A Bigger Pool, the Same Design
Anthropic will spend $5 million funding independent, open-source research into how Claude affects user wellbeing, covering companionship-seeking behavior, mental-health-crisis guidance, refusal calibration, and risk that escalates across long conversations. The company published the program on August 25, 2026, with applications due September 21 and grantees notified by October 5. Grantees keep full independence over their findings and publish everything as open source that any developer can use, a genuinely stronger transparency commitment than most lab-funded safety research carries.
The pool is 2.5 times the size of a similar program OpenAI ran eight months earlier. OpenAI's AI and Mental Health Grant Program opened December 1, 2025 and closed December 19, capping the total pool at $2 million and individual awards at $5,000 to $100,000 each. For a UK-based deployer weighing both numbers, that is roughly GBP 3.9 million against roughly GBP 1.6 million, a real difference in size that leaves control of the underlying data exactly where it was.
What the Grantwashing Critique Already Said
Researchers J. Nathan Matias and Avriel Epps published a critique of OpenAI's program in Tech Policy Press on December 18, 2025, calling it grantwashing, eight months before Anthropic announced its own. Their central comparison was blunt: the median mental-health research grant awarded by the US National Institute of Mental Health in 2024 was $642,918, more than six times OpenAI's entire top individual award and well over a hundred times its floor.
Money was only part of the complaint. Matias and Epps argued that labs hold the one resource that would let outside researchers actually measure real-world harm - the usage logs and interaction telemetry showing how people talk to these systems day to day - and that neither program hands any of it over. Their proposed fix was structural: they called on AI companies to commit 3 to 5 percent of their annual research budgets to genuinely independent safety science, the same order of magnitude the Human Genome Project set aside for ethical research alongside the science itself.
Why EU and UK Compliance Files Need More
European regulators built a binding process for exactly this kind of harm claim before Anthropic's grants existed. On July 10, 2026, the European Commission issued a preliminary finding that Instagram's and Facebook's addictive design breaches the Digital Services Act, naming infinite scroll, autoplay, push notifications, and personalised recommenders directly. The Commission's complaint centers on Meta's failure to adequately assess and mitigate the wellbeing risk those features create for users, including minors.
That is the standard a $5 million grant program gets measured against once an EU or UK company cites it in a DSA risk assessment or an AI Act Article 5 compliance file. Anthropic still picks who gets funded, still controls model and telemetry access, and still defines what counts as rigorous - the same arrangement Matias and Epps described eight months earlier as built to forestall binding rules rather than satisfy them.
Read next: 1,134 Lab Staff Say The Brake Does Not Exist | Anthropic's Safety Report Lists Its Own Failures
