On LLM Bullshit

As someone who not too long ago published a paper extending Harry Frankfurt’s concept of “bullshit”, and who is known to think about AI, I’ve been asked to review several papers discussing the claim that the outputs of Large Language Models are all bullshit, and have read several of the others that have appeared in the literature, and I thought I’d write a bit about them here.

The start of this literature seems to be “ChatGPT is Bullshit”, by Michael Hicks, James Humphries, and Joe Slater, published in Ethics and Information Technology, from June, 2024. As far as I can tell, everything written on the topic cites this one. Its central point is the obvious one - LLMs don’t have intentions, and thus they don’t have the intention to speak truly or falsely, so their assertions qualify as “bullshit” on Frankfurt’s account. On this basis, they also dismiss the use of the terms “hallucination” and “confabulation” for certain kinds of errors, saying all these concepts require a certain kind of anthropomorphism. However, they call the bullshit derived from mere lack (and impossibility) of intentions a kind of “soft bullshit”, and suggest that there may be a case that LLMs in fact generate “hard bullshit”, where there is an intention to deceive the audience about the nature of the interaction. The intention is not that of the LLM, but one the audience might read in, or perhaps one from the designer of the system.

However, the one paper in this literature I would most recommend people to read is “Cut the crap: A critical response to ‘ChatGPT is Bullshit’”, by David Gunkel and Simon Coghlan, published in Ethics and Information Technology in April, 2025. They begin with a detailed literature review of the various blog posts, news articles, and non-philosophical books that were already drawing the Frankfurt connection within days of ChatGPT being released (a blog post by Narayanan and Kapoor, who later wrote the influential essay on “AI as Normal Technology”) and in one case, even in response to GPT-3, two years before ChatGPT! But the more important part of the paper considers carefully the ways in which anthropomorphisms like “fabricate”, “hallucinate”, “confabulate”, and “bullshit” might sometimes be useful for describing the behavior of systems as different from humans as LLMs are. They note that Hicks et al. surely go too far, and in somewhat contradictory ways, when they insist that no anthropomorphic terminology at all is useful for describing the behavior of these systems, and also want to call these systems “bullshit machines”. (Gunkel has a September 2025 paper proposing it is better to understand them as “différance engines”, making a provocative pun on both Derrida and Babbage.)

Sarah Fisher also has a useful response, “Large Language Models and their Big Bullshit Potential”, in Ethics and Information Technology, in October, 2024. She points out that there surely is a way in which LLMs generate text that seems truth-apt, but in ways that are relatively unconnected from truth, that is usefully characterized as “bullshit”. But she notes that there are a variety of ways in which LLMs can avoid this sort of bullshit, either by implementing fact-checkers scrutinizing outputs that are likely to be taken as fact and blocking claims that seem too far from published responses, or by routing certain kinds of requests to other mechanisms, or to retrieval-augmented generation, or just refusing to answer certain kinds of requests. She notes that some of this was already going on by late 2023, and I note that a lot more started happening around the time her paper came out, when OpenAI released o1, their first “reasoning model” (trained partly by reinforcement in using its output to talk to itself in solving verifiable problems, in addition to just outputting the result of predicting the next word), using techniques that are now standard in all major LLMs (though the free versions sometimes economize and avoid using these abilities).

Daniel Tigard, in “On Bullshit, Large Language Models, and the Need to Curb Your Enthusiasm”, in AI and Ethics, in May 2025, argues in a different direction. He argues against Hicks et al.’s “soft bullshit” idea by noting that it means LLMs are so devoid of intention that not only do they fail to intend to speak the truth, but they fail to have any of the intentions characteristic of bullshiters. He asks us to instead consider the question of whether it is useful to call bullshit on certain LLM interactions, and concludes that if the interaction is useful (as he argues some factual queries and even mental health therapy sessions can be) then calling bullshit doesn’t help.

A similar response is offered by Jesse Fitts in “ChatGPT is Not Bullshit, Nor is it Not Not Bullshit”, in Ethics and Information Technology, in May 2026. But Fitts has much more discussion of how the use of language by LLMs interestingly parallels the use of language by children who haven’t yet developed full theory of mind (and thus neither mean things nor bullshit, despite using language that has meaning), and ends with a more negative assessment of the value of LLM outputs. (I think this direction Tigard and Fitts take is the wrong sort of direction to go - I think we really should pay attention to the things tying assertions to their usefulness or truth, even if we don’t think these qualify as “intentions”.)

Jimmy Alfonso Licon, “ChatGPT is Bullshit (Partly) Because People are Bullshitters”, in Philosophy & Technology, May 2025, argues that Hicks et al. miss an important element of the ways in which LLMs produce bullshit. In particular, he points out that through various kinds of social desirability bias, people tend to bullshit each other (and ourselves) when producing explanations of why we donated to charity or why someone broke up with our friend, and also in lots of other interpersonal interactions. Even if we included a fact-checker in an LLM, to eliminate the kinds of bullshit that Hicks et al. focus on, LLMs would still produce lots of bullshit in these other situations, if they faithfully copy human text.

But Humphries, Hicks, and Slater reply, in “LLMs Bullshit by Design: A Reply to Licon”, in Philosophy & Technology, May 2026, that Licon’s classification would end up saying that way too much human text is bullshit, and they want to preserve an important role for the LLM structure in the production of bullshit. (I definitely side with Licon in this debate - humans produce a lot of bullshit, in ways that are important to maintaining relationships.) Incidentally, Philosophy & Technology has published a few other reply papers to Licon, one in October 2025, and another in December 2025, but I won’t even try to summarize them because they are almost unreadable.

Robert Sparrow and Gene Flenady, in “Bullshit Universities: The Future of Automated Education”, published in AI & Society in April, 2025, use the idea that LLMs always bullshit, and perhaps can’t even truly “assert” anything, to argue against the idea of LLMs replacing university instructors. This paper makes a lot of important points about the role of education - that a lot of education is about learning skills rather than facts, that humans often find it harder to check machine outputs than to create correct outputs themselves, that part of the role of the instructor is to be a role model for the student. These points probably even point towards a stronger conclusion than the one they officially make - there are specific parts of the job of the university instructor that probably shouldn’t be turned over.

But I’m not sure why any of these conclusions need the strong assumption that AI can’t be moral agents, and thus can’t act and can’t testify, and thus can’t even use words meaningfully, because they can’t be held responsible for anything. And in fact, there is a response paper by William Tuckwell, Daniel Cohen, and Morgan Luck, “In Defense of Bullshit in Universities”, published in AI & Society in November, 2025, that uses my work on bullshit to argue that the mere fact that something is bullshit doesn’t entail that it’s bad or ineffective (though there might be more specific problems for specific uses).

Interestingly, Sparrow and Flenady make their argument that LLMs can’t mean or assert anything on the basis of a version of Robert Brandom’s inferentialist theory of meaning. One of the best papers I read recently is Ryan Simonelli’s “Sapience without Sentience: An Inferentialist Approach to LLMs”, which just came out in the Asian Journal of Philosophy, and which argues that specifically on Brandom’s inferentialist theory of meaning, LLMs can mean and understand things, and even be responsible for them in relevant ways - but that this adds up to a strange new type of creature that is sapient without being sentient. (He also argues this is probably for the best, because many of the biggest risks of AI come with sentience, in the sense of having actual goals and desires.)

Another interesting paper on LLM bullshit is “Chatbot Apologies: Beyond Bullshit”, by PD Magnus, Alessandra Buccella, and Jason D’Cruz, in AI and Ethics, in July 2025. They argue that separately from any thought about assertion, apology requires entering into particular social relationships, and that LLMs can’t do that, and thus can’t offer any more than a “rote apology”. However, they seem to be working with older models. They say:

One might hope that more powerful AI could use output phrased as apologies to convey information to the user. For example, if a chatbot were to apologize more profusely when a mistake is more grave, or if it were to rebuff promptings to apologize when it had not made a mistake, these behaviors could convey a world-oriented element to the user. They would let the user learn about the moral valence and gravity of the situation. For chatbots that do not have these capacities, though, pretending that their apologies are sincere speech acts focuses our attention on the fantasy about chatbots rather than on features of the world we care about.

I regularly have chatbots that do these things in interactions with me! (And they also don’t quote the phrase “that’s on me” that seems to me to be ubiquitous in LLM apologies.) I think they’re getting at important issues, but I’m not sure which ones are beyond current LLMs.

One last paper I’ll mention discusses LLMs in academic publishing. In “Publishing Robots”, by Nicholas Hadsell, Rich Eva, and Kyle Huitt (in Inquiry, February 2025), the authors argue that philosophy journals should be willing to publish “excellent papers” fully written by LLMs (though they argue for a “sponsorship model” to help journals deal with potential backlogs of huge numbers of LLM-written papers). I agree with basically all the points they make about the value of publishing (human or otherwise) and about potential costs of allowing AI publication. But I think they’re a bit too simplistic in thinking about authorship as either human or AI, rather than mixed. (And from my experience receiving largely AI-written submissions as an editor at journals, they’re also skipping over the part where people use AI inexpertly to produce lots of almost-good papers that definitely aren’t good.) They also are too optimistic about the “sponsorship model” properly addressing issues about authorial responsibility, and flooding the journals. They have a section discussing the objection that LLM-authored works are bullshit, but they point out that lots of publishing is already bullshit (many people are doing it just to get tenure!)

I think there’s a lot of interesting ideas coming out in these papers overall, but I suspect that there will be more interest in two different directions. One is to think more about what it takes for LLM text to have meaning or intentions behind it at all, and the other is to think even more surface-level consequentially about what makes the outputs good or bad independent of what’s going on inside the model. And many of these people have been working on some of these issues in their other work.