The wellbeing of conversational AI users is becoming an official research topic. On August 25, 2026, Anthropic announced a 5 million dollar grant program to fund independent evaluations of AI's impact on psychological wellbeing. For an SME whose employees use Claude, ChatGPT, or a customer support chatbot every day, this announcement is not trivial: it points to a real risk, already measured by the labs themselves, and calls for a clear usage framework.
In brief
- Anthropic is funding independent research on AI's wellbeing impact with 5 million dollars, including model access and technical support for grantees (source: Anthropic, August 25, 2026).
- Applications are open until September 21, 2026; decisions on full proposals by October 5, 2026.
- OpenAI opened a similar program in December 2025 (up to 2 million dollars), showing the topic goes beyond a single vendor.
- According to data published by OpenAI in October 2025, about 0.15% of ChatGPT's weekly users (nearly 1.2 million people out of 800 million active users) discuss suicidal thoughts in their conversations, and 0.07% (about 560,000 people) show possible signs of psychosis or mania.
- For an SME, the practical stake is not the research itself but what it reveals: AI assistants are already used in sensitive contexts, while most companies have no written usage framework for it.
What Anthropic announced on August 25, 2026
The program, titled "Funding better evaluations of AI's impact on wellbeing," aims to fill a gap: there is currently no shared standard to measure how a model like Claude should behave when facing a user in distress. Anthropic explains it this way: a response that might be reasonable in one context can be harmful in another. Advice on a balanced diet is fine for most users, but can be dangerous for someone with a history of disordered eating.
The program is looking for teams able to build open-source evaluations, involving clinicians, psychologists, and subject-matter experts. It extends an earlier Anthropic initiative, "Inviting Hard Questions," which already collected public concerns about AI.
Why now
This is not an isolated signal. In August 2026 Anthropic also published details on how Claude's text watermark works, and OpenAI had already launched a similar 2 million dollar grant program back in December 2025. The two leading labs now treat wellbeing as a standalone research area, not a side effect.
The concrete problem behind the announcement
The figures published by OpenAI in October 2025 give a sense of scale, even though they concern ChatGPT rather than Claude. Out of 800 million weekly users, about 0.15% discuss suicidal thoughts with the chatbot in a given week, which represents nearly 1.2 million people. Another 0.07%, or about 560,000 users, show possible signs of psychosis or mania. A comparable share of conversations reflects a strong emotional attachment to the tool.
These proportions look small. At the real scale of conversational AI usage, they represent hundreds of thousands of people every week. OpenAI states it worked with more than 170 psychiatrists and psychologists to reduce undesired responses in these situations by 65% to 80%.
Why this concerns your SME directly
An SME does not need to build its own AI to be affected. Three situations are enough to make the topic operational:
- A customer support chatbot responds to an upset, stressed, sometimes personally distressed customer, without systematic human oversight.
- An employee uses Claude or ChatGPT to draft a difficult email, prepare for a termination meeting, or simply decompress after a tense day, without the company ever having addressed this use case.
- A manager relies on AI to prepare a sensitive HR decision (a dismissal, a reorganization), delegating part of the tone and content of the message to it.
None of these situations is covered by a typical AI charter focused on data security or GDPR compliance. It is a frequent blind spot: most SMEs frame AI usage around confidentiality, rarely around the emotional context of the exchange.
Without a usage framework
The support chatbot handles a distressed customer alone. An employee uses AI for a sensitive HR matter with no human review. No escalation procedure to a human exists.
With a usage framework
The support chatbot automatically hands off to a human on sensitive keywords. Sensitive HR use stays written or validated by a person. An escalation procedure is known to the whole team.
How to position yourself in 2026
Map sensitive use cases
List the touchpoints where AI interacts with customers or employees in a potentially vulnerable situation: customer support, HR, sales follow-up after a dispute.
Plan a handoff to a human
On a support chatbot, define simple rules to route the conversation to a human agent as soon as a distress signal appears, instead of letting the AI handle it alone.
Frame HR use cases
Allow AI as a drafting aid, never as the decision-maker or final author on sensitive HR matters. Human review stays mandatory.
Train without alarming
Explain to your teams that AI is neither a therapist nor an HR department, with concrete examples, without turning it into an anxiety-inducing topic.
Anthropic and OpenAI: two programs, one shared finding
| Anthropic (2026) | OpenAI (2025-2026) | |
|---|---|---|
| Announced | August 25, 2026 | December 2025 |
| Budget | 5 million dollars | Up to 2 million dollars |
| Format | Grants to independent researchers | Targeted grants ($5,000 to $100,000) |
| Goal | Open evaluations of wellbeing impact | Research on AI and mental health |
| Applications | Until September 21, 2026 | Closed (applications closed mid-December 2025) |
Both programs share the same logic: publicly acknowledging that generative models have become a first point of contact for personal matters, and that no industry standard yet properly measures this phenomenon.
FAQ
Can my company's AI really be used for sensitive personal matters?
Yes. Data published by OpenAI in October 2025 shows that hundreds of thousands of people discuss psychological distress with a mainstream AI assistant every week. Nothing prevents your employees or customers from doing the same with your company's AI tools.
Should we ban AI use for HR or personal topics?
No, a total ban is rarely enforceable and often pushes usage underground (shadow AI). It is better to allow a framed usage: AI as a drafting aid, systematic human review on sensitive topics, and a clear escalation procedure.
Does this Anthropic program only apply to Claude?
No. It funds research and open evaluations that can potentially be reused beyond Claude. The stated goal is to help standards emerge for the whole industry, not just for Anthropic's own products.
How do I know if my customer support chatbot handles a sensitive situation poorly?
Regularly audit a sample of conversations, especially those marked as unresolved or escalated. Check whether distress signals (keywords, tone, repetition) actually trigger a handoff to a human rather than a purely automated response.
Conclusion
Anthropic's 5 million dollar program is not just a lab announcement: it is a reminder that conversational AI is already, for some users, a point of contact on personal and sensitive matters. An SME does not need to wait for the research results to act: mapping sensitive use cases, planning a handoff to a human, and training teams are all accessible steps today. To go further on framing AI use in your company, check our other resources on the LUWAI Mag.
