Some thoughts about artificial stupidity and artificial dumbness

2.1. A terminology note

This work is based on the reflections of French researchers (the authors of this article), with the objective of studying the notion of bêtise artificielle, which we translate here as “artificial stupidity”. However, the French term bêtise does not correspond exactly to the English term stupidity. The English and French vocabularies are both very rich in terms to express various forms of impairments with respect to the behavior normally expected of a person who reasons and makes decisions. Indeed, in French, the two main words bêtise and stupidité have slightly different meanings, namely: stupide, the corresponding adjective, usually refers to behaviors where wrong conclusions are drawn (while the right ones are easy to derive), while bête, the adjective counterpart of bêtise, rather applies to situations where people cannot conclude anything (although conclusions follow rather straightforwardly). Thus, bêtise rather corresponds to English terms such as dumbness or cluelessness. It is also worth noting that bête in French also means beast which suggests a comparison with animals (implicitly assumed to lack reasoning capabilities); in English, if we except donkey (in French, âne) wrongly reputed to behave stupidly in both languages, there is no word referring to animals for expressing some form of stupidity. Foolishness is another term that refers to the quality of being unwise, stupid, or not showing good judgment, and has a rather broader meaning that encompasses craziness.

2.2. An anthropocentric viewpoint

Let us start by trying to understand the meaning of the term “artificial intelligence”. In such an attempt, one is faced with the tricky question of defining what “intelligence” means. To do this, we most often get around the difficulty by making a reference to the human being, which presents the form of intelligence that is probably the most accomplished, or at least the most polymorphous.

The Turing test [39] goes in this direction of such an imitation game: a process is seen as “intelligent” when the observer cannot predict whether he/she is interacting with a human or a machine.

Thus, the adjective “artificial” does not pose any specific concern in the elaboration of a definition (“artificial” only means that the intelligent process being analyzed is produced by a machine). Only the noun “intelligence” is questionable.

It can be seen that the comparison to humans to characterize the term “intelligence” existed before the Turing test. In particular, for René Descartes [11, fifth part], “intelligence” (in the sense of reason) is the prerogative of human: It is what separates humans from animals. Therefore, stupidity would be a contrario characteristic of being an animal. The position of Descartes is debatable, of course: reason is just one form of intelligence and, from a biology viewpoint, humans are just animals among others. Today, in the field of AI, when bio-inspired algorithms implementing a form of collective intelligence are designed and studied, most of the time the inspiration comes from animal societies (such as colonies of ants [7]), not from human societies. Is it because the latter have too complex behaviors to be modeled in a satisfactory way or because the value added by the collective is too limited to be significant? Everyone can have a point of view on this issue.

2.3. To be dumb

Stupidity (in the sense of the ability to be stupid or rather of the inability to not be stupid) has often been defined as a deficiency of intellectual capabilities (in particular, as the etymology of “intelligence”3

For qualifying an agent (whether human or artificial) as stupid, it is obviously not enough that it produces a stupid answer from time to time (otherwise, every humans would be stupid), but that this agent persists in her/his/its errors, even when a continuous interaction with the environment provides her/him/it enough good arguments so that, in principle, these errors should not be repeated. From this viewpoint, artificial agents have a higher predisposition to stupidity than human agents, simply because the interactions with the environment is typically more reduced (some of them have neither sensors nor effectors) but also because their strength is to be able to repeat without getting tired (and very quickly) the same treatments and, consequently, the same errors. It can be noted that this idea of stupidity linked to the persistence in error is one of the points studied in detail in the work of François Rollin cited in the introduction, even if it is hardly put forward in the chapter on Artificial Stupidity of his book.

Stupidity takes very varied forms; it is probably vain to want to characterize precisely what it is to be stupid. However, this idea of persistence in error can be retained as one of its characteristic features. The immutability of beliefs takes us out of the human condition and of its capability to question them whenever necessary. It brings us closer once again to the animal, in particular, the ostrich (whose head in the sand policy could be expressed as “see nothing, know nothing”), or even to the super-human, since, as the famous Latin quotation suggests, errare humanum est, perseverare diabolicum. Leaving the last word to Eugène Marbeau: “Foolishness believes itself very clever and doubts nothing”5

Defining what is a stupidity is probably simpler than defining what being stupid means. So, let us embark on an attempt to identify stupidity!

A stupid answer can be seen as the result of a process of reasoning or decision making that has gone awry; this result is simply wrong or inadequate, because it is based on information that is also wrong, because it is generated by using fallacious rules of reasoning or by taking shortcuts in the reasoning, or because it is based on an irrational decision model. Assuming from the start the existence of an “intelligent” process that might have been appropriate allows us to separate stupidity from nonsense: the former would be a deviation from the result of the intelligent process whereas the latter would be produced by a totally irrelevant process. Thus, to the question “How much is 12345 times 67890?”, an answer like “Stéphanie de Monaco” would not only be stupid, but absurd.

To illustrate the production of stupid conclusions as a result of an inadequate intelligent process, consider the following two examples (one related to reasoning, the other to decision making):

Cats are mortal, Socrates is mortal, therefore Socrates is a cat. 6

Eugène Ionesco, Rhinoceros.

Here, the information used is correct, it is the inference that is not: in this case, it is abductive, not deductive, and therefore can lead to erroneous conclusions (as is the case here).

Here is another example (a dialogue between two friends):

I read in a magazine that my alcohol consumption (one bottle of Bordeaux per day) is excessive and could seriously damage my health.

Deciding whether the result is incorrect or inadequate requires comparison to a standard. Depending on a certain “distance” to a reference point, the result in question may or may not be considered appropriate and the process that produced it will be considered (more or less) “smart” or “dumb”. The distance in question may then arouse various feelings in the person measuring it, from anger to laughter, or even wonder.

Note that the standard used to measure the adequacy of the result produced by the reasoning or decision-making process under consideration may be universal (mathematical logic allows, for example, the separation of valid reasoning from invalid reasoning) or it may not be (there is no single, consensual model to characterize what reasonable decision making is). Indeed, this norm is often context-dependent (thus non universal), in particular for the person who receives the message and decides to classify it as stupid; and this is a function of his or her own abilities, so that what is stupid for one individual is not necessarily so for another. To take an example for which there is no universal norm, consider the problem of preparing a glass of Scotch whiskey. Serving it “on the rocks” (i.e., with ice) is viewed as stupid by some and as normal by others.

Let us take an example for which there is a universal norm. It is again the question of multiplying 12345 by 67890. There is only one true result: the product is 838102050. The result ‘ − 1 ’ produced by individual A could be considered stupid by anyone with little mathematical background, the result 838102051 produced by individual B, also erroneous, will be considered as a calculation error by some and as stupidity by others (we can observe that as 67890 ends with 0, the product sought must also end with 0), and the result 838012050 produced by individual C, just as false, will probably be perceived as less stupid. It is this “distance” from the expected result (which may also be incorrect) that causes a simple error to acquire the status of stupidity.

It can be noticed that for this kind of tasks, humans, because of their cognitive limitations, are often dumber than simple machines such as a simple pocket calculator. It has been known for a long time that the capability of humans to correctly reason with numbers (and, in particular, with probabilities) is reduced. In his book [28], the American mathematician John Allen Paulos has introduced the neologism “innumeracy” to designate this inability (innumeracy is to numbers as illiteracy is to letters). An illustration of this idea in the context of the pandemic period that started in 2020 can be given. During this period, some people were refusing to be vaccinated because of the (proven) fact that there are more severe cases of COVID among the vaccinated ones than among the non-vaccinated ones. This fact had been seen as a (fallacious) argument that would support the idea that the probability of contracting a severe form of COVID when being vaccinated would be higher than the probability of contracting a severe form of COVID when not being vaccinated (obviously, the error comes from the fact that the numbers of vaccinated and non-vaccinated persons are not the same!). Such errors of reasoning can unfortunately lead to tragic outcomes, as it happens in other settings such as the legal framework (see the fascinating book [37]).

More generally, it has been shown that human decision making frequently appears as irrational. Indeed, although rationality postulate systems [36] are sets of postulate that are widely accepted as reasonable in decision making, it has been shown that most human decision makers have made choices that deviate from these postulates in experimental settings [8].

Obviously, if machines largely surpass humans when it comes to making calculations, their intelligence is much less polymorphic: they do not know how to do anything else! In fact, the intelligence that is manifest at the execution of these programs comes more from the humans that have designed these programs than from the machine. In particular, a pocket calculator has no more conscience of the operations that it computes and of what is a number than a clock has of the time, or than a facial recognition system has of what is a face and what are relevant parts of a face.

One should also keep in mind that humans can perfectly consider as stupid some reasoning processes that are not, since they correspond perfectly to a universal norm, when the latter exists (e.g. for deductive reasoning). Thus, the validity of an inference depends only on whether its conclusion is true in all cases where its premises are true and absolutely not on the truth of the premises and conclusion involved in the inference. In particular, when the premises are inconsistent, any proposition can be deduced from them. The great logician Bertrand Russell gave this famous example: “If 2 + 2 = 5 then I am the Pope.” Indeed, it is known that the same value can be subtracted from the two members of an equality, while preserving its validity, and this can be applied to the value 3 and the equality 2 + 2 = 5 . Therefore, from premise 2 + 2 = 5 , it can be deduced that 1 = 2 . Now, the Pope and I are two persons. So, if 1 = 2 , then we are the same person: I am the Pope! The formal validity of this reasoning in classical logic can be established as follows: first, every logical consequence of any statement α must hold when α holds; second, α ∨ β is a logical consequence of α whatever β; third, when α is a contradiction, α ∨ β is equivalent to β. Overall, every β holds under any contradiction α. Accordingly, the reasoning followed by Bertrand Russell is valid, it respects the canons of classical logic and yet it appears as meaningless, in particular because its premises α ( 2 + 2 = 5 ) and its conclusion β (“I am the Pope”) have no connection.

3.1. A nice example of human stupidity

There is a story going around about a simple math problem that a student would have solved in a surprising, creative, but wrong way (a nice piece of stupidity). This problem (after some reformulation) was the computation of lim x → 0 1 5 x 2 It seems that the student has remembered the solution of a similar problem: lim x → 0 1 8 x 2 = + ∞

In order to give the reader the opportunity to imagine the student’s solution, the latter is given in Appendix: the remainder of this section is based on the assumption that the reader has read this Appendix.

Now, the following question can be raised: what competences an AI system requires to generate an answer such that this one in an “unintentional” way (i.e., an AI system that is not designed to generate stupid answers but generates some despite of its designer). We imagine that a system capable of generating the student’s stupidity must at the same time work on a representation of mathematical expressions (by a tree structure, for example) and on a visual representation, with links between these representations. From our perspective, building such an AI system would not be simple!

3.2. Stupid AIs

An AI program, or simply an AI, does only what it is programmed to do. One easily understands that a sophisticated program can do things that will be considered impressive, with regard to problem solving or image recognition, showing some kind of (artificial) “intelligence”. However, even very simple programs can fool us. This is the case, in particular, with the program ELIZA [40] which was able, in the mid-1960s, to simulate an interview with a therapist by simply rephrasing the “patient”’s sentences into questions, and then restarting the dialogue with preconceived sentences appearing to be related with the patient’s words. It is clear that when ELIZA was stating “I understand” because it was unable to provide a better answer in the context of the dialogue, it was bluffing its interlocutor, even if it was actually a quite stupid system.

This question of AIs that may seem “realistic” to their users, while relying on very simple (or even simplistic) techniques, has been relevant for a long time in video games. As stated in [42, Section 1.2.2], in the first era of video games, AI games were mainly based on predefined scripts. This limits the possible behaviors of non-playable characters (NPCs, which are the characters of the game controlled by the machine). The result of this is that, despite the huge progress made in the domain of AI for video games, there are still many players who complain about the stupidity of NPCs,8

The Dong Mingzhu case, or stupid and mean AI. Dong Mingzhu is a Chinese businesswoman, leading an important company in the field of air conditioning. To promote her company, she has organized a publicity campaign. Her portrait has been printed on city buses. However, in the People’s Republic of China, video surveillance is generalized. The large number of cameras and AI algorithms dedicated to face recognition allow the authorities to detect citizens who commit infractions or incivilities, such as crossing a road when it is forbidden to pedestrians because vehicles are passing. A social credit system is used, and when some incivility is noticed, the person who has committed it loses points; after losing too many points, this person can be denied some rights such as using public transportation. Dong Mingzhu quickly lost her points because each time a bus with her face on it was going in front of a security camera, the AI system concluded wrongly that Dong Mingzhu had crossed the road while it was forbidden.10

When natural language processing gets wrong. To be able to perform automatically (i.e., algorithmically) and in a satisfactory way language processing tasks, such as dialogues based on animated conversational agents (or other bots), summarizing or translating from one language to another, requires an ability to understand, which is not acquired by AI systems that remain devoid of common-sense knowledge about the world in which we live.

Solving anaphora (in particular, satisfying pronominal references) is a difficult problem that requires knowledge of the world (and the ability to reason about it) in order to be solved correctly. This problem is on the basis of Winograd schemas [41], which can be used as variants of the Turing test.

Ask your personal assistant (a robot): “Could you please go shopping for me at the book store and buy the last book by François Rollin, and if they have pencils, get 10.” If a short time later your robot comes back with 10 copies of the last book by François Rollin, you should not be so surprised. Indeed, it is probable that they had pencils. The mistake here comes from the fact that, for you, “10” refers to the pencils and not to the last book by François Rollin, while your robot solves the anaphora differently: for it, “10” refers to the number of copies of the last book by François Rollin to be bought, and not to the pencils.

Beyond assuming common-sense knowledge about the world, speech acts, as they are achieved by human beings, are also impacted by a number of conventions ruling the way communication between people usually takes place in talk exchanges. A chatbot that neglects those conventions may easily lead to provide stupid statements. As a matter of illustration, let us consider the following dialog:

Although the response provided by the chatbot is perfectly correct from a logical point of view, it is clearly off the mark. The problem comes from the fact that focusing on the locutionary content of the question is not enough. The question has an illocutionary function: what you really want to know is the birth sex of the baby. Answering “Yes” also does not comply with one of the Gricean maxims of conversation [19] that describe specific rational principles observed by people who follow the cooperative principle in pursuit of effective communication. Indeed, it can be assumed that everyone knows that the birth sex of any person is male or female. Thus, answering “Yes” is at odds with Grice’s maxim of quantity, which requires that the response should give as much information as is needed, but no more.

The automatic translation tools that are available online are wonderful tools for shallow translations. They are not stupid at all, but it is possible to have fun using them as stupidity generators because of anaphoras that are wrongly solved, but because of other problems as well.

For instance, let us try to translate in French the following sentence (which gives an argument whose justification is quite questionable) with an automatic translation tool:

When we translate it into French, we obtain:

This sentence means that the French word “début” starts with a “b”, and the word “fin” starts with a “e”, which is obviously not the case. After translating it, the justification of the argument is thus even more questionable. This comes from the fact that the AI algorithm does not distinguish between de re and de dicto understanding of the sentence, despite the quotation marks.

Of course, translating is a very difficult task, even for human beings (see, e.g., [15]); as a general rule, for a somewhat complex text, no ideal, perfect translation exists, and thus the nonsense that is produced does not reflect a great distance from a universal standard, but rather a difference from what is generally expected.

Among the sentences that are hard to translate are ambiguous sentences, especially those containing syntactic ambiguities. Syntactic ambiguity, also called structural ambiguity (the scholarly term is “amphibology”), is a situation in which a sentence may be interpreted in more than one way due to the ambiguous structure of the sentence. Thus, the sentence “British left waffles on Falklands” (a classic of its kind) is ambiguous, since its meaning is not at all the same depending on whether one considers that the verb is “waffles” or that the verb is “left”. Did the British political left rambles indecisively about Falkland Island policy, or did the British physically leave waffles on the Falkland Islands? Taken outside of a context that makes it possible to eliminate one of the two possible hypotheses, this sentence is not really translatable since its very meaning is unknown. It is funny to watch how machine translation tools do with these kinds of sentence. For this, once translated into French, the result is surprising: “Gaufres britanniques à gauche sur les Malouines”. This result does not correspond to any of the two meanings presented above, but to a third one (at least as unlikely), which (back to English) would be given by “British waffles on the left on the Falkland Islands”.

Many more examples of translations failing to preserve what is expected when achieved using automatic translation tools could be found. Since AI algorithms do not understand the texts they process, the possible hidden meaning of these texts remains inaccessible, even when taking them into account is essential, such as in puns or other jokes based on words. For example, the pun “How do you cut the ocean in half? With a see-saw!” once translated into French becomes “Comment couper l’océan en deux ? Avec une balançoire !” The homophony between “see” and “sea”, which is at the core of the pun, is fully lost in translation. In the process, the sentence loses its status as a pun, as one can hardly identify what could be funny in the answer. Once translated into French, it becomes mainly non-sensical, just as if the answer in English had been “With a banana!”

Here, it is important to highlight the fact that the errors produced by natural language processing systems, in particular for translation, can be considered as stupid only according to a non-universal norm (as discussed in Section 2.4), in this case, from the context of human understanding the natural language used (or natural languages for the case of translation) as well as puns in this (or these) language(s).

4.1. In order to pass the Turing test

As we have seen, the Turing test aims at verifying whether an automatic system could fool humans by making them believe that it is human itself, under certain conditions of interaction [39]. If such a system is too clever when answering a question, it may fail the test. To take an example from Section 2.4, if a machine interlocutor answers the question “ 12345 × 67890 = ? ” very quickly and correctly, he would fail in its deception. An answer that is far enough from the result without being completely nonsense could fool the human examiner.11

Therefore, with the aim of deception induced by the Turing test, in order to develop a computer system that would pass this test, it would have to include both AI and AS. Therefore, accepting the Turing test as a characterization of intelligence involves accepting stupidity as a part of intelligence.

One could imagine an AI system that would deliberately produce nonsense, perhaps in the spirit of [22], which notices that video games players usually find less enjoyable a game where the opponent (i.e. the AI algorithm controlling the NPCs) is “too intelligent” and plays optimally, which makes it very hard (or even impossible) to defeat. A slightly stupid opponent may then be more realistic, and more enjoyable (of course, a very stupid opponent is also less enjoyable since then winning the game would become too easy). Finding the good level of stupidity in this context is challenging.

One can imagine at least two other reasons for conceiving AS systems:

These are obviously hard challenges for AI! Understanding and automatically generating funny stories is a very complicated problem since humor is a matter of belief revision (the punchline forces listeners of a joke to revise the context set up at the beginning of the joke) [13,33]. It is also a matter of incongruity and violation of norms (incongruity often makes people laugh) [14,26]. Saying stupid things deliberately to make people laugh requires to be smart enough. Obviously a program able to do such things would be a very special AS system, able to say stupid things on purpose.

4.3. Creative stupidity

If separating the true from the false is the Holy Grail of the scientific process and is indispensable to walk with confidence in this life (to paraphrase René Descartes), one should not be afraid to make mistakes. Foolishness is fruitful. We learn by doing errors.

Mistakes are, in fact, a source of creation. History is full of scenarios where blunders have led to successes, such as the “bêtises of Cambrai”, the tarte Tatin, or leaving the culinary field while remaining in the realm of serendipity, the discovery of the vulcanization process by Charles Goodyear in the xixth century. Of course, in order for an error to be fruitful, intelligence is then required to recognize the existence of a potential to be exploited in the situation created by the error. Thus, an error in a mathematical proof will be fruitful if its discovery can lead to a new approach to the problem.

In humans, since early childhood, learning goes through phases of trial and error, through abusive generalizations that education, teaching, and reasoning allow (sometimes) to correct. Reinforcement learning implements this idea in AI: it goes through the exploration of situations that are deemed unsatisfactory or not. Observe that to be creative, reinforcement learning should probably allow a certain permissiveness in the estimation of the unsatisfactory character of a situation.

The same is true for the machine: inductive generalization, like abduction, analogy, or other forms of common-sense reasoning, does not preserve the truth. Learning, when it is not limited to rote learning (which does not offer the ability to take into account new situations), presupposes the risk of being wrong. The algorithms implemented in machines to learn aim at minimizing this risk in an empirical way, by relying on the available situations. They cannot completely eliminate it.

In the machine, it should be noted that the error committed can be more significant when the data is too small or of poor quality. This is the question of bias in the data used to learn, which can lead AIs to reproduce reprehensible behaviors,12

For stupidity to be creative, in humans and machines, the risk of being wrong must be accepted, but it must also be modulated and controlled by the implementation of mechanisms that allow us to question our beliefs and the decisions they lead to. This is essential in order to avoid the aforementioned perseverare diabolicum and thus strive not to remain stupid. Thus, a lot of work in AI has turned over the last 40 years to this multifaceted problem, including nonmonotonic reasoning (where conclusions change according to the hypotheses made, whether implicit or not), belief revision (roughly speaking, how to question one’s beliefs in order to change them sufficiently for restoring consistency without changing them too much?), or even learning techniques based on the decreasing of an error rate (such as the backpropagation algorithm for neural nets).

5.1. Reasoning foolishly

To commit errors in reasoning is undoubtedly to reason foolishly (even if it is within everyone’s reach). Deductive reasoning is largely based on the following scheme p → q p q (if p is true and if p implies q, then q is true). Allowing other schemes, such as, for example, p → q q p (which has an abductive flavor), or p → q p ¬ q , is more risky! According to the first, which does not offer guarantees of deduction, the observation of a consequence of p suggests that p is true, which is not stupid! Whereas the second scheme leads to incoherence, for sure. One can consult [6] for a study of fallacious arguments. A bad analogy, based on an overly hazardous parallel, likewise leads to absurdity: “The hedgehog is like a toothbrush, it has a stiff coat, and the toothbrush curls when scared”.

Improperly using deduction also leads to absurdities:

Cheap apartments are rare.

In the above situations, stupidity is due to the production of erroneous conclusions. On the contrary, dumbness may lead to the lack of conclusions from premises that would lead to obvious conclusions for most people. Drawing wrong conclusions and failing to conclude (when it is simple) are two different types of deficiencies that reveal unintelligent behavior.

If an AS system wants to look silly, producing false conclusions is not enough! It has to make it perceptible how it arrives at them. The top of silly reasoning is perhaps arriving at a correct conclusion by false inferences…unless this is a sign of intellectual virtuosity!

In a similar way, a perfectly correct reasoning process could be perceived as stupid simply because it is not clever, because it is too complex, because it requires complicated calculations, whereas there are much simpler approaches (as an example, we can think of the famous problem of the cyclists and the fly, see, e.g., https://www.othot.com/blog/the-complexity-of-simplicity and [20]).

Finally, if stupidity is also, as has been said, an inability to see, to create links between facts or statements of those facts, then, one can use ideas of systems with limited reasoning capacity [16] in the design of an AS system. For example, the system may be aware that p and p → q are true, without being able to form their conjunction and conclude q, e.g. [30].

5.2. Learning stupidities

A supervised learning system is based on examples where an example is an ordered pair ( x , y ) ∈ P × S with P , the space of problems (or inputs), and S , the space of solutions (or outputs).

In order to simplify the forthcoming discussion, it is assumed that the relation “has for solution” is functional and the underlying function is denoted by F : the non-noisy examples ( x , y ) of a supervised learning system verify y = F ( x ) . Thus, the objective of such a system is to approximate the function F : if TS denotes the training set and F TS the function learned by a given supervised learning system, the error is computed by estimating the difference between the functions F and F TS .

In order to build a supervised learning system for AS, one idea is to apply an AI technique of supervised learning, using stupid examples. So, following this idea, the issue that must be addressed is how to constitute TSS , the training set of stupid examples. Now, it is assumed that a necessary condition for ( x , y ) to be a stupid example is that ( x , y ) is an error, that is, y ≠ F ( x ) . This raises two questions:

Is the condition y ≠ F ( x ) sufficient to learn stupidity?

First, consider the binary classification (i.e. | S | = 2 ) and let y ‾ denote the element of S that is different from y ∈ S . Then TSS can be built from a training set TS (that is meant to be used by an AI system): TSS would be the set of ( x , y ‾ ) where ( x , y ) ∈ TS . So, the binary classification system will learn a function F TSS that, provided that the learning is successful, would be wrong most of the time, meaning that the function F TSS ‾ : x ∈ P ↦ F TSS ( x ) ‾ ∈ S would give a correct learning function:13

When the size of S grows, things may get different. Let us consider the case of regression, where P = R m and S = R n , a setting used in many numerical machine learning techniques, such as multilayer perceptrons. Given ( x , y ) ∈ TS , how will ( x , y ′ ) ∈ TSS be chosen with the sole condition y ′ ≠ y ? If this is through a one-to-one correspondence φ : y ∈ R n ↦ y ′ ∈ R n , this would amount to a situation similar to the one of binary classification as mentioned above ( φ − 1 ∘ F TSS could be an approximation of F as good as F TS ). If the computation of y ′ is done randomly and regardless of the value of y then F TSS would only be representative of the random generator on S : this falls back to the hypothesis of François Rollin of an AS system that answers any random thing to a given question (as exemplified in the introduction). If the computation of y ′ is done randomly but stays close to y (e.g., for n = 1 , y ′ is chosen randomly in [ y − C , y + C ] ∖ { y } , where C > 0 is a constant that is relatively “small”), this amounts to the framework of learning with noisy examples, with additive noise.

These few abstract examples illustrate the fact that the necessary condition “to be stupid is to be wrong” is not always sufficient to obtain a learning system in AS that is neither a “reverse” AI learning system nor a system that just gives nonsense answers.

SBR would be case-based reasoning (CBR) producing stupid solutions. CBR consists in solving (or trying to solve) a target problem with the help of a case base, where a case is a representation of a problem-solving episode [31,32]. In many applications, a case consists of an ordered pair ( x , y ) and thus coincides with the notion of example in supervised learning and is sometimes associated with additional pieces of information about the relation between x and y . For many applications, a session of CBR consists in selecting in the case base a case that is considered to be similar to the target problem (case retrieval step) and to modify it in order to solve the target problem (case adaptation step).

Two complementary ways to build an SBR system can be considered.

First, stupid supervised learning, as briefly studied in the previous section, can be a source of inspiration for building a base of stupid cases. This meets the recent issue of CBR using positive and negative cases. These negative cases can be compared to (even be assimilated to?) stupid cases. Negative cases can be acquired during the use of a CBR system: these are the cases corresponding to failures of this system. Using positive and negative cases for the learning of adaptation knowledge has improved the results of carrying out this learning with only positive cases [24]. This tends to suggest that if both positive and negative experiences are memorized, with a distinction between the two kinds, this can be useful for improving adaptation in CBR.

A second way to turn CBR into SBR would consist in changing the reasoning steps (as this is the case in the general framework of Section 5.1), for example by selecting from the case base the case that is least similar to the target problem: if the adaptation manages to modify the case-recipe of the gratin dauphinois14

It should be noted that these ideas are in line with the one presented by François Rollin of an AS system that would abstain from sorting (here: making a stupid case retrieval) or adapting (here: making a stupid adaptation).

Although we have already mentioned the issue of reasoning foolishly in Section 5.1, we now turn to the topic of formal argumentation, which is an instantiation of the general question of reasoning. In AI, formal argumentation is a field dealing with the representation of conflicting information and the definition of non-trivial reasoning approaches in presence of this conflicting information. Classically, reasoning is based on notions of logic and rhetoric: an argument is made of a set of reliable pieces of information (the support) and one piece of information that can be deduced from this support (the claim). If an argument a contradicts an argument b, and nothing contradicts a, then we consider that a is acceptable and b is not. Argument-based reasoning approaches are based on various principles, so it seems quite easy to define a system of argument-based stupidity as an argumentation system that would fail to satisfy some of these basic principles.

Let us start with abstract argumentation [12]. In this context, we only use the relations between arguments to determine which ones are acceptable, without any need to focus on the exact nature of the arguments (in particular, their internal logical structure). An abstract argumentation framework is then a directed graph F = ⟨ A , R ⟩ , where the nodes A represent arguments, and the edges R ⊆ A × A represent the attacks, i.e. the notion of “counter-argument”. Reasoning with such a graph is done thanks to the notion of extension, which are sets of jointly acceptable arguments, corresponding to a possible solution of the problem represented by the argumentation graph. There are various ways to define an extension, depending on the properties expected from this set of arguments. Among the classical approaches, we focus here only (as a matter of exemplification) on the stable semantics: a set of arguments E ⊆ A is a stable extension if and only if

it is conflict-free (formally, ∀ a , b ∈ E , ( a , b ) ∉ R );

Charles Lytton stole the jewel! (a)

If we do not only focus on the relations between arguments, but also on the method used to build arguments from structured knowledge (generally using some logical formalism [5]), then there are also several ways to argue stupidly. In summary, there are two steps in the building of structured argumentation frameworks from logical knowledge: the identification of arguments and the identification of attacks towards them. Assume that we are facing someone stating the following sentence: “I like chocolate, so I am the King of England.” This is equivalent to considering that ( { a } , b ) is an argument (where a represents the fact that I love chocolate and b the fact that I am the King of England), even if there is no reason to assume that b is a consequence of a. It is already quite stupid to claim to be a monarch just because one loves chocolate, and this kind of reasoning can obviously lead to more severe problems if argumentation is applied to more sensitive subjects. Even in the case where arguments are “correctly” defined, a way to build a system of argument-based stupidity is to fail the second step, that is, the identification of attacks. Ignoring some attacks or, on the contrary, adding some attacks that should not exist, will obviously change the outcome of the reasoning.

One can reasonably16

All fields of AI could be investigated under the lens of AS. Below is presented a short and non-exhaustive inventory of “from AI to AS” exercises.

For decision-making, the following question could be raised: “How to make a stupid decision?” For example, in a multi-criteria decision, one can decide from an irrelevant criterion: buying a car because it is red when it has been decided to buy it mainly on ecological criteria, on price, and on performance, can appear to be stupid.

For unsupervised learning, a question could be “How to stupidly categorize a domain?” For example, humans can be separated into two categories according to the number of vowels of the first name of the mothers of their fathers. If such categories appear to be completely useless, this would confirm that this unsupervised learning would be stupid.

In the framework of multi-agent systems, a question would be: “How to make stupidity emerge from a group of agents that are individually non-stupid?” In this context, the five fundamental laws of human stupidity proposed by Carlo Cipolla can be pointed out [9]. This theory states, in its third law, that a stupid (human) agent is an agent who harms another agent (or a group of other agents) without benefiting from it or, even worse, by being harmful to herself/himself. Figuratively, this could be seen as someone who shoots her/his own foot while shooting someone else! These laws have given birth to experiments aiming at estimating the validity of Cipolla’s theory in the context of simulations using artificial agents [38], but also to work that aims to give a biophysical interpretation to this theory [29].