Measuring the Bluff Rate

For two years, executives have asked me the same question, and for two years I've told them it's the wrong one. What's the hallucination rate? It sounds like due diligence. It's actually a way to avoid the harder question underneath it.

There is no single hallucination rate for an AI model, for the same reason there is no single "failure rate" for a hospital, a law firm, or a newsroom. Failure depends on the task, the pressure, the review process, and the cost of being confidently wrong. Ask for one number, and you get a number that describes one narrow situation — usually the most flattering one.

This is the third time I've written about those numbers. What changed in 2026 is not that the picture got cleaner. It's that the distance between the flattering number and the honest one is now measured, published, and hard to wave away. The bluff rate didn't go away. It grew up.

The Nice Number Comes From the Leash

Start with the good news, because there is real good news.

When a model is handed a document and asked to summarize it faithfully, the best systems now stay in the low single digits. Vectara's HHEM leaderboard, updated May 11, 2026, puts its top model at 1.8%, with gpt-5.4-nano at 3.1% and Gemini 2.5 Flash Lite at 3.3%. Stanford's 2026 AI Index, drawing on that same HHEM leaderboard, reports its top 15 models clustered between 1.8% and 5.4%, most of them in the 4-to-5% band.

Read those numbers for what they measure. This is document-grounded summarization.

The model already has the text. Its job is not to know the world; its job is to avoid adding anything the source didn't say.

That is a genuinely useful result, and it maps to genuinely useful work: contract digests, call-note summaries, research abstracts, document review — anywhere the answer sits close to a provided source, close to retrieval, close to a human who can check it.

The number is real. The leash is what makes it look good.

Off the Leash

Now take the leash off.

AA-Omniscience, built by Artificial Analysis and reported in the same AI Index, asks a different kind of question: open-ended factual knowledge across 6,000 questions in six domains, from law and health to software engineering and math. No source document. No retrieval.

Just — do you know this, and do you know whether you know it?

The benchmark is scored to punish exactly the behavior the summarization test can't see. Correct answers earn points, wrong answers lose them, and refusing to answer costs nothing. It is built to reward a model for saying "I don't know." And even so, across 26 models, hallucination rates run from 22% to 94%. The best model in the set is wrong more than a fifth of the time. The worst is wrong on nearly everything it's asked.

Same generation of models. Same year. A seventy-two-point spread, depending entirely on whether the model was handed the answer or asked to supply it. This is the bluff rate the summarization number hides — and unlike its 2024 ancestor, it's now measured on a scoreboard built so that bluffing can't win.

Guessing Still Wins

Why does this survive better models? Because the product around the model rewards it for guessing.

OpenAI's own SimpleQA breakdown is the cleanest illustration, and it's worth sitting with, because it's OpenAI indicting its own incentive structure. Two models, same test. The older o4-mini abstained on 1% of questions, answered almost everything, and was wrong 75% of the time — while posting 24% accuracy. The newer gpt-5-thinking-mini abstained 52% of the time and cut its error rate to 26% — at 22% accuracy.

Look at what a product team sees. The reckless model is slightly more accurate and answers nearly every question. It keeps the conversation alive.

It never hands the user that cold little moment where the assistant admits it doesn't know. But that cold little moment is the safety feature, and the model that offers it scores worse on the one metric most leaderboards actually track.

This isn't OpenAI's problem alone, and you don't have to take its word for it. HalluLens, Meta's 2025 benchmark, found the same tradeoff in a different shape: GPT-4o hallucinated on 45% of the answers it attempted while almost never refusing, while Llama-3.1-405B-Instruct held its hallucination rate to the lowest in the set — and got there by refusing to answer more than half the time. Precision and coverage, pulling against each other, every time.

The peer-reviewed version of OpenAI's argument, published in Nature in 2026, states the mechanism plainly: accuracy-based evaluation incentivizes hallucination. When a scoreboard gives credit for a lucky guess and nothing for restraint, the model learns the lesson any student learns. Silence looks like failure. Fluent guessing looks like effort. A model doesn't need to be stupid to be unsafe. It only needs to be asked to perform certainty it doesn't have.

Liability Is Learning the Shape of the Machine

The law doesn't need new vocabulary for any of this. It already has the buckets — defamation, negligence, consumer protection, professional responsibility, data accuracy. What's happening now is that courts and regulators are deciding which bucket each failure falls into.

Two 2025 cases show the range. In Walters v. OpenAI, a Georgia court granted OpenAI summary judgment in a defamation suit over a ChatGPT output that falsely tied a radio host to a fraud and embezzlement complaint. The win was real but narrow: it turned on specific facts — the warnings the user saw, what a reasonable reader in his position would have believed, the public-figure malice standard, the absence of proven damages. It is not a permission slip.

The NOYB complaint filed in Norway the same year points the other way. There, ChatGPT generated a fabricated account in which a named user had murdered two of his children and attempted to murder the third. NOYB framed it not as defamation but as a data-accuracy violation under Article 5(1)(d) of the GDPR — the argument being that once a model produces false personal information, the question is no longer whether the user was warned that mistakes happen. The case is still open; as of mid-2025, the Irish DPC took over as lead authority. That is the harsher governance lens, and it's the one agentic deployments should worry about.

The argument over whether hallucinations exist is finished. The next one is about duty: who was obligated to anticipate them, constrain them, label them, test them, and keep them from reaching a person as if they were fact.

The Fix Is Still Boring

The mitigation playbook hasn't changed, which is exactly why it gets ignored. Ground answers in retrieved sources. Make the model expose its uncertainty instead of smoothing it over. Penalize confident errors harder than abstentions. Verify claims before they leave the system. Keep high-risk outputs under human review.

None of that is technically hard. The resistance is cultural.

Product teams like completion. Users like confidence. Executives like a clean metric they can drop into a board deck. A model that frequently says "I'm not sure" satisfies none of those appetites, even though it is the safer system.

The version of this technology worth building treats every factual answer as a set of claims, each carrying a status — supported, inferred, uncertain, unsupported — instead of one smooth paragraph that pretends every fact in it is equally load-bearing. That would make AI feel less magical. Good. Magic is the wrong design language for a system that can invent court citations, corporate policies, and criminal allegations.

So here is the honest answer to the executive's question, the one I've been giving in some form for two years.

A low hallucination number almost always means a tightly bounded task. A high one almost always means the model was asked to know the world, and the test was honest enough to measure whether it could admit that it couldn't. Deploy according to which of those two situations you're actually in.

The bluff rate never disappeared. In 2026, it finally became measurable — which means the next excuse is gone too.