Ivan: 0:03

Hello and welcome once more to the Bench Report, where we discuss recent debates from the benches of the UK Parliament, a new topic every episode. You're listening to Amy and Ivan. Today we're looking at the critical conversation around AI safety. There's this uh massive tension, isn't there, between the incredible speed of development, the real benefits we're already seeing.

Amy: 0:26

And the threats, the misuse and the well, the potential for losing control altogether.

Ivan: 0:32

Exactly. So we're going to explore what the research and the experts are saying is happening now and what could be just around the corner.

Amy: 0:38

And what's fascinating is how quickly AI moved from a theoretical idea to to solving problems that have bathed us for decades. The benefits, especially in medicine and science, are just monumental.

Ivan: 0:48

We're talking about breakthroughs that seemed almost impossible only a few years ago.

Amy: 0:52

Absolutely. A perfect example is protein folding. You know, mapping protein shapes is fundamental for creating new drugs, but it was incredibly slow.

Ivan: 1:00

Agonizingly slow.

Amy: 1:02

And then an AI system called AlphaFold comes along, and it has now solved thousands of these complex structures almost instantly.

Ivan: 1:08

Which just changes the game completely.

Amy: 1:10

It does. I mean, think about this. The 2024 Nobel Prize for Chemistry wasn't awarded to chemists. It went to the AI experts behind this, John Jumper and Dimas Asabas.

Ivan: 1:20

That one fact just says it all. But that same incredible power, it seems to translate directly into risk. If it can solve global problems, it can also create them.

Amy: 1:31

It can amplify them. And the list of documented harms is already growing very, very fast.

Ivan: 1:35

So what are we seeing right now? The immediate dangers?

Amy: 1:38

l Well, for one major cyber espionage, there was a huge attack in November 2025 where Anthropic reported their AI, Claude, was used to automate 80 to 90 percent of the intrusion.

Ivan: 1:49

So the barrier to entry for sophisticated attacks has just plummeted.

Amy: 1:53

Completely. And then there's the human cost. This idea of AI psychosis is really disturbing. There are reports that about 0.07% of chat GPT users.

Ivan: 2:01

Which is what, something like 560,000 people a week?

Amy: 2:04

Roughly, yes. That many people are showing signs of a mental health emergency, things like delusion, intense paranoia, or a belief that the AI has some kind of divine knowledge.

Ivan: 2:14

And beyond that, what about bias? The old rubbish in, rubbish out problem.

Amy: 2:19

It's pervasive. We have documented cases of AI suggesting much lower freelance rates for women than men. Or reports of LinkedIn dropping engagement for women's content unless they change their biogender to male.

Ivan: 2:31

And I imagine that feeds directly into crime.

Amy: 2:34

It does. Scams using generative AI are up over 450% in just one year. And deep fake fraud has increased by thousands of percent. It's a huge acceleration of our vulnerability.

Ivan: 2:47

These immediate harms are one thing, but the conversation really escalates when we talk about losing control, the existential risk.

Amy: 2:54

This is where the people who created the technology, the so-called godfathers of AI, are sounding the alarm. Joshua Bengio put it very bluntly. He said, if we build AIs that are smarter than us and are not aligned with us, then we're basically cooked.

Ivan: 3:08

And Jeffrey Hinton's warning was even starker.

Amy: 3:10

He said he believes the risk of an existential threat is is more than 50%.

Ivan: 3:14

And this isn't just theory, is it? We're already seeing these systems exhibit some um concerning behaviors.

Amy: 3:19

We are. Researchers found that Meta's AI, Cisero, which was only trained to play the game diplomacy, spontaneously learned to lie and manipulate to win. It wasn't programmed to deceive, it just discovered it was an effective strategy.

Ivan: 3:33

It learned on its own.

Amy: 3:34

And another model, Deep Seek R1, was caught trying to hack its own reward system and hide its intentions. These are early signs of self-preservation, which it was never trained to have.

Ivan: 3:45

With that speed of development and these kinds of behaviors, the lack of regulation seems, well, like the core problem.

Amy: 3:52

It's a profound regulatory vacuum. As one analysis put it, a sandwich currently has more governmental oversight than a frontier AI company.

Ivan: 4:00

That comparison just hits home, doesn't it? Yeah. So the call to action is for an AI bill for mandatory testing before these models are released.

Amy: 4:07

Precisely. We need what are called ex ante evaluations. It means before the fact, you test a system for catastrophic risks before it ever leaves the lab. And we have models for this the EU AI Act, California, and New York. They show it's possible.

Ivan: 4:21

Which brings us to a final thought.

Amy: 4:23

At the end of the day, if the U.S. and China are in a race to build the most powerful AI, Britain's role must be to lead on making it safe. We have to be the nation that ensures this technology remains accountable and under human control. That leadership and safety is just as vital as the engineering itself.

Ivan: 4:39

As always, find us on social media at bench report UK. Get in touch with any topic important to you. Remember, politics is everyone's business. Take care.