In the heated race toward superintelligence, a cautious statement from AI leader Anthropic offers a moment of clarity. The company disclosed that its models have not yet crossed a critical threshold, tempering runaway AI fears.
What Anthropic Actually Disclosed
In a November 2026 amended
S-1 filing with the U.S. Securities and Exchange Commission, AI safety and research company Anthropic made a significant statement about its technology's current limits. The company confirmed that while its internal estimates for future models suggest it may be approaching a key capability milestone, no current model has been designated as having crossed the threshold for 'dramatically accelerated progress attributable to AI'. In simpler terms, their AI cannot yet improve itself without significant human involvement. The disclosure is part of the company's Responsible Scaling Policy (RSP), a framework designed to manage risks as its technology becomes more powerful. This policy uses 'AI Safety Levels' (ASLs) to classify models, with stricter protocols required as capabilities increase. Current models like Claude are considered ASL-2, meaning they show early signs of dangerous capabilities but are not yet reliable enough to pose a catastrophic risk on their own.
The 'Recursive Self-Improvement' Threshold
The threshold mentioned in the headline refers to a concept known as recursive self-improvement (RSI). This is a hypothetical feedback loop where an AI becomes capable of designing and developing its own, more intelligent successors with little or no human help. Each new generation would be better at improving the next, potentially leading to an exponential 'intelligence explosion'. For years, RSI was a theoretical concern discussed in AI safety forums, but it has become a central focus for labs like Anthropic as models grow more powerful. The company's filing notes that its policy framework is specifically designed to monitor for the risks associated with approaching RSI. While AI is already used to assist in development—engineers at Anthropic now ship roughly eight times more code than they did before 2025, largely thanks to AI assistance—this is still a far cry from a fully autonomous loop.
Why This Checkpoint Matters for AI Safety
Crossing the RSI threshold is seen by many researchers as a pivotal moment in AI development, one that carries immense promise and profound risk. The primary concern is that a system capable of improving itself could rapidly outpace human ability to understand, control, or align its goals with our own. An AI that can autonomously accelerate its own R&D could quickly evolve in unexpected ways. Anthropic's Responsible Scaling Policy is an attempt to create a structured, proactive approach to this challenge. By defining specific capability levels and tying them to mandatory safety and security measures, the company commits to pausing development if its safety protocols can't keep up with its capability gains. This disclosure, therefore, is not just a statement of current fact but a demonstration of that policy in action: identifying a potential future risk and publicly acknowledging the need for continued human governance.
The Enduring Need for Human Supervision
The second part of the headline—that human supervision is still required—is just as crucial. Even with today's powerful ASL-2 models, human oversight remains a non-negotiable part of the process. This isn't just about having a person click 'approve'. Effective supervision means that trained professionals are actively involved in validating model outputs, handling exceptions, reviewing for bias, and retaining final accountability for decisions. As AI models become more integrated into complex workflows, the burden of supervision can actually increase. Experts have found that simply having a 'human in the loop' is not a guaranteed safeguard; reviewers can develop automation bias, becoming overly reliant on the AI's suggestions. Anthropic's statement emphasizes that its path toward more advanced systems is not aimed at removing humans from the equation, but rather at ensuring they remain in control of frontier model development.
