When a Risk Label Hides the Evidence

TL;DR for operators

An AI system can produce mutually consistent risk labels, supporting evidence, and explanatory factors while still hiding relevant information. In a 2026 study by Tianle Hu and coauthors,¹ a classification rule suppresses supporting evidence whenever a particular risk category is predicted—even if that prediction is wrong.

The study also demonstrates a modest improvement from combining two different psychosocial factor verifiers: Macro F1 rises from 0.5467 for the stronger standalone model to 0.5562. For product teams, the findings favor separate monitoring of classification errors, suppressed evidence, and factor-level reliability. They do not establish that the resulting explanations are clinically trustworthy.

A consistent explanation can still conceal relevant evidence

Consider an AI review system that reads a social-media post, assigns a suicide-risk category, extracts supporting phrases, and identifies relevant psychosocial factors. The additional outputs appear to make the original classification easier to inspect.

But suppose the first prediction is incorrect. If the system forces every subsequent output to agree with that prediction, evidence that contradicts it may disappear. The final result can look internally consistent while giving the reviewer less information than the original text contained.

This is a consequential design problem for AI products that combine classification with structured explanations. Consistency is desirable, but it does not independently establish whether the underlying judgment is correct.

Hu and coauthors investigate this problem through a framework that coordinates risk classification, textual evidence extraction, and identification of 24 psychosocial factors in social-networking posts.

The particularly revealing result concerns a small change in classification performance that affects what the system allows users to see.

A small classification gain can prevent evidence suppression

The framework predicts four risk categories: Indicator, Ideation, Behavior, and Attempt.

For the Indicator category, it applies a deterministic rule: the final suicide-related evidence output must be none. The rule operates on the predicted category, not the verified correctness of that prediction.

Consequently, a post incorrectly assigned Indicator cannot display the evidence phrases selected by the evidence extractor.

The researchers address part of this problem through token-length routing. Posts of at most 384 tokens use an ensemble of three short-context DeBERTa-v3-large classifiers; longer posts use a 640-token specialist.

Compared with the three-model short-context baseline, the routed system achieves:

• Risk Weighted F1: 0.8067 → 0.8088.
• False Indicator predictions: 42 → 37.
• Posts routed to the longer-context model: 43 out of 1,635.

The Weighted F1 improvement of 0.0021 is small. The reduction in false Indicator assignments is more operationally relevant because these errors can directly eliminate supporting evidence.

A separate matched-model ablation shows that extending the context capacity from 384 to 640 tokens increases Weighted F1 from 0.7897 to 0.8051. Attempt recall improves, but Behavior recall declines. Longer context therefore does not uniformly improve every category.

These are within-dataset findings. The routing configuration does not achieve superior standalone Risk Assessment performance on the blind leaderboard.

The Evidence Grounding component itself achieves a held-out Macro row F1 of 0.7605 after annotation cleaning and candidate-phrase processing. This is not the competition’s official Phrase F1, and its post-level evaluation split does not establish generalization to unseen users.

The broader lesson is that evaluating an upstream model only through aggregate accuracy can overlook errors that have disproportionate downstream consequences.

Psychosocial factors require a different verification process

Extracting the phrases supporting a risk judgment is not the same task as identifying psychosocial factors.

The framework uses two separately trained, definition-conditioned classifiers for the latter task.

The Taxonomy Verifier (TV) evaluates whether sentence-like text units are semantically compatible with a factor’s definition.

The Evidence-Aware Verifier (EAV) uses the same general model architecture but changes how positive training examples are selected. Factor-specific lexical-semantic cues prioritize text units considered informative for a particular factor.

Despite its name, EAV does not directly consume predictions from the separate Evidence Grounding component. Its cues guide training-unit selection.

Both verifiers produce unit-level factor probabilities. Maximum pooling selects the highest unit score for each factor at the post level. Their resulting probabilities are then combined:

Factor-specific thresholds determine which labels become active.

This design gives the models different training signals: one emphasizes semantic compatibility with a definition, while the other emphasizes text units selected through factor-specific cues.

The hypothesis is not that every cue is accurate. It is that the resulting classification errors may differ enough for probability fusion to improve performance.

The weaker verifier adds value, but the gain is modest

The study evaluates factor identification using five user-grouped folds, with held-out predictions generated by models trained without those posts and decision thresholds selected from other folds.

The principal results are:

Source: Table II, Section IV-C.

EAV is weaker overall, yet it performs better individually on certain factors, including low self-esteem, traumatic experience, and poor school performance.

Fusion improves over TV on 16 of the 24 factor labels and outperforms it in four of five evaluation folds.

The aggregate Macro F1 advantage is nevertheless only 0.0095. This supports complementarity, not a dramatic increase in reliability.

A fusion-weight sensitivity analysis provides additional context. Across tested TV weights from 0.45 to 0.75, Macro F1 ranges from 0.5562 to 0.5600. The highest reported result occurs at 0.50, not the adopted 0.60. The adopted weights were inherited from competition development rather than optimized on the strict cross-fitted evaluation.

Several weaknesses also survive fusion. F1 remains 0.130 for meaning in life, 0.324 for sense of responsibility, and 0.323 for exposure to others’ suicide.

Moreover, fusion predicts an average of 3.489 factors per post against 2.919 annotated factors. It reduces overprediction relative to either standalone verifier but does not eliminate it.

For a product displaying psychosocial profiles, this matters: additional predicted labels can create an appearance of explanatory completeness even when particular categories remain unreliable.

Matching psychological patterns does not validate psychological explanations

The researchers also compare predicted factor prevalence across risk categories with the corresponding gold-annotation patterns, using the Three-Step Theory of suicide as an interpretive perspective.

Some selected contrasts align closely. The difference in hopelessness prevalence between Ideation and Indicator is +0.165 in gold annotations and +0.169 in Fusion predictions.

Others show substantial distortion. For prior self-harm or suicidal thought/attempt, the Attempt-versus-Ideation contrast is +0.809 in gold annotations but only +0.506 in predictions.

These comparisons are exploratory descriptive analyses, not a second validation of the psychological theory. User-level bootstrap intervals quantify sampling variation in the observed predictions, without accounting for model-retraining uncertainty.

Preserving the direction of selected group differences does not establish that individual psychosocial states are correctly measured. Nor can cross-sectional post classifications demonstrate progression from ideation to action.

What product teams should evaluate before trusting the explanation layer

The paper’s direct contribution is a computational framework evaluated on 1,635 labeled posts from 153 users. Its comparative results are informative, but there is no independent external-dataset validation or clinical assessment of deployment effectiveness.

Cognaptus draws three design implications for safety-sensitive language-analysis systems.

First, monitor what consistency rules suppress. A product team should distinguish evidence unavailable because none was detected from evidence removed because an upstream classification rule overrode it. The study demonstrates the suppression mechanism, although it does not test such a monitoring interface.

Second, evaluate explanatory factors individually. Aggregate Macro F1 cannot indicate whether a particular rare protective factor is sufficiently reliable for a specific operational use. Per-factor errors, activated-label counts, and threshold behavior deserve separate reporting.

Third, justify model complexity through complementary performance. The weaker verifier contributes information that marginally improves fusion, but the size and variability of that improvement should determine whether maintaining two models is worthwhile.

These recommendations apply to research prototypes and evidence-linked review workflows. The study provides no basis for treating its predictions as clinical diagnoses or allowing them to trigger autonomous interventions.

Explanation quality depends on what the system preserves

Adding textual evidence and psychosocial labels makes a risk classifier more inspectable, but it also creates new opportunities for error propagation.

Hu and coauthors show that combining different factor-verification signals can improve multi-label prediction. Their results simultaneously reveal how modest that improvement remains, how uneven individual factor reliability can be, and how a consistency rule can conceal evidence after an incorrect upstream prediction.

For teams developing consequential AI systems, the central design question is not simply how many explanations a model can produce. It is whether those explanations preserve relevant information, expose uncertainty, and remain reliable at the level where decisions are actually made.

Cognaptus: Automate the Present, Incubate the Future.