AnthropicがAIのウェルビーイングへの影響を評価する500万ドルの助成プログラムを開始Funding better evaluations of AI’s impact on wellbeing

Anthropic has announced a $5 million grant program to fund independent research into how artificial intelligence affects the wellbeing of the people who use it. The initiative matters because conversational AI systems are now embedded in the daily routines of millions of users, yet the industry still lacks widely agreed-upon, rigorous ways to measure whether those interactions leave people better or worse off.

According to the company, the program will provide grantees with three forms of support: direct funding, access to Anthropic's models, and technical assistance. In return, researchers are expected to build open-source evaluations, meaning standardized tests and measurement tools that the broader AI industry can adopt to gauge how systems influence users' wellbeing. By committing to release this work openly, Anthropic appears to be framing the effort as shared infrastructure rather than a proprietary advantage, which lowers the barrier for other developers and academics to reuse or extend the methods.

Evaluations, frequently shortened to "evals," have become a central concept in modern AI development. They are structured procedures for probing a model's behavior under specific conditions, and they underpin claims about capability, safety, and reliability. Most existing benchmarks focus on measurable tasks such as coding, mathematics, or factual accuracy. Wellbeing is considerably harder to quantify, since it involves subjective experience, long-term effects, and outcomes that vary across individuals and contexts. Building credible evaluations in this area is likely to require input from psychology and the social sciences, not only machine learning, which may explain why Anthropic is routing the work through external, independent researchers rather than handling it entirely in-house.

The wellbeing focus reflects a set of concerns that have grown more prominent as chatbots become more capable and more personal. Researchers and clinicians have raised questions about issues such as sycophancy, where a model tells users what they want to hear rather than what is accurate or helpful, as well as emotional over-reliance, the reinforcement of harmful beliefs, and the effects of prolonged use on vulnerable people. Independent, open evaluations could give companies, regulators, and the public a more objective basis for assessing these risks, though the program's ultimate influence will depend on whether the resulting tools are adopted beyond Anthropic itself.

The announcement fits within Anthropic's broader positioning as a safety-oriented AI developer. The company, founded by former OpenAI staff, has built its public identity around research on alignment and responsible deployment, including its Constitutional AI approach, in which model behavior is guided by a written set of principles, and its Responsible Scaling Policy, which ties capability increases to safety commitments. Funding external wellbeing research extends that stance into a domain that has received comparatively less structured measurement than raw capability or misuse.

The move also arrives amid wider industry attention to the human effects of AI companions and assistants. Other major labs have expanded their own safety and trust teams, and platforms offering AI chat experiences have faced scrutiny over how they handle sensitive conversations, minors, and mental-health topics. Grant programs are a familiar mechanism in this space: technology companies and philanthropies have previously funded academic work on AI safety, interpretability, and evaluation, in part to build external expertise and in part to shape emerging norms. Directing grants toward wellbeing evaluations specifically appears intended to seed a research community around metrics that currently do not exist in standardized form.

Several practical questions remain unaddressed in the initial announcement, including how grantees will be selected, what the size of individual awards will be, and how progress will be reviewed. It is also not yet clear how the open-source evaluations will be governed or maintained over time, which can be a challenge for shared research tools once initial funding ends. Independent measurement funded by a model developer inevitably raises questions of independence as well, and the credibility of the outputs will likely rest on transparent methodology and genuine researcher autonomy.

For now, the program represents a concrete financial commitment to a field that has been discussed more than it has been measured. If the resulting evaluations are adopted across the industry, they could help move debates about AI's effects on people from anecdote toward evidence, giving developers and outside observers a common yardstick for a question that has so far lacked one.