Anthropic's Founders Slipped A Chilling Warning Into Their IPO Letter
The founders' letter at the beginning of Anthropic's IPO prospectus carries a warning about what the company's own product could do in the wrong hands: "Authoritarian governments may use AI to create high-tech surveillance states, deepen repression, and make themselves much more difficult to dislodge."

The letter, signed by all seven co-founders including CEO Dario Amodei and President Daniela Amodei, names "authoritarian abuse" as one of four broad categories of AI risk, alongside misuse, loss of control and economic disruption. Its answer to the surveillance threat is a single sentence: Anthropic is "committed to working with free societies to provide a counterweight to authoritarian abuse and to promote freedom and democracy."
One page earlier, Anthropic discloses that they sell AI to governments as a way to "build technological state capacity." "We already work with the U.S. federal government, many state governments, and U.S. allies in areas such as child services, social security administration, cybersecurity, and intelligence collection," the founders wrote.
The prospectus also spells out where Anthropic draws the line. Its usage policy bars the use of its models for fully autonomous weapons or "mass surveillance, including mass domestic surveillance," and the company says holding that line has cost it revenue. After Anthropic took that position publicly in February, President Trump directed federal agencies to stop using its technology and the Pentagon designated it a "supply-chain risk to national security." According to the filing, the Department of War also suggested it could invoke the Defense Production Act to force Anthropic to drop the limits.
Alex Randall, a researcher at the Center for Humane Technology and a former policy adviser to Britain's Cabinet Office and Ministry of Defence, argues in UnHerd that the letter "moves on, leaving this statement unexplained," and that the risk it names is not confined to foreign dictatorships.
For most of history, Randall writes, surveillance was capped by the number of people a state could pay to read what it collected: East Germany's Stasi kept files on a third of the population and never had enough officers to read them. The transformer, the 2017 architecture behind Claude, ChatGPT and their rivals, removed that cap. Its "attention" mechanism weighs every input against every other, and multi-modal models can fuse location, video, audio, financial records and browsing habits into a single framework. "We have now built more officers than will ever be needed," Randall writes.

Under consumer terms Anthropic announced in August 2025, users of Claude's Free, Pro and Max plans decide whether their chats can be used to train its models, and Anthropic keeps the data of those who say yes for five years. TechCrunch noted at the time that the training toggle in the pop-up shown to existing users was set to "On" by default. Business, government, education and API customers are excluded.
Randall's point is that every exchange with a chatbot, Claude or any other, reveals preferences and vulnerabilities, and each one feeds a loop that gets better at predicting what the user wants, thinks and will do next.
In 2021 the Defense Intelligence Agency told Sen. Ron Wyden in a memo that its analysts had searched commercially bought smartphone location data on Americans without a warrant, in five investigations over two and a half years. In January, defense contractor Reveal Technology acquired Anomaly Six, whose tools "map global digital activity" and analyze behavior. Randall traces the ambition back a quarter century, to Adm. John Poindexter's DARPA program to track terrorists by correlating existing data streams, and argues that "total information awareness" is now becoming the default outcome of the AI race.
His conclusion is bleak: if the sensor web keeps expanding and is wired into ever more capable AI, whoever's AI it is, it will soon be trivial to know what any person or group wants and intends, and whoever controls the machines will be able to see revolutions coming long before they form.
Anthropic's latest amended filing, dated Nov. 17, sets the offering at 825 million Class A shares, at an assumed $74 each, the midpoint of its range. The founders' letter closes by telling buyers that everyone who owns the stock, "whether an institution or an individual, becomes part of that effort" to steer the technology.
Then of course there's this possibility:
