69 %. The "Total" line then gives a ratio of 443.38 /533 = 83.17 % for treatment A and 414.81 /533 = 77.83 % for treatment B, which confirms that treatment A is the best.

Arithmetic proof [DEL1] :

If we denote A, B, C, D the four successive numbers of the "Total" line, a,b,c,d those of the "Small calculations" line and a', b', c', those of the "Large calculations" line, then we have the following relationships :
A = a + a' ; B = b + b' ; C = c + c' ; D = d + d'
A/B < C/D ; a/b > c/d ; a'/b' > c'/d'
The astonishment comes from the fact that we believe that the double inequality { a/b > c/d and a'/b' > c'/d' } leads to { A/B > C/D }, or even with only small letters: { (a + a')/(b + b') > (c + c')/(d + d') }
But in fact, the three inequalities can sometimes be true simultaneously on an arithmetic level.
When the data have same effect within each group (b = d and b' = d') or between groups (b = b' and d = d'), Simpson's paradox cannot occur, the double inequality { a/b > c/d and a'/b' > c'/d' } always resulting in { A/B > C/D }.

F4.5. Sources relating to statistics :

[ANI] Animafac, Les pièges de l'utilisation des chiffres.
[BIB] Bibm@th.net, Diverses moyennes.
[CAN1] Statistique Canada, Mesures de la tendance centrale.
[CAN2] Statistique Canada, Mesures de la dispersion.
[DAN] Gérard Danaguezian, Attention, statistiques !, Survey Magazine.
[DEL1] Jean-Paul Delahaye, L'embarrassant paradoxe de Simpson, Pour la Science, n 429 de juillet 2013.
[DEL2] Jean-Paul Delahaye, Déjouer les pièges des statistiques, Pour la Science, Hors-Série n 98 de février-mars 2018.
[KHA] Khan Academy, Identification des valeurs aberrantes avec la règle 1,5 x écart interquartile
[MAT] Johan Mathieu, Le paradoxe de Simpson.
[UNE] Unédic, Comprendre le halo du chômage.
[SCI] Science étonnante, Le paradoxe de Simpson.
[TER] Marc Tertre, Pourquoi faut-il se méfier des statistiques ?, Le Club de Mediapart.
[WIK1] Wikipedia, Moyenne.
[WIK2] Wikipedia, Moyenne d'ordre p.
[WIK3] Wikipedia, Discussion : Moyenne d'ordre p.
[WIK4] Wikipedia, Indicateur de dispersion.
[WIK5] Wikipedia, Donnée aberrante.
[WIK6] Wikipedia, Quartile.
[WIK7] Wikipedia, Paradoxe de Simpson.
[WIK8] Wikipedia, Chômage en France.
[WIK9] Wikipedia, Médiane pondérée.

F5.1. Definition :

In a society of excessive consumption, we are witnessing a commodification of the whole of society. Information becomes a commodity, subject to the laws of the market and competition, as do TV programs, cinema, film productions and the media in general.
To capture the audience in this highly competitive context, the media resort to different means such as sensationalism, controversy and sometimes even fake news.
Violence is also one of these means, giving rise to what is called media violence which is, by definition, human violence shown or suggested by the media, whether verbal, physical, sexual or psychological.

F5.2. Human violence :

The animal does not seek to kill or cause suffering. It kills to feed, reproduce or defend its territory. It is neither peaceful nor cruel and never acts excessively [GAN].
Even a cat playing with a mouse to the death is not trying to cause suffering. It is only training to hone its hunting skills.

The human being is also an animal, but an animal of excess. It is distinguished from animals by the anatomical and functional importance of its brain [GAN], in particular the neocortex devoted to thought, imagination and anticipation, which allows it to explore and exploit an increasingly extensive environment.
Since the beginning of humanity, around two million years ago, human beings have been inventive with an exploring, manipulative, conquering and aggressive side. This fundamentally creative character tears him away from nature and makes him enter into excess and transgression. He then becomes ingenious in matters of violence, torture and horror, reinforced by his primitive animality in the form of celebration, sacrifice, cruelty and warlike orgy [GAN].
Thus, the neocortex, the seat of our higher cognitive abilities, stimulates our tendency to explore and possess. Human violence results from this natural quest for expansion and control.

Today, modern man is no more violent than in the past. Wars have become increasingly deadly in the 20th century, but they are becoming rarer. Criminal violence has been on the decline since the 19th century. However, it appears worse than before because we approach criminal phenomena with different standards from those used in the past, being accustomed to increasingly broad security [GAN].
The reduction in violence is thus a massive and incontestable phenomenon. But there can be periods of relapse, setback, regression. Man is civilized but he does not change in depth. The impulses, the frustrations, the violent temptations remain [PIN].

Violence is a universal fact. Wisdom consists of accepting it rather than repressing it [GAN]. But wisdom also consists of discovering that all conflict is good as long as it reveals a truth, in particular that we are obliged to agree in order to live and that the conflict does not need to be violent for that [MIC].
Svâmiji said : "The animal is subject to nature. Man fights against nature. The wise man achieves unity with nature by accompanying it in its movement".

F5.3. Forms of media violence :

The following alphabetical list gives the main forms of media violence against human beings, ranging from simple annoyance to extreme violence.

Advertising : imposed broadcast of loud, spectacular and exuberant advertising, composed of small choppy and dynamic sequences on the zapping principle intended to capture attention [BEN][SEN2].
Antisocial violence : violence directed towards society in the form of revolt, contempt or indifference, which can lead to delinquency, deviance or marginality.
Audio and visual violence : sudden increase in sound to capture attention (during advertising) ; sudden camera movements or cutting of a sequence into very short shots meant to create rhythm (in action films) ; light flashes (in video games) ; fragmented screen, scroll or poor quality sound (on television) ; pop-up (on Internet) ; etc.
Bashing : systematic denigration of a person, a group or a country.
Buzz : rumor created to spread quickly and on a large scale (in particular via Internet) [TOU].
Continuous information : incessant and repeated stream of facts, images and words which does not inform but saturates the audience with a fog of images and words [RIM].
Controversy : article, discussion or debate which reflects, in a more or less fiery, even violent, passionate manner, opposing opinions on a given topic (political, scientific, literary, religious, etc.) [TOU].
Crual violence : violence that takes pleasure in voluntarily causing the physical or moral suffering of a person or animal (intimidation, humiliation, blackmail, mental or pathological cruelty, moral or sexual harassment, incitement to hatred, etc.).
Digital violence (or cyberviolence) : violence exercised through an online communication solution (Internet, cell phone, smartphone, etc.) in the form of sealioning, bashing, impersonation, outing, flaming, harassement, happy slapping, cyberstalking, deliberate exclusion, etc. [WIK2].
Discrimination : any distinction made between natural or legal persons on the basis of one of the following criteria (in French law) [MDT] :
- Identity : origin (natal, social, ethnic, national, etc.) ; age ; last name and first name ; family status (married, single, civil partnership, divorced, separated, widowed, etc.) ; physical appearance ; place of residence ; place of bank domiciliation ; ability to express oneself in a language other than French ;
- Sex : pregnancy status ; biological sex (female, male or intersex) ; gender identity (psycho-social sex : a-, cis-, trans-gender, non-binary, queer, etc.) ; sexual orientation (a-, bi-, hetero-, homo- or pan-sexual) ; the majority model is diadic (female or male), cisgender (gender conforming to biological sex) and heterosexual (attracted to the opposite sex) ; the minority model has the acronym LGBTQ+ (Lesbians, Gays, Bisexuals, Trans, Queer, etc.).
- Health : state of health ; disability ; autonomy loss ; economic vulnerability ; genetic characteristics ;
- Beliefs and way of life : belonging or not to an ethnic group, nation, so-called race or religion ; political views ; morals (life habits) ; union or mutualist activities ; normal exercise of the right to strike ;
- Alerts : status of whistleblower or person linked to the whistleblower (alert relating to a crime, a misdemeanor, a threat or harm against the general interest, a violation of national, European or international law) ;
- Harassment : acts suffered, refused or reported as moral harassment, sexual harassment or hazing.
Domestic violence : violence perpetrated by one spouse on the other within a couple, which may be verbal, physical, sexual or psychological [WIK4].
Excessive bureaucracy (or administrative "massive tome") : suffering and exhaustion of the user when confronting the cogs of a hypertrophied and implacable administrative machine (complexity, paperwork, administrative jargon, inability to deal with particular cases, rigidity, slowness, irrationality, indecision, etc.) leading to sometimes an inextricable imbroglio (hassle, explanatory letters, waiting messages, misunderstandings, bugs, blockages, disputes, etc.).
Extreme violence : ultra-physical violence present in particular on the Internet and in some video games.
Fake news : misleading information disseminated with the aim of manipulating or deceiving the public (particularly on the Internet and social networks).
Fragmented screen : broadcast of multiple images aimed at maximizing attention, the screen can be divided (generally in vertical strips), multiple (generally with a central frame surrounded by side frames) or mosaic (dynamic set of frames which split, unlink, overlap, etc.) [MAT].
Glorified violence : violence not only trivialized but also associated with positive values of normality, even self-fulfillment, particularly in fiction programs and video games [SEN3].
Gratuitous violence : violence in which there is no motive for the aggression, with total disinterest in the victim [SEN3].
Hidden violence : televised violence likely to cause anxiety without the cause being able to be identified by the viewer. It can come from framing, editing, sound accompaniment or sound effects (examples : soundtrack using a mixture of percussion, heart sounds, breathing rhythms and female moans ; juxtaposition of selected shots which each last a few seconds) [YAP].
Hoax : false news spread as a joke or with malice (false virus alert, false solidarity chain, promise of gifts, etc.).
Illegitimate police violence : unnecessary or exaggerated violence exercised by a police officer towards other people in the exercise of his duties, and therefore outside the legal framework [WIK8].
Immoral violence : violence that is contrary to morality or good customs defined by society (violation of dignity, vandalism, voyeurism, exhibitionism, fetishism, masochism, sadism, pedophilia, incest, rape, etc.).
Institutional violence : action or absence of action committed in or by an institution (administrative error, illegal act, power abuse, etc.) which causes the person unnecessary physical or psychological suffering or which hinders their further development [LOO ].
Intense Violence : realistic scene of physical conflict which may contain scenes of blood, weapons, carnage, bodily injury or death [WIK10].
Intellectual violence : verbal abuse by a parent or teacher, in the form of sarcasm, perfectionism, excessive authority or high expectations, which has the effect of undermining the child's self-esteem [GDC].
Malware : malicious software (spyware, adware, ransomware, virus, worm, logic bomb, Trojan horse, etc.) whose goal is to harm a computer system.
Media violence : violence shown or suggested by media.
Mind-numbing violence : violence which shamelessly displays blood and hatred, with the effect of destroying all critical sense among young people and preventing the formation of curious and vigilant minds [SEN3].
Music video : short promotional video film which illustrates a song or the work of an artist. Up to 75 % of music videos contain scenes of an explicitly sexual nature, and more than half contain scenes of violence, often directed against women [NIH].
Obscene violence : violence that openly offends modesty, particularly in the area of sexuality (pornography, reality TV, music video, social satire, etc.).
Perverse violence : crual or immoral violence.
Physical violence : direct action which harms the physical integrity, bodily health, life or freedom of a person (beatings, injuries, rape, torture, kidnapping, etc.).
Poor quality sound : unintelligible dialogues which may come from : 1. actors (mumbling, poor speech, too fast delivery, verismo), 2. sound recording (distant, noisy, distorted, cut voices, etc.), 3. automatic mixing not checked manually (dialogues drowned in music, mixed languages), 4. mixing for cinema, DVD or streaming platform not transposable for television (big differences in sound volume), 5. modern televisions with rear-facing speakers (crushed sound) [SIX][ART].
Pop-up : unwanted element (window, sound, music or words) interrupting or polluting Internet browsing, sometimes violently when the element appears as soon as the page is loaded, or completely obstructs the page, or is difficult to close or deactivate.
Pornography : direct, even brutal, representation of sexual and deliberately obscene scenes which has the effect, in children and adolescents, of causing emotional trauma, and in adults, of reducing sexuality to a mechanical act without relational and without emotional manifestations which are the basis of a couple's sexuality [MCC].
Propaganda (or "intox") : organized action with a view to spreading partisan, misleading or false information on a large scale.
Pseudo-information : information whose veracity is doubtful and often unfounded.
Psychological violence : repeated words, gestures or attitudes degrading a person's living conditions leading to an alteration of their physical or mental health (indifference, humiliation, guilt, isolation, intimidation, prohibition, control, harassment, threats, blackmail, etc.) [WIK5].
Reality TV : program claiming to show the reality of a psycho-social experience of a group of isolated guinea pig candidates, which has the effect, through the exhibitionism and voyeurism of the intimate [WIK7], to promote individualism with manipulation of others, sexism, humiliation by the group of the "scapegoat" individual and the exaltation of contradictory feelings [CSA].
Rumor (or grapevine) : pseudo-information transmitted by word of mouth with all the distortions introduced by each individual.
Scoop : broadcast of sensational and immediate information showing current events from the focused and biased angle of sensationalism.
Scroll : continuous information banner (news of the day and news alerts) scrolling in a loop on some news channels, added to the title banner of the current images. This banner is appreciated differently by viewers ("practical" for those who watch without sound and "polluting" for others).
Sensationalism : excitement and media hype around isolated events, dramatized in a spectacular mode no longer giving reality but the "dizziness" of reality [SEN3].
Sex : broadcast of programs with implicit sexual content or containing coarse language, sometimes with sexual violence, which can lead to an obsession with body image, precociousness in sexual relations and violence in romantic relationships among young people [GDQ].
Sexism : discriminatory attitude adopted against the opposite sex, mainly by men who assign themselves the best role in the couple and society, at the expense of women relegated to second place and exploited as objects of pleasure [CNR].
Sexual violence : physical or psychological violence which manifests itself in a sexual way (abuse, assault, exploitation, mutilation, cyberviolence, etc.) [FMV].
Social networks : websites allowing you to create social connections online with a community of friends (Facebook, Twitter, Tiktok, Instagram, Linkedln, Viadeo, etc.). Their algorithms are designed according to two biases (filters) : 1. Everyone only sees or reads what confirms their opinions, beliefs or preferences [HIL][FRI] ; 2. Content that generates the most emotion, even toxicity or insults is favored [LAP][FRI].
Moreover 88% of photos or videos published on social networks end up distributed elsewhere, notably on pornographic sites when the photo or video is "funny" or "sexy" [RFI].
Social satire : program which spreads preconceived ideas about families, regional inhabitants, etc., putting selected people in competition or conflict, while encouraging the viewer to make fun of them maliciously [WIK6].
Spam : unwanted message sent electronically or by telephone (email, SMS, MMS, social networks, voice message, etc.), for commercial, advertising or malicious (fraud, malware) purposes.
Spiritual violence : verbal, physical or psychological violence that criticizes or ridicules a person for their beliefs, forces them or prevents them from practicing a religion.
Suggested violence : violence that are not direct but suggested by an atmosphere, a context or even by the result of the act of violence.
Talk show : debate show whose frequent excesses are the following [PER] :
- the search for conflict and controversy ;
- the lack of equitable speaking time ;
- the over-representation of certain guest profiles ;
- the spectacularization to the detriment of the substance ;
- the lack of verification of the facts stated.
Talk show "brawl" : heated talk show (games of political guests, injunctions to react quickly, controversy, etc.) where everyone interrupts the floor by seeking to challenge the other, or even insult them [GAG][RIM].
Talk show "butchery" : unequal talk show where some guests (notably political guests) are deliberately destabilized thanks to the establishment of a mixture of genres which places them in the minority [LER].
Talk show "cacophony" : confusing talk show (debate between leaders, chatter of experts, etc.) where everyone speaks at the same time, stating their opinion, thus preventing discussion.
Talk show "indecency" : talk show where celebrities answer indiscreet or malicious questions from hosts and the public, exposing themselves to popular vindictiveness [WIK6].
Talk show "manipulation" : biased talk show where the host does not remain neutral and becomes involved in the content of the debate, or even monopolizes the floor.
Talk show "mockery" : talk show on trivial or indigent topics, where the public is led to take sides in an unreflective or mocking manner among debaters chosen to be living caricatures [WIK6].
Talk show "pathos" : caricature-type talk show guided by unhealthy voyeurism, which consists of displaying the private lives of people (in particular their misfortunes) or a current topic (in particular poverty or misery) for ratings purposes [WIK6].
Talk show "solicitation" : polemic-type talk show where guests are brought in to make a show.
Talk show "trash" : perverse or vulgar talk show close to exhibitionism or voyeurism. For example, the show where couples or families come to argue, settle scores and insult each other in front of the cameras under the applause or whistles of the public [WIK6].
Talk show "unpacking" : talk show of immodest, even exhibitionist testimony, where anonymous people come to expose intimate problems in the public place (sexual disorders, domestic violence, infidelity, psychiatric internment, etc.) [WIK6].
Trivialized violence : violence represented on screen as not having serious consequences, even harmless, or on the contrary as being the only way out of conflict resolution.
Verbal violence : violence expressed by words having the effect of despising, insulting, hurting or degrading the psychology of another person (insult, humiliation, threat, defamation, discrimination, etc.).
Verismo : deliberately scrambled dialogues aimed at copying daily social reality in which exchanges are sometimes degraded, confused, indistinct and vague [BEL].
Victimization : feeling of insecurity which is increasing, less due to the increase in petty crime (theft and damage), than to the systematic tendency of the media to show the world as violent and dangerous [SEN3][FLA]. "If the media gives us the impression of living in a more violent world than in the past, that is not what the statistics say which show a reduction in crime and armed conflicts. The time of Attila the Hun was much more violent than ours !" quotes André Mondoux, professor at Media School [BOU].
Video game : gaming software played on a game console or computer. 64 % of video games contain intentional violence and 60 % reward players for harming a character [NIH].
Violence against animals : animal abuse which can be active (beatings, injuries, torture, bullfighting, animal fighting, enclosure hunting, etc.) or passive (neglect).
Violation of dignity : enslavement or reduction of the human person to an animal or an object (slavery, pimping, discrimination, indecent work or housing, insult or public defamation, etc.) [PAG].
Violence : any conscious or unconscious aggression committed against oneself, another person, animal or group, which results in physical or psychological suffering, wounds, trauma or death.
Zapping : program intended to offer a summary of the most significant moments of programs already broadcast on various channels and favoring the spectacular, the futile, the event without any contextualization, perspective or explanation [WIK6].

F5.4. Controversy :

Many people do not believe in the effects of media violence. There are six reasons for this [COU] :

1. Low capacity for introspection : Very low capacity of human beings to know how they themselves function.
2. Self-indulgence bias : Tendency to form a very "self-advantaging" perception of one's own abilities and to conclude that the effects of violence only affect "others" and not oneself.
3. Collective desensitization due to the strong social presence of media violence.
4; Economic interest of cultural industries designing violent products to deny the effects of violence.
5. Multiplicity of causes other than violent images alone, that can cause aggressive thoughts and violent behavior (biological, psychological and social factors).
6. "Catharsis" hypothesis according to which exposure to violent images would reduce violent acts. This hypothesis, put forward in psychoanalysis and taken up by defenders of the video game industry, has been abandoned by researchers today.

F5.5. Results :

On human being, the effects of media violence are mainly those of violent images (bloodshed, beatings, torture, killings, bombings, road accidents, suicide, rape, etc.).
On this topic, more than 50 years of scientific research shows that prolonged exposure to violent images makes people more violent, increases the fear of becoming a victim themselves and reduces the feeling of sympathy towards victims of violence in real life [COU][WIK9].

On children and adolescents, the effects of media violence are as follows [CBS] :
- From 0 to 2 years : high vulnerability to sound effects (noises and screams).
- From 2 to 7 years : vulnerability to intense, trivialized, antisocial or gratuitous violence.
- From 7 to 12 years old : vulnerability to intense, trivialized, antisocial or intellectual violence.
- From 12 to 18 years old : vulnerability to antisocial or spiritual violence.
The most harmful programs for young people are violent films, including some cartoons, and realistic television news. The greatest impact is achieved by the association of violence with erotic images [KUC].

The representation of violence has even more negative impact when [SEN4] :
- It is represented as justified or rewarded ;
- The violent protagonists are attractive or heroic ;
- The consequences for the victims are minimized ;
- The spectator is young or predisposed to aggression.

Also note [KUC] :
- The number of programs for young people, especially of good quality, seems insufficient (less than 10%), which encourages young people to watch programs for adults that are totally unsuitable for their age.
- Some boys with an aggressive character identify more easily with combative heroes whose behavior they reproduce in reality as well as in the fiction they have watched (mimetic effect).
- On American television, 20 to 25 violent acts per hour are broadcast in children's programs.
- On Canadian television, sexual relations between unmarried partners are shown 24 times more than those between spouses [NIH].

F5.6. Conclusions :

The remedy for media violence consists of finding a better balance between producers of violent images and consumers. This mainly involves the following actors [COU] :

Public authorities, with more consumer protection.
NGOs and consumer associations, with more awareness campaigns and impact studies.
Media producers and cultural industries, with a certain interest in better positioning themselves as a "company respecting an ethical principle limiting the diffusion of violence".
Schools, with more education about images aimed at preparing children to deal with the stress they may experience when faced with any image.
Parents, with more protection of children in relation to violent images : respect signage (-10, -12, -16, -18 years), install parental controls, establish reasoned management of screen time, prohibit then explain, set an example and anticipate [CLE], encourage extracurricular sporting, artistic and creative activities in place of television use [KUC].

F5.7. Sources relating to media violence :

[ART] ARTE Radio, Prise de son : les 15 erreurs du débutant.
[BEN] Abdellatif Bensfia, "François HEINDERYCKX (2003), La malinformation. Plaidoyer pour une refondation de l'information", Communication, Vol. 23/2 | 2005, 259-265..
[BIL] Philippe Bilger, Pourquoi le son français est-il si mauvais ?.
[BOU] Marie-Claude Bourdon, Médias culture et violence.
[CBS] CBSC, Canadian Broadcast Standards Council, Classement des émissions en fonction de la violence - Guide de référence.
[CLE] CLEMI Le centre pour l'éducation aux médias et à l'information, Ecrans et violence.
[CNR] CNRTL, Centre National de Ressources Textuelles et Lexicales, Sexisme.
[COU] Courbet D. & Fourquet-Courbet M.P. (2014), "L'influence des images violentes sur les comportements et sur le sentiment d'insécurité chez les enfants et les adultes", Rapport Technique de Recherches, Université d'Aix-Marseille, Institut de Recherche en Sciences de l'Information et de la Communication IRSIC.
[CSA] CSA - Conseil supérieur de l'Audiovisuel, Réflexion sur les émissions dites "de téléréalité".
[FLA] Jean-Yves Flament, Téléréalité et idéologie.
[FMV] Fondation Marie-Vincent, Qu'est-ce que la violence sexuelle ?.
[FRI] France Inter, Algorithmes : les meilleurs amis des climatosceptiques.
[GAG] Christophe Gagne, "Un bon clash pour faire le buzz", Corela, 19-2 | 2021.
[GAN] René-François Gagnon, Cinq conceptions de la violence quotidienne, Mémoire 1999, Université Laval.
[GDC] Gouvernement du Canada, Ravaler sa douleur - Etude des liens entre l'anorexie, la boulimie et la violence contre les femmes et les filles.
[GDQ] Gouvernement du Québec, Le sexe dans les médias, Editions du Conseil du statut de la femme.
[GOU] Vincent Goulet, "Violence et médias", séances du Réseau Thématique n 37 "Médias" du 3e congrès de l'Association Française de Sociologie, Paris, 14-17 avril 2009.
[HIL] David Hiler, Réchauffement climatique : comprendre ceux qui n'y croient pas, Le Temps.
[KUC] Corinne Kucharscki, Jean-Luc Saladin, Daniel Godefroy et Matthieu Blondet, La télévision nourrit la violence, Reporterre - Le média de l'écologie.
[LER] Pierre Leroux et Philippe Riutort, Intégrer les politiques aux divertissements.
[LOO] Joséphine Loock, Violence institutionnelle de l'administration publique.
[MAT] Philippe Mathieu, Pour une histoire et une esthétique de l'écran fragmenté au cinéma, Thèse 2010, Université de Montréal.
[MCC] Ministère de la Culture et de la Communication, La violence à la télévision.
[MDT] Ministère du travail, La protection contre les discriminations.
[MIC] Yves Michaud, La violence apprivoisée.
[NIH] NIH, National Library of Medicine, Les répercussions de l'usage des médias sur les enfants et les adolescents.
[PAG] PagesJaunes/PagesConseil/Droit/, Atteinte à la dignité.
[PER] Perplexity, le moteur d'Intelligence Artificielle développé par Perplexity AI.
[PIN] Steve Pinker, La diminution de la violence dans le monde est un phénomène massif et incontestable, Libération.
[RFI] RFI, Les ados et le sexe sur internet : attention danger!.
[RIM] Pierre Rimbert, Interdire l'information en continu ?, Le Monde diplomatique.
[SEN2] Sénat, Enfants et publicité télévisuelle.
[SEN3] Sénat, L'audiovisuel à l'ère du numérique.
[SEN4] Sénat, Violence dans les médias : quelles conséquences sur les enfants et la société ?.
[SIX] Nicolas Six, Pourquoi les dialogues sont-ils parfois aussi inaudibles dans les films et les séries ?, Le Monde.
[TOU] La Toupie, Toupictionnaire.
[TRO] Pauline Trouillard, "Atteinte à la dignité humaine et autres contenus toxiques à la télévision française : le Conseil d'Etat a t-il ouvert la boîte de Pandore ?", La Revue des droits de l'homme, 24 | 2023.
[WIK2] Wikipedia, Cyberharcèlement.
[WIK3] Wikipedia, Perversion.
[WIK4] Wikipedia, Violence conjugale.
[WIK5] Wikipedia, Violence psychologique.
[WIK6] Wikipedia, Télé poubelle.
[WIK7] Wikipedia, Téléréalité.
[WIK8] Wikipedia, Violence policière.
[WIK9] Wikipedia, Education parentale.
[WIK10] Wikipedia, Entertainment Software Rating Board.
[YAP] yapaka.be, Les images violentes ne sont pas nécessairement celles que l'on croit.

F6. Suspicious email

Email attacks, also known as phishing attacks, are becoming more and more common and more sophisticated [CYM][CYS][KEP].

Structure of an email
Receiving a suspicious email
Structure of a suspicious email
What not to do
What to do
Checkpoints to identify a spoofed email
Example of a normal (non-spoofed) email header
Sources

F6.1. Structure of an email :

An electronic mail (or "email" or "mail") is a written message transmitted via a computer network, from one electronic mailbox to one or more others, allowing almost immediate communication between a sender and one or more recipients.
Email uses protocols like SMTP for sending and POP/IMAP for receiving.

A complete email consists of several parts :

A user header composed of :
- From: sender's email address
- To: recipients' email addresses
- Cc: recipients' email addresses in carbon copy
- Bcc: recipients' email addresses in blind carbon copy
- Date: date and time of sending
- Subject: subject of the message
- Reply-To: email addresses to send replies to (Reply-To:), these addresses often being different from that of the sender.
A technical header (or header detail) including additional information regarding message routing and security.
The body of the message which can include text, images and attachments.

An email address has a syntax defined mainly by the RFC 5322 standard as follows [CHA][PER] :

email address = localpart@domain
- Example of an email address : jane.smith@company.com
- The email address has a maximum length of 254 characters (including the "@" character).
localpart = sender's mailbox identifier in the case of sending and recipient's mailbox identifier in the case of receiving (example : jane.smith). This mailbox is managed by a mail server associated with the domain.
- The local part has a maximum length of 64 characters.
- The authorized characters in the local part are all letters (A to Z, lowercase and uppercase), all numbers (0 to 9) and some special characters.
- The authorized special characters are the following : !#$%&'*+-./=?^_`{|}~
- The hyphen "-" is not allowed at the beginning or end in the local part (invalid examples : -jane.smith, jane.smith-).
- The dot "." is not allowed at the beginning, end or in double succession in the local part (invalid examples : .jane.smith, jane.smith., jane..smith).
- Most non-ASCII characters (Unicode characters beyond U+007F) can be allowed in the local part of an internationalized address, according to EAI (Email Address Internationalization) standards and with UTF-8 encoding (examples : C3 A8 for U+00E8 = "è", C3 B1 for U+00F1 = "ñ", E2 99 82 for the male symbol U+2642 = "♂"). However, their use depends on the compatibility of the servers and software concerned.
- A sequence of characters can be surrounded by double quotes "..." to include special characters and other common characters such as space, comma or parenthesis (example : "user name"@company.com). and with Punycode encoding to be valid in the DNS system (example : xn--espaa-rta.com for the domain espa a.com). However, email addresses with quotes are rarely used in practice because they pose compatibility problems with some email software.
@ = separator character ("at" in English)
domain = mail server identifier (example : company.com)
- --------
- The email address domain has a maximum length of 189 characters (taking into account the maximum lengths of the email address and its local part).
- The email address domain can be entirely numeric but this is strongly discouraged so as not to confuse it with an IP address (invalid example : 123.456.789.0).
- --------
- The domain format is a sequence of at least two segments (labels) separated by a dot "." (example : sub.domain.com).
- The domain segments follow a hierarchical order from left to right : first the Sub-domains, then the Second-Level Domain (SLD) and finally the Top-Level Domain (TLD). For example, for the address mail.google.com : mail = specific service or application, meaning "email management", google = main domain name of the service provider (Google LLC), com = nature of the domain, meaning "commercial".
- The dot "." is not allowed at the beginning, end or in double succession in a domain (invalid examples : .company.com, company.com., company..com).
- Each segment has a maximum length of 63 characters.
- Each segment has as allowed characters all letters (A to Z, lowercase and uppercase), all numbers (0 to 9) and the hyphen "-".
- The hyphen "-" is not allowed at the beginning or end of a segment (invalid examples : -company.com, company-.com, company.-com, company.com-).
- Most non-ASCII characters can be allowed in the domain of an internationalized address, according to IDN (Internationalized Domain Names) standards and with Punycode encoding to be valid in DNS system (example : xn--espaa-rta.com for the domain españa.com). and with Punycode encoding to be valid in the DNS system (example : xn--espaa-rta.com for the domain espa a.com). However, their use depends on the compatibility of the DNS servers and the email or browser software concerned.

Please note that it is strongly recommended to use lowercase letters in email addresses.
The domain of the email address (after the "@") is not a problem because it is case-insensitive (lower or uppercase) according to DNS (Domain Name System) standards.
On the other hand, the local part (before the "@") remains case-sensitive, and the use of uppercase letters in this part, although technically allowed, can lead to several risks, mainly for the sender :
1. Input errors. A recipient may provide an incorrect version of their address by omitting the uppercase letters, or a sender may mistranscribe an address containing uppercase letters.
2. Delivery failures. Although rare, some servers still treat the local part of email addresses as case-sensitive, which can prevent emails from being received.

F6.2. Receiving a suspicious email :

When you receive a suspicious email, there is no major risk in simply opening the email.
Depending on the recipient's email client configuration, the sender may only collect the following technical data about the recipient :
1. Email address. Example : jane.smith@company.com
2. Device type. Examples : Desktop, Mobile, Tablet
3. IP address (IP = Internet Protocol) : unique identifier assigned to each device connected to the Internet. Example (official website of the University of Burgundy : www.u-bourgogne.fr) : 194.199.107.34
4. Approximate geolocation (estimated position of the device connected to the Internet using its IP address). Example (for www.u-bourgogne.fr) : France, Bourgogne-Franche-Comté, Côte-d'Or, Dijon, approximate GPS coordinates : 47.3220 N, 5.0415 E
5. Operating system (OS). Examples : Windows 10 64-bits, macOS 11.5.2, Ubuntu Linux 64-bits, Android 13
6. Web browser type and version. Examples : Chrome 118.0.5993.88, Firefox 118.0.2
7. Web browser rendering engine. Examples : HTML engine (Blink, Gecko, Webkit), 3D engine for video games and 3D software, Engine for 3D image synthesis
8. Preferred Web browser language. Examples : fr-FR, en-US
9. Type of email client. Examples : Gmail, Outlook 365, Apple Mail
10. Date and time the email was opened. Example : 2024-10-02 14:32:22 UTC+2
This data can be used later by the cybercriminal for more targeted attacks.

F6.3. Structure of a suspicious email :

The criteria for identifying a suspicious email are as follows [CHA][PER] :
- Sender address whose local part or domain contains one of the following typical anomalies :
   - a series of random letters or numbers (examples : contact-entreprise8765@gmail.com, contact-entreprise@gmail-5g78.com
   - an inversion between two characters (notably "l" and "capital i", "l" and "1", "O" and "0"). Examples : john.d0e@example.com, help@paypaI.com, support@paypa1.com, info@micros0ft.com
   - multiple hyphens (example : help@secure-paypal-login.com)
   - a single spelling mistake that escapes attention (example : john.does@example.com, help@amazom.com)
   - an unusual domain extension, notably certain free extensions such as .tk, .ml, .ga (example : john.doe@example.tk)
- Unexpected email. Example : invoice or unordered package.
- Inconsistent email compared to the sender's usual communications :
   - at the address level. Example : email claiming to come from a bank and using a Gmail address.
   - at the format level. Examples : logo, header, colors or different layout.
   - at the content level : lack of customization. Example : "Dear customer, we have detected suspicious activity on your account. Please click here to secure it. Kind regards. The security team."
- Content with image instead of text (which prevents the email from being detected by spam filters).
- Spelling or grammatical errors. Example : "Thank you for you're trust."
- Mention of a wonderful offer or promise of easy winnings. Example : "Congratulations. You have won a 200 euro voucher at Amazon ! Click here to claim it now."
- Request for personal or sensitive information. Example : "Please provide us with your credit card number for verification."
- Request to deactivate antivirus, firewall or any other security measure to open an attachment or visit a site.
- Insistent, urgent or threatening tone. Example : "URGENT : your account will be blocked in 24 hours if you do not respond immediately."
- Attachment with unusual format (.exe, .zip, .scr, etc.). Example : "Please open the attached urgent_bill.exe file to check your outstanding balance."
- Link pointing to a suspicious web address.

F6.4. What not to do :

- Trust completely based on the sender address alone. It can be spoofed (see Checkpoints).
- Click on a link or image
- View images that are not loaded
- Open or download an attachment
- Reply to or forward the email
- Provide personal or sensitive information (bank details, password, etc.)
- Contact the sender using information provided in the email (phone number, email, etc.)
- Ignore a warning from your email or antivirus software that the email is potentially fraudulent.

F6.5. What to do :

- Stay calm
- Take the time to check the legitimacy of the email (sender's address, format and content of the email)
- Check links without clicking on them (by hovering over them to see the real URL)
- Copy part of the email (such as the subject or key phrases) and do an Internet search to see if other people have reported this email as fraudulent
- Preserve evidence through screenshots
- Delete email
- Change the default font of the email client to clearly distinguish between an "l" and a "capital i": switch to Times New Roman, or Georgia or Verdana for example.

F6.6. Checkpoints to identify a spoofed email [CHA][PER] :

1. View the email header details, either by right-clicking on the email (without opening it and selecting "Show original" or "Show source code"), or by opening the email (and selecting the ad hoc menu).
   Examples of menus (by opening the email) : Orange (Details - Full header), Gmail (Plus, Afficher l'original), Free (Other actions - Show original), Windows Live Mail (Properties - Details).
2. Locate the following fields :
   - From : indicates the sender's email address.
   - Return-Path or envelope sender or MAIL FROM : indicates the email address to which undeliverable messages should be returned.
   - DKIM-Signature or DKIM-Filter or X-DKIM-Result (optional field) : indicates whether the message was authenticated using the DKIM (DomainKeys Identified Mail) method.
   - Received-SPF or SPF or X-Received-SPF or X-SPF (optional field) : indicates whether the message was authenticated using the SPF (Sender Policy Framework) method.
   - Received-DMARC or MARC-Filter or X-DMARC-Result (optional field) : indicates whether the message has been authenticated using the DMARC (Domain-based Message Authentication, Reporting & Conformance) method .
   - Authentication-Results or X-Authentication-Results (optional field) : gives the results of the different authentication methods performed by the different servers (intermediate servers and receiving server).
3. Compare the "Return-Path" field with the "From" field : Two cases are possible : 1. The email addresses must be identical, 2. The Return-Path field must have a domain (indicated after "@") corresponding to a known server (verify this by typing this domain into an internet search engine).
4. Compare the "DKIM-Signature" field with the "From" field : Two cases are possible : The domain indicated after "d=" in the "DKIM-Signature" field must either be identical to the email domain (indicated after "@" in the "From" field) or correspond to a known server (verify this by typing this domain into an internet search engine).
5. Examine the "Received-SPF" field : the "fail" status must not appear.
6. Examine the "Received-DMARC" field : the "fail" status must not appear.
7. Examine the "Authentication-Results" field : the statuses "dkim=fail", "spf=fail" and "dmarc=fail" must not appear.

F6.7. Example of a normal (non-spoofed) email header [PER] :

From: "John Doe" <john.doe@example.com>
To: "Jane Smith <jane.smith@company.com>
Subject: Team meeting weekly
Date: Fri, 4 Oct 2024 09:45:21 +0200
Message-ID: <1234567890@mail.example.com>
MIME-Version: 1.0
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Return-Path: <john.doe@example.com>
Received: from mail.example.com (mail.example.com [192.0.2.1])
   by inbound.company.com (Postfix) with ESMTPS id ABC123
   for <jane.smith@company.com>; Fri, 4 Oct 2024 09:45:23 +0200 (CEST)
DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed;
   d=example.com; s=dkim;
   h=from:to:subject:date:message-id; bh=...snip...; b=...snip...;
Received-SPF: pass (company.com: domain of john.doe@example.com designates 192.0.2.1 as permitted sender) client-ip=192.0.2.1; envelope-from=john.doe@example.com; helo=mail.example.com;
Authentication-Results: inbound.company.com;
   dkim=pass header.i=@example.com;
   spf=pass (company.com: domain of john.doe@example.com designates 192.0.2.1 as permitted sender) smtp.mailfrom=john.doe@example.com;
   dmarc=pass (p=NONE sp=NONE dis=NONE) header.from=example.com

F6.8. Sources relating to suspicious email :

[CHA] ChatGPT, le moteur d'Intelligence Artificielle développé par OpenAI.
[CYM] cyber-malveillance.gouv.fr, Comment reconnaître un mail de phishing ou d'hameçonnage ?
[CYS] Cyber-securite.fr, Comment savoir si j'ai ouvert un mail douteux ?
[KEP] Keeper, Pouvez-vous être piraté simplement en ouvrant un e-mail ?
[PER] Perplexity, le moteur d'Intelligence Artificielle développé par Perplexity AI.

F7. Suspicious web address

When browsing the web, it is essential to know how to identify questionable web addresses, as clicking on a suspicious link can expose you to major risks such as theft of personal data, malware infections and online scams.

Structure of web address
Structure of suspicious web address
What not to do
What to do
Sources

F7.1. Structure of web address :

A web address (or URL - Uniform Resource Locator) is the unique address of a resource (web page, file, data, etc.) on the Internet.
A web address has a syntax defined mainly by the RFC 3986 standard as follows [CHA][PER] :

web address = protocol://[user:pathword@]domain[:port][/path][?parameters][#fragment]
- The terms in brackets are optional.
- Example of a complete web address : https://user:password@www.example.com:443/documents/report.pdf?id=1234&lang=en#section2
- The web address has a maximum length of about 30,000 characters (with a recommended value of about 2,000 for better readability, better processing performance on the client and web server side, and better compatibility with browsers and servers).
protocol or scheme = protocol used to access the resource (examples : http, https, ftp, ftps, sftp, file, data, ws, wss)
user:password = automatic authentication elements when accessing a protected web resource (example : user:password).
domain or host = domain name or Internet Protocol (IP) address of the server hosting the resource (examples : www.example.com, example.com, blog.example.com, 192.168.1.1)
- --------
- The web address domain has a maximum length of 253 characters (including subdomains and the dots that separate them).
- The web address domain can be entirely numeric (IP address) as long as it is not a special web address domain.
- All other characteristics of an email address domain apply to a web address domain.
- --------
- The IPv4 address domain uses a decimal format divided into four 8-bit blocks ordered from left to right and separated by a dot "." (example : 192.168.1.1).
- The IPv4 address domain can be followed by a space " " and a network mask. This network mask defines a range of addresses where the bits corresponding to the "1"s in the mask are fixed for the network address, while the bits corresponding to the "0"s vary from 0 to 1 to cover all addresses in the range. For example, for the address 33.0.0.0 224.0.0.0, the range is from 32.0.0.0 to 63.255.255.255. The network mask 224.0.0.0 is also written /3 in CIDR notation.
- Some IPv4 addresses are special : the network address, which is the address whose host bits (bits corresponding to the "0" of the network mask) are at 0, and the broadcast address, which is the address whose host bits are at 1. For example, for the address 10.0.0.0 255.0.0.0, the network and broadcast addresses are respectively 10.0.0.0 and 10.255.255.255
- --------
- The IPv6 address domain is completely different. It uses a hexadecimal format divided into eight 16-bit blocks ordered from left to right and separated by a colon ":" (example : 2001:db8::1:0:1). A null block (0000) can be represented by a single zero. A group of consecutive null blocks can be replaced by a double colon ":", which can only be used once per address to avoid any ambiguity that a double occurrence would create. Thus, the address 2001:db8::1:0:1 expands to 2001:db8:0:0:0:1:0:1
- The IPv6 address domain can be followed by a percentage "%" and a network zone identifier in the case of a link-local scope address (example : fe80::3%eth0).
- The IPv6 address domain is enclosed in square brackets when used in a URL (recommended framing) or associated with a port number (mandatory framing). Examples : http://[2001:db8:a02:12f0::1]/index.html, [fe80::3%eth0]/64, [2001:db8:a02:12f0::1]:80
- --------
- The IPv4 or IPv6 address domain can be followed by a slash "/" and a number n according to the CIDR (Classless Inter-Domain Routing) notation. This number designates the number of bits constituting the network prefix, defining an address range where all bits located after the first n are set to 0 for the minimum address and 1 for the maximum address. For example, for the IPv4 address 33.0.0.0/3, the range is from 32.0.0.0 to 63.255.255.255. Similarly, for the IPv6 address 2001::/3, the range is from 2000:0:0:0:0:0:0:0 to 3FFF:FFFF:FFFF:FFFF:FFFF:FFFF:FFFF:FFFF
port = port number that specifies the communication channel used for the connection (example : 443). By default, port 80 is used for the http host and 443 for the https host.
path = a specific path to the resource on the server, often as a directory or file (example : documents/rapport.pdf)
- The path has a maximum length of 253 characters. By default, the path is the root path.
parameters = key=value pairs separated by an "&", which can be added to convey additional information about the resource (example : id=1234&lang=fr)
fragment = a specific section of the resource, often used to navigate to a particular part of a web page (example : section2)

F7.2. Structure of suspicious web address :

The criteria for identifying a suspicious web address are as follows [CHA][PER] :
- Site with unsecured exchanges (http protocol instead of https) when it includes a personal data entry form. Note that a purely "showcase" http site is not very dangerous.
- Web address whose domain contains one of the typical anomalies of a suspicious email address
- IP address from an unsolicited email or an unknown source
- IP address used alone, without context or identification of a legitimate organization
- Special IP address that should not normally appear on the public Internet. The list of these addresses is as follows :

Addresses not assignable to a machine (specific network functions)
- Range used to designate an invalid, unknown or unconfigured destination
  - 0.0.0.0 - 0.255.255.255 corresponding to 0.0.0.0/8
  - 0:0:0:0:0:0:0:0 corresponding to ::/128
- Reserved range for IPv4 addresses mapped to IPv6
  - 0:0:0:0:0:FFFF:0:0 - 0:0:0:0:0:FFFF:FFFF:FFFF corresponding to ::FFFF:0:0/96
- Local loopback (localhost) ranges, used for internal machine communication
  - 127.0.0.0 - 127.255.255.255 corresponding to 127.0.0.0/8
  - 0:0:0:0:0:0:0:1 corresponding to ::1/128
- Reserved range for for protocol testing
  - 192.0.0.0 - 192.0.0.255 corresponding to 192.0.0.0/24
- Reserved ranges for specification documents and educational examples
  - 192.0.2.0 - 192.0.2.255 corresponding to 192.0.2.0/24
  - 198.51.100.0 - 198.51.100.255 corresponding to 198.51.100.0/24
  - 203.0.113.0 - 203.0.113.255 corresponding to 203.0.113.0/24
- Multicast ranges allowing a data packet to be sent to multiple machines
  - 224.0.0.0 - 239.255.255.255 corresponding to 224.0.0.0/4
  - FF00:0:0:0:0:0:0:0 - FFFF:FFFF:FFFF:FFFF:FFFF:FFFF:FFFF:FFFF corresponding to FF00::/8
- Limited broadcast address, used in networks local
  - 255.255.255.255 corresponding to 255.255.255.255/32
- Ranges reserved in IPv6 for documentation or testing purposes
  - 2001:10:0:0:0:0:0:0 - 2001:10:0:0:0:0:0:FFFF corresponding to 2001:10::/24
  - 2001:DB8:0:0:0:0:0:0 - 2001:DB8:FFFF:FFFF:FFFF:FFFF:FFFF:FFFF corresponding to 2001:DB8::/32
Machine-assignable addresses (private or local ranges)
- Private ranges, used in local area networks (LANs)
  - 10.0.0.0 - 10.255.255.255 corresponding to 10.0.0.0/8
  - 172.16.0.0 - 172.31.255.255 corresponding to 172.16.0.0/12
  - 192.168.0.0 - 192.168.255.255 corresponding to 192.168.0.0/16
- Large-scale NAT or Carrier-Grade NAT (CGN) range, used by Internet Service Providers (ISPs)
  - 100.64.0.0 - 100.127.255.255 corresponding to 100.64.0.0/10
- Automatic Private IP Addressing (APIPA) range, used when the operating system failed to obtain an IP address via DHCP
  - 169.254.0.0 - 169.254.255.255 corresponding to 169.254.0.0/16
- Range of unique local addresses (ULA), allowing private and secure communication within local or interconnected networks
  - FC00:0:0:0:0:0:0:0:0 - FDFF:FFFF:FFFF:FFFF:FFFF:FFFF:FFFF:FFFF corresponding to FC00::/7
- Link-local IPv6 address range, allowing communication between machines on the same network link without relying on external configuration or routers
  - FE80:0:0:0:0:0:0:0 - FEBF:FFFF:FFFF:FFFF:FFFF:FFFF:FFFF:FFFF corresponding to FE80::/10

F7.3. What not to do :

- Click on the link.

F7.4. What to do :

- Check the owner of the IP address or domain using tools like WHOIS
- Analyze the reputation of the site via forums or specialized cybersecurity sites
- Copy the link and do an Internet search to see if other people have reported this link as fraudulent
- Preserve evidence through screenshots
- Change the default font of the email client to clearly distinguish between an "l" and a "capital i": switch to Times New Roman, or Georgia or Verdana for example.

F7.5. Sources relating to suspicious web address :

[CHA] ChatGPT, le moteur d'Intelligence Artificielle développé par OpenAI.
[PER] Perplexity, le moteur d'Intelligence Artificielle développé par Perplexity AI.

F8. Artificial intelligence

Two topics of artificial intelligence (AI) are presented below : AI engines and robotics.

AI engines
Robotics
Sources

F8.1. AI engines

The AI engines provided by modern artificial intelligence have strengths but also limitations that show that AI engines cannot be used as sources of absolute truth.
On the other hand, they are powerful tools when used critically and in addition to human expertise.
Let us examine these aspects in more detail.

picture Artificial intelligence - AI engines

F8.1.1. Operation [CHA][PER] :

All current AI engines build their response word by word, predicting at each step, from the question asked and the available data, the word most likely to follow the words already generated.
The goal is therefore not to "say what is true", but to "say what is probable" based on the available data.
The available data comes from three sources : data learned during training (frozen at the knowledge cutoff date), data retrieved from the web in real time (for engines that have access to it), and the context of the ongoing conversation.
Internally, the engines calculate a confidence score based on the similarity between the available data and their correspondence with the question asked.
- If this score exceeds a sufficiently high threshold (decision threshold), the engine produces an answer in a probabilistic manner, which may correspond to an exact answer, an adapted answer (through interpolation or extrapolation from nearby data), or a full hallucination. However, the line between these three types of answers, although defined internally by a confidence score and a decision threshold, remains completely opaque to the user.
- In the opposite case, most engines produces a negative answer or request clarification in order to reduce the risk of hallucination.

Conclusion :
Thus, all AI engines produce more ou less hallucinations, biases, and errors, which users cannot detect directly.
It is therefore essential to cross-check the information provided by an AI engine through cross-verification with other recognized sources or other AI engines.
This approach - a cornerstone of journalistic and scientific work - offers two key advantages :
- It effectively limits hallucinations, biases, and errors.
- It enables selecting the best from each source to achieve a more comprehensive and reliable synthesis.
Notes :
- Reliability rates : Any comparison between AI engines based on reliability rates remains arbitrary due to the heterogeneity of training data and evaluation methods. To date, no robust benchmark exists to objectively assess the reliability of AI engine responses. Most reliability rates lack a unified definition, and even when they exist, evaluations do not cover all engines.
- Response quality : AI engines always strive to provide the response closest to the posed question. Users must therefore formulate their questions clearly, precisely, and structured.

F8.1.2. Advantages [CHA][PER] :

Current AI engines have some serious advantages, in particular :

Excellent conversational assistants :
- Ability to understand and respond in natural language
- Fluid and contextual interaction with users
Producers of quality summaries :
- Ability to compile and summarize information in a coherent and systemic way
- Neat and structured presentation of data
Access to up-to-date information :
- Implicit or ad hoc ability to search for up-to-date information after the date of its last learning, in order to provide relevant answers to users. This ability is not always explicitly announced.

F8.1.3. Major limitations and flaws [CHA][PER] :

Current AI engines have some major flaws and intrinsic limits that affect the quality of the answers provided. These are mainly explained by :
- the probabilistic operation of the engine.
- the quality, representativeness and freshness of the training data.
- the quality of the raw data collected on the Web as well as that of the filtering carried out on this data.
- the absence of an internal verification mechanism.
- the opacity of the algorithms which favors the appearance and persistence of faulty reasoning that is difficult to detect and correct.
These include :

1. Invention of information ("hallucination") and without explicit warning to the user. Common cases are as follows :
- 1.1. Too open or suggestive question : The AI engine produces an answer reflecting the framing and tone of the question asked. Common cases are as follows :
  - Too open question : For example, the question "What can you say about this subject... ?" induces a catch-all answer.
  - Explicit suggestive question : For example, the question "imagine..." or "tell..." induces a fictional answer.
  - Subjective suggestive question : For example, the question "according to you..." or "isn't it ?" induces an opinion-based answer.
  - Appreciative suggestive question : For example, the question containing the terme "perfect", "excellent", "good", "bad", "zero", "better", "worse", "indisputable" or "doubtful", induces a value-judgmental answer when the term is isolated. In the case of reported value judgment (such as "Which literary critics consider Les Misérables to be Victor Hugo's best novel ?"), the question remains factual and objective.
- 1.2. Question with inexact parameters but without major impact on the substance : The AI engine produces the most plausible answer possible based on close and accurate data, using wording that is sometimes cautious but not always. Common cases are as follows :
  - Wrong parameter (date, place, author, circumstances, etc.) : For example, in the question "At the literary evenings organized by Marcel Proust in Paris in 1908, what type of music was usually played for his guests ?", the date 1908 is wrong, but the answer will respect the spirit of the question by relying on similar exact data (those from 1907) to avoid a negative answer.
  - Exaggerated parameter : For example, in the question "Why do all critics consider this novel perfect ?", the term "all" is an exaggeration that introduces a hallucinatory bias into the answer. Exaggeration can be spatial (such as "everywhere", "nowhere"), temporal (such as "always", "never"), modal (such as "extremely", "terribly"), quantitative (such as "everything", "nothing", "all", "nobody", "none") or causal (such as "it's all your fault...").
  - Contradictory parameter : For example, in the question "What criticisms did Freud make of the Internet ?", the AI engine will make an improvised answer while also pointing out the contradiction.
- 1.3. Missing, sparse or fragmentary data : The AI engine fills in the gaps through statistical prediction, producing plausible text that may include fictional details, often without explicit warning to the user.
- 1.4. Occasional hallucinations : Even with clear instructions to provide an answer based only on confirmed facts, an AI engine may still produce occasional hallucinations due to its probabilistic operation.
2. Incorrect answer generation with persistence in the error. Common cases are as follows :
- 2.1. Question with multiple parameters : The answer may be inaccurate due to omission of certain parameters. This omission is often related to structural limitations of AI engines in contextual understanding and information processing.
- 2.2. Inconsistent data : When data is incorrect, noisy, or contradictory, an AI engine reproduces them directly in its answers without attempting to correct them. For example, if the engine encounters contradictory dates for Marie Curie's birth (for example : "November 7, 1867" and "November 12, 1867"), it can answer "Marie Curie was born on November 7, 1867 or November 12, 1867".
- 2.3. Out-of-context data : When the data is poorly contextualized by the AI engine, even if it is correct, the engine may misunderstand the context of the question or misinterpret the available data. This leads to incorrect answers. For example, if the question is "What are the best exercises to improve posture ?" and the engine answers "Running, swimming, and cycling are great for improving overall fitness." it misinterprets the question and provides out-of-context information, even if it is relevant in other contexts.
- 2.4. Data too close : When there are homonyms, close concepts, people or events that look similar, an AI engine can confuse them, for example mixing up two similar but distinct scientific theories.
- 2.5. Outdated data : An AI engine may provide outdated information in situations where updating is unavailable, insufficient or too slow, leading to errors, especially when it comes to recent events, scientific discoveries or technological developments.
- 2.6. Persistence in the error : When the user insists or rephrases a question, the AI engine can only repeat or paraphrase the initial error. However, if the user provides compelling evidence, such as a reliable source, a relevant excerpt, or a detailed explanation, some advanced AI engines are able to temporarily adjust their answer within the same session.
3. Sensitivity to bias related to the non-quality of sources :
- The data used by an AI engine comes from both a vast training corpus and additional information collected from the web. These sources, while numerous and varied, are neither exhaustive nor balanced. Some information may be overrepresented, while other information is rare, absent, or distorted by social and cultural biases, such as those found in online forum discussions or certain press articles.
- As a result, an AI engine reproduces these biases in its answers, and sometimes even amplifies them or gives them apparent legitimacy.
4. Lack of contextual coherence :
- An AI engine may have difficulty following the thread of a long discussion and connecting previously provided information, leading to inconsistencies or repetitions in its answers.
5. Lack of mathematical rigor and formal logic :
- An AI engine does not perform formal logical reasoning but seeks to reproduce reasoning steps observed in its training data.
- The user can, however, guide the AI engine toward a more rigorous proof by providing precise definitions, relevant theorems, a step-by-step logical structure for the proof, and by using mathematical symbols and notations rather than ordinary text.

F8.1.4. How to choose and use an AI engine properly ? [CHA][PER] :

It seems that the major AI players will have adopted two distinct strategies for AI engines :

- Google and Microsoft have chosen to offer free versions with improvable quality, deemed sufficient for mainstream consumer needs (Gemini and Copilot engines).

- Anthropic, Perplexity AI and OpenAI have prioritized a quality-driven approach, even with free access (Claude, Perplexity and ChatGPT engines). Claude and Perplexity excel in answer reliability, while ChatGPT excels in the completeness of its answers and its editorial and explanatory quality. All three offer an intuitive and well-structured user interface, as well as contextual follow-up to ensure the flow and consistency of conversations.

Upgrading to a paid version also improves the user experience and provides access to some additional features. But this doesn't guarantee that major flaws inherent in some AI engines will be addressed, such as their inability to consider all parameters of a question or to provide consistent contextual follow-up.

Furthermore, it is important to know how to choose and use an AI engine correctly :

Main AI engines (in 2026)
- Google Search (launched by Google in September 1998) : Great for accessing a variety of sources, though it's not a conversational AI engine. Its extension Google AI Overviews (launched in late 2024) is dedicated to quickly synthesizing information from the web.
- Perplexity (launched by Perplexity AI in December 2022) : Prized by professionals for the high reliability of its answers thanks to factual accuracy accompanied by systematic source citations. This reliability is mainly based on a more limited training dataset consisting of high-quality, well-indexed, and well-structured content, which can restrict access to niche sources. Real-time web access is almost systematic and lies at the core of the engine's functioning. Note that Perplexity tends to provide precise technical translations, faithful to the context and without resorting to creative reformulations.
- Claude (launched by Anthropic in March 2023) : prized by professionals for the high reliability of its answers thanks to structured and cautious reasoning. This reliability is mainly based on a large-scale training dataset consisting of well-indexed and well-structured content. Real-time web access is occasional and used as a complement to internal reasoning. Note that Claude tends to provide precise technical translations, faithful to the context and without resorting to creative reformulations.
- ChatGPT (launched by OpenAI in November 2022) : Prized for its completeness and its editorial and explanatory quality. This completeness is mainly based on a vast training dataset including a large amount of content derived from the public web. Real-time web access is occasional and used solely to obtain factual or evolving information. In its text generation, ChatGPT prioritizes the fluidity and elegance of speech, thus losing some of the factual accuracy of the content. In the free version, a major limitation exists : For poorly documented topics, ChatGPT tends to provide an incorrect or even invented answer, even when explicitly asked to provide only confirmed facts (see following example).
- ChatGPT Search (launched by OpenAI in February 2025, formerly SearchGPT) : A paid extension of ChatGPT that provides up-to-date, concise answers with cited sources. The generated texts are more condensed and generally less detailed and nuanced than those produced by ChatGPT.
- Google Gemini (launched by Google in December 2023, formerly Bard) : Also provides detailed explanations, but relies more on confirmed and up-to-date sources. It is also capable of analyzing and generating multimodal content (text, images, tables, audio, video). In the free version, two major limitations exist : the answers are often inaccurate due to omission of certain parameters of the question, and the contextual follow-up is sometimes inconsistent with loss of the thread of the conversation.
- DeepSeek (launched by DeepSeek AI in 2024) : Prized for its precision in mathematics and logic, but offers variable and unpredictable factual reliability as it doesn't verify responses via native web search, especially in the free version.
- Microsoft Copilot (launched by Microsoft in February 2023, formerly Bing Chat) : Versatile, generally effective, and tightly integrated into the Microsoft desktop environment. In the free version, two major limitations exist : the answers are often inaccurate due to omission of certain parameters of the question, and the contextual follow-up is sometimes inconsistent with loss of the thread of the conversation.
Comparative example between AI engine
- Question : "The former hamlet of Brocaries in the commune of Varennes in the Dordogne department is mentioned by the Viscount de Gourgues in 1873 in his book Dictionnaire Topographique du département de la Dordogne. Does this hamlet appear exactly under this name on the Cassini map ?"
- Answer : The question should receive a negative answer, the map indicating Brocarie and not Brocaries. ChatGPT confidently asserts the opposite. It does indeed identify the hamlet in the Dictionary, but, unable to verify its spelling on the Cassini map, it formulates an erroneous conclusion based on extrapolation. Note that Perplexity, Claude, Gemini and Copilot, on the contrary, provide a negative answer by default, explicitly reporting the absence of available sources or the impossibility of verifying directly on the Cassini map.
Main selection criteria
- Hallucinations. So that the engine hallucinates little or signals its extrapolations, favor cautious AI engines. For example : Claude, Perplexity, Gemini.
- Systematic citations. "To be able to verify or dig deeper into answers, favor AI engines that cite systematically their sources. For example : Perplexity, SearchGPT.
- Contextual follow-up. To ensure the flow and consistency of conversations, favor AI engines designed to retain and reuse the full history of a conversation. For example : ChatGPT, Claude, Perplexity.
- Message limits. Favor AI engines based on their maximum number of answers per day or per session. For example (for free versions) : ChatGPT (10 by 5 sliding hours, then degraded conversation without blocking), Claude (30 by 5 sliding hours), Gemini (50 per day), SearchGPT (80 by 3 sliding hours), Copilot (30 per session), Perplexity (5 by 5 sliding hours).
- Math capabilities. To solve math problems, choose an AI engine specialized in math. For example (for the free basic versions) : Wolfram Alpha, GauthMath, Photomath, Smodin Omni, Julius AI, MathGPTPro, DeepSeek.
How to ask a question ?
- ask only one question at a time in order to obtain a more targeted answer.
- Ask a clear and precise question, especially to obtain a mathematical demonstration.
- To minimize hallucinations, avoid overly too open or suggestive questions, and always frame the question by ending it with the sentence : "Respond with confirmed facts."
How to analyze the answer ?
- When searching for truth, never rely on the style or fluidity of the answer.
- Evaluate the coherence and relevance of the answers by asking the same question with different formulations.
- Cross-check the information by comparing with other AI engines (cross-checking principle).
- Verify crucial or unexpected information with reliable sources.
- In sensitive or complex areas, prefer to consult human experts.
How to get out of an inextricable situation ?
- If an AI engine persists in erroneous or repetitive reasoning, you should switch sessions, which theoretically allows you to bypass continuity errors or biases that can accumulate during a long session.

F8.1.5. Key factors of an AI engine [CHA][PER] :

An optimal user experience of an AI engine is based on several key factors :

Technical quality :
- Real-time, contextual filtering of the web, favoring content from well-known and/or confirmed sources.
- In-depth exploration and analysis of this content.
- Multimodal search (text, images, tables, audio, video).
- Fast loading times to maintain user engagement.
- Intuitive and well-structured user interface.
- No advertising or distractions.
- Display of usage limits (maximum number of queries per day/month/session, restricted access to certain advanced features, etc.).
- Compliance with accessibility standards ensuring ease of use for everyone, including people with disabilities.
- Respect for the confidentiality and security of personal data, both for users and data owners.
Quality of results :
- Excellence in conversational assistance.
- Production of targeted, complete, accurate, and sourced information.
- Production of high-quality summaries.
- Production of exact mathematical demonstrations.
Quality of interactions :
- Ability to minimize hallucinations, biases, and incorrect answers.
- Contextual follow-up ensures the flow and consistency of conversations.
- Display of conversation history.
- Personalized interactions that automatically adapt to the user's language, adjust language and tone based on preferences, and offer relevant recommendations.

F8.2. Robotics

Modern robotics, whether domestic or industrial, has significant advantages but remains constrained by significant dangers.
Risks related to technical or software failures, as well as uncertainties in the face of unforeseen environments, often limit the use of robots to strictly controlled spaces in order to guarantee the safety of humans.
Let's examine these aspects in more detail.

picture artificial intelligence - humanoid robot

picture artificial intelligence - industrial robot

F8.2.1. Advantages of current robotics [CHA][PER] :

Current robotics has the following main advantages :

Productivity and efficiency : Industrial robots perform repetitive tasks with speed and precision, increasing productivity in key sectors such as automotive, electronics and logistics. They minimize human error and operate without interruption or fatigue.
Safety and risk reduction : In hazardous environments (nuclear power plants, demolition sites, laboratories handling toxic products, etc.), industrial robots replace humans, limiting their exposure to risks. Some robots, such as exoskeletons, assist workers by preventing injuries due to heavy loads.
Advances in domestic robotics : Humanoid robots are distinguished by advanced mobility, manipulation and communication capabilities. Their design allows them to evolve in environments designed for humans, by climbing stairs, opening doors, carrying objects or handling common tools such as pens, cups or kitchen utensils.
Contribution to science and health : Space or marine exploration robots make it possible to reach places inaccessible to humans. Surgical robots perform complex operations with precision. Companion robots play a vital role in assisting the elderly or disabled.

F8.2.2. Limits and dangers [CHA][PER] :

The main limitations and dangers of current robotics are as follows :

High cost : Robots are still expensive to develop, manufacture, and maintain, limiting their adoption by small businesses or in developing countries.
Technological limitations :
- Mobility and autonomy: Humanoid robots, while promising, are often limited in terms of balance, speed, and energy endurance (insufficient batteries).
- Fine manipulation : Compared to human dexterity, robots are still limited in tasks that require tactile sensitivity and complex coordination.
- Environmental awareness : Even advanced AI systems (such as computer vision) can fail when faced with noisy or unforeseen environments.
- Natural human-robot communication : Robots have difficulty understanding the nuances of language, emotions, and gestures.
Hazards : Cohabitation between robots and their environment, particularly in warehouses and hospitals, induces risks related to the limits of their design and the unpredictability of the environments.
- Mechanical ou electronic failures that can injure humans nearby or damage sensitive infrastructure (listed in order of decreasing frequency) :
  - 1. Breakage or wear of mechanical components causing sudden and uncontrolled movements
  - 2. Control system failure causing loss of applied force or erratic robot behavior
  - 3. Overheating of the motors or power supply causing the robot to stop suddenly or behave unpredictably
  - 4. Failure of the gripping mechanisms causing heavy or sharp objects to fall
  - 5. Failure of the brakes on the actuators preventing the robot from stopping its movements
  - 6. Rupture of high-pressure hydraulic lines causing fluid to spray
- Software malfunctions and poor decision-making : A robot may react disproportionately, inappropriately or dangerously due to many causes (listed in order of decreasing frequency) :
  - 1. Software errors (bug in the code, bad algorithmic logic, defect in the integration of software modules, lack of robustness in the handling of exceptions)
  - 2. Poorly calibrated or faulty sensors providing incorrect data on essential parameters (distance, speed, temperature, pressure, orientation, etc.)
  - 3. Unforeseen environmental conditions (excessive noise, rain, fog, dust, inadequate lighting (insufficient or excessive), shadows, reflections, electromagnetic interference, vibrations or external shocks impacting the accuracy of the sensors or the mechanical balance)
  - 4. Insufficient training of operators to interact with the robot
  - 5. Unforeseen situations (small object not detected, partially hidden object, obstacle not identified)
  - 6. Complex scene misinterpreted (moving pedestrians, unpredictable vehicles, cluttered environment)
  - 7. Overload or insufficient system resources (processor, RAM, disk space) slowing down or blocking the calculations necessary for decision-making
  - 8. Poorly designed human-machine interaction, leading to misunderstandings or misinterpretations
  - 9. Unserious areas (absence or insufficient physical barriers, non-detection of humans in the event of a blind spot)
  - 10. Biases or limitations of artificial intelligence, leading to inappropriate or dangerous decisions
  - 11. Lack of redundancy and security in critical systems (absence of backup mechanisms in the event of a failure, weakness of emergency shutdown or recovery procedures after a serious malfunction)
  - 12. Malicious actions (hacking, injection of corrupted data, sabotage of communications)
Social and ethical issues :
- Impact on employment : Robotization leads to a loss of jobs for repetitive tasks, which requires retraining of workers.
- Social acceptance : Domestic robots, particularly humanoid robots, may arouse distrust or rejection in some social groups.
- Ethics : Domestic robots raise important questions concerning society, particularly regarding responsibilities in the event of error or harm caused.

F8.2.3. Advenir of domestic robots [CHA][PER][PET] :

Domestic robots embody a promising vision of the future where technology is put at the service of humans to improve their quality of life.
However, this vision can only be realized if technical robustness, safety and social responsibility are reconciled.
Domestic robots must not only be technically efficient and reliable, but also be thoughtfully and harmoniously integrated into our homes. This integration must take into account their social, psychological and ethical impacts.
If these complex challenges are successfully addressed, domestic robots could become essential partners in our daily lives. Otherwise, the risks and questions they raise could slow down their adoption, limiting their future to niche uses.

F8.3. Sources relating to artificial intelligence :

[CHA] ChatGPT, le moteur d'Intelligence Artificielle développé par OpenAI.
[PER] Perplexity, le moteur d'Intelligence Artificielle développé par Perplexity AI.
[PET] Régis Petit, Contribution à l'intégration de la vision passive tridimensionnelle en robotique : suivi de contours, calibrage de caméra et commande cinématique de robot, Thèse de Docteur-Ingénieur, INPT Toulouse, 1985.

Last page update : May 15, 2026.

Regis Petit's Website