Truth versus plausibility

There is no agreed definition for artificial intelligence as a whole, nor generative AI in particular. A useful placeholder could be to say that artificial intelligence is ‘the capability of computer systems or algorithms to imitate intelligent human behavior’ [sic.] (Merriam-Webster 2026). As Melham (Reference Melham, Zou, Poncibò, Ebers and Calo2025, p. 3) points out, one of the virtues of such a definition is that it ‘reminds us that this is technology that generates outputs that look like – that mimic – the intellectual productions of humans, from everyday speech to creative literary works’ (my emphasis). It should be noted that this renders AI a form of mimesis by definition. It is imitative, and the thing it is imitating is human intelligence.

To successfully imitate is, in some way and to some extent, to be able to pass one thing off as something it is not. Plato himself points this out: one cannot imitate a general if one is a general (cf. Republic 394e–395b). This is where plausibility comes into play. It was plausibility on which Alan Turing famously played in his proposal for the Turing Test (Reference Turing1950). In this test, a human subject interacts with a text-based computer system and another human subject, not knowing which is which. Turing’s test hinges on whether a human subject can correctly identify whether she is communicating with a machine; in other words, whether it can plausibly communicate as though human.

In the specific case of generative AI, plausibility is crucial to how the technology works. The language models behind tools such as ChatGPT are probabilistic. They produce unique responses to text inputs by being trained on a large dataset of language. Given a large enough dataset, generative AI can very effectively predict the next word or part-word in a sequence, and hence produce text ‘of such a high quality and apparent meaningfulness that it is often difficult to tell if it has been written by a human or a machine’ (Melham Reference Melham, Zou, Poncibò, Ebers and Calo2025, pp. 3–4, my emphasis).

Bender et al. (Reference Bender, Gebru, McMillan-Major and Shmitchell2021) call on those working in the field to consider the implications of systems that mimic human thinking, using the metaphor of a parrot to describe how generative AI works:

In summary, generative AI employs processes entirely focussed on the plausibility of the tokens of speech, and not at all on the content of what is uttered.

Comparisons with the ‘eristic’ teachers portrayed by Plato are clear. As with a generative AI chat interface, the ‘game’ (as it is described in the Sophist 234a–b) of the rhetoricians and sophists is to sound plausible, with very little regard for the content of what they are saying. Kamtekar (Reference Kamtekar and Fine2011) notes that the focus of the education offered by these figures was the ability to speak persuasively. In the Gorgias, it is commented that ‘Oratory doesn’t need to have any knowledge of the state of [its] subject matters’, as is the case with generative AI, ‘…it only needs to have discovered some device to produce persuasion in order to make itself appear to those who don’t have knowledge that it knows more than those who actually do have it’. (Gorgias 459c, my emphasis). In the Phaedrus, it is underlined that the ‘rhetorician has no need to know the truth…They only care about what is convincing. This is called ‘the likely’, (Phaedrus 272d–e). Socrates distinguishes between ‘…two types of persuasion, one providing conviction without knowledge, the other providing knowledge’, and later concludes that ‘…evidently oratory produces the persuasion that comes from being convinced, and not the persuasion that comes from teaching’ (Gorgias 454e).

Anyone who has played with generative AI tools will have seen that it is liable to invent answers when there is a gap in its dataset. The term ‘hallucination’ has come to be used for this phenomenon, where a generative AI interface fabricates plausible answers to questions posed, such as inventing new Shakespeare texts or claiming emphatically that the word ‘strawberry’ has but two ‘rs’ (cf. Cosma et al. Reference Cosma, Ruseti, Radoi and Dascalu2025). Xu et al. (Reference Xu, Jain and Kankanhalli2024) argue that hallucination is inevitable because of the very way in which generative AI operates. A language model operates in a ‘formal world’, which is not perfectly isomorphic with the ‘real world’. When it formulates an output, its reference point is the ‘formal world’ of the dataset it has, not reality. So, whilst there is a lot of overlap with truth in its outputs, the mechanism is actually about a fit within the ‘formal world’ of its dataset, not any relation to the ‘real world’. The Gorgias sums it up in a way that seems prescient: ‘rhetoric is a producer of persuasion. Its whole business comes to that, and that’s the long and short of it’ (453a, my emphasis).

The example of the spelling of ‘strawberry’ highlights another similarity between the eristic teachers of Socrates’ day and generative AI. In both cases, an utterance (if we can call it that) not only has truth as a relatively low priority but also is not understood by its utterer. In the Meno (71c, 73c, and 76a–b), Meno proudly asserts that his expertise in virtue has come from memorising speeches on the subject. As Scott (Reference Scott2006, p. 13) points out, ‘Meno thinks’, having been a student of Gorgias, ‘that he has learnt to speak well about virtue – not only in the rhetorical sense, but also in the sense that he has actually gained knowledge of what it is’, but it is soon discovered that he is mistaken. He can mimic, but he does not understand. As Irwin notes, ‘the issue with rhetoric is that [its practitioners] do not understand the content they are presenting’ (Reference Irwin and Kraut1992, p. 68), or as the Apology has it, they speak ‘without any understanding of what they say’ (22c).

In the Phaedrus, the conversation moves further. Until now, we have seen that plausibility is prioritised over truth for these rhetoricians, orators, and sophists. Now, we see not just a lack of interest in truth, but a conscious disregard for it, should truth hinder the ability to make a plausible argument. ‘No one in a lawcourt’, it is observed, ‘cares at all about the truth of such matters. They only care about what is convincing’ (272d–e). Truth gets in the way of persuasion, and hence winning the case.

Perhaps the most interesting thing about this observation, though, is that the citizens seem to be in on it. ‘No one…cares’, not just the rhetoricians. This leads us to consider the relationship between the plausibility of their speeches and those apt to find them plausible.

Consensus and algorithmic bias

Plato, of course, had good reason to be suspicious of the opinion of the crowd. In the Crito, Socrates urges his interlocutors not to ‘think so much of what the majority will say about us’, if one is uttering ‘truth itself’ (Crito 48a). However, this exhortation is tinged with danger, as in the same breath, he reminds us that someone might say, ‘the many are able to put us to death’ (ibid.).

Believing the truth is often portrayed in the Platonic corpus in contradistinction to siding with the majority. Truth is truth ‘above all…whether the majority agree or not’ (Crito 49b), but what is plausible, what is likely, is simply ‘what is accepted by the crowd’ (Phaedrus 273a–b). Indeed, ‘what else’ could it be, asks Phaedrus (ibid.).

In a way, this absolves the sophists and rhetoricians of some of the blame. As Kamtekar notes, ‘[t]he sophists are not the source of confusion but are merely reflectors of popular opinion, and that real source of confusion is the opinion of the crowd’ (Reference Kamtekar and Fine2011, p. 352). In the Republic, we hear ‘[n]ot one of those paid private teachers, whom the people call sophists … teaches anything other than the convictions that the majority express when they are gathered together’ (Republic 493a). They sound plausible to the crowd because they reflect back to the crowd their own views, their own prejudices, their own mistakes. Irwin notes that this is an observation made widely in the Platonic corpus: for example, Protagoras’ position is ‘designed to show that the views that appear true to the many are true’ (Reference Irwin and Kraut1992, p. 65) (cf. Theaetetus 167b–c). This means that, although rhetoricians and sophists are not mere fabricators, they ‘do not attempt to found their views on any rational basis that goes beyond the unexamined beliefs and prejudices of the majority’ (Irwin Reference Irwin and Kraut1992, p. 65).

So, Socrates has unveiled a further problem with the sophists and rhetoricians. Yes, they are just trying to sound plausible, but more than this, they are grounding what is plausible in what we might call a bad dataset. The people are unreliable; they are trying to say things reflective of the people, and hence they are unreliable.

Exactly the same issue is the cause of a problematic bias called ‘negative legacy’, which is evident in generative AI. LLMs (large language models) are trained on vast datasets. The plausibility of their utterances comes from their ‘fit’ with this dataset. What this means is that mistakes and biases found within the dataset will be replicated in the outputs created by generative AI.

In the context of education, Hedlund (Reference Hedlund2025) has demonstrated that many generative AI platforms will produce biased results when asked to explain the same concept in science to a male or female student. Similar results have been shown in representation of women of colour (cf. Hauduc et al. Reference Hauduc, Jarmy, Nanji and Hedlund2025) in generative AI outputs, or indeed the failure to represent such groups.

Importantly, when generative AI tools produce biased outputs, such as assuming a professor is white and male, or that a parent is a mother, or that a couple is young and heterosexual, this has more to do with a legacy of public opinion, which has made it into the dataset, rather than a stance the AI is taking.

Mimesis

We have already talked about how the ‘whole business’ (Gorgias 453a) of rhetoric comes down to persuasion, and we have just considered the role of the crowd in forming the notion of what is likely to persuade. What makes the interactions persuasive or plausible, though, is not only about what is said. Just as important is the ability to act ‘as though’ expert. In other words, to imitate expertise. This mimetic act, Plato argues, intrinsically runs counter to any real expertise. It makes it, Plato thinks, logically impossible. As Griswold (Reference Griswold, Zalta and Nodelman2024) notes, ‘[i]mitation is itself something one does, and so one cannot both imitate X (say, generalship) well and also do the activity X in question’ (cf. Republic 394e–395b).

A kind of mimesis is exactly what is going on when we train a chat interface to behave as a tutor (or, for that matter, any other kind of expert). We describe the way in which teachers interact with students; we determine the ‘rules of engagement’.

There are many who are optimistic about AI tutors, seeing a future of high-quality, low-cost, personalised education for all. For example, on 5 February 2025, the founder of Microsoft, Bill Gates, was interviewed for NBC’s The Tonight Show by Jimmy Fallon. In it, he claimed that until now, intelligence has been rare, but that generative AI makes it ‘commonplace’ and ‘basically free’. He predicted that over the next decade, the work of teachers will be replaced with ‘great tutoring’ (The Tonight Show 2025). More recently, the UK Secretary of State for Education, whilst not making the same claims about the replacement of teachers, has also hailed the advent of 1:1 AI tutoring as a way to tackle disadvantage and improve educational opportunities for all (Department for Education and Department for Science, Innovation and Technology 2026).

Taking these comments in light of what Plato says about mimesis and rhetoric provides a helpful lens through which to respond. We have already heard Plato’s observation that when undertaking the work of a general, a general cannot also be pretending to be a general. Mutatis mutandis, we might say that (1) generative AI mimics the interactions of a teacher or tutor, but (2) you cannot both imitate teaching and also be teaching. If it is true that the operation of AI tutors is a mimetic act, where ‘real’ teaching is what is being imitated, then it seems to follow that AI tutors cannot also be really teaching. If AI tutors cannot teach, then the claim that ‘real’ teaching can be so easily replaced must come under serious scrutiny.

However, drawing a clear line between human teachers, who ‘really teach’, and AI tutors, whose interactions are purely mimetic of human teaching, may be more complex than is first imagined. Indeed, defining ‘real teaching’ at all is, as Jackson (Reference Jackson1986) notes, challenging. Activities that encourage learning from others, that might seem imitative, are often crucial to the early stages of teacher training, for example. Trainees typically begin by observing lots of classes before they get to teach, picking up on techniques, routines, and behaviours of the teachers around them. At the next stage, they ‘try on’ some of those behaviours for themselves, imitating, to an extent, what they have seen. Could it be that a discomfiture with an AI interface learning through imitation is purely chauvinistic, and there is no significant difference between how human teachers and how AI ones are trained?

The key difference between what is going on in the case of the trainee, and what is going on with the AI tutor (and the eristic teachers of Socrates’ day), is the inherent emptiness of understanding in the latter. Whilst our trainee teacher is in a sense learning through imitation, Jackson (Reference Jackson1986) would see her as an initiate in a complex network of practices, attitudes, and dispositions that form the context of teaching. Though a novice, she is responding to the reasons of others for certain practices, and can give her own reasons for what she assimilates into her own teaching, and what she herself is trying to achieve.

Teaching, for Jackson, must be understood not as a set of practices or activities that one can understand from the outside by looking purely at a set of behaviours. Rather, they are wired into a context where a community has an understanding of the endeavour and, crucially, the reason we do certain things. As Jackson points out, observing a nursery school teacher bending at the knees and bringing her whole torso down to the level of the student is manifestly better than someone craning over them, but it is only manifest to those understanding enough of the context of values and intentions that make up what teaching is about. The kind of imitation that trainee teachers take part in during the early stages of initiation is not, therefore, empty in the way Plato depicts the imitation of the eristic ‘teachers’, who employ imitation as ‘some device to produce persuasion in order to make itself appear to those who don’t have knowledge that it knows more than those who actually do have it’ (Gorgias 459c). In the Sophist, we hear that ‘imitation is a sort of production, but of copies and not of the things themselves’ (Sophist 265a–b). As Socrates addresses Ion, ‘you turn out to be the representative of representatives’ (Ion 535a). The only expertise that these figures have is to imitate, and this is not the expertise the trainee teacher acquires.

The mimesis an orator or rhetorician enters into finds its place, of course, within a much wider context in Platonic thought to do with truth, imitation, art, appearance, and reality. As Irwin notes, ‘Plato equates what they do with appearances or images’ (Reference Irwin and Kraut1992, p. 65, cf. Republic 515a5–6; Sophist 232a–236d). For Plato, as Asmis notes, ‘language is used correctly when it is used for true communication’ (Reference Asmis and Kraut1992, p. 360). If it is true that one cannot both be X and imitate X, in the sense we have discussed, then the orator is not engaging in true communication. It is clear that this incorrect use of language carries with it a moral burden. Socrates describes the place of oratory in politics as ‘a shameful thing’ (Gorgias 463d). In the Sophist, Socrates concludes that ‘…we have to regard him as a cheat and an imitator’ (Sophist 235a). If we were to apply this thinking back to the case of the AI tutor, the focus on the plausibility of utterances and not their truth, as well as the mimetic nature of the interactions, might lead us to claim that these cannot be conversations, since they do not aim for ‘true communication’(Asmis, ibid.).

In Gates’ comments, it is taken as read that generative AI can be a teacher/tutor. The fact that the mechanism for how an AI tutor works is purely imitative is ignored. As far as Gates’ comments go, AI is not acting ‘as’ a teacher or tutor, it is seen as a teacher or tutor in its own right. As Griswold notes, ‘[i]t is as though the fictionality of the persona is forgotten’ (Reference Griswold, Zalta and Nodelman2024).

‘Psychic dangers’

For Socrates, the nature of the interaction between an eristic teacher and a student (and by implication, an AI tutor and a student) renders it harmful almost by necessity, given imitating X entails not being X, and given that communication should be used for truth-telling. There are, however, additional dangers of engaging sophists, orators, and rhetoricians mentioned in the Platonic corpus that might shine further light on the employment of AI tutors.

It is not merely the eristic teachers’ lack of interest in truth that is of concern, but the influence that training students in a particular way of using language could have on the soul. In the Protagoras, learning influences who you become. ‘You cannot carry teachings away in a separate container. You put down your money and take either damage to your soul by having learned it, or you go either helped or injured’ (314b). As Kamtekar notes, ‘our poor epistemic condition makes us dependent on persuasion’ (Reference Kamtekar and Fine2011, p. 356). This makes the young very susceptible to the influence of others. As Kamtekar further notes, ‘[i]t is dangerous to study with a sophist, not just because you might be throwing away your money but because you might end up with a damaged soul’ (p. 351).

In the Republic (605a–b), the eristic teachers are said to strengthen the non-rational, appetitive part of the soul: ‘he arouses, nourishes, and strengthens this part of the soul and so destroys the rational one’ (605b). In the Phaedo, we hear about how this cultivates an indifference to truth. Those who learn with these ‘teachers’ come to believe ‘that there is no soundness or reliability in any object or in any argument’ (90c).

Education, on the other hand, is not like this. It cultivates the best part in us, our rational capacity, and overcomes the indifference to truth. As Kamtekar (Reference Kamtekar and Fine2011, p. 357) notes, ‘Reason’s powers are not content-neutral’. More than this, education is holistic. It does not simply shape what we know, but who we are. In Book 7 of the Republic, as Storey (Reference Storey2022) notes, education is depicted (1) as a turning of the soul and (2) as the turning of the whole soul.

The question of whether the use of an AI tutor corrupts young and hence vulnerable learners is what we philosophers like to call ‘an empirical matter’, and this work is still relatively nascent. The work of Kosmyna et al. (Reference Kosmyna, Hauptmann, Yuan, Situ, Liao, Beresnitzky, Braunstein and Maes2025) has given certain early indications that ‘convenience [comes] at a cognitive cost’ (p. 153) when ChatGPT is used in the context of writing an essay. More recently, a report from the OECD (Li and Hu Reference Li and Hu2026) highlighted that when using AI tutors, ‘current evidence points to a high risk of cognitive offloading if “guardrails” are absent’ (p. 78).

Despite the fact that the empirical work is in its early stages, there are still some theoretical questions that it is helpful to address. First, there is a question of the attitude of an AI tutor to truth. We have said that generative AI prioritises plausibility, replicates mistakes in its dataset, and is liable to hallucinate false, but plausible responses. One key test is whether it is ‘content neutral’, to borrow Kamtekar’s happy phrase, and hence indifferent to the truth of what it is teaching. It should be noted that the AI tutor shouldn’t be wholly indifferent to the truth (or at least to the correctness) of what it is teaching, since an AI tutor that is liable to hallucinate will not be a success, if those mistakes are picked up. It will therefore be in its interest to get things right. For Socrates, this falls short of real truth-telling, since its utterances are not made in the belief they are true and for the sake of being true, but for the sake of being successful, which really only sounds like an extension of the aim to be plausible.

Second, if AI is indifferent to truth, or at least only cares about the success of its interactions rather than the truth of what it says, there follows a question about the impact of this for young people. Teaching involves acting in an epistemically responsible way, and inducting students into epistemically responsible practices (Jarmy Reference Jarmy2025). In using AI, do students, as Socrates fears in the Phaedo, become indifferent to truth? If your teacher doesn’t know or care whether something is true, why should you? Whilst this would need to be established empirically, there is potentially an interesting area to unpick here theoretically, perhaps using Anscombe’s notion of ‘believing’ someone (cf. Anscombe Reference Anscombe and Delaney1979; Wanderer Reference Wanderer2013), and what changes about this when engaging with an AI tutor.

Third, and relatedly, there is the question of whether the mimetic nature of the interaction (the fact that the tutor is imitating a human-to-human conversation) is damaging to students’ attitudes towards truth, learning, and the value of their education. Not only does the AI imitate how a teacher interacts, but the student becomes a participant in an exchange which is mimetic of conversation.

Fourth, and finally, there is the question of the holistic model of education that Plato outlines. When Bill Gates talks about ‘great tutoring’ (The Tonight Show 2025), it seems that this might refer to quite a narrow part of what happens in teaching, and education more broadly. An AI tutor could conceivably be good at helping students through problems, analysing their answers, and offering feedback in a way that prepares them for certain kinds of assessments. This, however, seems far removed from the turning of the ‘whole soul’ as described in the Republic, the making of who you are. It is clearly far beyond the remit of this paper to adjudicate on which, if either, is a better vision of education, but we can note for now that there is a big gap here in what is being meant by ‘education’.

These questions are weighty, and cut to the heart of what we think we are doing when we educate. The vulnerability of the young, their susceptibility to negative influences, and the fact that, as Mary Donaldson so memorably put it, ‘they are conscripts’ (Reference Donaldson1979, p. 5), with no choice about whether schooling is something they do, makes this particularly important. It recalls Socrates’ exhortation in the Protagoras to be cautious. ‘Do you see what kind of danger you are about to put your soul in?… Please don’t risk what is most dear to you on the roll of a dice, for there is far greater risk in buying teachings than in buying food’ (313a–314a).

Hype and economic pull

Buying is, of course, an apposite verb, here. One of Plato’s concerns was an economic one: that the eristic ‘teachers’ charged their impressionable students top dollar for their services, whilst Socrates, the true educator, would never take money for what he did. A similar point has been made about the adoption of AI-driven products in education, which has been referred to as a ‘gold rush’ (Educate Ventures 2024), with companies all hurrying to cash in.

The hype and concomitant economic pull that AI has generated in education in recent years recalls Socrates’ warning to his young interlocutor at the beginning of the Protagoras.

Schools can feel pressure to adopt AI solutions quickly and uncritically, fearful of being left behind other schools, and seen to be behind the times. This is no different to what goes on when a new rhetorician appears in Athens, as Kamtekar notes, ‘The merchandise is potentially dangerous and the eager buyers are but poor judges of the value of what they are getting’ (Reference Kamtekar and Fine2011, p. 351).

In the Platonic corpus, the indifference to truth that eristic teachers have is seen to spill over into their economic behaviour. In the Protagoras, for example, we are exhorted to beware ‘or the sophist might deceive us in advertising what he sells, the way merchants who market food for the body do’ (313d).

AI education products will increasingly be big business, and not just for companies that have always specialised in education. Microsoft, Google, and OpenAI are amongst the large global companies now vying for business in this area. The comparisons with ancient Athens are clear. As Griswold (Reference Griswold, Zalta and Nodelman2024) notes, ‘the… rhetoricians [sell] their products to as large a market as possible, in the hope of gaining repute and influence’.