Aug 25, 2026 Announcements Funding better evaluations of AI’s impact

What happened

I can see the article slug is /news/wellbeing-research-grants. Let me fetch that directly. I now have all the details from the actual article. Here is the article:

Anthropic on August 25, 2026 announced a $5 million grant program to fund independent research into how AI affects the people who use it. The program will provide direct funding, access to Anthropic's models, and technical support to grantees who build open-source evaluations designed to help the broader AI industry measure how models influence user wellbeing. All work produced under the grants will be published as open-source projects available to any developer.

This piece covers what Anthropic is offering, why the company says evaluating wellbeing is structurally harder than evaluating factual accuracy, which researchers and practitioners stand to benefit most, and what builders shipping AI products should understand about the grant criteria before applying. The deadline is September 21, with selected applicants notified by October 5 for a full proposal.

How it works

Anthropic is committing $5 million to a grant program explicitly focused on wellbeing evaluations, an area it says has lagged behind other evaluation work. The Safeguards team at Anthropic has published accompanying guidance — a publicly available PDF — that defines what makes a wellbeing evaluation rigorous enough to build on and catalogues the common failure modes that render existing evaluations too narrow to be useful in practice.

Grantees will work fully independently of Anthropic. They will receive funding, model access, and technical support, but will own and publish their results as open-source projects. The independence provision is structural: Anthropic describes it as core to the program, not merely aspirational. This design means the resulting evaluations will be available to competing labs, enterprise developers, and researchers equally.

Before this program, Anthropic says the industry had no clear standards for how models should behave in high-stakes conversational scenarios — such as when a user begins seeking companionship from a model or uses AI to navigate a mental health crisis. The core difficulty is that wellbeing cannot be assessed by examining a single response. A model might give reasonable dietary advice to a general user, yet that same advice could be harmful if the user has demonstrated a history of disordered eating in the same conversation.

Why it matters

Anthropic's Safeguards team guidance specifies five criteria for evaluations it considers credible: a clear definition of what constitutes a pass or fail; involvement of clinical and subject-matter experts in design and validation; testing of both overcompliance and overrefusal risks; scenarios that model multi-turn conversations where risk escalates over time; and graders validated against real subject-matter experts. These five criteria are the substantive bar applicants will need to meet.

Clinicians, psychologists, and methodologists are explicitly named as the kinds of experts Anthropic wants to bring into evaluation design. Independent researchers who have found AI wellbeing work underfunded, and academic labs with expertise in mental health measurement or conversational dynamics, are the primary targets of the program. Teams that have previously found it hard to access frontier model APIs for safety-oriented research now have a direct path to both funding and model access.

Who is affected

Builders shipping products with emotional support, companionship, or mental health adjacent features should pay close attention to what comes out of the program. The open-source evaluations produced by grantees will be the most concrete, publicly available tools for testing whether a Claude-powered or comparable product is behaving appropriately in these sensitive conversation types. Operators who are not themselves applying should plan to use the resulting benchmarks once they are published.

Applications are open now via a Google Form linked from the Anthropic Newsroom post at anthropic.com/news/wellbeing-research-grants. The application deadline is September 21, 2026. Applicants selected to submit full proposals will be notified by October 5, 2026. Anthropic has not published grant size ranges or the number of grantees it plans to fund, so applicants should structure their proposals around the guidance document rather than any assumed budget tier.

The guidance PDF from the Safeguards team is publicly accessible without applying and is worth reading before filling out the form. It is the most specific signal Anthropic has released about what methodological standards it considers sufficient, and reading it in full before drafting a project description is the clearest way to align a proposal with the stated criteria.

What to watch next

The October 5 notification date means the first cohort of projects should be publicly identified before the end of 2026. Because all outputs are required to be open-source, the evaluation frameworks and benchmarks produced will be accessible from the moment grantees publish them. Watching the GitHub repositories and publication channels of selected grantees will show whether multi-turn conversation modeling and expert-validated grading actually become standard in wellbeing evaluation, or whether the field continues to rely on single-turn proxies.

Anthropic has been publishing related safety work in parallel — including research into how people use Claude for emotional support and companionship, and separate work on safeguards for biology-related queries in Fable 5. Whether the grant program produces evaluations that Anthropic subsequently adopts into its own internal testing pipeline, or whether they remain external reference tools, is the key open question for anyone tracking how industry wellbeing standards develop over the next 12 months.