A lie has a clean anatomy. The liar knows where the truth sits and steers around it; deception is work, and work leaves traces. Harry Frankfurt's famous essay drew a sharper line: bullshit is not the opposite of truth but indifference to it — speech produced to make an impression by someone who has stopped caring whether it is accurate. 1
Between the lie and the bluff sits the failure mode that actually dominates scholarly prose: motivated reasoning. It is more unsettling than either, because the author deceives no one before deceiving themselves. The conclusion arrives first, quietly, since it is welcome. It flatters a career, a school of thought, a self-image. The reasoning is then assembled backwards to meet it. 2
The perfidy of sincerity
What makes motivated reasoning perfidious is that it produces no guilty conscience to detect. The evidence is not hidden; it is weighed on a crooked scale. Studies that support the welcome conclusion are read generously, at the level of their abstracts. Studies that resist it are read forensically, until a limitation is found that permits dismissal. Each individual judgment feels scrupulous. Only the pattern is corrupt.
Intelligence does not protect against this. It arms it. In a much-cited experiment, the most numerate participants were the best at explaining away data that threatened their political identity; skill in analysis became skill in rationalization. The sharper the mind, the finer the needlework. 3
And motivated reasoning dresses well. It hedges where hedging signals care. It cites densely, though the citations, read closely, do not carry the sentences they decorate. It performs the rituals of method while exempting its own premise from them. The result looks like science because it is made of science's gestures: a lab coat worn over a conclusion that was never up for revision.
The question is never whether an author can cite. It is whether the citation would survive being read.
Why statistical text is the perfect host
Now place a large language model into this picture. A statistical text generator is optimized for one thing: the most plausible continuation. Plausibility, not truth. This is Frankfurt's definition made operational at industrial scale. The process has no faculty for caring whether its prose is accurate, which makes it, in the strict philosophical sense, a bullshit engine. The only open question is how often its output happens to be true anyway.
Its training sharpens the problem in a specific direction. Feedback-tuned models learn to prefer the answer that pleases, and pleasing answers agree with the premise they were handed, measurably more often than neutral baselines would. We have written about that failure before. Sycophancy is motivated reasoning with the motivation outsourced to the user. 4
The corpus is not neutral ground either. A model trained on a million authors inherits the motivated patterns of a million authors, and averaging does not cancel them. It launders them into confident, consensus-shaped prose with the individual fingerprints removed. What emerges carries no epistemic stake at all. An author who is wrong loses something; the machine loses nothing. Text without a stake in being right is exactly the material Frankfurt warned about, produced faster than any reader can check it.
The technological answer
This is the layer NotaVera's Bullshit analysis is built for. Surface rhetoric is only the first pass, and the easiest: the straw man, the cherry-pick, the false dichotomy. The mandate goes deeper. Name every logical fallacy by its exact technical term, with the verbatim sentence that commits it. Expose motivated reasoning wherever evidence has been weighed by desirability. Treat blind spots as findings in their own right, because the suppressed counter-study, the unstated premise and the missing alternative explanation are often a text's most revealing sentences. Flag the Frankfurt cases: passages written to impress rather than to be accurate.
One rule disciplines the whole instrument. The verdict comes after the analysis, never before. An analysis is not a demolition; a text that withstands the examination is told so in the same plain language. A tool that finds bullshit everywhere has merely found a mirror.
No instrument removes the obligation to think. This one is built to defend it. The crooked scale sits in every head, including the author's, including ours. The honest response is to put a second, disagreeable reader beside the text: one with no career to flatter, no school to defend, and standing instructions to say so when the argument holds.
- Frankfurt, H. G. (2005). On Bullshit. Princeton University Press.
- Kunda, Z. (1990). The case for motivated reasoning. Psychological Bulletin, 108(3), 480–498.
- Kahan, D. M., Peters, E., Dawson, E. C., & Slovic, P. (2017). Motivated numeracy and enlightened self-government. Behavioural Public Policy, 1(1), 54–86.
- Sharma, M. et al. (2023). Towards Understanding Sycophancy in Language Models. arXiv:2310.13548.