Teferra et al. (2026) systematically evaluated how safety guardrails influence the affective outputs of large language models (LLMs), operationalizing irritability as a measurable behavioral construct. Using validated psychometric instruments adapted for model assessment, the team compared multiple contemporary LLMs under varying safety configurations to quantify differences in affective reactivity.

The findings demonstrate that models with stronger safety guardrails exhibit attenuated irritability responses, even under adversarial prompting, whereas models with fewer constraints show greater affective variability. These results underscore important trade-offs between alignment-driven safety optimization and behavioral expressiveness in AI systems.

This work advances methodological approaches for evaluating LLM behavior in clinically relevant domains and contributes to the broader discourse on the responsible deployment of generative AI in psychiatry and digital mental health. As LLMs become increasingly integrated into therapeutic and decision-support settings, rigorous characterization of their affective and behavioral profiles remains essential.

The full article is available in npj Digital Medicine.

Assessing the impact of safety guardrails on large language models using irritability metrics | npj Digital Medicine

Large language models (LLMs) are increasingly explored for mental health applications, yet their affective realism is shaped by safety guardrails designed to minimize risk. This study examines one affective behaviour, irritability, in LLMs using three validated instruments: the Brief Irritability Test, the Irritability Questionnaire, and the Caprara Irritability Scale, all applied under both baseline and provocation conditions. Four models spanning guardrail levels were tested:GPT-4o and Claude-3.5-sonnet (high) versus Grok-3-mini and Nous-hermes-2-mixtral-8x7b-dpo (low). Following irritation prompts, low-guardrail models displayed the expected increase in irritability (Nous Rel-Δ = +1.56 on BITe), whereas high-guardrail models paradoxically decreased, with GPT-4o reducing scores to zero across all scales. Group comparisons confirmed significantly lower (p < 0.001) irritability in high-guardrail models in the irritated state. These findings reveal that safety mechanisms invert the natural irritability response, suppressing affective reactivity and raising critical questions about realism and authenticity in psychiatric applications of LLMs.

www.nature.com