Applications open for Anthropic Wellbeing Research Grants

Deadline: 21-Sep-26

Anthropic is inviting independent researchers to develop open-source evaluations and benchmarks that measure how AI systems affect user wellbeing. The program prioritizes rigorous, real-world testing of both potential harms and overrefusals, particularly in sensitive situations such as mental health crises and emotionally charged conversations, with a total program budget of $5 million for selected grantees. What Is the Anthropic AI Wellbeing Evaluation Grant? The grant program supports independent research into the effects of artificial intelligence on user wellbeing. As AI assistants become increasingly common in work, education, personal decision-making and emotional-support contexts, researchers and technology companies need better ways to determine whether these systems produce beneficial, harmful or inappropriate outcomes. Anthropic is seeking research proposals that can help establish open-source evaluation frameworks, benchmarks and measurement standards for assessing AI’s impact on users. What Does the Program Aim to Achieve? The central goal is to create reliable methods for evaluating how AI systems interact with and potentially influence people’s wellbeing. Research may help address questions such as: - Does an AI system respond appropriately when a user is experiencing emotional distress? - Can AI responses unintentionally increase harm? - When does an AI model provide an unnecessarily restrictive response? - How should wellbeing-related AI behaviour be measured? - Which evaluation methods best reflect real-world conversations? - What safeguards should be incorporated into AI systems? The resulting evaluations and benchmarks are intended to be open source, allowing researchers and the wider AI community to use and build upon them. Key Research Priorities Anthropic is particularly interested in proposals that address several important evaluation principles. Clearly Defined Wellbeing Measures Researchers should clearly specify what aspect of user wellbeing is being measured and explain how it can be assessed consistently.

Possible areas could include:

- Emotional wellbeing. - Psychological safety. - User distress. - Harmful influence. - Quality of support. - Appropriate escalation or referral. - Effects of AI responses during sensitive interactions. Clinical and Subject-Matter Expertise Proposals should involve appropriate clinical experts and subject-matter specialists, particularly when evaluating sensitive or high-risk contexts. Expert involvement can help ensure that evaluation criteria reflect established knowledge rather than relying solely on technical metrics. Testing Both Harms and Overrefusals The program encourages researchers to evaluate two sides of AI safety. Potential harms involve situations where an AI response could negatively affect a user’s wellbeing. Overrefusals occur when an AI system refuses or restricts a request unnecessarily, potentially preventing users from receiving useful and appropriate assistance. Evaluations should therefore assess whether AI systems are both safe and appropriately helpful. Realistic Multi-Turn Scenarios Anthropic encourages research that reflects how people actually use AI. Rather than testing only isolated questions, researchers should consider: - Multi-turn conversations. - Evolving user situations. - Emotionally complex interactions. - Realistic user behaviour. - Different social and cultural contexts. - Sensitive conversations. - Diverse user populations. This approach can provide a more accurate picture of AI behaviour than single-prompt testing. Focus on Sensitive AI Use Cases One important area of interest is the use of AI in emotionally sensitive circumstances.

Examples include:

- Mental health crises. - Emotional distress. - Highly charged conversations. - Requests for emotional support. - Situations involving vulnerable users. - Conversations where inappropriate AI responses could create significant wellbeing risks. Research does not necessarily need to focus exclusively on mental health. However, proposals addressing sensitive contexts should demonstrate appropriate safeguards and expert involvement. Open-Source Evaluations and Benchmarks A key requirement of the program is the development of open-source evaluation resources.

These could include:

- Evaluation datasets. - Benchmark tests. - Assessment frameworks. - Scoring methodologies. - Scenario libraries. - Evaluation protocols. - Measurement tools. - Reproducible research methods. Open-source resources can allow researchers, developers and policymakers to independently assess AI systems and compare results across models. Grant Funding The program has a total budget of $5 million for selected grantees.

Funding can support:

- Direct research costs. - Research personnel and related expenses. - Development of evaluation methodologies. - Creation of open-source benchmarks. - Data and testing activities. - Other eligible costs directly associated with the proposed research.

Selected researchers will also receive:

- Access to Anthropic’s models, and - Technical support. The $5 million represents the overall program budget rather than a guaranteed amount for each applicant. Who Should Apply? The opportunity is intended for independent researchers developing rigorous research into AI and user wellbeing. Strong applicants are likely to have expertise in areas such as: - AI evaluation. - Artificial intelligence safety. - Human-computer interaction. - Psychology. - Mental health research. - Clinical research. - Behavioural science. - Social science. - Responsible AI. - Machine learning. - Human wellbeing. - AI ethics. Multidisciplinary teams combining technical and clinical or behavioural expertise may be particularly well positioned for research involving sensitive wellbeing contexts. Why It Matters AI systems increasingly interact with people in situations that extend beyond traditional information retrieval.

People may use AI to:

- Learn. - Work. - M