A Financial Times analysis found Anthropic used risk words like "harmful" and "dangerous" at roughly eight times OpenAI's rate in 2026. The first concrete restriction: a US bar on foreign nationals using Anthropic's newest models, Mythos and Fable.
Anthropic spent 2026 warning the world that powerful AI models were dangerous. Last week, Washington used that same vocabulary to act against Anthropic.
A Financial Times word-frequency analysis of 2026 statements, social posts, and articles by Anthropic and its chief executive, Dario Amodei, found roughly 5 risk, regulation, or restriction terms per 1,000 words. The list included words like "harmful," "dangerous," and "misaligned." The equivalent rate for OpenAI and Sam Altman was 0.6 per 1,000, about one-eighth as frequent (Ars Technica, summarizing the FT analysis).
That rhetorical asymmetry now sits next to a concrete policy fact. Washington last week barred foreign nationals from using Anthropic's newest frontier models, named Mythos and Fable. These are Anthropic's most capable AI systems. The company's own framing for them, that they sit at the cutting edge of model risk, is the same framing the US action now applies to Anthropic itself.
Yann LeCun, Meta's former chief AI scientist, publicly blamed the export bar on Amodei's "ridiculous fear-mongering" in a social post about a week before publication, adding that "one reaps what one sows" (via Ars Technica). LeCun's critique is a named external reaction, not a neutral expert consensus, and the stronger claim that Anthropic's rhetoric caused the ban is interpretive. The documented fact is the word-frequency gap. The ban followed.
Anthropic is described as a $965 billion AI group. The dispute has alarmed parts of Europe and Silicon Valley, where observers describe the US action as an early test of how Washington will oversee increasingly capable AI systems. Critics say the company's repeated AI-risk warnings, especially around Mythos, helped build the political conditions for restricting the very class of systems Anthropic was selling.
The mechanism is what makes this story more than a single-company anecdote. Anthropic's advocacy framed powerful AI as a regulatory problem. Regulators, including the US government, appear to have listened. The first concrete restriction they imposed landed on Anthropic's own newest models, not on a competitor's. The policy outcome the company warned about, pointed at someone else, may now be pointed at it.
That raises a durable field-level question the next round of AI policy will have to answer: whether loud AI safety advocacy can survive contact with policymakers without triggering self-defeating regulation. If warning loudly about frontier AI risk also helps justify restrictions on the frontier AI that one company sells, the cost-benefit calculus of public risk advocacy changes for every actor in the field, including the critics of any one company.
What to watch next: whether Washington publishes the legal instrument behind the foreign-nationals restriction on Mythos and Fable, and whether Anthropic's primary communications shift away from risk vocabulary in the second half of 2026.