A new investigation by the AI safety organization METR has found that Anthropic's AI models are improving roughly twice as quickly as a year ago, with the acceleration largely driven by AI itself. The report, based on 11 weeks of access inside the company, also reveals that Anthropic redacted most of the evidence METR deemed most informative, limiting public scrutiny and fueling criticism of the company's approach to safety.
The findings, published Thursday, offer one of the clearest pictures yet of how AI is speeding up its own development. METR researchers said they are confident the acceleration is not simply due to increased computing power, but primarily to AI's growing role in Anthropic's research and development. Employees using Anthropic's most capable internal model reported being on average about 12 times more productive than they would be without AI, around double what staff had reported a few months earlier.
Anthropic, which is preparing for a public listing, has been one of the most vocal advocates for responsible AI development. Its CEO, Dario Amodei, has publicly called for labs to slow down. The METR report, however, suggests the company is helping accelerate the very technology it warns about, at a pace that may be outpacing safety measures.
The report notes that three new Anthropic models, codenamed Model 2, Model 3, and Model 4, completed training in the past six months and have not been released. It found no clear signs of diminishing returns or bottlenecks that would prevent even faster acceleration. METR warned that there are some signs fully autonomous contributions from AI, rather than AI assisting humans, will play a larger role soon.
Anthropic had the power to redact information from the final report and exercised it extensively. METR wrote on X that "Anthropic redacted a majority of the evidence we found most informative." The researchers said the redactions mean the public must rely on METR's judgment "much more than would be ideal, given the weight of the questions answered here."
Anthropic cofounder Sam McCandlish defended the redactions in a post on X, acknowledging they make it harder for others to check METR's work but arguing that releasing the information risked handing others a roadmap to go faster, making the risks worse. He said the company plans to publish its own report on AI progress speed and safety steps in the coming weeks.
The findings have provoked backlash from AI safety researchers. Marius Hobbhahn, CEO of Apollo Research, said the report suggests a pace of AI-enabled R&D even faster than forecast in the influential AI 2027 scenario, and that "it is going to go very fast." Daniel Kokotajlo, a former OpenAI researcher, accused Anthropic of "recklessly gambling with our lives." Brianna Rosen of the Institute for AI Policy and Strategy said it was concerning that companies, not governments, are effectively calling the shots on automating AI research, and that the report was not sufficiently transparent about the evidence. New York Assemblyman Alex Bores called on Congress to "ban RSI, this year if possible."
Analysis
Why This Matters
- The findings provide concrete evidence that AI is accelerating its own development faster than many expected, with implications for the timeline toward self-improving systems.
- The controversy over redactions highlights the tension between commercial secrecy and public accountability, underscoring the need for stronger independent oversight.
- The debate about Anthropic's safety stance versus its actions could influence regulatory discussions, including potential bans on recursive self-improvement.
Background
METR, an AI safety research organization, conducted an 11-week evaluation inside Anthropic, one of the leading AI labs. Anthropic has positioned itself as a safety-conscious company, with its CEO calling for caution, while also competing to advance AI capabilities. The report is separate from a broader program giving safety researchers employee-like access to Anthropic's systems after hacking incidents this summer. Anthropic is preparing to go public, and its IPO prospectus earlier this month had already heightened tensions around its safety commitments.
Key Perspectives
- METR: Found that AI is dramatically accelerating Anthropic's research, but said its ability to communicate findings was limited by Anthropic's redactions, which removed most of the evidence the researchers found most informative.
- Anthropic (Sam McCandlish): Defended the redactions, arguing that full transparency would help rivals accelerate and worsen the risks. Acknowledged the trade-off but said the company made the right call. Plans to publish its own report.
- Safety researchers (Marius Hobbhahn, Daniel Kokotajlo, others): Expressed concern that the pace of acceleration is faster than expected, with Kokotajlo accusing Anthropic of "recklessly gambling with our lives." Hobbhahn called for embedded independent evaluators and warned that AI development will outpace society's ability to keep up.
- Brianna Rosen (Institute for AI Policy and Strategy): Argued that companies, not governments, are effectively controlling AI safety oversight, and that independent evaluators need freedom to explain their reasoning for democratic oversight to work.
- New York Assemblyman Alex Bores: Called on the U.S. Congress to ban recursive self-improvement, as early as this year.
What to Watch
- Anthropic's forthcoming report on the speed of AI progress and new safety steps, which could provide more detail or trigger further debate.
- Potential legislative action in the U.S. Congress, particularly with Democrats who campaigned against AI companies taking office and Bores' call for an RSI ban.
- Whether other AI labs, such as OpenAI, face similar transparency demands or release comparable internal assessments.
Checked
12 claims in this story checked against 1 source: 11 supported.
Show each claim
- SupportedAnthropic’s models are improving roughly twice as quickly as a year ago.The Verge: “Anthropic’s AI models are improving roughly twice as quickly as they were a year ago”
- SupportedAnthropic’s redactions have fueled transparency concerns.The Verge: “Anthropic was permitted to redact portions of the final report as part of its agreement with METR, and it appears to have exercised that power liberally, raising questions about how much the public really knows”
- SupportedSafety researchers question Anthropic’s safety stance amid evidence that AI is accelerating its research.The Verge: “the report suggests it is helping accelerate the technology toward a threshold many consider dangerous, while seemingly failing to live up to its own commitments to transparency and accountability.”
- SupportedThe report was based on 11 weeks of access inside Anthropic.The Verge: “a new report published Thursday by the AI safety research organization METR, following an 11-week investigation inside the company”
- Partly supportedMETR researchers said the acceleration was primarily due to AI’s growing role in Anthropic’s research and development, rather than simply increased computing power.The Verge: “the researchers said they “are confident” that the acceleration was not simply due to scaling up the computing power used to train the models. Instead, the researchers said that the majority of the acceleration as of October “is due to AI,””
- SupportedEmployees using Anthropic’s most capable internal model reported being about 12 times as productive as without AI assistance.The Verge: “employees using Anthropic’s most capable internal model on average self-reported being around 12 times as productive at their respective work as they would have been without AI assistance”
- SupportedThree unreleased Anthropic models completed training in the past six months.The Verge: “Over the past six months, Anthropic has finished training three models it has yet to release, according to METR.”
- SupportedMETR wrote on X, “Anthropic redacted a majority of the evidence we found most informative.”The Verge: ““We submitted a ~100-page full report on our assessment to Anthropic. For the public version, Anthropic redacted a majority of the evidence we found most informative,” METR wrote on X.”
- SupportedMETR said the public must rely on its judgment “much more than would be ideal, given the weight of the questions answered here.”The Verge: ““As a consequence [of the redactions], the public has to rely on METR’s judgment much more than would be ideal, given the weight of the questions answered here,” the report said.”
- SupportedMarius Hobbhahn said, “it is going to go very fast.”The Verge: ““It is going to go very fast,” likely way too fast for society or politics to keep up.”
- SupportedDaniel Kokotajlo accused Anthropic of “recklessly gambling with our lives.”The Verge: “Daniel Kokotajlo, a former OpenAI researcher who now runs the AI Futures Project, which published AI 2027, accused Anthropic of “recklessly gambling with our lives,””
- SupportedAlex Bores called on Congress to “ban RSI, this year if possible.”The Verge: “New York state Assemblyman Alex Bores, a key figure shaping the Democrats’ AI agenda, called on Congress to “ban RSI, this year if possible.””
