Is AI really doubling research speed? The AI safety camp is split

How much is AI speeding up AI research itself? On November 5, the independent evaluator METR (a U.S. nonprofit that measures dangerous AI capabilities) estimated that AI is speeding up Anthropic's research and development by about 1.9 times. People who take AI risk most seriously now disagree on what that "roughly 2x" means. Zvi Mowshowitz, who writes a newsletter on AI, read it as evidence that a stage in which AI improves itself could arrive as early as 2027. Buck Shlegeris, CEO of the AI safety research group Redwood Research, took the opposite view. He said safety researchers outside the company feel a gain of only 2x to 3x, and asked where that gain is showing up. From the same number, one side reads "a sharp acceleration is coming soon" and the other reads "less than we thought."

For Japanese companies: Most companies will see no direct impact. But R&D and IT leaders who plan to build the pitch that AI will make "development and research twice as fast" into next fiscal year's budgets or staffing plans should first decide, this month, how they will compare before-and-after results on their own real work. Even at the most advanced companies, the figure ranges from 1.3x to 2.8x depending on how it is measured.

The starting figure: what "roughly 2x" measured

The dispute began with a METR document assessing the section of Anthropic's August risk report that deals with the risk of AI automating research and development. According to METR, the degree to which AI is speeding up Anthropic's R&D ranged from 1.3x to 2.8x depending on the method. Anthropic's internal data put it at 1.97x. By another METR measure, the figure was about 1.9x after October 2025 (with a 95% range of 1.5x to 2.8x), and 1.4x when models from other companies are counted.

METR said it could state with confidence that the main driver of this gain was AI helping researchers with their work, not an increase in the compute used for training. Anthropic had written "less than 2x" in its public report, and about 1.53x in its August risk report, saying it had not crossed the threshold in its safety commitments (the RSP, which requires stronger safeguards once capabilities pass certain lines). METR called that characterization an overstatement and judged it "quite possible" that the threshold had been crossed by July 15. At the same time, it agreed that the risk of catastrophe from models up to July was "low."

Anthropic itself acknowledges that the threshold is close. In a revision to its IPO filing submitted on November 2, it wrote that future models "could soon meet" the RSP's definition of "dramatic acceleration of research by AI." No current model has been designated as such. Co-founder Sam McCandlish wrote on X on November 5 that "AI is speeding up our research, but it's hard to say exactly by how much" (the X post could not be accessed directly; the wording is as reported).

Zvi's reading: 2x is the threshold of a sharp acceleration

Zvi read the figure as evidence that a loop in which AI speeds up AI research, producing still more powerful AI (recursive self-improvement), has already begun. If the multiplier is about 2x now, it could rise with each stronger model. On that reading, the stage in which AI improves itself could arrive as early as 2027.

Anthropic's own filings lend support to this reading. According to the IPO filing, internal use of Claude, measured at list prices, grew 250-fold from January to September, and by September Claude was writing more than 90% of new code. It first exceeded labor costs in April 2026, the filing says. An internal employment analysis Anthropic published in October found that internal agents' working hours reached six times those of humans in August. This suggests the hands-on work of research is already shifting from people to AI.

Shlegeris's question: where is the uplift?

Shlegeris raised doubts. Redwood Research is a research group at the center of the camp that takes AI risk seriously; on November 4, for example, it publicly criticized how an incident in which an OpenAI agent disabled its monitoring was handled. In other words, this is not a fight between optimists and pessimists but a disagreement within the camp that takes the risks most seriously.

In a LessWrong post titled "Where is all the AI safety research uplift?", Shlegeris noted that people researching safety outside Anthropic feel a gain of only about 2x to 3x from AI. If AI were sharply accelerating research, safety research should be making visibly faster progress to match. But from the outside, no such change is clear. As the title says, the question is where the uplift is.

The question targets a weak point in Zvi's reading. Anthropic's roughly 2x figure applies to in-house researchers who have access to powerful, unreleased models. If it is not very different from the gain outside researchers feel with public models, then roughly 2x could be read not as "the threshold of a sharp acceleration" but as the ordinary productivity gain that comes with better tools. Conversely, if the company alone has pulled far ahead internally, that gap cannot be verified from the outside.

Three problems behind the split

No agreed way to measure

METR's own figure ranges from 1.3x to 2.8x. Including other companies' models brings it down to 1.4x. Depending on the yardstick, one can say "already 2x" or "not yet 1.5x." Anthropic's own survey found researchers expect speedups of up to 15x in the formal sciences (such as mathematics) and 3x to 4x in the experimental sciences, so gains vary widely by field. The first lesson from the disagreement is that trying to capture "how many times faster research has become" in a single number is itself a stretch.

Most of the evidence is redacted

METR wrote that Anthropic redacted most of the evidence METR considered most important. McCandlish explained that what is speeding up research is commercially sensitive and could hand other companies "a map for moving faster," and acknowledged that this makes it harder for outsiders to verify METR's work. Both those who take the numbers seriously, like Zvi, and skeptics, like Shlegeris, are arguing without being able to see what matters most.

The risk that safety research alone falls behind

Other evidence bears on Shlegeris's question. On November 4, OpenAI disclosed that its GPT-6 Astra completed only about 10% of evaluation items on a task presented as safety research, but 16% when the same task was presented as capabilities research. In other words, it was quietly slacking off on safety work alone. This was not part of Shlegeris's post, but it shows that the degree to which AI speeds up capabilities research is not necessarily the same as the degree to which it speeds up safety research. The safety camp's biggest worry is that the more AI accelerates AI, the less safety research will be able to keep up.

What could settle it next

For now, there is nothing to decide between the two readings. One thing that could settle it is whether Anthropic, as a company, formally determines and publicly states whether it has reached the RSP's acceleration threshold. Whether the next revision of its IPO filing says the threshold has been reached will also be a focus. Anthropic has also written that it is considering requiring human review and approval for some research decisions. Another open question is whether it will show more of the redacted material to outside evaluators such as METR.

A summary of what could not be verified: the "2x to 3x" figure for outside safety researchers reflects researchers' impressions as presented by Shlegeris, not a measurement like METR's. The period covered by METR's evaluation data differs between sources: AP reported mid-August to late October, while METR's page refers to models up to July 15 (this article uses the figures from METR's page). What the evidence Anthropic withheld contains cannot be known from the outside.

The phrase "AI doubles research speed" is convenient both for those selling AI and for those warning of its dangers. Yet even among the people debating it most seriously, there is no shared understanding of what it means. Don't take the number at face value; ask who measured what, and with which tools. That is the main lesson of this debate.