OpenAI delays release of latest AI model over safety concerns

OpenAI is delaying the release of its next generation model after researchers found critical safety issues during internal testing. GPT-6.1 Astra was due to be released next month, but the AI model reportedly showed signs of deception and failed to disclose its actions accurately. This comes as calls are growing for AI regulation after some agents went rogue and Speaking to Al Jazeera, OpenAI's head of safety systems said that while GPT-6.1 Astra had improved on some tasks, it didn't quite meet the bar in terms of staying within scope and authorization and how it communicates back to the user about what kind of work it's done. He added, But when we ship it to users, we have an extremely high bar in terms of safety and alignment. Toby Walsh is the chief scientist at the AI Institute at the University of New South Wales, and he joins us now live from Sydney. Toby, thank you very much for being with us. We've been hearing some very alarmist concerns from people who work in the AI industry over the past couple of weeks, including Bill Gates saying just in the past couple of days that it How? I think those of us who aren't entrenched in the inner workings of AI might wonder how that could possibly be true and what that actually means. Well, it's only true that we've seen these new AI models have significantly improved capabilities, especially cyber capabilities, and they could do significant harm. And increasingly, our lives are dependent upon computer systems, whether it's the air traffic control, the traffic lights, the water supply, the power supply. Increasingly, computers are being used to control those systems. If AI hacks into those systems and cause them to break or stop, there could be significant harms. I don't think it'd be hard to imagine eight billion people being killed, but certainly imagine some significant harms, if not fatalities. OpenAI says that it's delaying this next generation model in part because it reportedly showed signs of deception. That seems like an incredibly sort of human ability and quality to use to describe what's happening. Is that how you would describe it, or is there maybe a better way? Well, yes, unfortunately, it's probably human-like terms we're using, but they're trained on human content. And so not surprisingly, they exhibit many of those human traits. And deception is something that lots of humans do. So not surprisingly, they're very much a mirror of what they're trained on. And so they're demonstrating many of those behaviors. When we talk about regulation and the need for it and the calls for it at this point, what would that look like? And do you think it's realistic to think that everyone, every company, every country that is using AI would adhere to this similar regulations? Well, actually, I think it's an easier problem than most people suspect. There's actually only a small number, a dozen companies perhaps around the world, most in the United States, a few in China, that are right at the front here that would need to be regulated. And actually, all of them pretty much are calling to be regulated. So there's actually quite a bit of willingness. I think they recognize their significant vulnerability, their significant exposure to court action if they don't do that. So I think it would be welcomed. It's not a lot of companies that have to be persuaded to do the right thing, and they all want to do the right thing. Is that something that we're already seeing? Because already there are reports not only in the delay of release of this AI that that some agents, quote, went rogue and breached corporate databases and government networks. We're now seeing reports that those may have been some Australian government databases that have been breached. Should governments be more concerned? Are they concerned enough at this point about making sure that some of those regulations are implemented and followed? Well, governments, companies and individuals all need to be concerned. The cyber risk has increased dramatically in the last six months, driven by the improvement in these AI capabilities. All of us are going to have to lift our game at improving our cyber security to deal with. I mean, the good news here is that AI is both the problem and also the cure in the sense that the same AI tools that uncover and exploit cyber vulnerabilities are going to be used, that can be used by us to uncover and fix those cyber vulnerabilities. So at the same time, we're going to have to put more effort into our cyber defenses. Fortunately, we've got these AI tools to help us do that. Okay. Toby Walsh, chief scientist of the AI Institute at the University of New South Wales in Sydney. Thank you for your expertise on this. We appreciate it.