What happened
Anthropic told investors on Monday that its future AI models may soon meet the company's own definition of recursive self-improvement, the point where AI systems get good enough at building better versions of themselves that progress speeds up sharply. The warning appeared in new risk disclosures in the company's revised S-1A filing, the updated paperwork behind its planned stock market listing. The filing also showed revenue rising fast, alongside rising expenses.
Recursive self-improvement, often shortened to RSI, describes a loop in which AI helps design and train its successors. Supporters say that loop could unlock rapid scientific discovery and productivity. Critics warn it could push capabilities past human control before safety measures catch up. Anthropic's filing treats both views seriously, and it puts the risk in plain terms: "Using AI to help build more capable AI could also pose a considerable safety risk."
The company added two striking admissions. It said it has used AI to develop new models whose jumps in capability it did not recognize at the time. And it said its internal estimates for future models, rather than its current models, indicate it may imminently meet the threshold in its Responsible Scaling Policy for "dramatically accelerated progress attributable to AI."
Why it matters
The disclosure matters because it comes from the lab itself, not from an outside critic. Anthropic already uses AI heavily in its own research and development, and its Claude models have made large gains this year in coding, math, and cybersecurity, according to TechCrunch's report on the filing. If the lab's own systems start improving faster than its safety checks, the gap between capability and oversight is the central risk.
That context also explains why investors and regulators pay close attention to Anthropic's filings. The company recently went public in a structure that [left public investors with little control](/article/anthropic-ipo-leaves-public-investors-little-control), and it has been signing huge computing deals, including [another $45B compute arrangement with Nscale](/article/anthropic-may-double-down-nscale-another-45b-compute-deal). The broader AI spending boom is already showing strain, as [big tech burns through cash](/article/ai-spending-boom-shows-cracks-as-big-tech-burns-cash). A claim about near-term self-improving AI therefore lands in a market that is both excited and nervous.
For US readers, the filing feeds directly into the American debate over who sets the rules for frontier AI. Anthropic is a US company, its models are used by American businesses and developers, and Washington has been wrestling with how to oversee systems that change this fast. A formal disclosure that self-improvement may be near gives policymakers, customers, and rival labs a concrete document to point to, rather than speculation.
Skeptics have a fair point too. An S-1 filing serves two purposes at once: it must warn investors about real risks, and it also markets the company. A claim of being close to AI that can build itself could fall into the second category. Readers should weigh the warning alongside that incentive, and alongside the fact that Anthropic itself says the threshold has not been crossed yet.
What happens next
Anthropic said it has not designated any current model as exceeding its RSI threshold, and it continues to assess the underlying data. It also said it does not expect to be able to safely use Claude to fully design and build successive generations of frontier models without meaningful human supervision, if that ever becomes technically possible.
The company said it is considering updates to its Responsible Scaling Policy to keep humans in control of frontier model development. Watch the next filing updates and any policy revision for the clearest signal of whether Anthropic believes it has moved closer to, or further from, that line.
