Can a depression AI be ‘correct’ when the dialogue never supplied the evidence?
An audit of 2,744 symptom cells found most lacked sufficient textual evidence. In a smaller model test, many benchmark-correct predictions were still unsupported by the dialogue.
The short version
- What happened
- EviDial contains 2,744 symptom cells from 343 clinical participants across Chinese PDCH and English E-DAIC interviews; 82.8% lacked sufficient textual evidence under the authors' working reference.
- Who it affects
- If benchmark accuracy is mistaken for clinical readiness, people could face assessments that sound decisive even when the conversation did not supply the required evidence. The study supports better evaluation, not deployment.
- What remains unproven
- The work is an unreviewed preprint and does not test diagnosis, treatment, patient outcomes or clinical deployment.
Keep in mind
The work is an unreviewed preprint and does not test diagnosis, treatment, patient outcomes or clinical deployment.
The model audit used 87 development cells, lightweight models and single runs, so it cannot support model rankings.
E-DAIC provided 79% of the main reference cells; language, source and clinical instrument are confounded.
The working reference was generated from overlapping LLM judge families and remains measurement-dependent despite a 120-cell human validation.
Source analysis · 1 linked source. Primary analysed source date: 20 October 2026.
Read the full report · 3 min The evidence, methods, implications and limitations
Matching the recorded label was not the same as being justified
Depression-dialogue systems are often scored against symptom labels derived from clinical scales. EviDial asks a different question: did the transcript shown to the model contain enough evidence for that symptom judgment? The preprint covers 99 Chinese PDCH interviews from 72 participants and 271 English E-DAIC interviews from 271 participants, producing 2,744 participant–symptom cells. Twelve simulated TalkDep dialogues were analysed separately from the main clinical reference.
The authors classified each cell as having absent, partial, conflicting or sufficient textual evidence after checking who a statement concerned, its time window, scope and any retraction. Because there is no direct gold standard for evidence sufficiency, they combined 5,240 valid judgments from three language-stratified LLM judges using a random-effects latent-class model. Under that working reference, 82.8% of clinical cells lacked sufficient evidence, with a 95% confidence interval of 82.1% to 83.5%.[1]
The model audit found unsupported correctness—not clinical failure
On 87 development cells from 76 participants, the researchers tested Gemini, GPT and Qwen families with four prompting methods. Among predictions counted as correct by the native benchmark label, 69.2% to 96.9% lacked sufficient dialogue support. A prompt that required a quotation did not remove the problem; a sufficiency gate reduced the unsupported-correct rate but often abstained, leaving coverage between 34.5% and 67.8%.
The result warns researchers against treating label accuracy as proof of evidence-grounded clinical judgment. It does not show that a participant's clinical label was wrong, reveal how a model reasoned or establish diagnostic safety. E-DAIC supplied 79% of reference cells, its transcripts were reformatted by another model, and the automated reference disagreed with plurality on 13.3% of jointly labelled cells. A blinded 120-cell human check broadly preserved the pattern, but broader human annotation, repeated tests with stronger models and prospective clinical evaluation would change the assessment.[1]
The distinction that matters
Two questions a benchmark should keep separate
A prediction can match the stored label even when the visible dialogue cannot justify it.
Label match
Compare prediction with the recorded score
Measures agreement
Text support
Check eligible dialogue evidence
Measures justification
Clinical validity
Test with people and outcomes
Not established here