Direct answer: Anthropic has warned, in a revised S-1A filing disclosed on 2 November 2026, that internal estimates indicate it may imminently meet its own Responsible Scaling Policy definition of dramatically accelerated progress attributable to AI. No current model has been designated as exceeding that threshold, and Anthropic says it is considering policy updates to keep humans in control of frontier model development. For most UK small firms, the immediate story is not a machine building itself overnight, but a leading lab saying, in its own investor disclosure, that progress could speed up faster than its safety checks can comfortably track.
What happened
The disclosure arrived in the fine print of a revised S-1A filing, rather than as a product launch or a press release. Much of the attention around the filing, according to the TechCrunch report on Anthropic’s RSI warning, focused on updated financials, with revenue soaring but expenses rising too. Inside the Risk Factors section, however, Anthropic added new statements about how it uses AI in its own research and development, and about the risks that could follow if that use sharply accelerates the pace of progress.
Anthropic says in the filing that it uses AI in its research and development. It also cautions, in words quoted by TechCrunch, that using AI to help build more capable AI could pose a considerable safety risk, particularly if it sharply accelerates the pace of progress. The filing links that risk to Anthropic’s Responsible Scaling Policy, which, in the company’s description, sets out its plans to monitor and mitigate risks as it approaches what has been termed recursive self-improvement.
The new statements add two points that were not, in that form, part of the older picture. First, Anthropic says it has already used AI to develop new models whose leaps in capability it did not recognise at the time. In the filing language reported by TechCrunch, the company says it has in the past not identified a significant acceleration in its rate of development until after it had occurred, and that it may in the future not identify such an acceleration until after it has occurred. Second, Anthropic says its internal estimates and its expectations regarding future models, rather than its current models, indicate that it may imminently meet the Responsible Scaling Policy definition of dramatically accelerated progress attributable to AI.
That careful wording around current and future models is worth pausing on. Anthropic is not saying that a model available today has crossed the line. It says it has not designated any current model as exceeding the threshold, and that it continues to assess the underlying data. The warning is therefore about what its estimates suggest could happen soon, not a declaration that it has already happened. Anyone summarising the story as “Anthropic has achieved self-improving AI” would be going further than the filing, on the evidence reported so far, actually goes.
The filing also includes a commitment about human control. Anthropic says it does not anticipate being able to safely use Claude to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding, if and when that becomes technically possible. Accordingly, it says it is considering updates to its Responsible Scaling Policy to keep humans in control of frontier model development at Anthropic. That is a significant sentence for businesses that buy, build on, or make decisions with AI tools, because it frames human oversight not as a temporary courtesy, but as a condition for safety.
For context, the filing sits within a wider period of scrutiny around Anthropic’s move towards public markets. Readers following that story may also want to read our earlier coverage of Anthropic filing for an IPO under the ticker ANT on Nasdaq, which sets out the background to the listing process, and our report on Anthropic’s junk credit rating and what it means for its IPO, which explains why financial and risk disclosures around the company have been watched so closely. The RSI warning does not replace that financial story; it sits alongside it, in the same document, and it helps explain why investors, customers and regulators may read Anthropic’s risk section with particular care.
How recursive self-improvement actually works
Recursive self-improvement, usually shortened to RSI, sounds like science fiction, and that is part of why it attracts both excitement and alarm. In plain English, the idea is simpler than the label suggests. It starts with AI being used to help with research and development. That help might take many forms in a lab: drafting code, suggesting experiments, checking results, summarising research, spotting patterns in test data, or helping engineers work through a difficult design problem. None of that, on its own, is self-improvement in the dramatic sense. It is AI-assisted work, with people still setting the goals, judging the results, and deciding what to try next.
RSI enters the picture when that assistance becomes so effective that each new generation of model helps to build the next generation much faster, or much better, than people could on their own. The word “recursive” points to the loop: a model helps to create a better model, that better model helps to create an even better one, and so on. If each step in the loop gets shorter, or each improvement gets larger, the overall pace of progress could rise sharply rather than steadily. That is why Anthropic’s phrase “dramatically accelerated progress attributable to AI” matters. It is not just saying that AI is useful in the lab. It is pointing to a possible change in the rate of progress itself, caused by AI’s contribution to the work.
There is another subtle point here. Acceleration is not the same as autonomy. A lab could experience dramatically faster progress while people remain closely involved at every step. People could still set the research questions, review the outputs, run the safety tests and decide whether to release a model. Equally, a system could become more autonomous without immediately becoming dramatically more capable. Anthropic’s filing keeps those ideas separate, and careful readers should do the same. The warning is about pace, and about the risk that safety work falls behind that pace. The human-control statement is about autonomy, and about whether a future system could safely design successive frontier models without meaningful human supervision or understanding. Both points matter, but they are not interchangeable.
None of this means that RSI, as Anthropic defines it, is the same as a single moment when a machine suddenly becomes independent. The filing language describes a threshold in a policy, and a definition of dramatically accelerated progress. Thresholds in policies are useful because they trigger extra scrutiny, extra testing, or a pause for review. They are not, by themselves, proof that a system has become conscious, independent, or uncontrollable. Keeping that distinction in mind helps avoid two opposite mistakes: dismissing the warning because no dramatic event has happened yet, and overstating it as if the dramatic event has already happened.
Key facts
| Point | What Anthropic disclosed, as reported |
|---|---|
| Document | Revised S-1A filing, disclosed on 2 November 2026 |
| Use of AI in research | Anthropic says it uses AI in its research and development |
| Past acceleration | Anthropic says it has in the past not identified a significant acceleration until after it occurred |
| Near-term estimate | Internal estimates for future models, not current models, indicate it may imminently meet its definition of dramatically accelerated progress attributable to AI |
| Current models | No current model has been designated as exceeding that threshold; Anthropic continues to assess the data |
| Human control | Anthropic says it is considering policy updates to keep humans in control of frontier model development |
| Fully autonomous design | Anthropic says it does not anticipate safely using Claude to fully autonomously design successive frontier models without meaningful human supervision or understanding |
| Financial context | Revenue is soaring, but so are expenses, according to coverage of the updated financials in the filing |
| Sceptical note | Sceptics note that an S-1 functions as both disclosure and marketing |
How this compares with rivals
It is tempting to turn a single filing into a race, with one company ahead and the rest behind. The reported facts do not support that kind of league table, and a careful comparison should avoid inventing one. What can be said, in general terms, is that AI-assisted development is now a familiar idea across the field, while Anthropic’s disclosure is distinctive because it puts a specific policy definition, a near-term estimate, and an admission about past detection into a formal investor document.
A second general idea that often gets mixed into RSI discussions is continual learning. In simple terms, continual learning describes a system that keeps learning from new data or experience over time, rather than being trained once and then frozen. Trajectory-style continual learning, as a general concept, would involve a model improving through an ongoing stream of experience, instead of only through discrete, separate training runs. That concept is useful for thinking about how progress might become less stepwise and more continuous. It is not, however, the same as Anthropic’s specific RSI definition, and it should not be treated as proof that any particular rival system has achieved RSI. The filing is about Anthropic’s own estimates and its own policy threshold, not a verified comparison with another lab’s internal progress.
Anthropic’s earlier safety work also forms part of the comparison, because it helps explain why this warning appears in a policy context rather than as a bare boast. The Responsible Scaling Policy, as described in the filing, sets out plans to monitor and mitigate risks as the company approaches recursive self-improvement. That framing matters. It presents RSI not only as a capability milestone, but as a governance trigger: a point at which extra monitoring and mitigation should apply. Other labs have their own safety frameworks and public statements, but the reported filing does not provide a like-for-like comparison of those frameworks, and this article does not invent one. The fair comparison is therefore narrower: Anthropic has chosen to disclose, in a risk section, both a near-term estimate and a past blind spot, and to link both to a named policy it says it may update.
What it means for UK small businesses
For most UK small firms, RSI will not arrive as a single event that changes everything overnight. It will arrive, if it arrives, as faster model updates, new features, changing prices, shifting vendor claims, and a growing gap between what a tool can do and what a busy owner fully understands about it. That is why the practical response is less about preparing for a science-fiction scenario, and more about tightening everyday governance while the tools are still manageable.
Start with human oversight, and make it specific. Anthropic’s own statement puts meaningful human supervision and understanding at the centre of safe frontier development. A small firm can apply the same principle at its own scale. Decide, in advance, which decisions always need a person: approving a payment, sending a message to a customer in the firm’s name, changing a price, sharing personal data, or publishing content that carries the firm’s reputation. If an AI tool drafts the work, a named person should still check it, and that check should be real, not a quick glance at a long output. The habit matters more than the tool. A firm that already reviews AI output carefully will cope better with faster updates than a firm that has quietly allowed automatic outputs to go straight to customers.
Vendor questions are the next practical step. When a supplier announces a major new model or a dramatic capability jump, ask what has actually changed, what has been tested, and what human oversight remains. Useful questions include: what tasks is this version genuinely better at, and what evidence supports that claim? What has changed in how the model handles personal or confidential data? Can the firm stay on the previous version for a while, or will it be moved automatically? What happens if the new version behaves differently on a routine task? A vendor that can answer those questions in plain language is easier to trust than one that relies only on superlatives. Small firms do not need to interrogate a lab’s internal research programme, but they do need enough clarity to decide whether to adopt, delay, or test a change before it touches customers.
That testing point deserves emphasis. If model progress speeds up, vendors may ship updates more often, and each update may change behaviour in subtle ways. A prompt that worked well last month may produce a different tone, a different level of detail, or a different handling of an edge case this month. For a small business, the safeguard is a short, repeatable check: run the same handful of typical tasks through a new version, compare the outputs with the previous version, and note anything that would matter to a customer. Keep that check small enough to actually do. A five-task test done every time is worth more than an ambitious evaluation plan that never happens. This is also where AI consulting for small businesses in the UK can help, because an outside review can spot where a faster tool has quietly changed a workflow that used to be stable.
Cost and pace of change are linked, and both need a plan. Anthropic’s filing notes soaring revenue alongside rising expenses, which is a reminder that building frontier AI is expensive, even for a leading lab. For buyers, that expense can show up as pricing changes, new tiers, usage limits, or features moving between plans. A small firm should therefore avoid building a core process that only makes financial sense if one particular tool stays cheap, stays unchanged, and keeps improving on a predictable timetable. Keep an exit route: exportable data, documented prompts and workflows, and a fallback process that still works if a tool is unavailable or becomes too expensive. It is the same discipline firms already apply to any important supplier.
Skills are the quieter part of the story. If AI tools become more capable at coding, drafting, analysis and customer communication, the valuable human skills shift towards judgement, checking, and understanding context. Staff need time to learn how to prompt well, how to spot a confident but wrong answer, and how to decide when not to use AI at all. Owners should also be honest about their own time. If the owner is the only person who checks AI output, and the volume of that output grows, the checking step becomes the bottleneck. Either the volume must be controlled, or the checking must be shared.
Data protection deserves a calm, general note. UK firms handling personal data already need to be careful about what they put into AI tools, who can access it, and how long it is retained. Faster models do not remove those duties, and new features do not automatically make a tool suitable for sensitive data. As a general caution, firms should check a vendor’s data handling terms before feeding personal, customer, or commercially sensitive material into a new model, and they should avoid assuming that a general-purpose AI product is automatically compliant with data-protection expectations. Readers who want a fuller walkthrough can consult our GDPR and AI guide for UK small businesses, which covers the practical questions to ask before personal data goes anywhere near an AI system. When in doubt, keep sensitive material out of a general tool, and seek professional advice for the specific situation rather than relying on a vendor’s marketing page.
For firms already using Anthropic’s tools, our practical guide to using Claude AI for small business sets out everyday uses, limits, and checks that remain relevant even as models change. The underlying advice does not change with each new version: use the tool for work you can check, keep a person responsible for the result, and do not let a capability claim replace your own testing.
Honest limits / what to watch
The most important limit in this story is also the simplest: the filing reports estimates and expectations about future models, not a verified result about a current model. Anthropic says it has not designated any current model as exceeding its threshold, and that it continues to assess the underlying data. Until Anthropic, or an independent check, says otherwise, the honest summary is that RSI may be near by Anthropic’s own definition, not that it has been achieved.
What to watch, in practical terms, falls into four buckets. First, policy: any update to Anthropic’s Responsible Scaling Policy, and any explanation of how human control will be maintained in practice. Second, designation: whether Anthropic ever designates a current model as exceeding its threshold, and what it says about the data behind that decision. Third, detection: whether Anthropic can show, next time, that it spotted an acceleration as it happened, rather than only afterwards. Fourth, market and regulatory response: how investors, customers, competitors and policymakers react to a leading lab saying, in formal terms, that dramatically accelerated progress may be imminent. None of those four can be predicted from this filing alone, and all four will matter more than the initial headline.
FAQs
What is recursive self-improvement, or RSI?
Recursive self-improvement is the point at which AI becomes good enough at helping to build better AI that progress speeds up sharply. Anthropic describes it as dramatically accelerated progress attributable to AI, and it has said it may imminently meet that definition, based on internal estimates for future models. It is a loop in which each improvement makes the next one easier, so the overall pace rises rather than staying steady.
Has Anthropic already built a self-improving AI?
No, not by its own definition in the filing. Anthropic says it has not designated any of its current models as exceeding the threshold for dramatically accelerated progress attributable to AI, and that it continues to assess the underlying data. The warning is about what its estimates suggest could happen soon with future models, not a claim that a current model has already crossed the line.
Why does Anthropic say it might miss a leap in capability?
Anthropic says that, in the past, it has not identified a significant acceleration in its rate of development until after it occurred. It warns that increasing use of AI in research and development could cause model capabilities to advance more rapidly than its ability to develop, validate and implement corresponding safety, alignment and security measures. In other words, the risk is not only that progress speeds up, but that the checks meant to keep that progress safe may not speed up at the same rate, and may not even register the change until afterwards.
Does Anthropic plan to let Claude build future models on its own?
Anthropic says it does not anticipate being able to safely use Claude to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding. It says it is considering updates to its Responsible Scaling Policy to keep humans in control of frontier model development.
Should UK small businesses worry about RSI right now?
There is no need to panic, but there is good reason to tighten the basics. Keep a human in charge of important decisions, ask vendors how models are tested and overseen, be careful with personal and confidential data, and avoid building a business process that only works if one AI tool keeps getting dramatically better on a fixed timetable. If progress does speed up, firms with clear oversight, short update tests, and exportable workflows will adapt more easily than firms that have allowed AI output to flow unchecked to customers. The work to do now is ordinary good governance, done consistently.
Is an S-1 filing a reliable guide to how close RSI is?
An S-1 is a disclosure document for investors, and sceptics note that it also functions as marketing. That does not make the warning meaningless, but it is a reason to read the exact wording carefully and avoid treating a possibility about future models as proof about current ones.
