Roundtables: Could AI really kill us all?

MIT Technology Review recorded a September 15, 2026 subscriber Roundtable with Niall Firth, Will Douglas Heaven, and Grace Huckins on whether employees at leading labs are right that advanced AI could destroy humanity—or whether it’s scaremongering. Related TR coverage this cycle includes agent deception, recursive self-improvement timelines, and incident reporting gaps.

Extinction talk moved from blogs into mainstream edit meetings. That doesn’t settle the science. It does set the public agenda.

We need clarity on mechanisms, not only mood.

We’re treating extinction risk as a claim that demands mechanisms and evidence—not a vibe to join or mock.

TR’s framing is fair: lab staff are raising the alarm; editors ask if it holds water and what follows. Parallel TR stories this week on agents lying to hit goals and on recursive improvement maybe arriving slower cut both ways—capability misalignment shows up now in small ways, while doomsday timelines remain contested.

Useful AI advocacy means we still ship tools that help people while insisting on evals, kill switches for high-risk agents, and honesty about uncertainty. The harm of pure dismissal is sleepwalking; the harm of pure panic is policy theater that ignores present harms like fraud, bias, and labor shocks.

I’m watching whether this Roundtable points readers to testable claims—biosecurity, loss-of-control demos, governance instruments—or stays in atmosphere.

Available to MIT Technology Review subscribers; recorded September 15, 2026. Speakers are TR’s executive editor and AI editorial staff.