GenAI Must Augment, Not Replace, Psychometric Standards in Mental Health Measurement

A team of researchers has published a framework for how artificial intelligence should be used in mental health measurement—and their central argument is deliberately cautious: AI can help, but it cannot replace the rigorous standards that have governed psychological testing for decades.

The paper, published in PLOS Mental Health in September 2026, addresses a growing problem in the field. Generative AI systems, particularly large language models, are increasingly being used to support psychological assessment and digital mental health measurement. They can generate candidate test items, refine wording, classify open-ended responses, extract information from clinical notes, and help integrate data from smartphones, wearables, chatbots, and other digital sources. The appeal is obvious: these systems are fast, scalable, and can process vast amounts of unstructured data. But speed and scale, the authors argue, are not the same as validity.

The researchers—Villarreal-Zegarra, Paredes-Gonzales, and García-Serna—propose a lifecycle framework that maps where AI can enter the measurement process without weakening the evidentiary standards that psychometrics demands. The framework covers nine stages: construct definition, item generation, content and response evaluation, piloting and calibration, validation, fairness testing, scoring and interpretation, documentation, and post-deployment monitoring. At each stage, AI may assist, but human judgment and empirical evidence must remain in control.

The distinction they draw is sharp and important. Digital traces from smartphones, wearables, chatbots, and social media should not be treated as psychometric measures simply because they are quantifiable. They should remain raw data or candidate indicators until their construct interpretation and intended use are theoretically specified, technically verified, empirically calibrated, and validated with actual human participants. An AI system might generate a hundred candidate test items in minutes, but those items are not valid until they have been reviewed by experts, tested on real people, and shown to measure what they claim to measure. Linguistic fluency is not validity. Predictive performance alone is not sufficient evidence of construct validity. Apparent neutrality is not fairness.

The authors identify ten critical failure modes that can undermine measurement when AI is deployed without sufficient oversight. Construct drift occurs when AI-generated items appear coherent but gradually shift the target construct—a particular risk in mental health, where neighboring constructs like distress, depression, and anxiety overlap semantically but have distinct clinical meanings. Face validity without structural validity describes the risk that AI-generated items sound clinically plausible but fail to organize into the internal structure the measurement model proposes. Synthetic respondents—using AI systems as stand-ins for human test-takers during development—may help identify obviously poor items, but they compress variability and cannot replace calibration in the actual population. Algorithmic bias is perhaps the most consequential risk: because large language models are trained on vast corpora that encode social biases related to gender, race, age, language, disability, and culture, AI-generated items and AI-assisted scoring systems should never be assumed to be fair simply because they appear neutral. Measurement inequivalence across groups—the failure to measure the same construct in the same way for different populations—is a first-order requirement for valid comparison, not a secondary analysis.

Other risks include prompt sensitivity (small changes in how you ask an AI to do something can alter its output), model drift (when AI providers update their systems, the instrument may change without the researcher knowing), automation bias (the tendency to over-rely on algorithmic outputs even when they are uncertain or wrong), and data security failures that can compromise the confidentiality of sensitive mental health information. The authors also warn against the flattening of subjectivity and disagreement: when open-text responses, clinical narratives, and expert annotations are aggregated into single labels or consensus scores, individual-level variability and meaningful human disagreement may be lost.

The paper distinguishes two complementary uses of AI in this domain. GenAI for psychometrics describes cases where AI supports the development, scoring, or interpretation of human measures—generating items, refining wording, classifying responses, extracting information from clinical text. Psychometrics for GenAI describes the reverse: using psychometric methods to evaluate how AI systems themselves behave when they become part of measurement workflows. If an AI system is used as a rater or scorer, its outputs should be evaluated with standards at least as strict as those applied to human raters. This distinction matters because the same AI system may function both as a tool for measurement and as a potential source of error or bias.

The authors propose five research priorities for the field. First, fairness and measurement invariance should be treated as first-order requirements, not afterthoughts. Second, future studies should quantify how model updates, prompt variations, and provider-side configuration changes affect generated items, scoring outputs, and psychometric parameters—establishing thresholds for when AI-assisted measures require recalibration. Third, researchers should determine the boundary conditions for when AI respondents are useful for preliminary screening and when they distort latent distributions. Fourth, the field needs consolidated psychometric benchmarks for evaluating AI behavior itself. Fifth, as AI is increasingly applied to multimodal data—text, speech, images, wearable signals, smartphone interaction—research must clarify how these heterogeneous signals can be combined into valid constructs and composite indices.

The central message is conservative by design. AI should be treated as an augmentation layer across the psychometric lifecycle, not as a replacement for psychometric science. Its responsible use requires transparent documentation of model versions, prompts, system instructions, and human edits. It requires validation with human data, not just synthetic data or predictive performance. It requires explicit fairness testing across demographic groups. It requires secure data governance and procedures for human review when outputs are ambiguous or clinically consequential. It requires continuous monitoring after deployment to detect construct drift, model changes, or emerging biases. GenAI does not reduce the need for psychometrics; it makes psychometrics more central, more demanding, and more consequential.