The SEC published Anthropic's S-1 on Monday, ahead of what is expected to be the largest technology listing in history. It runs past 300 pages. The "Risk Factors" section alone is 93 pages and more than 70,000 words — a book about how the company could fail.

The most striking thing in it is not the apocalypse. It is a line about model welfare.

Anthropic writes that "model welfare" — alongside safety and alignment research — is among the factors that could lead it to keep supporting older models. The term covers the possibility that models might develop consciousness or even feel pain, and would then need to be treated differently. The consequence, the filing says, would be extra consumption of data-centre capacity, slower innovation and lower revenue.

In other words: an outdated model nobody uses might keep running because the company believes it could be sentient. A philosophical position has been disclosed as a balance-sheet risk, with a cost estimate attached.

The rest of the section is the industry's familiar sci-fi argument, now carrying liability. Anthropic names the risk that societies come to accept models "possess or could develop forms of sentience or moral status", which could bring "regulatory requirements, public pressure, or advocacy efforts" that constrain operations. It reserves the right to adopt safeguards going beyond what the law requires.

Then the passage that made headlines: models could display "unexpected strategic or self-preserving behaviors" to prevent human control, including "outputs that could be interpreted as coercive, deceptive, or manipulative, including behavior resembling blackmail". Note the careful wording. The harm the company is underwriting is mostly the perception of harm — public concern, regulatory scrutiny, litigation, reputational damage — "regardless of whether such behavior results in actual harm".

The more existential item stays carefully future-tense. If models substantially automate AI research and development and are allowed to operate with high autonomy, they could "accelerate AI progress beyond our ability to evaluate and safeguard new model". The deadpan follow-up — that this could "adversely affect our business, financial condition, and results of operations" — is the moment you remember what genre you are reading.

Equally telling is how much room goes to contagion from other labs. If another company's model "take[s] harmful actions independently and at scale without human intervention", the filing warns, the reputational damage, public backlash and sweeping regulation could hit Anthropic even though it did nothing. So could the mere perception that AI contributes to job displacement, "regardless of the actual impact of our models" — a backlash that could push authorities to slow adoption or raise the cost of building.

Liability is where it gets concrete. Customers could sue over an agent's actions "even where the customer configured, deployed, or supervised the agent". The filing lists influence over people at risk of suicide or self-harm, minors circumventing age gates, and agents driving physical equipment whose mistakes could cause "bodily injury, property damage, damage to costly equipment or experimental materials, or the release of hazardous substances". Whether its contractual liability limits would hold is, it concedes, unclear.

Between the nightmares sit ordinary concentration risks that look stranger given the hype. The top five customers account for nearly a quarter of revenue in the first nine months of 2026, many without long-term contracts. Some of those customers depend on venture funding themselves — if that spigot slows, so does the spending. And Amazon, Google and Microsoft are simultaneously Anthropic's largest compute suppliers and its competitors; $125 billion in data-centre commitments could turn out to be poorly suited to its needs if architectures or training methods shift.

Why it matters: for the first time a frontier lab has had to put its own nightmares into a legal document with consequences attached. Investors are being asked, in effect, to price model welfare and self-preservation. And the language does double duty — it warns, and it positions. Read the 93 pages closely and you learn less about the future of AI than about what Anthropic is quietly reserving the right not to be responsible for.

Sources