The Audit Gap: Why Anthropic’s AI-Driven Acceleration Defies Human Oversight
AI-generated image · Bay Street Wire
A METR report reveals that AI is now the primary driver of progress at Anthropic, creating a productivity loop that threatens to outpace the ability of human researchers to verify the results.
The prevailing narrative around artificial intelligence usually focuses on the speed of the treadmill—how fast we are moving toward some undefined horizon of capability. But for those of us watching the machinery, the real story isn't the velocity; it is the breakdown of the verification mechanism.
As first reported by The Verge, a new investigation by the AI safety research organization METR into the internal operations of Anthropic suggests we are entering a phase where the AI is no longer just a tool for the researcher, but the primary engine of the research itself.
According to the METR report, Anthropic’s models are currently improving roughly twice as fast as they were a year ago. The data is startling: METR researchers found that Anthropic achieved approximately two years of historical progress within a single year. While the instinct of the industry is to credit this to 'scaling'—simply throwing more compute at the problem—METR explicitly stated they are confident that the majority of this acceleration as of October is due to AI, not just increased computing power.
From a practitioner's perspective, this creates a systemic risk. If the acceleration is driven by AI, the resulting optimizations may be based on patterns and efficiencies that are opaque to the human researchers overseeing them. We are moving toward a 'black box' development cycle.
***
**Opinion: The Collapse of the Human-in-the-Loop**
In my view, the industry is sleepwalking into a verification crisis. The METR findings highlight a productivity explosion that is fundamentally unsustainable from an auditing standpoint. When Anthropic employees reported via Slack polls that using the company's most capable internal model made them roughly 12 times as productive as they would be without AI, they weren't just describing a faster way to write code or document findings. They were describing a shift in the cognitive load of development.
If a researcher is 12 times more productive because an AI is handling the heavy lifting of optimization, that researcher is no longer auditing the process in real-time; they are auditing the output. When the output is another AI model, the gap between 'it works' and 'we know why it works' widens.
***
This acceleration is happening in the shadows. The Verge reports that Anthropic has completed training on three models that have not yet been released to the public, codenamed Model 2, Model 3, and Model 4. These models were completed in rapid succession: Model 2 in July, Model 3 in August-September, and Model 4 in September-October. METR also identified an older, unreleased model from 2022 referred to as Unreleased 2022.
This rapid-fire release cycle suggests a development cadence that leaves little room for the slow, methodical safety auditing that CEO Dario Amodei has publicly advocated for. There is a glaring tension here: Amodei has called for AI labs to 'pump the brakes,' yet the internal machinery at Anthropic is accelerating.
Furthermore, the transparency of this process is questionable. The Verge notes that Anthropic exercised its power to redact portions of the METR report extensively, despite the fact that METR researchers had access to confidential research, employee interviews, and model performance data. While METR acknowledged that Anthropic shared more information than initially scoped, the company did not grant the researchers the 'employee-like' unrestricted access it has promised other safety researchers in separate programs.
The most alarming takeaway from the METR probe is the absence of a ceiling. The researchers found no observable bottlenecks or signs of diminishing returns that would prevent even faster acceleration. While current gains are largely attributed to AI assisting humans, METR warned that signs point toward fully autonomous contributions playing a larger role in the near future.
Once the loop closes—once AI is not just assisting the researcher but autonomously contributing to the next iteration of the model—the human-in-the-loop becomes a formality. The danger isn't that the AI will fail; it's that it will succeed in ways we can no longer explain, and we will be too dazzled by the 12x productivity gains to notice the loss of control.