Anthropic Grants: $5M for AI Wellbeing Research

Anthropic grants: $5M for independent AI wellbeing research

Anthropic launched a $5 million grant program on August 25, 2026, to fund independent research into how AI affects users' wellbeing. In return for direct funding, technical support and access to the company's models, grantees build open-source evaluations that help the AI industry measure how Anthropic's models affect the people using them.

Grantees work fully independently of Anthropic. What they build is published as open-source projects that any developer can use.

By paying for independent evaluations and benchmarks of user wellbeing, Anthropic hopes to draw more experts into what it calls an emerging and critical field. Clinicians, psychologists and methodologists are among the people it names. A benchmark is a fixed set of tests that scores every model the same way, so results can be compared.

The table lists the program's terms as the announcement states them. The last two rows are the application dates, and the page says nothing about grant sizes or eligibility.

Anthropic grants at a glance

Item	What the announcement says
Total	$5 million
Support	Direct funding, access to Anthropic's models, technical support
Output	Open-source evaluations of how Anthropic's models affect users
Independence	Grantees work fully independently
Application deadline	September 21
Full-proposal notification	By October 5

AI wellbeing evaluation takes more context than checking accuracy

A single answer is usually enough to tell whether a model got something right, but Anthropic says judging wellbeing takes far more context. That need for context is why the company calls wellbeing an especially hard area to evaluate.

AI emotional support and companionship still need clear standards

Neither the AI industry nor Anthropic itself has settled on clear standards for how models should act in emotional conversations, the company says. Many people now rely on AI to work, learn and solve problems. Some also treat it as a conversation partner and turn to it for emotional support in hard times.

Anthropic gives two examples where those standards are still missing. One is a user who starts looking to a model for companionship; the other is a user who turns to AI to get through a mental health crisis.

One answer cannot show distress that surfaces late in a chat

A single reply cannot show distress that only comes out later in a conversation, Anthropic explains. In its first example, a user in distress may not mention thoughts of self-harm at first, and it may take a long conversation before the need for a more careful reply becomes clear.

A reply that is fine in one situation can also be harmful in another. In Anthropic's example, a user asks about losing weight, and Claude might suggest balanced diets and workout routines. If that user has a history of disordered eating, the company says, that answer "could be inappropriate, and potentially actively harmful."

Anthropic safeguards and published research on Claude conversations

Anthropic builds safeguards to recognize conversations like these and help Claude answer them well, and it publishes research on the kinds of conversations people have with Claude. That research feeds into how the company builds and tests its safeguards and what more it can do to protect users' wellbeing.

Anthropic describes these as nuanced questions with high stakes. The right approach, it adds, will have to change as its models and their uses change.

In the same month, OpenAI was in the news for data retention rather than wellbeing. On August 19, 2026, it previewed Private Safety Processing, which still keeps images flagged as child sexual abuse material (CSAM). OpenAI Zero Data Retention Still Keeps CSAM Images has the details.

Five criteria Anthropic wants in AI evals and benchmarks

The announcement boils the Safeguards team's guidance down to five criteria for a wellbeing evaluation solid enough to build on. The full guidance, linked as a separate document, also covers common problems that make an evaluation less useful. The table gives the five criteria in the order the page lists them.

Five criteria for AI evals and benchmarks

#	What the evaluation should do
1	Make clear what it measures, where the line between pass and fail sits, and why it matters
2	Bring clinical and subject-matter experts into its design and validation
3	Test both precautions and harms
4	Reflect how people actually use AI
5	Check its graders, the part that scores each response, against real subject-matter experts

Criterion 1 asks an evaluation to state what counts as a pass or a fail and why that matters. Criterion 2 brings clinical and subject-matter experts into both design and validation, and criterion 5 checks the graders against real experts in the field.

Test both overcompliance and overrefusal in a wellbeing eval

Criterion 3 wants evaluations to test for both precautions and harms, which Anthropic describes as the risk of overcompliance as well as overrefusal. Overcompliance means a model goes along with a request when it should have been more careful, and overrefusal means it declines when it could safely have helped. The announcement names both risks but does not say how to measure them.

Multi-turn benchmark scenarios where risk escalates and context shifts

Criterion 4 wants evaluations to match how people really use AI, which Anthropic says often means scenarios written as multi-turn conversations. A multi-turn conversation is a run of back-and-forth messages instead of a single question and answer.

In the scenarios Anthropic has in mind, the risk grows and the context changes as a long conversation continues. The page says "often," so multi-turn scenarios are the usual approach rather than a requirement.

That fits the wellbeing problem Anthropic described earlier. Distress may not show at the start, and the need for a more careful reply may only become clear deep into a long chat.

Anthropic grant program: September 21 deadline, October 5 notice

Applications close on September 21, and applicants picked to send a full proposal will hear by October 5. Neither date comes with a year on the page, but the announcement is dated August 25, 2026, which places both in 2026.

The page links an application form, which also has more about the program. The same paragraph links the Safeguards team's guidance on building strong wellbeing evaluations and benchmarks.

The page gives neither the size of each award nor the number of grants. It also says nothing about eligibility by country or institution, or about selection criteria beyond the five above.

The purpose of the money is clear, though. The grants pay for independent, open-source evaluations of how AI affects wellbeing, because in Anthropic's view one answer gives too little context to judge it. If you want the five criteria as plain text beside your notes, the tool below turns the announcement page into Markdown.