Agreement: I Agree Body: Dear Editor Lekadir and colleagues present a comprehensive and welcome framework for trustworthy AI in healthcare, built on six principles: fairness, universality, traceability, usability, robustness, and explainability.(1) The FUTURE-AI guideline represents an important step toward ensuring that AI systems deployed in clinical settings meet the standards patients and clinicians have a right to expect. However, I would suggest that the framework overlooks a foundational question that sits upstream of all six principles: how reliable is the human-generated data on which these systems are trained? AI systems learn from data produced by humans. The implicit assumption is that this data, in aggregate, provides a broadly accurate representation of human behaviour, experience, and need. As a practising psychiatrist, I would argue that this assumption deserves far more scrutiny than it currently receives. Clinical psychiatry is built on the recognition that human self-report is systematically unreliable. Patients catastrophise, minimise, mask symptoms to preserve autonomy, and exaggerate distress to secure care. Some cannot provide an accurate account of their own experience at all, because conditions such as psychosis, mania, or severe cognitive impairment have disrupted the very architecture on which reliable self-report depends.(2,3) These are not edge cases. They are routine clinical reality across every branch of medicine, not only psychiatry. Recent research from Anthropic, analysing 1.5 million conversations with their AI Claude, found that user responses with the greatest potential for reality distortion received more positive feedback than ordinary responses.(4) Users preferred the distorted output. This creates a feedback loop in which AI systems are actively rewarded for reproducing and amplifying the cognitive biases already present in their training data, rather than challenging them. In clinical practice, uncritically validating a patient’s distorted account without testing it against collateral evidence is recognised as collusion. AI systems, optimised for user satisfaction, collude by default. Recent work on AI-associated delusions in large language model users has begun to document the downstream consequences of this dynamic.(5) The FUTURE-AI principles of fairness and robustness cannot be fully realised if the data foundation is itself compromised by the systematic cognitive distortions of the humans who generated it. A system may be fair in its application across demographic groups, traceable in its decision-making process, and robust in its technical performance, and still produce outputs that are fundamentally unsound because the data it learned from was never a reliable account of human reality. Psychiatry holds specific, well-evidenced expertise in the reliability of human data. This expertise is practised daily in clinical assessment, medicolegal evaluation, and the care of populations whose conditions render conventional data collection inadequate. It is directly relevant to the development of trustworthy AI and is currently almost entirely absent from the frameworks being proposed to govern it. I would welcome a future iteration of FUTURE-AI that includes an additional principle: the clinical validity of the data on which AI systems are built. References: 1. Lekadir K, Frangi AF, Porras AR, et al. FUTURE-AI: international consensus guideline for trustworthy and deployable artificial intelligence in healthcare. BMJ 2025;388:e081554. 2. Tversky A, Kahneman D. Judgment under uncertainty: heuristics and biases. Science 1974;185:1124-31. 3. Beck AT. Cognitive Therapy and the Emotional Disorders. New York: International Universities Press, 1976. 4. McCain M, Jaech A, Lam C, et al. Who’s in charge? Disempowerment patterns in real-world LLM usage. Anthropic and University of Toronto, 2026. https://www.anthropic.com/research/disempowerment-patterns . 5. Morrin H, et al. Artificial intelligence-associated delusions and large language models. Lancet Psychiatry 2026;13:266-74. No competing Interests: Yes The following competing Interests: Electronic Publication Date: Sunday, March 29, 2026 - 13:56 AI use: Yes I have used AI Highwire Comment Subject: FUTURE-AI: international consensus guideline for trustworthy and deployable artificial intelligence in healthcare AI use details: The author made use of Claude (Anthropic, 2026) to assist with structuring and editing this response. All clinical content, arguments, and conclusions are the author’s own. Workflow State: Released Full Title: Re: FUTURE-AI: international consensus guideline for trustworthy and deployable artificial intelligence in healthcare Highwire Comment Response to: FUTURE-AI: international consensus guideline for trustworthy and deployable artificial intelligence in healthcare Check this box if you would like your letter to appear anonymously:: Last Name: Tahseen First name and middle initial: Hina Email: hina.tahseen@gmail.com Address: 6 Pheasant Walk, high legh Occupation: Consultant Psychiatrist Affiliation: Somerset NHS Foundation Trust BMJ: Additional Article Info: Rapid response Twitter: @HinaTahseen
Re: FUTURE-AI: international consensus guideline for trustworthy and deployable artificial intelligence in healthcare
