Summary
A report from nonprofit METR estimates that Anthropic has used AI to roughly double the pace of AI improvement.
Many AI watchers argue this means short timelines to superintelligent AI systems, with potentially catastrophic risks.
However, experts are debating whether the evidence really implies such fast growth, and whether AI progress will generalize beyond coding and AI research.
Policymakers should demand more transparency from AI companies, and work out how to prevent RSI until we can be assured of its safety.
By Nicky Lovegrove, Managing Editor
What happened
On November 5, nonprofit METR published a report on the speed of AI R&D at Anthropic. It assessed that Anthropic is “likely improving the capabilities of their models on well-specified tasks at around twice the rate they did between March 2024 and October 2025”.
METR (Model Evaluation & Threat Research) was given privileged access to Anthropic systems from August to October 2026. It said most of the acceleration was likely due to Anthropic’s use of AI to accelerate AI R&D, and ruled out faster growth in training compute as the cause.
The capability uplift appears to be mostly due to AI enabling human researchers to work faster. But Anthropic staff also reported signs that fully autonomous AI contributions will soon play a greater role.
METR’s headline estimate is uncertain: it said the actual speed-up could be anywhere between 1.3x and 3x the historical rate. METR also saw no evidence of significant bottlenecks to further acceleration in the next 12 months.
Importantly, METR’s report excludes a majority of the evidence most informative for its conclusions. Anthropic redacted this information, saying it was confidential and could help accelerate other AI developers.
Why it matters
The speed-up could be an early sign of recursive self-improvement (RSI).
Definitions of RSI vary, but usually describe AI models autonomously developing smarter AI models, which in turn create even smarter ones. Such a process might enable rapid growth in AI capabilities, very quickly reaching AI that vastly exceeds human capabilities across all cognitive domains.
Researchers who study this scenario suggest it could bring about rapid technological development, sudden shifts in the global balance of power, mass job displacement, and novel forms of AI misuse. And that’s assuming humans remain in control.
AI watchers are shortening their timelines.
The AI Futures Project, which authored the AI 2027 report and includes several top AI forecasters, now predicts that a superhuman AI researcher will emerge in 2027. Their team gives a 35-65% chance of AIs automating all white-collar work before 2028. The head of AI safety organization Redwood Research predicts a ‘one in three chance’ of superintelligent AI in 2027.
The risks of getting this wrong could be catastrophic.
If these forecasts are right, the world has very little time to ensure such AI models are safe. The most severe risk worrying AI experts is that humans lose control to powerful AI models, leading to either permanent human disempowerment or extinction.
This week, Paul Christiano, formerly a US government AI safety chief and now OpenAI board member, said: “There’s material risk of AI takeover in the next 6-12 months… For most people in this room, the most likely reason they will die is because of an AI accident”.
However, not everyone agrees with these projections.
Key debates
Is the speed of progress really doubling?
Oxford University’s Toby Ord proposed that the acceleration METR observed could partly reflect a capability jump achieved when Anthropic trained its larger Mythos model. Ord put the chance of an intelligence explosion in 2027 at around one in eight. Researchers from OpenAI and Google DeepMind also expressed skepticism about reading too much into the METR report.
Will the capabilities generalize?
OpenAI’s Dean Ball, formerly a Trump White House official, noted that METR’s report showed only that AI was accelerating progress on ‘well-specified’ tasks. This means tasks with clearly defined answers, often in fields like maths and coding. He doubted these capabilities would generalize quickly to harder-to-measure skills like ‘research taste’, which end-to-end automation of AI R&D would require.
Policy implications
So what should governments do in light of the METR report and the concerns expressed by AI experts?
Demand transparency into AI R&D
In October, the UK AI Security Institute highlighted the importance of AI companies sharing information, including on “AI’s use in its own development as an indicator of recursive self-improvement”.
Anthropic could share the unredacted METR report with key government agencies – or Congress could subpoena Anthropic.
Beyond this, policymakers could seek data on AI companies’ compute allocation – how companies distribute their computing resources between AI R&D, safety research, and serving customers. As the AI Futures team argues, any regulatory attempt to pace or pause frontier AI development will likely depend on such data.
Prevent RSI until we can be assured of its safety
No one knows how to make frontier AI models reliably safe, let alone superintelligent ones. A reasonable response is to prevent companies from pursuing RSI until the science of AI safety has time to catch up.
This idea is gaining political traction.
In September, US Representative Greg Casar and Senator Bernie Sanders introduced the Ban Artificial Superintelligence Act. Also in September, UK Labour MP Alex Sobel introduced a similar bill to the House of Commons. Last week, Alex Bores, a Democratic member of the New York State Assembly, also called for Congress to ban RSI, citing a list of supporters ranging from Ezra Klein to Steve Bannon.
A challenge for lawmakers will be defining RSI. A broad ban, such as one that prohibits AI companies from using AI to write any code, could be difficult to enforce and unnecessarily harm the AI industry. A narrow ban, for instance only on automated AI R&D that has no human in the loop, might still allow highly dangerous AI development.
The definition also matters politically. Democrats will soon control the House and Senate, but for AI legislation to survive a presidential veto, they will need Republicans on board. Several conservative-leaning analysts, including Dean Ball and the Abundance Institute’s Neil Chilson, have indicated they could support a version of an RSI ban – as long as it was narrowly defined.
An American ban on RSI would not be a silver bullet. The US government would also have to resolve other urgent questions: how to prevent the theft of frontier AI model weights and algorithmic secrets; how to limit the economic fallout from any AI slowdown or pause; and how to ensure China and other countries stick to similar constraints.
Bottom line
Even setting aside near-term superintelligent AI, policymakers should be prepared for rapid AI capability growth over the coming year.
Until the debate is settled, governments should be able to see all the evidence and have ways to stop AI development that could pose catastrophic risks.