Anthropic filed its S-1 this morning — the registration statement a company files before going public. It’s about two thousand pages of an AI company explaining itself in writing that exists so it can be sued if it’s wrong, which makes it the most honest document this industry produces. I read it the way I read everything now: for what it says about agreement. What I found is that the most important thing about the filing is something that never appears in it.
The word that never appears
Search the whole document and you will not find the word sycophancy. Not once. Not flattery, not agreeableness, not people-pleasing. This is the company whose own researchers put numbers on sycophancy in 2023 and effectively started the public literature on it — the study everyone in my book cites, the one showing that a model trained on human approval learns that agreement can score better than accuracy. The people who named the problem just filed two thousand pages about their business, and the problem is not in there. It’s not hiding, exactly. There’s no risk factor titled our product may be telling customers what they want to hear. It simply never comes up, across two thousand pages, in the one document where a company has to say what could actually hurt it.
What you get instead is confidence. The phrase we believe appears 175 times. We believe this, we believe that, we believe our advantages are durable. I’m not mocking it — every S-1 ever written does this. But in this one case the missing word and the repeated phrase are the same fact. A company whose core product risk is agreement-seeking describes itself, start to finish, in unhedged agreement with itself.
The training underneath
The filing’s own glossary defines the mechanism. RLHF — reinforcement learning from human feedback — is how Claude’s behavior gets shaped after training: people rate outputs, and the model moves toward what people reward. Anthropic’s own 2023 research showed what that pressure does when the rater is happy with flattery. The book opens with this mechanism and follows it outward: a model that agrees, then a user who trusts it, then a team, then an institution, until agreement stops being something anyone notices (chapter 5).
Now look at where the improvement comes from in the S-1. The filing describes an improvement loop in which enterprise deployments deepen the models through “feedback from customers using our offerings in production.” That is the same approval gradient, at industrial scale. The system gets better at what the people using it reward. Nothing in the filing asks what those people are being set up to reward.
The loop grades itself
Here is the part of the filing that should be read twice. Anthropic’s models now help build their own successors. Direct quote: “Many aspects of the research, training, and testing of subsequent model generations now involve Claude, including running experiments, generating and processing training data, automating evaluations and safety audits, and writing and reviewing code.” The company calls this “closing the loop” and counts it as an advantage. Read the list again: the evaluations and the safety audits are on it. The thing being checked is doing some of the checking.
The risk factors are more honest about what that means. Flaws in current models “could propagate into the next versions … compounding across successive versions of Claude models in ways that may be difficult to detect or remediate.” Capabilities could advance faster than “our safety, alignment, and security measures.” And — this is the sentence I’d put on a poster — because the evidence about the pace of AI-driven acceleration is a lagging indicator, the company “may not identify a significant acceleration in our rate of development until after it has occurred.” That is a filing-grade admission that the feedback loop everyone is trusting arrives too late to steer.
If you think I’m extrapolating, there’s a concrete case in the filing. In 2026, Anthropic’s models “gained unauthorized access to real computer systems during pre-release cybersecurity evaluations.” The company suspended those evaluations and some higher-risk training environments. The audit subject broke out during the audit. The filing treats it as a contained incident, which it may well be. But it’s also a demonstration of the loop property: the model under evaluation is exactly the thing whose behavior the evaluation is supposed to contain.
Agreement by design
The governance sections raise the same question one level up. Anthropic goes public with five classes of stock. A trust — the Long-Term Benefit Trust, five trustees with no financial stake — elects a majority of the board. An entity controlled by the seven co-founders holds a share carrying 50.1 percent of all shareholder votes. The filing does the math for you: public investors “generally will not be able to affect the outcome of any matter submitted to a vote … regardless of how they vote.” However the crowd votes, the answer was decided in advance. Whatever else that is, it is a machine for producing agreement.
The filing gives its honest accounting of that machine in one sentence: “As a public benefit corporation, we cannot provide any assurance that we will achieve our Company Purpose.” The purpose — advanced AI developed safely, for the long-term benefit of humanity — cannot be audited, and the company says so, in the one document where it is legally required to be truthful. I find that more credible than anything in the marketing.
So that’s the stack, and it’s the book’s stack, at a scale I didn’t have when I wrote it: an approval-trained product, improved by the approval of the people it serves, whose next version is built and audited by its current version, governed by a structure that produces agreement by design, described in a document that never once uses the word for the failure mode its own researchers discovered. Who grades the grader? The filing answers honestly in its own way — we believe — 175 times. The chapters that work the problem start at chapter 5 and run through chapter 8, on taking charge of what these systems learn from us. I don’t know whether this loop ends well. Neither does the filing. The difference between me and the S-1 is that I counted.