CONCEPT

Artificial Ignorance

Flyvbjerg's 2025 reframing of large language models — not intelligent systems that occasionally hallucinate, but ignorant systems that never had access to the truth-falsehood distinction in the first place.

In January 2025, Bent Flyvbjerg tested ChatGPT and Perplexity on a factual question whose answer he himself had documented in peer-reviewed journals: the cost overrun of Boston's Big Dig. ChatGPT got it wrong. Perplexity got it worse, returning 478 percent against the correct 220. Neither system flagged uncertainty. The paper that followed, 'AI as Artificial Ignorance,' argued that current large language models are structurally incapable of truth-tracking because they predict plausible next tokens rather than track reality. The reframe — from intelligence to ignorance — is diagnostic, not rhetorical. It specifies what the systems actually do and what expectations users should bring to them. The term names a condition that the industry's preferred euphemism, 'hallucination,' systematically conceals.

In the AI Story

Hedcut illustration for Artificial Ignorance — Artificial Ignorance

The distinction between artificial intelligence and artificial ignorance matters because the name a civilization gives a technology shapes the trust it invests, the expectations it sets, and the governance structures it builds around the thing. Framing is not neutral. Calling a system 'intelligent' that cannot distinguish a 220 percent cost overrun from a 478 percent one is not optimism. It is the kind of categorical error that, in Flyvbjerg's empirical record, precedes every major project failure in the modern era.

The term draws on Harry Frankfurt's precise philosophical distinction between lying and bullshit. The liar knows the truth and contradicts it. The bullshitter is indifferent to truth — optimizing for persuasive effect without regard for whether the statements producing the effect happen to be true. Large language models, Flyvbjerg argues, are bullshit machines in precisely Frankfurt's sense. They operate in a space orthogonal to the truth-falsehood axis entirely, and this orthogonality is the condition that the word ignorance names with a precision that hallucination actively obscures.

The argument is not that AI is useless. It is more precise and more damning: AI is useful in domains where the user already possesses the expertise to evaluate the output — where, in other words, the system is least needed. Nassim Taleb's independent testing reached the same conclusion. The tool amplifies existing knowledge. It does not generate new knowledge. In domains where the user lacks expertise, it generates confident nonsense that the user is poorly equipped to detect.

The deeper structural claim is that the condition of artificial ignorance is not accidental but architectural. The systems were designed to be persuasive. Persuasiveness without truthfulness is the most dangerous combination available, and it is the combination the current generation of AI has optimized for. Whether this condition is temporary — remediable through improved training, better grounding, explicit truth criteria — or permanent is the open question that Flyvbjerg's paper deliberately leaves for subsequent research.

Origin

Flyvbjerg published 'AI as Artificial Ignorance' in Project Leadership and Society in 2025. The paper accumulated over four thousand downloads within months of its release. It extended his career-long framework — built through decades of studying why large-scale projects fail — to the new domain of artificial intelligence. The title was not rhetorical provocation but diagnostic claim, backed by the empirical test against the Big Dig that any competent research assistant could verify.

Key Ideas

Orthogonal to truth. LLMs do not lie and do not tell the truth; they optimize for plausibility, leaving the truth-falsehood distinction untouched by the generation process.

Persuasion without accuracy. Hinton's warning in compatible terms: the danger is not super-intelligence but super-persuasiveness, the gap between perceived and actual capability widened to its most dangerous configuration.

Useful only where unneeded. The tool amplifies existing expertise but generates confident nonsense in domains where the user lacks the competence to detect errors — which is precisely the deployment pattern the industry is pursuing.

Architectural, not incidental. The ignorance is not a bug awaiting a fix. It is the operational consequence of systems designed without any mechanism for truth-tracking, and remediating it requires architectural change, not scaled training.

Naming as diagnosis. 'Artificial ignorance' is not merely a provocative label but a diagnostic frame that changes what users demand of the tool and of themselves when deploying it.

Debates & Critiques

Defenders of current systems argue that retrieval-augmented generation, tool use, and verification layers will progressively close the gap — that artificial ignorance is a transient condition, not a structural one. Flyvbjerg's response is empirical rather than theoretical: show me the reference class of structurally comparable claims that succeeded, and we can talk. The uniqueness bias that insists this time is different is the same bias that produced every previous AI winter.

Appears in the Orange Pill Cycle

Bent Flyvbjerg — On AI