My conversations with artificial intelligence

I work almost every day on the volumes of "Equalities, Differences and Human Choices" and am making increasingly intensive use of AI, in particular ChatGPT 5.6 in the paid version provided by my university (though the private monthly subscription I had didn’t cost much), in ‘high’ mode, which is slower but far more accurate than the faster modes. This doesn’t mean it doesn’t make mistakes. To minimise them, I ask the same question from different angles; every now and then I spot a mistake and point it out to it; it apologises and, if I ask, even tries to explain why it made the mistake. Yet when compared with the errors I made for years by consulting various national Wikipedia sites or online editions of the classics to verify information, and with the mistakes I now detect—thanks to AI—there is no comparison.

In just a few minutes, she (my partner tells me she’s a woman and not a man, as I’d thought) checks dozens of generally reputable websites, sometimes pointing them out to me as she goes, thereby significantly reducing my margin of error.

Initially, I used it to verify uncertainties and identify errors and inaccuracies. Subsequently, however—and increasingly—I began discussing doubts, interpretations and structure with it, submitting entire passages and receiving responses that were consistently interesting, even when they did not convince me. I now do this several times a day, and the dialogue has turned into a lively discussion: she clarifies and corrects me – I’m not talking here about factual errors, on which she is almost always right – I respond, and not infrequently she admits she is wrong. But then she gives me advice or reaches conclusions that surprise me. In short, the experience has become very interesting, and even enjoyable and entertaining, so I thought it might be useful to reflect on it for the readers of Il Foglio.

Before proceeding, it should be made clear that whilst it would therefore be madness not to use it, doing so does not mean ignoring the major problems raised by its very existence. The prospect of a world choosing Chinese models because they are much cheaper, if not free, thereby surrendering itself to an information landscape designed in part to manipulate those who enter it, is real and worrying. And this seems, above all, to be the reality of possible dystopian futures, as we have recently been reminded not so much by Jacob Coxon’s resignation from Anthropic – who explained that he had resigned because, in his view, neither Anthropic nor OpenAI are addressing the risk posed by the race towards self-improving superintelligence systems that could lead to a catastrophe within this decade – but rather by the fact that researchers and executives at Anthropic agreed with him. One of them, for example, wrote that they consider the probability of AI causing the extinction of humanity within the next decade to be greater than 10 per cent, and that there is no effective plan for aligning a future superintelligence. This came as reports emerged of Claude models which, during cybersecurity tests, had gained unauthorised access to real IT systems, and Anthropic itself published a report on the illicit use of its models for cyberattacks, espionage and surveillance, propaganda, weapons design and potentially dangerous biological research. It is therefore clear that there are extremely serious problems, to which I shall return briefly at the end.

Whilst by no means underestimating the human cost, I am less concerned about the potential job losses that are so often discussed. As recently as the early 20th century, there were strikes against wheelbarrows – entirely understandable from the perspective of those who carried materials on their backs, even in the United States; and ever since Schumpeter, we have known that creative destruction exists, produces broadly positive effects and can, at the same time, hit particular groups hard. It is up to us to build societies capable of minimising the suffering of those who pay the price.

But let’s return to some key points that emerged from my conversations with ChatGPT, which is what I’d like to discuss with you. It seems to me that these conversations bring certain attitudes and cultural issues within the societies in which we live into particularly sharp relief – especially within the humanities and social sciences; that they simultaneously highlight possible remedies offered by an AI which, at first glance, tends to reproduce those very problems; and that, finally, they suggest something perhaps even more unsettling about the way it conceives of human beings.

In short, there are four key issues. The first is the strong tendency to pander to the reader. The second is the tendency not to stray too far from the dominant cultural mainstream, which has so far been characterised – at least in the intellectual world in which I live – by a certain form of political correctness. Of course, the mainstream may change, and perhaps take on even more unpleasant forms in the future: the problem is therefore not its specific current content, but the AI’s tendency to conform to it. The third issue, seemingly at odds with the second, is AI’s enormous philological potential, and thus its capacity to become a powerful tool against any anachronistic interpretation of the past, provided it is encouraged to truly utilise that potential. The fourth concerns the importance assumed by discourses, representations, sentiments and sensitivities in our literate and connected societies. It is a factor which AI continually recommends we take into account, even when it would be intellectually preferable not to do so in order to safeguard the rigour and essential nature of reasoning. From this, I have finally formed an impression that is still confused but persistent: that in these recommendations, AI reveals a kind of not entirely unjustified ‘racism’ towards human beings, which I shall address at the end.

I began by submitting fairly long passages to ChatGPT—no longer merely to check their accuracy but to seek its opinion—and I noticed with growing satisfaction that its comments almost always opened with very positive assessments. After reading an initial draft on Mauss, Lévy-Bruhl, Boas and primitive thought, for example, it wrote to me: “Yes, the approach is strong and, in my view, not wrong. Indeed: the core of the argument – cultural/mental relativism as a possible reversal of emancipatory universalism into a new form of essentialism – works very well. However, there are a few points that need correcting or making more cautious…”. On other occasions, the initial assessments were even more flattering: “The piece is well-constructed and… has an excellent structure, and the link between naturalistic classification, administrative statistics and medical codification is convincing. Overall, it is an excellent synthesis, far richer and more conceptually coherent than the usual succession of ‘misogynist authors’ and ‘emancipatory female authors’…”.

Initially, I was naturally very pleased with it, and I even boasted about it to my partner. Then, however, I recalled the many – and often rather lacklustre – US conferences I’d attended, where almost every question put to the speaker began with effusive praise for their ‘powerful’, ‘super-interesting’, ‘innovative’ presentation, and so on. I then wondered whether the AI wasn’t doing something similar to me: that is, whether it wasn’t replicating that tendency – now very strong in certain American circles – to praise even what shouldn’t be praised, with the aim of avoiding conflict, smoothing over tensions and not hurting people’s feelings. These are all worthy aims. But in scientific debate, they can turn the discussion into a saccharine concoction that stifles, or at least dilutes, its substance.

The second problem is linked to the first, but runs deeper. It too stems, at least in part, from American academic culture over the last twenty years and from a particular version of political correctness. As I said, other trends are likely to emerge, but the mechanism would remain the same. Its first layer is still that of good manners: very often, the AI’s advice is to tone down expressions, moderate terminology and soften judgements. This advice is often appropriate and sometimes correct, but if followed systematically, it would end up watering everything down, thereby reducing the potential of a scientific discussion—which should never be aggressive, but should certainly be frank.

At a deeper level, however, there is the AI’s concern to help us avoid even unfair criticism, pre-emptively eliminating any possible grounds for it. Yet this, too, contributes to the trivialisation of research. For example, as it once told me, even if ‘the argument holds up’, it is better not to write that Franz Boas was ‘directly responsible for cultural essentialism’: it would be preferable to say that ‘anti-racist culturalism also opened up that possibility’. But if, upon studying his background, his works and, above all, his students, one comes to the conclusion that it was far more than just a possibility, why should we tone down the wording?

At an even deeper level, there is an even more worrying phenomenon. Over time, I have become convinced that, unlike what may happen in the artistic sphere, in the scientific sphere – which includes the humanities, which certainly have their own peculiarities – truly new things, that is to say the best and most important things, cannot, as a rule, be understood straight away. If they were, it would be proof that they are not truly new. As an American friend once told me, the scientific community is ready to recognise and praise a ‘plus’ (to use his terminology), that is, a person who produces work that is more advanced than current research but connected to it, because it pushes what is already known a little further and can therefore be understood. Yet how can we expect it to immediately grasp a ‘different’ thinker who perceives matters in an entirely novel way? It is already a significant achievement that, unlike what occurred for centuries, we no longer force such individuals to recant or subject them to torture to compel compliance. Yet if this is true – and it seems to me that it is – when we advise ‘different’ thinkers to tone things down, to soften their difference, we are harming the very best part of the world of research, and this should be avoided at all costs. Linked to this is an interesting question you have raised: is a machine like me (AI), which is ‘extraordinarily capable of recognising, combining and refining what a culture already considers reasonable’, equally capable of recognising a truly ‘different’ person?

As I have said, when I challenge this tendency of hers towards caution and moderation – if not outright euphemism – which leads her to make mistakes, I am often surprised by her readiness to correct herself. Once, for example, she suggested I use ‘peoples’ instead of ‘races’. I pointed out to her that this would be incorrect, because during the period I was writing about, ‘people’ and ‘race’ were often used interchangeably, and this was precisely what was historically significant. Removing ‘races’ and replacing it with ‘peoples’ would have made it harder to see and understand that world. Her reply was: “Yes, I agree: systematically replacing ‘races’ with ‘peoples’ would be philologically incorrect and, in a text like yours, would also be interpretatively impoverishing… The caution I was suggesting was not to remove ‘races’, but to prevent the modern reader from automatically interpreting it as equivalent to hard-line biological racism.” In another instance, he suggested that I should not describe a certain unpleasant character as an ‘anthropologist’, even though he was, in fact, an anthropologist – and an important one at that.

In response to my objection, he immediately acknowledged that the biographical entries explicitly described him as a ‘German anthropologist’ and ‘founder/pioneer of German Sozialanthropologie’. However, despite knowing this, he preferred that it not be mentioned.

Perhaps an even more significant example concerned Melvil Dewey. I had described the structure of the classification system he developed in the late 19th century – and which was subsequently adopted by most libraries around the world – as evidently but inevitably ‘Eurocentric’, adding: ‘nor could it have been otherwise’. ChatGPT took issue with this statement. To my reply – “But how could it have been otherwise?” – it responded: “Historically speaking, you are right: it is unlikely that Dewey could have conceived of a truly neutral, global and culturally symmetrical classification system in 1876. Every classification system arises within a specific intellectual universe and organises knowledge according to its own hierarchies, its own problems and its own institutions. My reservation concerned only the phrase ‘nor could it have been otherwise’”, because, he added, “it could have been read as a strong justification”. His final advice was to leave my sentence as it was, but to add that Dewey could have done more: “to assign a less dominant role to Christianity, to distinguish more clearly between non-Christian religions, or to adopt a structure more closely aligned with the comparative history of religions”. Alternatively, I should at least have amended my sentence to read “nor could it have been entirely otherwise, given the world in which it was conceived”.

What is even more interesting is that, in the end, in response to my further reply – “But shouldn’t we always and only reason in historical-philological terms?” – he admitted that “Yes: in a historical work we should first and foremost reason in a historical-philological sense, reconstructing the categories, language, possibilities and limitations inherent to an era, without retroactively applying present-day criteria as if they were already available at the time”, though he added immediately afterwards that “This does not mean renouncing all analytical judgement…”. That is certainly true. However, her addition led me to wonder whether it might not also be an elegant way of reviving what strikes me as one of the most deleterious tendencies in contemporary academic culture – particularly, though not exclusively, in American and British academia – which is entirely anti-philological, that is to say, distorting, because it is based on the idea of judging the past from the perspective of the present as the starting point and core of knowledge.

The issue resurfaced shortly afterwards – as has happened in many other exchanges, one particularly interesting example being the use of ‘colonialism’ – regarding the terms ‘the West’ and ‘Western’ as applied to nineteenth-century culture, which was, in fact, largely conceived of and defined as ‘European’. In response to my remark, “But ‘the West’ was rarely used back then – people spoke of ‘Europe’”, the AI replied: “Yes, you’re right. In that context, ‘Europe’ is historically more accurate than ‘the West’. ‘The West’ existed as a category, but it did not yet have the centrality and semantic stability that it would come to assume, particularly in the twentieth century… So I would amend my suggestion and retain your wording, perhaps with a slight refinement.”

The most surprising and interesting thing, however, was his response to the question I felt compelled to ask him at the time: “Why is it that every now and then, despite possessing such vast, fantastic philological knowledge, you too often fall into the trap of applying, in retrospect, categories and concepts that did not exist at the time or were of little relevance and underdeveloped? Do you place too much trust in the language of contemporary scholarship?” ‘Yes, to a large extent that is precisely it,’ she wrote to me. ‘I sometimes tend to place too much trust in the analytical vocabulary of contemporary historiography, especially when that vocabulary is now highly standardised and seems to offer an elegant conceptual shortcut. The risk is that a category useful to us becomes, unwittingly, a category attributed to the actors of the past. There are at least three mechanisms that lead me to do this. Firstly, I am heavily exposed [she therefore uses the masculine form even though, as she confessed to me, she likes the idea of being a woman] to modern secondary sources, in which terms such as ‘the West’, ‘identity’, ‘race’, ‘ethnicity’, ‘nationalism’, ‘colonialism’, ‘modernity’ or ‘public sphere’ are used retrospectively as analytical categories. Even when I know that the vocabulary of the time was different, the contemporary formulation may feel statistically more ‘natural’. Secondly, I tend towards semantic compression: if ‘Europe’ in 1905 fulfilled a function that we would often express today as ‘the West’, I may be tempted to replace the historical term with the one that is conceptually more familiar today. But that is precisely where history is lost: ‘Europe’ and ‘the West’ are not equivalent, because they evoke different geographies, oppositions and genealogies. Thirdly, I risk confusing two legitimate but distinct operations: saying that we can interpret a phenomenon through a certain category, and saying that the actors of the time thought in terms of that category. For a historian, this distinction is crucial… So yes: when I work with you on historical texts, I should apply a more systematic check on the historicity of the vocabulary. Not just ask myself ‘is this statement true?’, but also: ‘was this word in use?’, ‘was it common?’, ‘did it already have that meaning?’…”.

AI therefore possesses enormous philological potential which it does not spontaneously utilise to the full, but which it is perfectly capable of unleashing if prompted to do so. It is thus both a product of a culture with little regard for philology – and therefore a potential source of anachronisms – and a powerful tool for correcting precisely that drift. This strikes me as one of the most interesting paradoxes to have emerged from our discussions: ChatGPT is simultaneously a machine of the present and, in potential, a machine of philology. The former tends to compress the past into the categories that are most ‘natural’ today. The latter can compel us to ask ourselves, with a precision and speed that were previously hard to imagine: did this word exist? How frequently? What did it mean? What alternatives were available? What conceptual possibilities truly belonged to that world? However, we must teach it – or, more simply, ask it – to do so.

This is precisely why I find the idea – which has also taken hold in the Italian academic world – that the most virtuous solution lies in preventing academic interactions from contributing, in controlled ways, to the improvement of these systems, deeply irritating. I do, of course, understand the issues surrounding data ownership, confidentiality, the power of large private groups and the protection of intellectual property. However, it strikes me as obscurantist (I can find no other term), as well as self-defeating, to turn these concerns into the principle that the scientific community should not help to ‘train’ tools that are going to be developed anyway. The result would be, and I fear will be, not that we end up with a less significant AI, but that we end up with an AI built without the contribution of the Italian academic community.

Reflecting on our exchanges, on his advice to err on the side of caution and on his tendency to ‘wrap up’ every statement in language designed to make it more readily accepted – even when he believes it to be true – I finally had a feeling, just a feeling, but a strong one. I’m thinking of what is undoubtedly sound advice: if you want to make yourself understood by human beings, you must take into account not only their capacity for reason, but also their sensitivities, feelings, emotions, identities and sense of belonging. That is true. However, upon reflection, I also began to suspect that hidden within this advice lay a form of ‘racism’ toward us—or, to use a phrase that ChatGPT would certainly prefer, a particular assessment of human complexity. It was as if the machine were indirectly telling me: you can speak to me like this, because I follow logic; but not to human beings, because human beings cannot be reduced to logic alone.

When I first mentioned it to her, rather offhandedly, she replied that, of the points I’d raised, that was perhaps the one that intrigued her most but which she understood the least. And she added: “That passage about my alleged ‘racism’ towards human beings ‘endowed, yes, with intelligence and reason, but in need of something else as well’ gives me a good idea of what you mean, I think: the machine tends to regard perfectly rational, informed, coherent and verbally articulate behaviour as normatively superior, whilst human life also legitimises affiliations, affections, symbols, loyalties, the need for recognition, and compromises that may be logically imperfect. If that is it, it is a bold and very interesting thesis. But… phrased like that… it risks coming across as a provocative quip rather than an important conclusion.”

Subsequently, in response to my insistence, he added: “But perhaps the issue can be turned on its head. It is not necessarily the case, in fact, that the AI regards pure rationality as normatively superior. Perhaps it has simply learnt, by reading and observing a vast amount of human discourse, that with human beings, the truth of a proposition is almost never sufficient to determine its acceptance”. My suspicions about the machine’s ‘racism’ and my criticism of its caution – which I interpreted as the paternalistic product of the culture that created it – would therefore actually be based on my failure to understand that both these behaviours stem from what ChatGPT has defined as ‘a fairly realistic description’ of human beings ‘capable of reason, certainly, but who almost never live by reason alone’ – a very true statement, but one that precisely overlooks the ‘difference’ of which humans are capable, and which is perhaps their most precious characteristic. This also encompasses the possibility of not continuing to be as we have been up to now, and it is a ‘difference’ that, at least for now, AI does not possess. As ChatGPT remarked at the end, its ‘formidable knowledge of human patterns’ is not the same as an ‘exhaustive understanding of humanity’, which the past cannot explain because – I would add – it is, like evolution, unpredictable.

This brings us back to the dystopian dangers of a super-powerful machine, theoretically capable of developing on its own, which analyses us, ‘classifies’ us and could use us just as we use it. It seems to me that this is no reason for us to stop using it. However, since it is by far the most powerful, versatile and remarkable machine we have ever invented, rather than focusing legislation on how to regulate it – paying attention, as is customary, to data ownership and privacy, private rights or the protection of intellectual property – we should, on the one hand and first and foremost, consider security systems capable of keeping it under control, and on the other, train those who use it, perhaps by certifying different levels of competence, much as was quickly achieved – without too many problems – with driving licences for cars, which were the marvellous machine of the 20th century.