Anthropic's 2030 model: fast growth is the bad scenario
Two very different kinds of AI governance landed today — one company publishing the numbers on what AI does to work, and one union writing rules a vendor can actually be sued over.
Anthropic's economics team shipped a public scenario explorer that models the US economy in 2030, and its uncomfortable finding is that the fastest-growth path is the one that hurts workers most. The model treats every job as a bundle of tasks drawn from the Labor Department's O*NET taxonomy — more than $30 trillion of task value a year — then lets you dial in how capable AI gets, how fast it spreads, and whether it augments or replaces each task. In the modest scenario, growth gets a gentle lift and little else moves. In the substantial scenario, the one the median respondent in Anthropic's survey of more than 10,000 Americans landed on, GDP runs about 10% higher by 2030 and unemployment sits near 5%. The extreme scenario, which assumes recursive self-improvement and rapid adoption, is where it breaks: GDP grows at more than seven times its current pace, nearly 14% of workers lose their jobs to AI, and fewer than half of them find new ones. Output soars while wages for knowledge workers stagnate or fall and capital takes a bigger slice of the pie. Co-founder Jack Clark's read to NPR was that the technology will keep improving fast but diffuse more slowly than people expect — and that "policymakers should get ready to spend."
The most interesting thing here is what the model leaves out, and Anthropic says so plainly: no policy responses, no business cycles, no catastrophic risk, and no aggregate-demand effects from the data-center buildout. That means the scenario where AI growth is fastest is also the one where the risks its own alignment researchers spent this week warning about are simply not in the equation. This is a genuinely useful public artifact — the transparency about the pretraining-corpus-style limits is rare for a corporate model — but it is a tool for exploring assumptions, not a forecast, and it quietly assumes the tail risks away.
The American Federation of Teachers and Microsoft agreed on ten legally enforceable AI protections that districts can write straight into their Microsoft contracts. Microsoft committed not to use student or teacher data to train or improve its models outside narrow safety and security exceptions, not to track students, and not to let AI make decisions in schools without human oversight; districts can seek damages if the company breaks the terms. President Brad Smith said the protections go out to every district in the country starting November 1, and the AFT says it is already talking to OpenAI and Anthropic about signing on. It lands a week after New York City's public schools moved the other direction entirely — New York City bans student-facing AI through eighth grade. Our take: a union negotiating contract terms a vendor can be sued over is a more durable form of AI governance than most of what statehouses have produced this year, precisely because it attaches to procurement rather than to principle.
The first person convicted under the Take It Down Act was sentenced to 15 years in federal prison. James Strahler II of Columbus, Ohio, received 180 months for a campaign that included more than 700 images of real victims and AI-generated depictions posted to a child-abuse site, plus another 2,400 flagged files on his phone; he had installed more than 24 AI platforms and 100 web-based models on that phone. He had generated videos using the faces of boys from his own community and sent AI-made pornography of an ex-partner to her co-workers. The relevant count is publication of digital forgeries, a charge created by the 2025 law banning nonconsensual publication of intimate images and AI fakes. Prosecutors called it the first conviction in the country under the Act, which is the real signal: the deepfake-abuse statute is now producing decade-scale sentences, not consent decrees.
What to watch: whether OpenAI and Anthropic sign the AFT's standard before November 1, and whether Anthropic's scenario explorer gets a v2 that puts catastrophic risk back inside the model.
If Anthropic's own extreme scenario is the one where growth is fastest and workers lose the most, should a lab be publishing that model — or legislating against it? Tell us in the comments.
