An independent investigation found that AI may be accelerating its own development. But the restrictions on publishing its findings reveal another problem: who can scrutinise the evidence, and who has the authority to act?
On 5 November 2026, independent research organisation METR published an assessment of Anthropic, the company behind Claude. Its researchers had spent eleven weeks examining whether artificial intelligence was helping the company develop increasingly capable AI systems at a faster rate.
The results attracted attention for an understandable reason. METR found that progress on certain measures of Anthropic’s model capabilities appeared to have accelerated substantially, with AI-assisted research likely responsible for much of the change.
But buried inside the technical findings was a statement that deserves equally serious attention.
Anthropic had given the independent evaluators extensive access to internal research information. Yet before publication, the company redacted much of the evidence the researchers considered most informative for their conclusions.
METR acknowledged that this made it difficult for the public to examine the underlying evidence or arrive at its own interpretation. Readers would have to place considerably more trust in the evaluators’ judgment than the researchers themselves considered ideal.
This raises an uncomfortable question. If the development of increasingly powerful AI may have consequences extending far beyond the companies building it, should those companies retain so much influence over what the rest of the world is allowed to examine?
The answer is more complicated than simply demanding that everything be made public.
What did the METR assessment actually find?
According to METR’s independent assessment, published on 5 November 2026, Anthropic’s progress on an internal index of AI capabilities had accelerated to approximately 1.9 times its earlier rate.
The researchers judged that AI assistance was probably responsible for much of the acceleration. AI tools were helping human researchers carry out parts of the work involved in developing new models, creating the possibility of further improvements in research productivity.
This is important evidence about the changing pace of AI development. It is not proof that Anthropic has created superintelligence or that AI is already improving itself without human involvement.
METR reported substantial uncertainty about its estimates. It also found that human researchers remained central to the most consequential stages of model development. While the evaluators did not identify clear bottlenecks likely to prevent continued rapid progress in the near term, they acknowledged that their understanding of those limitations was incomplete.
These distinctions matter because discussions about superintelligence can quickly become trapped between extraordinary promises and catastrophic predictions.
The evidence deserves serious attention without being exaggerated. And that makes the conditions under which the evidence can be scrutinised especially important.
Why was so much of the evidence redacted?
Anthropic’s explanation was that some of the information could help competing AI developers accelerate their own research.
That concern cannot simply be dismissed.
Frontier AI research involves valuable technical knowledge, commercial competition and potential security implications. Publishing sensitive details about model capabilities, training techniques or development processes could create risks of its own.
METR also confirmed that Anthropic had cooperated extensively with the investigation, providing more access than initially expected. The evaluators were responsible for their own conclusions and received no payment from Anthropic for conducting the assessment.
It would therefore be misleading to describe the investigation as meaningless or to assume that the redactions were intended to conceal wrongdoing.
Nevertheless, the arrangement illustrates a fundamental problem with relying on voluntary transparency.
An independent evaluator may obtain privileged access to important evidence while being unable to disclose enough information for other researchers, policymakers or members of the public to scrutinise the conclusions independently.
The report can be credible without being fully verifiable by its readers.
And credibility, however valuable, is not a substitute for accountability.
Imagine the same problem in a school
Consider a hypothetical situation.
A school is offered an AI system that promises to identify pupils who may need additional educational support. The provider presents an assessment by independent specialists, who have examined the system and found it sufficiently reliable for the proposed use.
But suppose the provider also controls which parts of the evaluation may be disclosed. The school can read the conclusions, but not all the evidence supporting them.
The assessment may have been conducted with complete integrity. The system may genuinely offer educational benefits. Its developer may have reasonable grounds for protecting some confidential information.
Yet the school still has responsibilities that cannot be transferred to the provider.
Can staff understand the limitations of the system? Has it been tested for different groups of pupils? Can a teacher question a recommendation? What happens when a family believes the system has made an error? Who can require additional testing or suspend its use?
This imagined school scenario is not equivalent to the risks associated with frontier AI development. Its purpose is to show something about the structure of responsibility.
A decision-maker needs more than reassurance that someone competent has examined a system. They need a reliable way to establish what happens when the evidence is disputed or the system’s behaviour causes concern.
The same principle becomes even more consequential when the technology is being developed for applications across entire economies and societies.
An independent assessment is only the beginning
AI governance discussions frequently concentrate on transparency, safety testing and independent evaluation. All three are necessary, but they address different parts of the problem.
An independent assessment can help establish whether a system possesses certain capabilities, whether safeguards work as intended and where important uncertainties remain.
Transparency helps other people understand the findings and examine the reasoning behind them. Where full disclosure would create legitimate security or confidentiality risks, trusted institutions may need special access to evidence that cannot responsibly be published in full.
Accountability concerns what happens next.
Who is authorised to challenge a company’s interpretation of its own safety thresholds? Who can require further testing when evidence is inconclusive? What authority exists to intervene when a technology presents unacceptable risks? And how can the people affected by these decisions influence the priorities that govern them?
These questions are not answered by publishing a safety report, however carefully it has been written.
They require institutions with defined responsibilities, access to relevant evidence and the capacity to act.
They also require something that receives considerably less attention in conversations about frontier AI: democratic legitimacy.
Who gets to decide which risks are acceptable?
Researchers can help us understand the capabilities of increasingly powerful AI systems. Their expertise is essential, especially when assessing technical risks that cannot be evaluated through ordinary public debate.
But scientific expertise does not automatically confer the authority to decide which risks other people must accept.
Those decisions involve competing values and interests. A technology may promise considerable benefits while distributing its risks unequally. A development that looks commercially attractive to one organisation may create dependencies, vulnerabilities or social consequences for institutions that had no influence over its design.
Citizens cannot be expected to resolve technical questions about model architecture or AI alignment. Nor should complex scientific findings be settled by popularity.
But people can contribute meaningfully to decisions about the purposes technology should serve, the protections it must provide and the responsibilities of the institutions introducing it into public life.
We already have examples of people participating seriously in such discussions. In 2024, Belgium convened a citizens’ panel on artificial intelligence, bringing together 60 people from different backgrounds to examine the technology and develop recommendations for policymakers.
The important question is what happens to those recommendations after the discussion ends.
A public consultation can be thoughtful, inclusive and intellectually valuable while leaving the actual decisions unchanged.
If public participation is meant to contribute to democratic accountability, institutions need to do more than collect opinions. They must be able to explain how those contributions were considered, which decisions they influenced and why other recommendations were not adopted.
That challenge becomes particularly difficult when the institutions making the most consequential technological decisions operate across national borders and possess information that governments and citizens may struggle to access.
The question is bigger than Anthropic
The METR assessment offers a valuable and unusually transparent account of the limitations of its own publication process.
That openness matters. It helps readers distinguish what researchers were able to establish from what the public cannot independently examine.
But the case also exposes a wider institutional challenge.
The organisations developing frontier AI are making decisions under conditions of scientific uncertainty, intense competition and considerable commercial pressure. Governments are attempting to establish meaningful oversight while the technologies themselves continue to evolve. International institutions are beginning to create opportunities for a wider range of countries and communities to participate.
The distance between having an opportunity to speak and possessing the ability to influence a decision remains substantial.
We should be concerned about that distance whether or not the most extreme predictions about artificial superintelligence eventually prove correct.
AI is already affecting education, employment, public administration, scientific research and access to information. Those changes make questions about evidence, authority and public accountability relevant now, not only in a hypothetical future involving machines more capable than humans.
The challenge is to build institutions that can understand technological risks, intervene where necessary and remain answerable to the people whose lives are affected.
That is a question of governance, certainly. But it is also a question about how much influence people retain over the conditions of their own future.
Beyond the assessment: who is speaking for the rest of us?
The Anthropic findings are one part of a much larger debate.
In our full essay, They Warn AI Could Kill Us All. Who Is Speaking for the Rest of Us?, we examine warnings from researchers inside frontier AI laboratories, the uncertainty surrounding superintelligence, and what earlier technological crises can teach us about responsibility and collective action.
We also explore an often-neglected part of the discussion: how ordinary people can participate meaningfully in decisions about technologies that may eventually affect everyone.
The history is not entirely pessimistic. Societies have sometimes allowed dangerous technologies to become deeply embedded before responding to their consequences. They have also demonstrated that scientific evidence, public participation and international cooperation can change technological trajectories.
What we make of that history matters enormously for the choices being made today.
The essay is part of REDefine’s public writing on responsible technology, democratic accountability and the systems shaping our everyday lives.
