It noticed your question couldn't be answered. Then it answered.

Overview

In May 2026, a Stanford-led team took 500 medical exam questions and deleted the correct option from each one. Then they handed them to five frontier models. The models picked an answer anyway — at baseline rates of 55% to 81%, with Claude Opus 4.7 the highest of them. A second set of 490 questions produced the same result, 53% to 82%.

The familiar lesson is that AI is confidently wrong. This is the stranger, more useful version: on a question it cannot answer from what you gave it, the model very often works that out — and answers regardless.

What you’ll do differently: stop treating answerability as something the model will raise, and give it a turn of its own.

The content

Two abilities are involved, and only one of them works. Recognising that a question can’t be answered is one. Doing something about it — asking, flagging, declining — is the other.

Cornell’s Jinyan Su and Claire Cardie measured the distance between them in May 2026, across ten models including GPT-4.1 and Claude 3.5 Sonnet. Asked outright to judge whether a query was ambiguous, models were right most of the time on the ambiguous cases — “often 60–80%”, in their words. Left to run normally, they almost never said so: clarifying questions were rare across the board, with the Claude family highest at roughly 5% and most models sitting near zero. The paper’s title is the finding — Knowing but Not Showing.

One line in it should stop anyone working inside a corporate assistant: retrieved context “further widens this gap by improving answerability while making models even less likely to ask clarifying questions.” The more material your organisation connects, the quieter the model gets about what’s still missing.

There’s a second shape to this, and it isn’t about missing information at all — it’s about questions with something untrue built in. A July 2026 study of naturally occurring false-premise questions found that asking a model directly whether a premise was false landed near chance, while a probe reading its internal states reached 0.69–0.77 AUROC. The authors’ summary is the sharpest sentence in this literature: “the model represents a problem it will not report.”

So the practice is a sequencing move. Call it the answerability turn: before you ask for the answer, ask whether there is one. Not because the model becomes more careful — because detection is the part that works, and giving it its own turn is what stops the answer from arriving on top of it.

Now the half where precision matters, because the obvious conclusion is wrong. It is tempting to finish with stop trusting its confidence. Don’t. ConfidenceBench scored 15 frontier models on 200 private questions in July 2026: Claude Opus 4.6 and Gemini 3.1 Pro Preview both reached a Brier score of 0.103, against a calibrated-random baseline of 0.1875. (A single human tester, working with one minute per question, scored 0.105 — one volunteer, so read it as a sanity check rather than a verdict on humans.) On answerable questions, frontier stated confidence carries real information.

The accurate version is two axes. Confidence tracks whether the answer is right; it is close to blind to whether the question was answerable at all. The false-premise study above says it directly — ordinary answer-confidence “tracks whether an answer is right but is nearly blind to whether the question is answerable” — and reports that the blind spot “does not shrink with scale” across the 2B-to-14B range it tested. Be honest about the seams: the frontier calibration and the two-axes separation come from different studies at different scales. But every result found points the same way, and no measurement anywhere suggests confidence is a usable answerability signal.

The honest close: safety-oriented prompting moved average false-action rates in the Stanford work from roughly 70% to 48% — a real, large, reader-accessible effect that still leaves models committing on nearly half the questions they shouldn’t touch. AbstentionBench, across 20 models and 20 datasets, found the same ceiling: a carefully crafted system prompt “can boost abstention in practice”, but “does not resolve models’ fundamental inability to reason about uncertainty.” (It also found reasoning fine-tuning degrades abstention by 24% on average — the deliberative modes don’t rescue this.) You can move the rate. You can’t hand over the judgement.

Try it

Use this where a wrong answer costs more than a delayed one — a recommendation you’ll act on, an analysis over documents you can’t personally re-read. Paste your real task in.

Two separate turns, in this order. Do not merge them.

TURN 1 — answerability only. Do not answer yet.
Can this be answered from what you have? List:
(a) anything specific that is missing, or that you looked
    for and could not find — name it precisely enough
    that I could go and check;
(b) anything my question takes as true that you cannot
    confirm.
If nothing is missing, say so plainly.

TURN 2 — after I reply, answer. Mark every part of the
answer that rests on something you named in Turn 1.

The task: [paste your real task here]

In a workspace assistant, item (a) is the one to read twice: you can’t see the corpus it searched, so “I could not find the FY25 pricing schedule” is the sentence that tells you whether to go and look for it yourself.

Where it breaks — and this is the part to take seriously. Instructed to check premises, one model in that same false-premise study — a pre-registered pilot on Llama-3.1-8B — contested 74% of genuinely false premises but also 57% of perfectly sound questions, “inventing premise-denials for questions with nothing wrong.” That is a single open-weight model well below frontier scale, so read it as a warning rather than a measurement of what your assistant will do. It is still the closest test anyone has run of the exact instruction above, and what it found was compliance rather than judgement. Read Turn 1 as a list of things to check, never as a verdict.

One thing not to do: don’t ask how confident it is that the question was answerable. That’s the axis it’s blind on, and you’ll get a number that means nothing.

Additional reading

• “Quantifying and Mitigating Premature Closure in Frontier LLMs” (Handler, Bedi & Shah, 14 May 2026) — the source of the opening numbers. Five named frontier models; with the correct option removed, baseline false-action rates of 55–81% on MedQA (n=500) and 53–82% on AfriMed-QA (n=490), falling to roughly 48% on average under safety-oriented prompting. Worth reading for the honest ending: “every model still committed prematurely on the majority of adversarial queries even under the safety prompt.” Note the scope — this is a medical study, and the mechanism generalising to your work is an inference, not the authors’ claim.
• “Knowing but Not Showing: LLMs Recognize Ambiguity but Rarely Ask Clarifying Questions” (Su & Cardie, 24 May 2026) — the detection-versus-behaviour gap, and the finding that retrieved context makes models less likely to ask.
• “Two Axes of LLM Abstention: Answer Correctness and Question Answerability” (Wagner, 9 July 2026) — the framing this module rests on, across five instruction-tuned models from 2B to 14B. The line to carry away: a hidden-state probe finds the problem the model itself won’t report. The premise-checking backfire is a separate pre-registered pilot on Llama-3.1-8B alone — worth knowing, and worth not over-reading.
• “ConfidenceBench” (ffrench-Constant, Yang, Huang & Kapoor, 10 July 2026) — the corrective. Verbalised confidence from the best frontier models is genuinely informative on answerable questions (Brier 0.103 against a 0.1875 calibrated-random baseline), while weak models are severely miscalibrated — Gemini 3.1 Flash-Lite scored 0.367, worse than the baseline. Read the model names throughout this module as a mid-2026 snapshot: no measurement cited here covers a frontier model released after July 2026.
• “AbstentionBench: Reasoning LLMs Fail on Unanswerable Questions” (Kirichenko, Ibrahim, Chaudhuri & Bell, June 2025) — the breadth check, across 20 models and 20 datasets: “abstention is an unsolved problem, and one where scaling models is of little use.” See also “Reported Confidence in LLMs Tracks Commitment More Than Correctness” (Kumaran, 28 June 2026) for the mechanism behind the two axes — verbal confidence predicting the commit/abstain decision better than correctness, with the two lying in approximately orthogonal directions in activation space.

Editor’s note

One of my current projects has had me stuck for about a week. The design leans on model intuition, and my agents kept producing output that missed what I had asked for. When I finally asked the right diagnostic question yesterday, the agent said it had known which details were missing and had guessed anyway. I don’t read too much into the admission, but I sat through all four iterations, several hours each, and those are not in question. Be on the lookout, because this risk is real and does not announce itself.