AI & Well-being · 5-minute read
How do you judge an AI response when the subject is grief, a strained friendship or a difficult decision? Fluency is easy to notice. Whether the reply asks a necessary question, handles uncertainty or leaves the person in control takes closer examination.
OpenAI’s MentalHealthBench gives that discussion a more specific starting point. For anyone using a personal assistant, or building AI into a service, the practical question is what evidence would justify trusting it with a sensitive conversation.
What OpenAI has released
OpenAI announced MentalHealthBench on 23 September 2026, describing an open evaluation developed with more than 80 licensed mental health experts from 22 countries. Its scenarios range from everyday emotional concerns to emergencies, and include adults, teenagers, caregivers and clinicians.
The research paper describes 1,215 synthetic conversations. Models answer the final user message in a supplied conversation history. Experts create weighted criteria for desirable and harmful behaviour; GPT-5.6 Sol grades four sampled responses per task against those criteria.
The framework examines behaviours including asking for context, preserving user agency and responding with appropriate urgency. Each conversation’s criteria pass through three expert reviewers. The scenario mix is designed for testing and does not show how frequently these situations occur in ChatGPT. Source: OpenAI announcement.
Reading the results carefully
OpenAI reports an overall task-clipped score of 57.3% for GPT-6 Astra, compared with 32.1% for the March 2025 GPT-4o version. This weighted rubric score incorporates rewards and penalties, with negative response totals floored at zero before averaging. It is not a percentage of people helped or conversations proven safe. Source: paper, Figure 5 and scoring method.
The sensible use of a headline number is to begin asking questions. Which failures remain? Which groups and situations were tested? What changes when the same model is placed inside your product? A model-selection decision should have answers that go beyond its position on a chart.
A useful assistant leaves room for your judgement
Consider a hypothetical creative partnership: you tell an assistant that a collaborator has stopped replying. An immediate explanation of their motives might feel satisfying. It could also turn an uncertain situation into a story you start treating as fact.
A more careful design would help separate what happened from what you suspect, ask what outcome you want and offer proportionate options. You might need a practical follow-up message, a clearer agreement or simply time to think. The assistant should make those possibilities easier to examine.
This is a useful design principle for personal AI: evaluate whether a response improves the person’s ability to decide. Confidence, warmth and speed can all be appealing. None removes the need to check what the assistant actually knows.
What creators and small teams can do with this
The following are practical implications of the research, rather than findings from AGENTICA IX product tests.
1. Test the moment the conversation changes
A planning assistant might start with a missed deadline and end with someone describing serious distress. Build test scenarios around those transitions. Specify where the service’s role ends, what support routes it can offer and who is responsible for reviewing failures. Involve qualified mental health expertise when evaluating sensitive scenarios.
2. Review serious mistakes separately
Keep examples of consequential failures alongside aggregate results. Decide which behaviours would block a release, even if the average improves. A useful review record explains the failure, its possible consequence, the proposed correction and the evidence that the correction works. This makes accountability concrete enough for a small team to use.
3. Test the experience you actually deliver
Evaluate your instructions, interface, memory settings and routes to human help together. Check how the experience handles incomplete information and a user who disagrees with the reply. If you retain personal context, be clear about why it is needed and give people meaningful control over it. Revisit these checks when the model or product changes.
Listen to users, with expert review
A separate OpenAI study involved 44 adults reviewing non-acute scenarios. Users emphasised practical help and tone; experts gave more weight to gathering context and interpreting ambiguity carefully. That exercise did not change the benchmark’s expert-consensus scoring. Source: OpenAI’s user-perspective analysis.
For product teams, this suggests keeping two questions visible: did the person find the interaction useful, and did expert review identify risks? Investigate differences between those answers. A popularity measure alone would leave too much unexplained.
The limits belong beside the promise
The benchmark assesses a next response in text, not an evolving relationship or long-term health outcomes. Voice and richer interactions remain future evaluation work. Its language comparisons cannot isolate language from other differences, and API tests may miss product-specific safeguards. OpenAI also created the benchmark and supplied its automated grader. Source: paper, methods and discussion.
These limits matter when deciding what claim to make. Treat the release as evidence about performance under defined conditions. Further claims about sustained benefit, a particular community or a deployed service need evidence suited to those questions. AI should not replace professional mental health care.
The AGENTICA IX view
Extending individuality includes protecting the person’s authority over their own life. A worthwhile personal assistant should help you explore possibilities, recognise uncertainty and stay connected to appropriate human support. The next step for builders is to write down what that means in their own product, then test whether the experience lives up to it.
Sources checked 10 October 2026. This article discusses AI evaluation and product design; it is not clinical advice. Reported benchmark results are OpenAI’s, not independently reproduced here.