Frankfurt's Bullshit
Harry Frankfurt, a Princeton philosopher, published 'On Bullshit' in 1986 and expanded it into a bestselling book in 2005. The essay drew a taxonomic distinction that had been conceptually implicit in ordinary usage but never formalized. A liar, Frankfurt argued, knows the truth and deliberately contradicts it; lying is parasitic on the truth it denies, because the liar must know what is true in order to say its opposite. A bullshitter, by contrast, is indifferent to truth. The bullshitter's goal is not to describe the world accurately but to produce an effect on the audience — to persuade, to impress, to sound authoritative — without regard for whether the statements producing the effect happen to be true. Forty years later, Flyvbjerg and Cambridge's Alan Blackwell applied the framework to large language models with diagnostic precision.
In the AI Story
The philosophical distinction matters because it specifies the relationship between a speaker and the truth-falsehood axis. The liar operates within the axis; the bullshitter operates orthogonal to it. Large language models are bullshit machines in precisely Frankfurt's sense — they do not lie because lying requires knowledge of what is true, and they do not tell the truth because truth-telling also requires such knowledge. They generate text optimized for plausibility, for the appearance of knowledge, for the rhetorical texture of authority, without any mechanism for determining whether the generated text corresponds to reality.
The distinction has legal and industrial consequences. Mercedes' chief technology officer Markus Schäfer articulated the operational concern: if you sit in a car and ChatGPT tells you something that is absolute nonsense, you might be exposed to product liability cases. The automotive industry, which has skin in the game in ways the technology press does not, has proceeded with a caution the broader culture has not matched. The word bullshit, despite its vulgar register, is the technically precise term for the category of utterance the industry must now manage.
Flyvbjerg's use of Frankfurt represents a rare case of philosophical terminology migrating intact into technical discourse. Artificial ignorance and bullshit machine are not rhetorical escalations but diagnostic claims, backed by the empirical finding that current systems produce persuasive text without any mechanism for truth-tracking. The smoothness that makes the output convincing is identical to the smoothness that conceals the absence of grounding — a single design choice producing both the utility and the hazard.
The philosophical lineage runs deeper than the term suggests. Frankfurt's essay drew on Wittgenstein's lifelong distaste for loose talk and on Augustine's theological analysis of mendacity. The insight — that indifference to truth is a distinct moral category, more corrosive than deliberate falsehood — has ancient roots. What Frankfurt added was the precision of analytic philosophy and the willingness to name the category in its plain vernacular form. The precision is what allows the term to function as a diagnostic in the AI discourse rather than a term of abuse.
Origin
Frankfurt, a Princeton philosophy professor, first published the essay in the Raritan Quarterly Review in 1986. The piece acquired a cult following in academic philosophy for two decades before Princeton University Press issued it as a slim standalone volume in 2005. The book became an unlikely bestseller, translated into dozens of languages, and established bullshit as a technical term in the philosophy of language.
Key Ideas
Indifference to truth. The defining feature of bullshit is not falsehood but unconcern with whether the statement is true — a moral stance orthogonal to the lying/truth-telling axis.
Parasitic on no truth. Unlike lying, which depends on a truth it denies, bullshit does not require any stance on reality — only a target effect on the audience.
Rhetorically optimized. The goal is to produce impressions, persuade listeners, project authority — and the content is selected for effect rather than correspondence to fact.
More corrosive than lying. Frankfurt's provocative claim: bullshit is a greater enemy of truth than lies, because the liar still operates within a framework where truth matters.
Perfect fit for LLMs. The definition maps onto the generative architecture of large language models with uncomfortable precision — plausibility optimization without truth grounding.
Debates & Critiques
Some philosophers have argued that Frankfurt's definition is too expansive, capturing ordinary rhetoric and polite convention alongside genuine epistemic pathology. Others have argued that applying the term to non-conscious systems is a category error — bullshit requires intention, and machines have none. Flyvbjerg's defense is pragmatic: the structural feature Frankfurt identified — optimization for persuasive effect without regard for truth — is present in the systems regardless of whether consciousness accompanies it, and the diagnostic value of the term does not depend on the metaphysical status of the generator.
Appears in the Orange Pill Cycle
Further reading
- Frankfurt, Harry G. On Bullshit. Princeton University Press, 2005.
- Frankfurt, Harry G. On Truth. Knopf, 2006.
- Blackwell, Alan F. 'Oops! We Automated Bullshit.' Cambridge Computer Laboratory, 2024.
- Hicks, Michael Townsen, James Humphries, and Joe Slater. 'ChatGPT is Bullshit.' Ethics and Information Technology, 2024.