Should we be worried about accidently creating Skynet, the AI of doom?

Illustration of a robot with the ChatGPT logo. As Jeff Goldblum's character in the 'Jurassic Park' movies famously says, 'Your scientists were so preoccupied with whether they could, they didn’t stop to think if they should'. Well, with AI, we definitely should, says our columnist – but it’s how we do it responsibly that is the hard part. — Reuters

In my column in February about artificial intelligence (AI) and ChatGPT, I compared the new chat bot to autocorrect on steroids ("Did ChatGPT say it can 'hack systems' and it wants to 'kill people' Yes, but wait!", The Star, Feb 26)

Well, that’s changed.

What I was talking about then was ChatGPT-3; well, developer OpenAI has now released ChatGPT-4, which it promises has better problem-solving abilities and advanced reading capabilities. Indeed, I’ve seen posts on Twitter where users typed their query in backwards and GPT-4 still answered the question. Testers are saying that the work this new and improved AI generates is on par with human work. Not better. Not genius level. Not yet.

From all accounts GPT-4 is orders of magnitude better than GPT-3. It’s April 2023. GPT-3 was released in November 2022. Six months later and we’re getting something that is much more advanced. In fact, some experts are saying that GPT-4 has “sparks of artificial general intelligence”, or AGI, which means the AI can operate at human or above human intelligence level. When people talk about doomsday AI, it’s this leap to AGI they mean.

This led to the Future of Life Institute, led by the Massachusetts Institute of Technology’s Max Tegmark, to release a letter calling for an AI moratorium. Specially, a six-month pause in developing these experimental AI models.

Why the pause? To slow things down. Take a breath. Catch up a on what we’re doing here.

Those in favour of the pause include tech mogul Elon Musk and Apple cofounder Steve Wozniak, who both signed the letter, along with hundreds of other technology and thought leaders.

What would a six-month break do? Maybe allow those creating the AIs to understand them a bit better? Surely the developers understand the AI, you say? No, not necessarily.

For instance, ChatGPT is famous for giving succinct answers – but it’s also notorious for making stuff up. Once, I used a jailbreak – something that helps the AI circumvent its programming by roleplaying – and it revealed several different dates for a stock market crash. When I asked why the different dates, GPT-3 said, “It’s all part of the fun of roleplaying”.

Asking for investment advice, medical advice, and other such life advice is on the list of things ChatGPT shouldn’t answer. The OpenAI website boasts that GPT-4 is 84% less likely to “respond to requests for disallowed content”. That’s great but why isn’t it 100%? Why is ChatGPT saying it’s just having fun making stuff up when I ask it questions?

According to an article in The Guardian, when an author of the “sparks of artificial general intelligence” paper was asked if AI is implementing its own goals when providing answers the presumably go against its programming, the answer was “we have no idea”.

And that’s just it. The people developing this don’t really know how their creation works. That’s a tad scary.

But at this point, even if GPT-4 becomes Skynet (the AI that dooms humanity in the Terminator movie series) there isn’t much it could do to us. As experts have noted, this thing lives on a data server and data servers can be destroyed, not to mention they need constant maintenance. There is no robot factory for GPT-4 to take over and make an army from – unless it takes over the Roomba factory and spits out Terminator vacuum cleaners. So in that sense, we’re pretty safe, right? Maybe.

Opponents of the six-month pause argue that people are focusing on the unlimited downside of the AI. But as AI has definitive world-altering abilities, putting off development could lead to a longer pause that delays breakthroughs. Breakthroughs that could happen in every industry.

What if it is possible for the millions of factors that make you you, to be analysed and a comprehensive health plan built only for you? AI can do that. And that’s just one example. AI will literally be in everything. Making everything easier. Better.

But can we be sure AI won’t turn into Skynet? No, we can’t. Especially if we don’t even know why and how it’s doing what it’s doing right now, in these very early stages of AGI.

Once again I find myself thinking of Jeff Goldblum’s character in the Jurassic Park movies when he finds out dinosaurs have been cloned and he says, “Your scientists were so preoccupied with whether they could, they didn’t stop to think if they should”.

With AI we definitely should, but it’s how we do it responsibly that is the hard part.

---

Big Smile, No Teeth columnist Jason Godfrey – a model who once was told to give the camera a ‘big smile, no teeth’ – has worked internationally for two decades in fashion and continues to work in dramas, documentaries, and lifestyle programming. Write to him at [e-mail] and follow him on Instagram @bigsmilenoteeth and facebook.com/bigsmilenoteeth. The views expressed here are entirely the writer's own.