Report Shows Anthropic Using AI to Accelerate Research, Raising Alarms

AI is dramatically speeding up Anthropic’s research, even as the company calls for slowing down AI development.

Anthropic is using AI to build more powerful AI, and this approach is proving highly effective. The company's models are improving roughly twice as quickly as they were a year ago, showing no signs of slowing down.

This conclusion comes from a new report published Thursday by the AI safety research organization METR. The investigation, conducted over an 11-week period, evaluated Anthropic's research and development processes. The probe is separate from a program offering safety researchers "employee-like" access following high-profile hacking incidents earlier this summer.

The findings provide some of the clearest evidence yet that AI is accelerating its own development, raising concerns about how quickly the technology might advance toward a future where AI systems can improve themselves with minimal human intervention. These insights arrive at a pivotal moment for Anthropic, which is preparing for an IPO even as its CEO, Dario Amodei, has publicly urged labs to slow down AI development.

While the report offers a rare look inside the world's leading AI labs, what it omits is equally significant. Anthropic was allowed to redact portions of the final report as part of its agreement with METR, and the company exercised this right extensively. This raises questions about the extent of public knowledge regarding the pace and risks of frontier AI development.

One of METR's most important findings is that Anthropic's AI models are improving roughly twice as fast as they were a year ago. Benchmarks suggest that "Anthropic has made roughly two years of historical progress over the last year." Crucially, METR researchers are confident this acceleration was not solely due to increasing computational power.

Instead, the researchers attribute the majority of the acceleration, which occurred by October, to AI itself. This finding was supported by internal Slack polls and interviews with employees, who reported that using Anthropic's most capable internal model made them about 12 times as productive as they would have been without assistance.

Currently, much of this acceleration stems from AI assisting human researchers rather than working independently, though METR warned that fully autonomous contributions are likely to increase soon. Notably, the researchers found no signs of diminishing returns or bottlenecks that could prevent even faster acceleration.

Anthropic's researchers utilize advanced internal models that are not yet available to the public, which accounts for much of the acceleration and productivity gains. Over the past six months, the company completed training three unreleased models—codenamed Model 2, Model 3, and Model 4—finishing in July, August-September, and September-October respectively.

Anthropic provided METR with access to confidential research, model performance data, and employee interviews, sharing more information than initially scoped. However, it did not grant unrestricted access, nor the "employee-like" access promised to other safety researchers, highlighting the reputational benefits companies can gain from assessments despite restrictions.

The restrictions were significant. Anthropic had the power to redact information from METR's final report and used it liberally. "For the public version, Anthropic redacted a majority of the evidence we found most informative," METR wrote on X.

METR reiterated these concerns throughout the report, noting that their information pipeline was narrow and that some data requests were denied. The redactions made it difficult for the public to interpret the findings and understand the uncertainties, forcing reliance on METR's judgment.

Anthropic justified the redactions to prevent competitors from benefiting, but METR noted the redacted information was very high-level. Anthropic co-founder Sam McCandlish acknowledged on X that redacting information makes it harder for others to verify the work. He stated the company determined that sharing the specific drivers of the speedup was commercially sensitive and risked handing out a roadmap.

The report has sparked fierce backlash among safety researchers, centered on the apparent contradiction of Anthropic advocating for safety while accelerating dangerous technology. Some believe the findings suggest AI recursive self-improvement is closer than anticipated.

Marius Hobbhahn of Apollo Research told *The Verge* the pace is even faster than forecast in the influential "AI 2027" scenario. He noted the report provides concrete evidence for a phenomenon that has been hard to quantify: how quickly AI is accelerating its own development.

The findings have fueled frustration over Anthropic's role in driving progress. The company has been a vocal proponent of responsible development yet appears to be racing ahead itself. Daniel Kokotajlo accused Anthropic of "recklessly gambling with our lives" for redacting information while continuing to accelerate.

The redactions have renewed calls for greater transparency and independent oversight of frontier AI labs. Brianna Rosen of the Institute for AI Policy and Strategy argued that for democratic oversight to work, independent evaluators must have the freedom to explain conclusions without heavy censorship.

Hobbhahn pointed to Anthropic canceling a portion of the assessment regarding compute allocation for safety work. He emphasized the need for independent safety researchers to be embedded more deeply within labs with significant access to internal systems.

In response, New York Assemblyman Alex Bores called for Congress to ban recursive self-improvement (RSI). With Democrats taking office, the prospect of stricter regulations is increasing.

As AI research shows no sign of slowing, Hobbhahn warned, "Brace for craziness," predicting the technology will advance too fast for society or politics to keep up.