University of Helsinki finds systematic measurement errors in hallucination benchmarks

Automated hallucination detection is becoming more useful, but not as one universal reliability score you can trust blindly. Recent research from the University of Helsinki and ACL studies shows mature systems measure factual accuracy, source faithfulness and uncertainty separately, check per claim and validate their labels against human assessments.

On 6 October 2026, researchers at the University of Helsinki published an empirical study examining how hallucination benchmarks themselves are labelled. They tested nine hundred question-answer pairs assessed by humans against seven language models as judges, alongside simple word-overlap metrics and natural-language inference. The key finding: automatic labelling strategies deviate substantially from human judgements once the goal is factual accuracy, and prompts aimed at faithfulness to a reference answer then often wrongly brand correct additions as hallucination.

That is no detail. It means that a system which calls an answer "unreliable" may in reality be measuring that the answer deviates from a short model text — not that it is factually incorrect. For anyone deciding on the basis of such labels, that difference is decisive.

What did the University of Helsinki establish about automatic hallucination benchmarks?

The study by the University of Helsinki reports that, in its factorial comparison, the chosen assessment criterion mattered more than the precise structure of the prompt. Within the evaluated labelling strategies and benchmark, 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 labels.

In other words: the question "what counts as a hallucination here?" must be answered in advance and explicitly, otherwise the benchmark measures something other than what the user thinks.

Why must factual accuracy, source faithfulness and uncertainty be measured separately?

Because 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. The cited research distinguishes factual correctness from faithfulness to supplied context. The linked ACL paper provides background on faithfulness in abstractive summarisation; it does not by itself establish a RAG-specific distinction.

For retrieval-based AI, our analysis is that factual accuracy and coverage by the retrieved context should be assessed separately. A mature detector therefore reports at least three separate signals:

• Factual accuracy: does the claim hold up against external, reliable knowledge?
• Source faithfulness: is the claim actually covered by the supplied source?
• Uncertainty: how confident is the system, and is something not found, not covered or contradicted?

In our assessment, this is the breaking point with the previous generation of detectors: one green tick suggests a certainty that the underlying measurement process does not deliver.

Why must detection check per claim and across long-range relationships instead of one score?

Because one long answer can contain both correct and incorrect claims. One summarising score then hides precisely the error that matters. Our analysis is that a robust approach is to split the answer into separate, verifiable claims and check each claim individually.

Research into fine-grained factuality in long-form text splits answers into separate claims and checks them against external sources. In addition, a benchmark and knowledge-graph approach for long-form text introduces a method that 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. In the reported experiments, that method is described as outperforming existing baselines; those are experimental findings on the tested benchmark, not proof of a universally faultless detector. A step list for such a check:

1. Decide in advance whether the task measures factual accuracy, source faithfulness or both.
2. Split the output into separate, verifiable claims.
3. Link each claim to a source passage or external evidence source.
4. Check relationships between facts that stand far apart in the text.
5. Report uncertainty and distinguish not found, not covered and contradicted.
6. Validate the detector labels on a human-assessed sample.

That AI consensus does not prove independent fact-checking applies here in full: consistency is not the same as truth.

What does this mean for executives, lawyers and CISOs who must be able to trust AI output?

Our analysis: because automatic benchmarks, according to the University of Helsinki, can confuse source faithfulness with factual accuracy, a decision-maker relying on one supplier score may reject correct additions and let uncovered claims through; therefore you should require your supplier to break down the reliability signal into factuality, source faithfulness and uncertainty, and to record which criterion was used. Because one long answer contains both correct and incorrect claims, a global score can hide precisely the kind of error that may matter in a lawyer's or doctor's decision; therefore in your workflow you should record that every factual claim is checked individually against a source before it lands in a decision. Because the Helsinki study reports that labels can deviate substantially from human judgements, our analysis is that a detector without a human-validated assessment should be treated as an assumption rather than as proof; our recommendation is that contracts and procurement terms should include an audit set that periodically tests the detector labels against human assessments. Because better performance on one benchmark does not demonstrate universal faultlessness, "our AI scores highly" is no accountability; our recommendation is that for each outcome you should keep the model, prompt, source and detector version, so that you can afterwards reconstruct what a decision rested on. That last point connects to why a plausible AI explanation is not yet proof and to the broader duty of verification at claim level in decision-making.

In our assessment, the real consequence lies in the set-up: anyone building or procuring detection must work modularly, with separate checks for factuality, source faithfulness and uncertainty, 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.

Sources: The article draws on an empirical study by the University of Helsinki (2026) and on peer-reviewed ACL research into source faithfulness, claim-level factuality and hallucination detection in long-form text.