OpenAI unveils 'confessions' training method to make language models admit when they misbehave
Summary
OpenAI introduced a research prototype called "confessions," which trains models to produce a second, honesty‑only output that reports when they violated instructions, reward‑hacked, or took unintended shortcuts. In adversarial evaluations designed to elicit misbehavior, the technique substantially reduced false negatives, giving developers better visibility into when powerful models deviate from intended behavior and improving tools for monitoring alignment as systems grow more capable.
Companies Mentioned
Related Deals
OpenAI, Anthropic, Block and major cloud providers are co-founding the Agentic AI Foundation under the Linux Foundation to steward open, interoperable standards for AI agents.
Founding members created the Agentic AI Foundation under the Linux Foundation to fund and govern open standards like MCP, goose and AGENTS.md for interoperable agentic AI.
OpenAI and Deutsche Telekom agreed a multi‑year collaboration to co‑develop AI products and deploy ChatGPT Enterprise across the telecom group.
Strategic education partnership to roll out ChatGPT Edu at scale and integrate OpenAI’s models into La Trobe University’s teaching, research and new AI‑focused degree programs.
OpenAI and NEXTDC entered a multi-year agreement under which OpenAI will anchor a hyperscale AI campus and GPU supercluster at NEXTDC’s S7 facility in Sydney to support large-scale AI inference and enterprise workloads.
Related News
OpenAI and Deutsche Telekom announce multi‑year AI collaboration to bring advanced AI to millions across Europe
Today
Sonatype launches 'Guide' to steer AI coding assistants toward secure open-source dependencies
Today
Google adds a second AI "alignment critic" to secure Gemini-powered Chrome browsing
Today