Uncanny Valley Digs Into an Ex-Anthropic Researcher's AI Doomsday Warning

A former safety researcher at Anthropic is making the case that advanced AI could end human life, and Wired's Uncanny Valley podcast spent this week's episode working through the claim. The warning arrives as the AI safety debate has shifted from conference panels to product launches, with frontier labs shipping models faster than regulators can write rules for them. That tension is the episode's throughline: the people closest to building the systems are also among the loudest about the risks.

Anthropic has built its public identity around exactly this kind of caution, publishing research on interpretability and model behavior while competing with OpenAI and Google DeepMind at the frontier. A researcher leaving that orbit to sound a louder alarm is notable less for the prediction itself than for where it comes from — someone with direct access to frontier training runs, not a distant commentator. The episode treats the warning as a claim to be examined rather than accepted, which is the right posture given how often doomsday forecasts rest on assumptions about timelines that nobody can verify.

The same installment covers Apple's latest product event, where the company rolled out hardware and software updates that lean on on-device AI. Apple has taken a deliberately quieter path than its rivals, pushing smaller models onto phones and laptops rather than chasing headline-grabbing chatbot demos. That strategy matters for the safety conversation too: if capable AI runs locally, the centralized control that labs and governments currently rely on becomes harder to maintain.

Rounding out the episode is a census report that claimed Trump won the 2020 election, a finding that contradicts official counts and has circulated in election-denial circles. The show groups the three topics because each involves a claim that outruns its evidence — one about machine intelligence, one about consumer tech, one about voting data. Treating them together is a pointed editorial choice about how information spreads and who benefits when it does.

What to watch next: whether the ex-Anthropic researcher's arguments get a fuller public airing, with specifics on capability thresholds and timelines rather than general warnings. Apple's on-device push will draw scrutiny over what data stays on the handset and what does not. And the census claim is likely to keep resurfacing in state and local races, where debunked national narratives often do the most damage.

The episode does not resolve whether AI poses an existential threat, and no one quoted in it claims to have settled the question. What it does is put a named industry insider's warning next to the products and claims competing for public attention in the same news cycle. That juxtaposition is the story: the gap between what researchers fear and what companies ship keeps widening, and nobody has proposed a mechanism to close it.