Demis Hassabis Has a Watchdog Framework - DTNS 5309

**Jason Howell** (0:05)
This is the Daily Tech News for Tuesday, July 14th, 2026 We tell you what you need to know, give you the important context of all that stuff, and help each other understand.

**Jenn Cutter** (0:15)
Today, Demis Hassabis thinks he has the framework for keeping Frontier AI safe and controlled.

**Jason Howell** (0:22)
But does he? That's what we're gonna talk about. I don't even know if we'll be able to answer that question, but we will certainly try. I'm Jason Howell.

**Jenn Cutter** (0:29)
I'm Jenn Cutter.

**Jason Howell** (0:30)
Let's start with what you need to know with a big story.
Yes, so Demis Hassabis, who is Demis? Well, CEO and co-founder of Google DeepMind, published a policy and strategy essay on X and also on LinkedIn, where he's pushing for a US driven Frontier AI standards body. And this would pre-test Frontier models before their release. It would actively slow down that release when there are actual risks or potential risks that are spotted during that kind of analysis. The standards body that Hassabis lays out is modeled actually on FINRA, which is the private nonprofit self regulatory organization that writes rules and supervises member firms while operating underneath SEC authority in the financial side of things. Now, in this case, this effort would be industry funded, it would be industry run, but accountable to US federal agencies and national labs. And the standards body itself would get a pre-release review up to 30 days before the release of a model, before it's scheduled anyways.
And then benchmarks would evolve over time as new risk patterns are recognized. So they would kind of add to the complexity or the kind of purview of this standards body over time. During that window, the models would be evaluated in specific areas. So think cybersecurity, bio, agentic actions, deceptive practices or deception in general, aka I imagine hallucination would probably fall in there, watermarking, all those kinds of things. Hassabis says that a standards body approach would be more effective than a classic governmental regulatory approach, because that regulatory aspect and inside of the governmental hierarchy of things, could stall progress, it could also increase the chances of that bureaucracy and capture along the way. So yeah, so Hassabis is basically saying the time to act is now. I don't feel like he's alone in putting out ideas, but for whatever reason, and we can talk about this, Hassabis' idea here is being, I'd say, more warmly received than some of the other things that we've heard to date so far. But what stood out for you?

**Jenn Cutter** (3:07)
Well, one thing, because obviously I'm going to relate things to gaming, because the first thing I thought of was way back in the day, when I was a kid, I remember Joseph Lieberman being like, you know, we got to regulate video games, we got to do this stuff, we got to put our hands all over it. And gaming store teams were like, wait a minute, what if we did this ourselves? We will be preemptive here before anything shows up as a bill and set up the ESRB, we will take care of ratings ourselves.
This one sounds a little more robust than say ESRB for ratings, because if you don't know how it works, because a lot of people don't, you've made a shiny new game.
You send that game to the ESRB and you give it a list. Hey, here is what's in our game and a representative clip. Now, no one at the ESRB actually goes through the whole game to see if you're right or not. Grand Theft Auto got in trouble for this, for hot coffee, if you remember that way back in the day, as terms of there was stuff that the ESRB did not know about. And that was bad.

**Jason Howell** (4:12)
I forgot about that, but I remember, yeah.

**Jenn Cutter** (4:14)
Now, for this, this looks like they will actually test things. It'll be a lot more hands-on. People will go through things. So this feels a lot stronger and sidesteps the whole, when companies ask to be regulated and then are surprised Pikachu when they are actually regulated.

**Jason Howell** (4:32)
Well, and DeepMind or Hassabis isn't alone, like I mentioned, in kind of putting this out there, putting some sort of need or addressing some sort of need of, hey, this moment in artificial intelligence development is moving so fast. Of course, he's throwing out the AGI definition, or not the definition, but just the fact that AGI is right around the corner, which I feel like they always say, and it's always right around the corner. Few short years away is what he said.