AI agreement on crisis labels is not the same as accuracy

A short social media post can leave important things unsaid. Expressions of distress may depend on context, community language or an implication that different readers interpret differently. That makes assigning crisis-severity labels difficult—even before artificial intelligence enters the process.

A November 2026 arXiv preprint by K M Sajjadul Islam and colleagues examines that difficulty using veteran-related tweets. Its central distinction is important: models agreeing with one another, models matching human annotations and models recognizing the most severe cases are different achievements. Success on one does not guarantee success on the others.

What the researchers compared

The team studied 149 tweets selected from a collection of approximately 4,500 veteran-related posts. Eight veteran subject-matter experts, recruited for their experience as peer mentors, independently annotated each selected tweet. These were judgments about social media content, not clinical assessments of the people who wrote it.

The annotation framework contained five ordered severity levels, ranging from no concern to acute crisis with suicidal ideation. Two additional categories covered material irrelevant to the research and an “other” outcome. Thus, although the abstract emphasizes five severity categories, the main annotation task retained seven possible labels.

For comparison with the models, the researchers assigned each tweet the label receiving the most expert votes. This did not necessarily mean that more than half the reviewers supported it. Nine tweets had tied leading labels; the researchers resolved those ties toward the more severe interpretation.

The resulting reference set was uneven: 100 tweets were labeled no concern, while just six received the highest-severity label. None received the category for self-identified crisis symptoms without seeking support. The reference labels therefore summarize expert judgments rather than establish an unquestionable answer for every post.

The researchers tested 15 locally deployed open-source models and 10 proprietary models across eight main prompt variants. Instructions ranged from label names alone to detailed definitions, worked examples and requests for reasoning. Additional experiments deliberately changed example labels or supplied explanations alongside them.

Strong agreement, but with whom?

Human reviewers often disagreed about exact categories. Under the original seven-label framework, none of the tweets received unanimous agreement. However, agreement was higher when the analysis accounted for the ordering of severity levels, suggesting that disagreements frequently involved nearby levels rather than opposite ends of the scale.

Six selected models showed substantially greater agreement with one another. Under the reviewer-matched baseline prompt, their ordinal Krippendorff’s alpha—a measure of agreement beyond chance that accounts for category ordering—was about 0.834. The corresponding human value was about 0.494. For that comparison, the two non-severity categories were combined in the human annotations; the selected models never used “other.”

These figures describe consistency, not clinical correctness. Models could converge on an interpretation that differs from the experts’ reference label. Their annotations also used a narrower range of categories than the human reviewers did.

Across the main experiments, proprietary models generally matched the expert-derived labels better than local open-source models. GPT-5.6-Terra had the highest reported mean Micro-F1, 0.783, and the lowest variation across its available prompt results. Some model–prompt combinations lacked usable results, so these averages did not all cover the same number of conditions.

Rare severe cases change the picture

Overall scores can obscure performance on uncommon categories, especially when most posts belong to the no-concern group.

In a separate table examining the six posts with highest-severity reference labels, Qwen2.5-7B assigned that label to all six, while GPT-5.6-Terra did so for five. Several other selected models assigned it to only two.

That does not establish Qwen as a reliable crisis detector. Six cases provide little evidence about general performance, and two were near-duplicate posts. Identifying those examples also does not reveal how often a model incorrectly assigns the highest-severity label elsewhere. Furthermore, some severe reference labels arose from tied expert votes, making the benchmark itself uncertain.

Why more elaborate prompts are not enough

Prompt changes affected models differently. Misleading demonstration labels reduced performance for many models, but others were comparatively insensitive. Adding explanations to examples did not consistently improve classification. A plausible explanation was therefore not a guarantee of a better label.

The study also evaluated each model–prompt–tweet combination only once, leaving repeated-run variability unmeasured. Its refined prompt incorporated earlier error patterns, so those improvements should not be treated as proof of performance on an independent new dataset.

This small, historical social media study did not test diagnosis, imminent-risk assessment or whether an intervention helps someone. Its useful finding is narrower: language models may assist annotation, but agreement among them cannot substitute for examining human disagreement, rare severe cases and the limits of the evidence.