Extinction is speculation, misuse is the risk: IDSIA director Rizzoli cools the AI doom debate

The director of a Swiss AI research institute has entered the doom debate with a firm shrug at the extinction scenarios. Andrea-Emilio Rizzoli, who leads the Dalle Molle Institute for Artificial Intelligence (IDSIA USI-SUPSI), says the warnings driving this month's headlines are highly speculative, and that the risks deserving attention are concrete: biological weapons, sabotaged infrastructure and oversight left to the companies selling the technology. His sharpest point is not about capability at all. Control of AI, he argues, is largely delegated to the very companies developing it, which have an interest in exaggerating what it can do to attract investment. Rigorous independent oversight, in his telling, is missing.

Watch Desk analysis

What happened

The debate he is answering was reignited by Jacob Coxon, the former Anthropic researcher whose resignation and warnings about future systems reignited the argument this month. Speaking on an RSI newscast, Rizzoli split the question into the pieces that rarely survive a social media exchange: capability, intention and use.

On the extinction scenarios: 'I would definitely say we do not need to be afraid. We are talking about highly speculative situations, with an undefined time horizon.' Today's AI, he said, 'has neither consciousness nor intention, and therefore cannot voluntarily decide to exterminate the human race', adding that he seriously doubts it would have the capability either.

That is not a claim that nothing can go wrong. 'We can imagine catastrophic situations, for example in the uncontrolled development of biological weapons or the deliberate sabotage of IT systems that regulate critical infrastructure, such as air and rail traffic or electrical power,' he said. 'These are genuinely concrete risks. That artificial intelligence might instead kill us all seems extremely difficult to me.'

He was equally candid about the spread of expert opinion. Geoffrey Hinton, the Nobel laureate, sits among the most concerned, while researchers who think catastrophic development is possible 'still represent a minority', though positions run across an extremely wide spectrum. Per USI's report of the interview, his prescription is worldwide regulation, and he notes that politicians in the United States are asking increasingly precise questions about control.

The central claims

• No extinction threat from today's AI
  Current systems have neither consciousness nor intention, so they cannot choose to exterminate anyone, and Rizzoli doubts they have the capability either.
• The real risks are misuse
  Uncontrolled biological weapons development and sabotage of the IT systems behind air, rail and power infrastructure are the concrete dangers he names.
• Oversight has a capture problem
  Control is delegated to the companies building the systems, which have an incentive to exaggerate capabilities to attract investment.
• Regulation needs to be worldwide
  Something is already moving, he notes, with US politicians asking increasingly precise questions about control.

Why it matters

Where you locate the risk decides what you do about it. If the danger is a system that wants to survive and outwit us, the answers are pacing pledges, coordination and kill switches. If the danger is people misusing a powerful tool, the answers are security, liability and control of who gets access. Rizzoli's framing pushes hard towards the second list, and he reaches for a familiar comparison: an issue 'that has accompanied all advanced technologies, from nuclear energy onwards'.

The governance critique is the part that should make uncomfortable reading in California. His complaint about oversight delegated to companies with a commercial stake in the hype lands in the same week Anthropic announced embedded evaluators and Amodei's pacing essay put antitrust waivers back on the table. He also points at transparency, contrasting closed models with open efforts such as Switzerland's Apertus. Whether oversight can sit inside the building and stay independent is precisely the argument his remarks sharpen.

Our read

The useful thing here is not the reassurance, it is the dissection. Doomers and dismissers alike tend to flatten the question into a single shout, and Rizzoli quietly takes it apart: no intention, doubtful capability, real misuse risk, captured oversight. You can disagree with his odds and still steal his structure.

His minority point cuts both ways, and he knows it. Hinton has a Nobel Prize and a congressional audience; Rizzoli has an institute and a broadcaster slot. Neither is a poll. The honest position is the one he occupies: a wide spectrum, a minority expecting catastrophe, and a governance gap almost nobody disputes. Treat the remarks as an invitation to argue about incentives rather than prophecies, because incentives are the thing policy can actually reach.

What to watch

• The counter-speech
  Whether other institute-level researchers follow Rizzoli's line, or whether the next intervention comes from the concerned camp again.
• Oversight independence
  Whether embedded evaluator schemes win funding and reporting lines that do not run through the labs being evaluated.
• The regulation track
  Whether those 'increasingly precise questions' from US politicians turn into bills with operative terms.
• The open-model argument
  Whether transparency cases like Apertus start appearing in oversight proposals rather than opinion pieces.

Discussion spark: Rizzoli says the sore point is oversight delegated to companies with an incentive to exaggerate their capabilities. Can oversight funded by the labs it watches ever count as independent, or does real independence require someone else's money?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.