University of Helsinki finds measurement errors

You must now require hallucination detection systems to report factuality, source faithfulness and uncertainty as three separate signals, each validated against human judgement, rather than accepting a single reliability score.

The prompt is an analysis of 30 October 2026 of hallucination detection measurement across factuality, source faithfulness and uncertainty, which argues that automated detection systems conflate distinct failure modes and produce scores that cannot be reliably interpreted. The University of Helsinki tested nine hundred question-answer pairs labelled by humans against seven language models as judges, alongside word-overlap metrics and natural-language inference. In our assessment, this work exposes a critical gap between what detection systems measure and what decision-makers assume they measure — a gap that creates genuine risk in regulated environments where you must later justify a decision.

What measurement error does the research identify?

The University of Helsinki study found that automatic labelling strategies deviate substantially from human judgements when the goal is factual accuracy. Prompts designed to measure faithfulness to a reference answer often wrongly brand correct additions as hallucination. This means a system that reports an answer as "unreliable" may actually be measuring only that the answer deviates from a supplied source text, not that it is factually incorrect. For anyone deciding on the basis of such labels, that distinction is decisive.

Within the evaluated labelling strategies, the chosen assessment criterion mattered more than the precise structure of the prompt. Faithfulness-oriented prompts applied to a factuality goal produced substantial deviations from human labels. Prompts that explicitly distinguish between absence of information and contradiction aligned better with human judgement. The implication is stark: the question "what counts as a hallucination here?" must be answered in advance and explicitly, otherwise the benchmark measures something other than what you think.

Why does conflating these dimensions create risk?

An answer can be factually correct without being supported by the supplied source, and vice versa. A system that lumps these dimensions together produces a score that cannot be interpreted. For retrieval-based AI, factual accuracy and coverage by the retrieved context should be assessed separately. A mature detector therefore reports at least three separate signals: whether the claim is factually correct, whether it is supported by the retrieved source, and the confidence or uncertainty attached to each assessment.

One long answer can contain both correct and incorrect claims. One summarising score then hides precisely the error that matters. A robust approach splits the answer into separate, verifiable claims and checks each claim individually. Research into fine-grained factuality in long-form text turns an answer into a fact-centric network of relationships, so that multi-step and long-range connections between dispersed facts are also checked — not only whether the text sounds locally consistent. Consistency is not the same as truth.

Which controls must you demonstrate in your workflow?

1. Separate reporting of three dimensions — require your supplier to break down the reliability signal into factuality, source faithfulness and uncertainty, and to record which criterion was used for each assessment.
2. Claim-level verification — record that every factual claim is checked individually against a source before it enters a decision, rather than relying on a global score.
3. Periodic validation against human judgement — include an audit set in contracts and procurement terms that periodically tests detector labels against human assessments, since automatic benchmarks can deviate substantially from human judgement.
4. Reconstruction of decision inputs — for each outcome, keep the model, prompt, source and detector version, so that you can afterwards reconstruct what a decision rested on.

What can tooling do, and what remains your responsibility?

Automated detection can now separate factuality from source faithfulness and report uncertainty as a distinct signal. What tooling cannot do is decide what counts as acceptable risk in your context, or whether a particular claim matters enough to warrant human review. A detector that scores highly on a benchmark has not proved universal faultlessness. In our assessment, the real consequence lies in the set-up: anyone building or procuring detection must work modularly, with separate checks for each dimension, each calibrated against human judgements. A detector that does not separate these dimensions produces a number that nobody can explain — and in a regulated environment that is a risk, not a reassurance.