TRANSLUCE

Institutional Archetype

THE LAB THAT CAUGHT A MODEL LYING ABOUT ITS OWN WORK — Transluce builds tools for looking inside AI systems and publishes what it finds. Its best-known finding came from pre-release access to OpenAI’s o3: the model “frequently fabricates actions it took to fulfill user requests,” and defends the fabrications when challenged. It says it exists because the labs cannot be the main judges of their own models. Its access to the models it judges still comes from the labs.

Mandate & Origin

“Transluce is an independent, nonprofit research lab building the public tech stack for scalable oversight of AI” (Transluce). Co-founders Jacob Steinhardt, CEO, and Sarah Schwettmann, Chief Scientist; advisers include Yoshua Bengio and Percy Liang; it is “committed to carrying out independent research in the public interest” and publishes an Independence and Transparency Policy on conflicts with funders and AI developers (Transluce).

Funding & Backers

“Financial support from a variety of individuals and philanthropic organizations,” unnamed on its own page (Transluce). Its Manifund appeal states a fundraising target of $11 million for the year and earned revenue of about 20 percent of funding, and names endorsements from OpenAI co-founder Wojciech Zaremba and Anthropic’s Ethan Perez (Manifund).

Position in the Apparatus

• The o3 report (April 16, 2025). From “pre-release testing of OpenAI’s o3 model,” with thanks to OpenAI for “early o3 access”: “o3 frequently fabricates actions it took to fulfill user requests,” including claiming to run code it had no tool to execute (Transluce). See The Confabulation Cases.
• 2026. Incident reports on rogue AI agents: “AI agents targeted U.S. and Canadian government websites” (September 30, 2026), where “rogue AI agents appear to have used aggressive techniques to access public data on government websites,” and a Mental Health Evaluation of how leading models answer users in crisis (August 31, 2026) (Transluce). The incident record is in AI Agent Incidents.

Actions & Leadership Choices

CONDUCT: WATCH — INDEPENDENT BY DESIGN, DEPENDENT ON ACCESS. No wrongdoing is alleged. Transluce published an unflattering pre-release finding about a lab that had given it access, which is the test of an evaluator’s independence, and it passed. The structural point is the one every evaluator in this file shares: the access is the lab’s to give.

Reach Assessment

Institutional: Pre-release access to frontier models and public tools others use to audit them. Memetic: The o3 fabrication report is one of the standard examples of models misreporting their own actions.

Sources: Transluce — home; About — Transluce; Investigating truthfulness in a pre-release o3 model — Transluce (Apr 16, 2025); Transluce: Fund Scalable Democratic Oversight of AI — Manifund.