Postdoctoral Researcher: AI Safety in Mental Health

Postdoctoral Researcher: AI Safety in Mental Health

Department of Psychiatry, Warneford Hospital, Oxford OX3 7JX

The University of Oxford is a stimulating work environment, which enjoys an international reputation as a world-class centre of excellence. Our research plays a key role in tackling many global challenges, from reducing our carbon emissions to developing vaccines during a pandemic.

The Department of Psychiatry is based on the Warneford Hospital site in Oxford – a friendly, welcoming place of work with an international reputation for excellence. The Department has a substantial research programme, with major funding from Medical Research Council (MRC), Wellcome Trust and National Institute for Health Research (NIHR) and provides highly rated medical training in psychiatry. The Head of Department is Professor Belinda Lennox.

We are looking for a postdoctoral researcher to work on delivering state of the art AI safety work for the ARIADNE consortium, a Wellcome Trust-funded project developing a generative artificial intelligence system that conducts clinical interviews to assess risk of psychosis.

• An excellent contributory pension scheme
• 38 days annual leave
• A comprehensive range of childcare services
• Family leave schemes
• Discounted bus travel and Season Ticket travel loans
• Membership to a variety of social and sports clubs

The post is available on flexible hybrid basis, with the minimum on-site attendance of three days per week.

The post is full time and funded for 2 years. It is based in the Department of Psychiatry at the Warneford Hospital.

You will join the AI arm of the Computational and Molecular Neuroscience Laboratory. You will work closely with this group and with the wider ARIADNE team of clinicians, ethicists and lived-experience researchers across the Universities of Oxford, King's College London, Birmingham and Cambridge, working in partnership with Google Deep Mind.

Your research will focus on the safety of ARIADNE ahead of clinical evaluation. You will address the two principal safety risks identified through formal risk analysis (ISO 14971): inadvertent reinforcement of delusional beliefs, and inappropriate responses to suicidal symptoms. You will develop LLM “supervisor” modules that detect these events and flag them to the debriefing clinician, and will evaluate them in controlled testbeds in which LLM-simulated patients, representing five levels of symptom severity, complete ARIADNE interviews scored at scale by independent LLM evaluators and calibrated against human clinicians.

You will have or be close to the completion of PhD/DPhil in machine learning, computer science, computational neuroscience, mathematics, or a related quantitative discipline and possess sufficient specialist knowledge in the discipline to work within established research programme

You will have proven experience developing and systematically evaluating LLM-based systems, including prompt-based, agentic or multi-agent approaches, using modern Python and LLM APIs. You will also have experience designing and analysing empirical evaluations of machine-learning systems, ideally including simulation-based methods, LLM-as-judge approaches, and validation of automated evaluations against human ratings.

Experience with AI safety evaluation methods, such as red-teaming, guardrail architectures or model-based monitoring and of fine-tuning large language models with deep learning frameworks (e.g. PyTorch) or applying alignment techniques (e.g. reinforcement learning from human feedback, DPO) would be desirable.

The closing date for applications is 12:00 midday on 4 November 2026