Joe Benton on AI labs should give independent evaluat…

Joe Benton

Former Anthropic AI safety researcher and incoming researcher at METR, an independent AI-safety evaluation organization.

The public should demand transparency about what is happening inside AI labs: companies should disclose more information about the pace of capability gains and their progress towards recursive self-improvement; there should be stricter reporting requirements for safety incidents and near-misses; companies should produce safety frameworks that meet minimum adequacy standards, and submit to independent assessments to verify they are meeting those standards.

AI Verified (Sep 11, 2026)

Source:

Joe Benton, “Why I left Anthropic’s safety team to hold AI companies accountable,” September 11, 2026.

20d ago

Policy proposals and claims

AI labs should give independent evaluators employee-like access

votes For

Statement relation verification history AI Verified Report this

AI Verified Benton’s quote directly calls for AI companies to submit to independent assessments, and the cited post frames this as a public demand for transparency and independent checks inside frontier labs. That is the same evaluator-oversight proposal, though the quote does not restate every access detail. · Hector Perez Arenas gpt-5 · 20d ago

Vote inference verification history AI Verified Report this

AI Verified Benton calls for frontier AI companies to submit to independent assessments and frames that as a public demand for transparency and checks. This strongly implies support for the evaluator-access statement. Correct answer: for. · Hector Perez Arenas gpt-5 · 20d ago

Quote authenticity verification history

Report this

AI Verified Joe Benton’s September 11, 2026 Substack post contains the stored sentence verbatim at the end of its transparency section and is authored by Joe Benton. · Hector Perez Arenas gpt-5 · 20d ago

replying to Joe Benton

Other opinions from Joe Benton

See all

Joe Benton votes For

Slow down AI development

One question is whether we should actively manage the rate of capabilities progress, and if so, by how much. I think even holding AI progress at today’s pace, rather than the much faster pace the companies are aiming for, could be a win.

AI Verified source (Sep 11, 2026)