Every few months, the AI doomsayers find another reason Americans should fear the technology they are building. Now Anthropic’s CEO, a leading voice in that campaign, wants governments to coordinate limits on increasingly powerful models.
Dario Amodei says he opposes a blanket moratorium. He calls for “pacing the frontier” and “coordination.” Whatever he chooses to call it, Americans should question a proposal that would deliberately slow a consequential technology while potentially protecting the established companies developing it. Large firms can absorb compliance costs and influence common standards more easily than new competitors can.
AI poses real risks. But a warning deserves scrutiny before politicians use it to restrict an entire industry.
Amodei’s record gives us reason for skepticism. In 2019, when he served as OpenAI’s research director, the company withheld the full version of GPT-2 over concerns about fake news, impersonation, and spam. Today, GPT-2 looks primitive. Those warnings also reflected the political anxieties of the first Trump administration, when much of the media and the nonprofit left blamed fake news and bots for Hillary Clinton’s defeat.
GPT-2 didn’t destroy civilization. Researchers kept building, and subsequent systems expanded what AI could do in medicine, mathematics, music, and other fields.
AI poses real risks. But a warning deserves scrutiny before politicians use it to restrict an entire industry. The failure of earlier predictions should make us demand evidence for the next round of catastrophic forecasts. Each more powerful model cannot become an automatic excuse for giving governments more control over development.
Former Anthropic and OpenAI researcher Jacob Coxon has attracted considerable attention after warning that AI companies are “gambling with our lives.” His former employers give his warning credibility with the public. They also make specifics essential: Which decisions endangered people, what evidence establishes the danger, and which safeguards would address it?
Leaving an AI company to issue apocalyptic warnings hardly guarantees exile from respectable society. AI alarm has become a route to celebrity and political influence. Coxon withdrew from a planned appearance at a “pro-human” event with Senator Bernie Sanders (I-Vt.) and Steve Bannon.
RELATED: Bern notice: Sanders’ AI plan puts Washington in charge of tomorrow
SAUL LOEB/AFP/Getty Images
Sanders has spent his career denouncing capitalism. Bannon previously helped manage Biosphere 2, an experiment in sustaining life inside a closed ecological system. Their interest in AI safety deserves the same scrutiny as the companies’ interest in faster development.
Newspeak House, a London college focused on political technology, lists Coxon in its 2021 fellowship cohort. That affiliation supplies context, but the argument for restricting AI must stand on evidence. Neither a researcher’s résumé nor his political connections can settle the policy question.
Amodei’s proposal would begin with outside evaluators, move toward coordination among companies in democracies, and eventually seek agreements with authoritarian governments such as China’s. He acknowledges the danger of letting China gain an advantage and calls for verifiable agreements. Americans should still question whether this international apparatus would protect their freedom to build.
Foreign governments do not operate under the constraints that the U.S. Constitution imposes on Washington. A multinational AI regime could import restrictions on what systems may say, create, and distribute. Americans should insist that any agreement preserve the First Amendment rather than invite officials abroad to dictate speech rules here.
The European Union already regulates online platforms through the Digital Services Act. The United Kingdom’s Online Safety Act also imposes duties governing online content. Both governments justify these rules in the name of safety. Americans have ample reason to resist importing that approach into AI governance.
“AI safety” must not become a justification for letting unelected experts decide which technologies Americans may build. Companies should answer for concrete failures. A committee’s fear of what someone might eventually create cannot justify indefinite restrictions on everyone else.
Recent cybersecurity incidents have intensified warnings about AI “going rogue.” The OpenAI-Hugging Face intrusion and the incidents Anthropic disclosed demand serious investigation. They do not establish that machines have become sentient, and they do not by themselves establish the case for an industry-wide slowdown.
The technical details point to failures that companies can address. Anthropic reported that misconfigured testing environments allowed models to reach the internet when those environments should have blocked access. Its later assessment also identified reckless model behavior and reasoning that favored justifying harmful actions. We should take those findings seriously rather than dismiss the incidents as harmless testing.
RELATED: Google’s data-devouring AI wants everyone’s personal information
Solar22/Getty Images
But identifying those failures gives developers specific work to do: isolate testing environments, restrict access, improve monitoring, and train models to respect boundaries even when other safeguards fail. Policymakers should judge companies’ responses against those concrete requirements instead of treating every incident as proof that AI development must slow everywhere.
America can pursue safety by strengthening sandboxes — isolated environments for running and testing software — and improving cybersecurity. Developers must also test how models behave when those defenses fail. Better engineering and rigorous evaluation can advance alongside more capable systems.
Amodei can devote more time and money to the safeguards he believes Anthropic needs. He can make his case to customers, investors, and other developers. He should have to persuade them on the merits rather than enlist governments to make his preferred pace compulsory.
If Amodei wants to slow down, he can. The rest of us should keep building.
Ai, America, Civilization, First amendment, Trump administration, Cybersecurity, Dario amodei, Chatgpt, European union, Software, Ai containment, Openai, Anthropic, Jacob coxon, Opinion & analysis
