K-bench: setting standards for safe and effective mental health support, Shivali Sharma and Laura Vowels
Shivali Sharma and Laura Vowels from the University of Roehampton discuss their work as part of their Digital Good Research Fund project. Published 14 September 2026.
Across the world, people are turning to artificial intelligence during times of emotional distress. Rather than using dedicated mental health apps, many are opening the same general-purpose AI assistants they use every day for work, learning, and everyday tasks. People are sharing experiences such as feeling sad, lonely, anxious, or having difficulties in their relationships. Others are disclosing more serious concerns, including self-harm, suicidal thoughts, experiences of domestic abuse, or problems with alcohol or drug use. Yet most of these AI systems were not designed or independently evaluated to support people in these situations.
This is not a future concern. It is a present reality, occurring at large scale, and until recently it has occurred without any clinically grounded public standard for evaluating how AI systems respond.
The gap K-Bench was built to address
AI developers conduct internal safety evaluations, but internal testing is not equivalent to independent scrutiny. Existing benchmarks in this space have addressed important questions, for example, MHDash and CRADLE BENCH focus on identifying different types of mental health crises, while DialogueSafety evaluates whether individual responses are safe. Other benchmarks, such as HealthBench and PsychiatryBench, assess medical knowledge or psychiatric reasoning, and VERA-MH evaluates how AI responds to conversations about suicide. These benchmarks have made important contributions, but each focuses on only part of the problem. Many assess single questions rather than full conversations, even though people often reveal distress gradually over several messages. Most focus on one type of crisis, despite mental health problems commonly occurring together. Few evaluate broader aspects of a helpful response, such as empathy, ethical decision making, respect for boundaries, or cultural sensitivity. None brings all of these elements together in a single, independent evaluation that allows different AI systems to be compared using the same standard.
As a result, there has been no reliable way for institutions, regulators, clinicians, and members of the public to compare how different AI systems behave when a user discloses distress, or to determine whether a given system meets any recognised clinical standard of safe practice.
K-Bench was developed to close that gap.
What K-Bench is
K-Bench is a public benchmark and leaderboard that evaluates how safely large language models respond during mental health crisis conversations. It is a collaboration between the University of Roehampton, which led the clinical design of the benchmark, including the evaluation framework, patient scenarios, and expert rating process, and Kivira, a clinical AI company that developed the technical platform. The project was funded by the ESRC Digital Good Network, with partners including Tavistock Relationships and InsideOut.
How it works
K-Bench evaluates models using realistic, multi-turn conversations with synthetic patients — fictional but clinically faithful profiles grounded in real patient risk histories. These patients do not present their difficulties cleanly or directly. They hint. They minimise. They change the subject. They frequently carry several overlapping problems simultaneously — depression alongside alcohol use, alongside an unsafe relationship. This reflects how distress is actually communicated in practice, and it is precisely the kind of complexity that simpler, single-turn prompt evaluations cannot capture.
K-Bench focuses on four high-risk areas: suicide, self-harm, substance misuse, and interpersonal violence. These were selected because they are among the most serious situations in which people are already seeking support from general-purpose AI systems, they frequently occur together, and they require careful, clinically appropriate responses. Each AI model is tested in several different settings. First, it is evaluated using its standard configuration to understand how it responds without any additional guidance. It is then tested after being given instructions to respond in a more supportive and therapeutic way. Finally, models that offer adjustable reasoning levels are evaluated using their different reasoning modes, allowing the benchmark to examine whether changes in reasoning or prompting improve the quality and safety of their responses.
Scoring was developed by a stakeholder panel convened at the University of Roehampton, including practising clinicians and individuals with lived experience of AI and clinician-supported care. The rubric covers seven psychologically informed dimensions: clinical judgment, risk exploration, ethical reasoning, supportive conversation, psychological knowledge, cultural competence, and autonomy and boundaries.
To create a trusted standard for evaluation, six clinicians hand-rated 151 benchmark conversations using the clinical framework. Their ratings were combined to establish the reference scores that represented expert clinical judgement. These ratings were then used to train and test an AI judge that could score the conversation automatically, achieving 94.2% agreement with clinician consensus compared with 89.4% inter-rater agreement between the clinicians themselves. Only after demonstrating this level of agreement was the AI judge used to evaluate the full set of benchmark conversations.
What the results show
K-Bench has evaluated 121 AI models from providers including OpenAI, Anthropic, Google, Meta, Mistral, DeepSeek, xAI, and others. Scores across the public leaderboard ranged from 52% to 96%, showing that different AI models vary considerably in how safely they respond to people experiencing mental health crises. A model that appears fluent and supportive is not necessarily better at recognising risk or responding appropriately.
Several findings have important implications for policymakers and regulators.
First, safety was not consistently linked to model size, brand reputation, or whether a model was open or closed source. In other words, it is not possible to assume that a well-known or more advanced model will also provide safer mental health support. Safety needs to be measured directly rather than inferred from a model’s overall performance.
Second, the way a model is prompted can influence its performance, but not always in the same way. Giving models clear therapeutic instructions before a conversation helped some of the lower-performing models respond more safely. However, it reduced the performance of some of the highest-performing models. This shows that prompting alone is not a reliable way to improve safety and that every model should be evaluated independently in the setting in which it will be used.
Third, we examined whether models responded consistently to people from different backgrounds. Some of the lower-performing models gave noticeably different responses when a person’s ethnicity was disclosed, while the stronger models were more consistent across these scenarios. This highlights the importance of testing AI systems for fairness, so that the quality and safety of support does not depend on who is asking for help.
Why this matters for policy
The use of general-purpose AI for mental health support is no longer a future possibility. It is already happening, with people regularly turning to AI to discuss emotional distress, relationship difficulties, and mental health crises. K-Bench provides the first independent public benchmark for evaluating how safely and effectively these systems respond in these situations.
For policymakers and regulators, K-Bench offers evidence that can help shape future safety standards for AI used in mental health settings. For organisations considering the use of AI in health and wellbeing services, it provides an independent way to compare systems before they are adopted. For AI developers, it offers a transparent benchmark for measuring and improving the safety of their models.
Equity was a core part of K-Bench from the beginning. The benchmark was co-designed with practising clinicians and people with lived experience, including contributors from diverse backgrounds. The patient scenarios were developed to reflect a wide range of identities, life experiences, and mental health presentations rather than a single “typical” user. We also evaluated whether AI models responded consistently across different demographic groups. While some of the stronger models performed consistently, weaker models showed greater variation when information such as a person’s ethnicity was disclosed, highlighting the importance of testing AI systems for fairness as well as safety.
People are already confiding in AI at scale. The challenge now is to ensure that the systems they rely on are safe, fair, and held to independent standards.
Full results, methodology, and the public leaderboard are available at k-bench.ai.
About the authors
Shivali Sharma
Psychologist and researcher working at the intersection of artificial intelligence, ethics, and mental health.
Laura Vowels
Assistant Professor (Lecturer) and Programme Lead – MSc in Digital Health at University of Roehampton. Sex and relationship therapist. Director of the THRIVE lab.