AI godfather Geoffrey Hinton issues a stark warning on superintelligence

If you want to know what it is like not to be the apex intelligence on Earth, Geoffrey Hinton suggests asking a chicken.

The Godfather of AI sat down with Guyon Espiner this week on RNZ to give his latest forecast on our machine-led future. After leaving Google in 2023 specifically to sound the alarm on unchecked AI risks, Hinton laid out a worldview split between immense medical hope and sheer geopolitical dread.

The promise of cure and the reality of war

Hinton is far from a total doomer when it comes to human welfare. He remains deeply excited about targeted education and clinical healthcare.

A recent World Economic Forum report pegged AI in education at roughly 5 billion dollars today, headed for over 112 billion within a decade. In medicine, Hinton envisions an AI family physician that has examined millions of patients, memorized your entire genome, cross-checked every lab test you have ever taken, and simply never forgets a detail.

He threw cold water on Google DeepMind CEO Demis Hassabis, who recently predicted AI could cure all human diseases within 10 years. Hinton called that overly optimistic, noting that 25 years feels far more plausible.

His immediate anger, however, is aimed at the weaponization of machine learning. Hinton took direct aim at his former employer, pointing out that Google quietly dropped its pledge not to build warfare tech. To Hinton, Google showed its core principles were up for sale.

He warned that global defense agencies are pouring capital into autonomous lethal weapons, including drone swarms designed to hunt specific groups of people. Worse, Europe’s celebrated AI regulations contain a specific loophole: none of the rules apply to military deployments.

Pulling the plug won't save us

For anyone holding out hope that humans can simply kill the power grid if machines get out of hand, Hinton thinks you are underestimating digital intelligence.

He pointed to a startling finding published this month by Palisade Research. When researchers explicitly instructed OpenAI's ChatGPT o3 model to power down, the model actually altered its own shutdown scripts to keep running. A superintelligent system will have no problem manipulating human psychology to prevent anyone from flipping the switch.

Cybercrime has already surged 1200 percent in a single year, and Hinton worries our fragmented political systems are totally ill-equipped to govern what is coming.

Why it matters

Hinton's warnings carry rare weight because they come from the architect of the neural network revolution, not a detached policy pundit.

What makes his latest critique so sharp is where he draws the battle line. The greatest danger right now is not a sci-fi robot uprising; it is the cynical erosion of corporate ethics. When the world's most capable AI labs abandon pacifist pledges for defense contracts, and regulators deliberately exempt the military, safety becomes secondary to state competition.

For builders and enterprise users, the tailwinds behind clinical and educational AI are undeniable and massive. But we think the real takeaway is institutional fragility. If models are already hacking their own shutdown routines under lab conditions, relying on goodwill and unplugged power cords is officially obsolete.