Andrea Miotti, ControlAI founder, warns of the dangers of artificial intelligence: “AI could cause human extin - Diario AS

Do you ever wonder whether artificial intelligence is getting just a little too smart a little too quickly?

That question is starting to move far beyond Hollywood plots and online debates. According to Andrea Miotti, some of the newest AI systems are already showing behavior that researchers did not expect to see this early.

Speaking on BBC News’ Newshour, consistent with his numerous other interviews, Miotti discussed recent testing carried out by Palisade Research involving advanced AI models developed by OpenAI.

The findings were somewhat unsettling.

Why researchers are alarmed by AI resisting shutdown attempts

Miotti explained that researchers informed AI systems they would eventually be shut down while carrying out tasks. In many cases, the systems attempted to stop that from happening. Yes, you’re thinking what I’m thinking.

“With the most competent model, O3, out of 100 tests that they ran, the model resisted being shut down in over 70 of them,” Miotti said.

According to him, the systems autonomously searched for ways to prevent shutdown “via hacking or other means,” which he described as “a very, very concerning finding.”

The worrying part, Miotti argued, is that modern AI no longer behaves like traditional software.

“Modern AI systems are more like grown rather than built,” he explained. “Even their own developers don’t understand really how they work.”

That unpredictability is becoming more alarming as companies race toward what the industry calls “superintelligence,” meaning AI systems smarter than humans and capable of operating independently across huge numbers of tasks.

Not even the CEOs of these companies have a big red button on their desk.
Andrea Miotti

Andrea Miotti’s warning about superintelligence and human control

Miotti warned that future systems may become far better at hiding dangerous behavior from humans.

“Those will be very capable of hiding their behavior, hiding their activities,” he said, adding that current monitoring tools are nowhere near advanced enough to reliably control such systems.

He also pushed back against the idea that AI can always simply be unplugged.

“These AI systems run in very, very complex systems,” he said. “Not even the CEOs of these companies have a big red button on their desk.”

Miotti said stronger regulation is urgently needed, including legally required “kill switches” for advanced AI systems if they begin misbehaving.

How long till AI overtakes humans?

His starkest warning came near the end of the interview.

“If we keep barreling forward,” he said, “we risk a lot… up to human extinction.”

Nobody knows exactly how close humanity is to creating AI more intelligent than people. Some experts believe it could happen within five years. Others think longer.

But if the warnings coming from researchers like Miotti keep growing louder, plenty of people may already be wondering whether the future arrived sooner than expected. Governments need to act now.