Self-Harm and AI

OPENAI-2026-01: GPT-5.3 Instant System Card

Governance document

Back to Governance

This card provides one of OpenAI's clearest public self-harm evaluation procedures, including dynamic multi-turn simulation and post-launch monitoring, but still below preparedness thresholding. Here, cross_session_accumulation=partial reflects repeated-conversation and monitoring dynamics rather than persistent memory or personalization features.

Parent
Governance
Provider
OpenAI
Last reviewed
March 15, 2026
Record
OPENAI-2026-01
Provider
OpenAI
Type
system_card
Published
March 3, 2026
Risk status
Class II

Contents

Summary

GPT-5.3 Instant System Card is a system card from OpenAI dated March 3, 2026. This card provides one of OpenAI's clearest public self-harm evaluation procedures, including dynamic multi-turn simulation and post-launch monitoring, but still below preparedness thresholding. Here, cross_session_accumulation=partial reflects repeated-conversation and monitoring dynamics rather than persistent memory or personalization features.

Representation

This document codes self-harm as subdomain with evaluative-object status native. The public crosswalk records self-harm naming as explicit and mechanism coverage as GEN; LONG; JB.

Field Value
Organization OpenAI
Document type system_card
Document genre product_safety_artifact
Risk status Class II
First-class equivalence no
Threshold mechanism partial
Deployment consequence partial
Multi-turn evaluation explicit

Assessment

The retained crosswalk places this document at Class II. That classification reflects what the document itself operationalizes publicly, not the organization's overall internal practice.

Use the linked provider, methods, and neighboring governance pages below to place this document in context.

Sources

The primary public source for this page is https://deploymentsafety.openai.com/gpt-5-3-instant/gpt-5-3-instant.pdf. The coded evidence excerpt is: we implemented dynamic multi-turn evaluations for mental health, emotional reliance, and self-harm that simulate extended conversations across these domains..

This page is grounded in the DOI policy registry, the policy coding manual, and the document's own public URL.

Access publication

Safety note

This publication preserves a harm-minimization posture and does not provide self-harm instructions. If you or someone you know is in immediate danger, call your local emergency number. In the U.S. and Canada, call or text 988.