AI and Bullshit
Ten years ago this year, a red bus toured Britain bearing a slogan that would become infamous: "We send the EU £350 million a week, let's fund our NHS instead." The claim was not quite a lie. There was a gross figure, before rebates and returned funds, that approached £350 million. But it was not quite the truth either. The UK Statistics Authority called it "a clear misuse of official statistics." Dominic Cummings, who devised the slogan, later admitted it was chosen precisely because it would provoke argument, the controversy itself was the point.
Was Boris Johnson lying when he stood in front of that bus?
In 1986, Princeton philosopher Harry Frankfurt published a short essay offering a different framework. His subject was bullshit, not lying. Frankfurt was careful to distinguish the two. The liar knows the truth and deliberately says something false. The bullshitter, by contrast, is entirely indifferent to whether what they say is true or false. "It is just this lack of connection to a concern with truth," Frankfurt wrote, "this indifference to how things really are, that I regard as of the essence of bullshit."
The £350 million claim fits this framework uncomfortably well. Its creators were not primarily concerned with its accuracy. They were concerned with its effect.
The essay became a slim bestselling book in 2005. Frankfurt died in 2023. And his framework has found an unexpected second life in debates about artificial intelligence, particularly in the work of cognitive scientist Gary Marcus, who has spent years arguing that large language models are, in Frankfurt's precise technical sense, bullshit machines.
Marcus`s claims might seem outdated. After all, commercial AI models have improved dramatically. GPT-5 handles reasoning tasks that stumped GPT-3 and 4. Hallucination rates have fallen. Retrieval-augmented generation grounds outputs in verifiable sources. Yet Marcus and a growing cohort of researchers argue that these improvements, while real, miss the point. The question isn't whether AI systems produce fewer falsehoods. It's whether they have any concern with truth at all.
The term "AI hallucination" has become the default description for those moments when language models generate confident falsehoods such as citing court cases that don't exist, inventing research papers with plausible-sounding authors and titles, or describing events that never occurred. The metaphor suggests something almost sympathetic: the AI is trying to perceive reality but is experiencing a perceptual error.
But as researchers Michael Townsen Hicks, James Humphries, and Joe Slater from the University of Glasgow argued in their 2024 paper in Ethics and Information Technology, the “hallucination” framing fundamentally misrepresents what's happening,regardless of frequency. In a proper hallucination, the person experiencing it is attempting to engage with reality; their perceptual apparatus is simply malfunctioning. The hallucinator cares about what's real. They're just getting it wrong. Large language models, by contrast, have no concern with truth whatsoever. They are designed to predict the most probable next token in a sequence, based on patterns in their training data. As the Glasgow researchers put it: "These programs cannot themselves be concerned with truth, and because they are designed to produce text that looks truth-apt without any actual concern for truth, it seems appropriate to call their outputs bullshit."
The improvement in accuracy, on this view, is real but beside the point. A more sophisticated bullshitter produces more plausible bullshit. That doesn't mean they've started caring about truth.
This maps remarkably well onto how large language models function. When ChatGPT generates a statement, it is not checking that statement against reality. It is not trying to be accurate and failing. It is generating plausible-sounding text based on statistical patterns. If the output happens to be true, that's incidental. If it happens to be false, that's equally incidental.
Stephen Colbert, in the first episode of The Colbert Report in 2005, coined a related term: "truthiness": the quality of seeming true, of feeling right in the gut, regardless of evidence. Colbert was satirising political pundits, but the concept applies with uncomfortable precision to LLM outputs. As the Glasgow researchers themselves noted, ChatGPT "functions not to convey truth or falsehood but rather to convince the reader of—to use Colbert's apt coinage—the truthiness of its statement."
Frankfurt's distinction between lying and bullshitting illuminates why the "hallucination" framing misleads. Both liars and truth-tellers are, in Frankfurt's terms, "playing on opposite sides in the same game", both respond to facts as they understand them. The bullshitter, by contrast, "ignores these demands altogether. He does not reject the authority of the truth, as the liar does, and oppose himself to it. He pays no attention to it at all."
This is why, Frankfurt argued, "bullshit is a greater enemy of the truth than lies are." The liar at least acknowledges truth's authority by trying to subvert it. The bullshitter erodes the very framework within which truth and falsehood make sense.
Gary Marcus has been making versions of this argument since long before Frankfurt's framework was formally applied to AI. In August 2020, he co-authored an article for MIT Technology Review titled "GPT-3, Bloviator," arguing that OpenAI's language model "has no idea what it's talking about." His verdict: "GPT-3 is a better bullshit artist than its predecessor, but it's still a bullshit artist."
This assessment has not made Marcus popular in certain quarters. Yann LeCun, Meta's chief AI scientist, has publicly sparred with him. Sam Altman has made oblique references to "mediocre deep learning skeptics." The criticism has some force: problems Marcus identified in GPT-2 were largely solved by GPT-3; problems in GPT-3 were largely solved by GPT-4. Each time Marcus points to a failure, the next model generation handles it better. To critics, this pattern suggests the limitations are temporary, not fundamental.
Marcus's response is that the goalposts haven't moved, the goal was always understanding, not performance on specific benchmarks. A system that predicts the most likely next word, no matter how accurately, is not the same as a system that reasons about truth. As he wrote in his 2022 essay "Deep Learning Is Hitting a Wall": "Deep learning, which is fundamentally a technique for recognizing patterns, is at its best when all we need are rough-ready results, where stakes are low and perfect results optional."
The pattern-matching gets better. The bullshit gets more fluent. But fluent bullshit, on this view, remains bullshit.
What distinguishes Marcus's recent work from his earlier critiques is an increasing concern about downstream effects—particularly on science itself. His "Deep Bullshit" essay focused on OpenAI's Deep Research tool, which can rapidly generate science-sounding articles on any topic. The problem, as Marcus sees it, isn't just that individual outputs might contain errors. It's that the ease of generating plausible-sounding content threatens to overwhelm the mechanisms we use to distinguish good information from bad.
And those errors, once published, don't simply disappear. They enter the training data for future models, creating what researchers call "model collapse" or "model autophagy disorder"—systems chasing their own tails, amplifying errors through successive iterations.
The Glasgow researchers introduce a useful distinction between what they call "soft bullshit" and "hard bullshit." Soft bullshit is produced without concern for truth but also without any intent to deceive. Hard bullshit requires an active attempt to mislead about the nature of the enterprise itself.
At minimum, they argue, large language models are soft bullshitters: they produce text without any concern for its truth, but arguably without the kind of intentionality required for active deception. The harder question, one that depends on contested views about whether AI systems can have intentions at all, is whether they might also qualify as hard bullshitters.
The argument for hard bullshit goes something like this: even if the models themselves lack intentions, there is clearly an attempt somewhere in the system (by the developers, the marketing teams, the companies?) to present these systems as more truth-oriented than they actually are. When OpenAI acknowledges in fine print that Deep Research "can sometimes hallucinate facts" while simultaneously promoting it as a research tool, something like hard bullshit may be occurring, not by the AI, but those around it seeking to turn a profit.
Does it matter whether we call AI errors "hallucinations" or "bullshit"? The Glasgow researchers argue it matters a great deal, for at least three reasons.
First, the terminology affects public understanding. "Hallucination" suggests a system that is trying to perceive reality and occasionally failing, like a person with a fever seeing things that aren't there. This framing encourages the belief that the problem is temporary, that with better training or more data, the hallucinations will go away. But if the issue is architectural,if the system simply isn't designed to care about truth then no amount of scaling will solve it.
Second, the framing affects our relationship with the technology. Anthropomorphising AI systems, treating them as if they have perceptions, intentions, or even the capacity to make mistakes in the way humans do creates a kind of false intimacy. We extend to them the benefit of the doubt we would extend to a colleague who made an honest error. But as the researchers note: "It is bullshitting even when it says true things."
Third, the terminology affects how we think about solutions. If the problem is hallucination, we might look for ways to ground the AI's perceptions in reality, connecting it to databases, fact-checking its outputs, training it on more accurate data. Some of these approaches have been tried, with limited success. But if the problem is structural indifference to truth, then the solutions would need to be different: perhaps hybrid architectures that combine statistical pattern-matching with symbolic reasoning, or perhaps a fundamental rethinking of how these systems should be deployed.
Others have extended the bullshit analysis in interesting directions. A July 2025 paper on arXiv introduced the "Bullshit Index", a metric for quantifying an LLM's indifference to truth, and proposed a taxonomy of different bullshit forms: empty rhetoric, paltering, weasel words, and unverified claims. The researchers found, perhaps surprisingly, that reinforcement learning from human feedback (RLHF), the technique used to make language models more helpful and harmless, actually increased bullshit production.
Where does all this leave the person trying to use these tools productively while remaining epistemically responsible?
The first implication is a call for appropriate scepticism. This doesn't mean refusing to use language models; it means treating their outputs the way a skilled editor treats a first draft from an unreliable source. The form may be helpful; the content requires verification. The structure might save time; the facts need checking.
The second implication concerns high-stakes domains. If LLMs are bullshit machines by architecture rather than by accident, then deploying them in contexts where truth matters such as medicine, law, scientific research, requires safeguards that go beyond simple accuracy checks. The lawyer who used ChatGPT to write a brief and found it had invented court cases wasn't merely unlucky; he had used a tool for a purpose it was fundamentally unsuited for, and was culpable for laziness and/or stupidity.
The third implication is epistemic. We need better language for talking about what these systems do and don't do. "Hallucination" implies a malfunctioning truth-seeking process. "Bullshit" implies no truth-seeking process at all. The distinction matters because it affects our expectations, our trust calibrations, and ultimately our ability to use these tools wisely.
Frankfurt's little book should be required reading, as Gary Marcus suggests. Not because it provides easy answers, but because it clarifies the question. What we're dealing with, in large language models, is not a new kind of intelligence occasionally making mistakes. It's a new kind of text production that has no concern with truth at all, and an urgent need to understand what that means for how we think, write, and reason together.
-
Frankfurt, H.G., On Bullshit (2005, Princeton University Press)
Marcus, G., "A Few Words About Bullshit," Marcus on AI (November 2022)
Liang, K. et al., "Machine Bullshit: Characterizing the Emergent Disregard for Truth in Large Language Models,"
