Dario Amodei was OpenAI’s research director in 2019 when the team (with his involvement) initially withheld the full GPT-2 model, calling it too dangerous to release due to risks like misinformation and misuse.
He specifically pushed the view that the world needed time to prepare. It was a staged, cautious rollout.
Yet Anthropic (and Dario) have aggressively developed and released increasingly powerful models like Claude, while simultaneously lobbying hard for regulations, transparency mandates, and export controls that disproportionately burden competitors or smaller players.
This is classic regulatory capture: using government levers to create a moat, positioning themselves as the safe frontrunner who gets to help write the rules, while racing ahead.
Recent events show the backfire
•Anthropic pushed for stronger government powers to audit, test, and even block unsafe frontier models (e.g., Dario’s recent calls for an FAA for AI with authority to halt releases).
•They highlighted risks in their own Mythos/Fable-class models.
•The US government then used export controls to force Anthropic to disable access to Fable 5 and Mythos 5 worldwide over jailbreak/cyber concerns, exactly the kind of intervention they invited.
This looks less like principled safety advocacy and more like a strategy to lock in advantages (e.g., via export controls favoring incumbents, state-level rules, or compliance burdens) that ultimately turned around and bit them.
OpenAI’s GPT-5.5 apparently dodged similar treatment, highlighting the selective and inconsistent application the original post complained about. It’s a textbook case of inviting the regulatory state in as a partner, only for it to act like a regulator when it suits broader (or political) priorities.
Dario’s early caution on GPT-2 was consistent with safety concerns, and the boy who cried wolf marketing, Anthropic playbook, safety rhetoric plus heavy policy engagement, has invited exactly this kind of blowback.
The good faith engagement narrative ignores how self-serving the lobbying often appears.
