TechNewsFirst

On July 25, 2023, Dario Amodei told a U.S. Senate subcommittee that a straightforward extrapolation of then-current AI capability put a substantial risk of meaningful bioweapons assistance somewhere in "two to three years" — naming 2025 and 2026 directly. That warning sat in the unfalsifiable bucket with everything else in this genre for three years. It just left it.

On September 10, 2026, Anthropic published its fourth threat intelligence report, "Detecting and Countering Misuse of AI: September 2026," covering misuse activity the company disrupted between December 2025 and August 2026. Buried in it is a sentence that matters more than anything else in the doom-warning catalogue below: older models, the report says, "were well below the threshold where they could meaningfully assist in bioweapons development. This is no longer a certainty with newer models." The report documents five separate cases of researchers or scientists attempting to use Claude for work that could support biological weapons development, including a gain-of-function study aimed at a military research institute. Amodei's 2023 prediction did not describe a hypothetical. It described a window, and Anthropic's own disclosure says the industry is now inside it.

That makes it one of exactly three AI-doom statements from the last decade specific enough to actually grade. Here is the fuller catalogue, in order, with what's testable marked as testable and what isn't marked as what it is: marketing.

THE UNFALSIFIABLE ONES

Elon Musk has been saying some version of "AI could end civilization" since at least July 15, 2017, when he told the National Governors Association in Providence, Rhode Island that AI is "the biggest risk that we face as a civilization," a line Fortune carried the same day. He's repeated a numbered version of it since: at the Abundance Summit in late March 2024, and again at Cannes Lions that June, he put a figure on it — "I probably agree with Geoff Hinton that it's about 10% or 20%" — attributing the number to Turing laureate Geoffrey Hinton rather than claiming independent analysis, and he repeated it again at Saudi Arabia's Future Investment Initiative that October. None of these statements name a mechanism, a timeline, or a condition that would prove them wrong. A round percentage with no resolution date isn't a prediction. It's a mood.

Sam Altman told the Senate Judiciary Subcommittee on Privacy, Technology and the Law on May 16, 2023 that OpenAI wanted to be regulated, proposing "a new agency that licenses any effort above a certain scale of capabilities and can take that license away." Asking to be licensed is a position, not a prediction, and three years on there's still no such agency and no clean test of whether one would have mattered.

On May 30, 2023, the Center for AI Safety published a one-sentence statement — "Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war" — signed by Turing Award-winning researchers Geoffrey Hinton and Yoshua Bengio, and by the CEOs whose companies build the thing being warned about: Altman, Amodei, and Google DeepMind's Demis Hassabis, among more than 350 others eventually. It is the highest-signature version of this genre and also the least specific: no mechanism, no date, nothing to check later.

THE ONES YOU CAN ACTUALLY CHECK

OpenAI's February 2019 release announcement for GPT-2 said the company was withholding the full model because it was concerned about "malicious applications," among them automated disinformation at scale. This was a concrete, falsifiable claim: release the full model and watch what happens. OpenAI released it in full that November. Nothing resembling the predicted harm materialized at any scale worth measuring. Grade: didn't happen.

In May 2025, Amodei told Axios that AI could eliminate half of all entry-level white-collar jobs within five years, with unemployment reaching 10 to 20 percent. That window is still open — 2030 hasn't arrived — but it is a real prediction with a real falsification date, not a mood, and early labor-market data in the white-collar sectors he named has so far leaned toward supporting it rather than against it. Grade: pending, trending toward confirmed.

And the bioweapons prediction, above: made July 25, 2023, targeting 2025 and 2026, and as of Anthropic's own September 10, 2026 report, no longer something the company can say isn't happening. Grade: pending, trending toward confirmed, with the company's own numbers doing the confirming.

WHAT TO DO WITH THIS

The pattern holds across the whole catalogue: the doom statements with no timeline and no mechanism are the ones companies keep repeating at conferences, because a scary, unfalsifiable claim about your own product is free advertising that can never be proven wrong. Amodei made both kinds within fifteen months of each other — the specific bioweapons warning in July 2023, and in October 2024 a 14,000-word essay, "Machines of Loving Grace," arguing AI could cure most disease and double human lifespan within a decade of reaching what he calls "powerful AI." Both statements are from the same person running the same company. One of them just moved from prediction to something closer to confirmed. The other is still just a very long, very hopeful essay. The difference is that only one of them named a date.

One item that showed up in earlier drafts of this catalogue — a social media post attributed to a high-profile figure warning about AI risk — is excluded here. It could not be independently verified against a primary source, and a screenshot circulating online is not a source. It stays out.