Brundage argues METR redaction harm to public understanding is overstated as transparency debate escalates

Miles Brundage, former VP of Policy at OpenAI, posted a series of threads on November 11 about the redaction controversy surrounding METR's evaluation report on Anthropic's frontier models. He argued that the public has generally overestimated how much the redacted content harms public understanding of AI progress, noting that many critics had advocated policies like slowing down (pacing) based on even less evidence before the report was released—so holding stronger views now with more evidence is logically inconsistent. Brundage also contended that requiring AI companies to publish "unredacted" safety reports is both unrealistic and contrary to historical experience: what actually updates public perception of AI tends to be safety incidents, public narratives, and firsthand experience, not technical data that only a few care about.

Confirmed

• Brundage said he has no inside information about this particular redaction, but based on METR's statements, at least some "over-redaction" is plausible, and he expects more information once Anthropic publishes the report
• He supports the public release of AI progress data, noting that OpenAI and Anthropic had shared such information before this report, and independent organizations like METR and Epoch have done excellent work in this area
• Attorney Nathan Calvin pointed out that AI Futures Project's (AIFP) latest forecasts were dramatically revised after the METR report, strengthening his belief that one of Anthropic's most important and beneficial actions would be to immediately disclose more unredacted information from the report, or be compelled to do so by government

Why it matters

• The dispute reflects a core divide over "transparency norms" in AI safety evaluations: one side holds that redaction hinders external oversight and forecast calibration, while the other argues unredacted disclosure is unrealistic and public perception is event-driven
• The METR report had a substantive impact on conclusions at professional forecasting organizations like AIFP, showing that redacted content—even when not aimed at the general public—directly affects the professional ecosystem for assessing frontier AI progress, and the controversy may further increase external pressure on Anthropic to disclose more

2026-11-11 ~ 2026-11-11 · 6 related posts

• METR report reshapes AIFP forecasts; experts urge Anthropic to un-redact more — eli_lifland · 2026-11-11
• [source] Ex-OpenAI policy VP Miles Brundage: METR report redactions hurt public understanding less than claimed — Miles_Brundage · 2026-11-11
• [source] Brundage: public AI beliefs update on incidents and narratives, not technical reports — Miles_Brundage · 2026-11-11
• [source] Ex-OpenAI advisor Brundage: no-redaction AI safety reports are unrealistic — Miles_Brundage · 2026-11-11
• Miles Brundage: METR report redactions may include at least some over-redaction — Miles_Brundage · 2026-11-11

1 near-duplicate retellings: zetalyrae