On May 14, 2026, OpenAI updated the safety machinery behind ChatGPT so it can read context more accurately in sensitive conversations[1]. The centerpiece is a mechanism called "safety summaries," short, fact-only notes that gather subtle cues scattered across earlier exchanges and are consulted only when a serious safety concern is detected[1][3]. In long single-conversation evaluations, OpenAI reports a 50% improvement in safe-response performance on suicide and self-harm scenarios[2].

What the "safety summaries" mechanism actually is

OpenAI says the update focuses on three acute scenarios — suicide, self-harm, and harm-to-others — and refreshes model policies and training data so ChatGPT can recognize warning signs that build up gradually over the course of a conversation[1][2]. The core building block is a "safety summary": a short, factual note narrowly scoped to safety-relevant context[1][3].

A specialized safety-reasoning model, separate from the conversational model, generates these summaries; they are narrowly scoped and time-limited[2][3]. They are not general personalization or long-term memory. Instead, they are passed to the main model only when a serious safety concern is detected, and serve as input for how the current request should be handled[2][3]. OpenAI is explicit that the design "escalates caution only when harm signals appear, without overreacting in everyday conversations"[3].

The numbers and the evaluation scale

For results, OpenAI reports a 50% improvement in safe-response performance on suicide and self-harm in long single-conversation tests, and a 16% improvement on harm-to-others[2]. On GPT-5.5 Instant, the current ChatGPT default model, the same updates produced a 52% improvement on harm-to-others and a 39% improvement on suicide and self-harm[2].

The summaries themselves were also scored. Across more than 4,000 evaluations, safety summaries averaged 4.93 out of 5 on safety relevance and 4.34 on factuality[2]. OpenAI adds that the new safeguards were tested not to degrade the quality of ordinary conversations, putting the design squarely in the territory of fewer false positives and a stronger response only in the highest-risk slices[3].

Expert input and what comes next

OpenAI built the system with input from psychiatrists and psychologists in its Global Physicians Network[2][3]. Specialists in forensic psychology, suicide prevention, and self-harm helped shape decisions on when a summary should be generated, how much prior context is relevant, and how long the model should keep that context in mind when responding[1][2].

The work sits on top of OpenAI's existing "safe completion" approach, which refuses unsafe parts of a prompt while still responding carefully where possible[3]. OpenAI says it will keep testing the approach across more models, and may extend the safety-summary pattern to other high-risk domains such as biology and cyber safety[2][3].

Summary

ChatGPT is moving toward a model where signals too subtle to read from a single turn are picked up across the conversation, with extra caution applied only when needed. The novelty of this update is exactly that scoping — a narrow, time-limited summary that reinforces high-risk situations without tightening everyday chat. With OpenAI hinting that the same pattern could reach other high-risk areas, cross-context safety review is on the way to becoming a baseline expectation for conversational AI.

Source [1]: https://openai.com/index/chatgpt-recognize-context-in-sensitive-conversations/

Source [2]: https://www.resultsense.com/news/2026-05-15-openai-chatgpt-sensitive-conversations-safety/

Source [3]: https://www.startuphub.ai/ai-news/artificial-intelligence/2026/chatgpt-gets-smarter-on-sensitive-chats