OpenAI Cuts Inadequate Distress Responses by Up to 80% in GPT-5 Update

OpenAI reduced responses that fall short of its desired behavior in sensitive, distress-related conversations by 65–80% with an update deployed on October 3. The figure comes from an addendum to the GPT-5 system card that the company published alongside a related blog post, offering baseline safety evaluations for the first time in this category.

The update improves ChatGPT's default model — also known as GPT-5 Instant — in how it handles moments of mental and emotional distress. OpenAI said it worked with more than 170 mental health experts to achieve the gains.

The stakes are straightforward. ChatGPT is one of the most widely used AI systems in the world, and its default model is what hundreds of millions of users encounter when they open the app. How that system responds when someone signals despair, crisis, or emotional pain is a safety question of a different order than ordinary chatbot quality issues. Regulators, researchers, and safety advocates have pressed AI developers for measurable evidence on exactly this point, and the addendum represents OpenAI's attempt to provide it.

What changed on October 3?

The October 3 update made three specific behavioral improvements, according to OpenAI:

• Recognition. ChatGPT can more reliably recognize signs of distress.
• Response. The model responds with care in those moments.
• Guidance. It directs people toward real-world support rather than leaving them with the conversation alone.

The headline metric — a 65–80% reduction in responses that fall short of desired behavior — covers this combined set of failure modes. OpenAI did not break the range down further in the addendum, so the exact per-category gains are not specified.

How does OpenAI measure 'falling short'?

The evaluations in the addendum compare two model versions:

• The August 15 version of ChatGPT's default model (GPT-5 Instant)
• The updated model launched October 3

This pairing gives the addendum its baseline. By publishing evaluations of the pre-update model alongside the post-update one, OpenAI is letting readers see where the model stood before the intervention — the baseline that the 65–80% improvement figure is calculated against.

The company framed the addendum as a continuation of commitments made at launch. "When we launched GPT-5, we noted in the system card that we were working to establish better benchmarks and to continue to strengthen model safety in areas related to mental and emotional distress," OpenAI wrote. The October 3 deployment, the company said, "reflected those efforts."

Why does this matter beyond one model update?

Safety documentation of this kind has become a contested area in AI. System cards — the technical reports OpenAI publishes with major model releases — are the primary public record of what a model can and cannot safely do. Critics have argued that such documents often lack measurable baselines, making improvement claims hard to verify. An addendum that publishes explicit before-and-after evaluations addresses that criticism directly, at least for this category.

The involvement of more than 170 mental health experts also signals how OpenAI is sourcing its safety judgments: not purely from internal red-teaming, but from clinicians and researchers with domain expertise in emotional distress. That number is unusually large for a single safety workstream, and it reflects the breadth of scenarios — from passing sadness to acute crisis — that a general-purpose chatbot must handle.

The commercial context matters too. ChatGPT conversations frequently touch on emotional topics, whether or not users intend them to. A default model that mishandles a distress signal can cause real harm; one that responds well and points to real-world support can play a constructive role. OpenAI's framing keeps that boundary explicit: the goal is recognition, care, and referral — not therapy.

What did the original system card promise?

When GPT-5 launched, its system card flagged mental and emotional distress as an area of active work. OpenAI committed to two things at that time:

• Establishing better benchmarks for distress-related behavior
• Continuing to strengthen model safety in those areas

The addendum closes the loop on both commitments, at least partially. It supplies the baseline evaluations that make benchmarking possible, and it documents the October 3 update as the concrete safety improvement the company had promised.

The related blog post OpenAI published alongside the addendum gives additional detail on the work, according to the company's announcement. Readers seeking the full methodology — how distress scenarios were constructed, how expert reviewers scored responses, and how the 65–80% reduction was computed — will find the primary material there and in the addendum itself.

What remains open?

Several questions sit outside the addendum's scope. The 65–80% figure describes improvement against OpenAI's own definition of "desired behavior"; the company has not published an absolute failure rate for the updated model. The evaluations cover the default model, GPT-5 Instant, and the addendum does not address other variants in the GPT-5 family. And the company has not said how frequently distress-related conversations occur in practice, which would determine the real-world scale of the improvement.

What the document does establish is a template: publish the baseline, name the experts, quantify the gain, and date the change. If OpenAI applies that template to other safety categories flagged in the GPT-5 system card, expect future addendum-style releases to become the standard way the company evidences post-launch safety work.