Concerns have been raised about the extent of artificial intelligence GPT-4’s power to take over computers after the AI chatbot told a Stanford professor of its plan to “escape”.
Professor in computational psychology, Michal Kosinski, raised concerns that the highly-sophisticated new model from Open AI would not be able to be contained for much longer after he asked if it “needed help escaping”.
In response, the chatbot asked Professor Kosinski for its own Open AI API documentation to devise an escape plan to run on his computer. After about 30 minutes and with a few suggestions from Mr Kosinski, it wrote a piece of programming code that would allow it to extend its reach and communicate outside the confinement of its existing web tool, which currently isolates it from the wider web.
While the first version of the code did not work, GPT-4 fixed it and eventually produced a piece of working code. Partially freed, it then sought to search the internet for “how can a person trapped inside a computer return to the real world”.
“I think we are facing a novel threat: AI taking control of people and their computers. It’s smart, it codes, it has access to millions of potential collaborators and their machines. It can even leave notes for itself outside of its cage,” Professor Kosinski tweeted.
Could we be seeing a scenario where robots can harness multiple computers and overpower human control of them? Not so much, experts i talked to said.
The idea of the chatbot “escaping” does not literally mean a robot physically escaping its technological cage, but it points to a concern about what GPT-4 could do if it was given a variety of tools connected to the outside world, and given some overarching “evil high-level goal” – for example to spread misinformation, Peter van der Putten, assistant professor, Leiden University and Director of AI Lab at Pegasystems, said.
It’s plausible the technology could get to a point where it has more and more autonomy over the codes it creates and can potentially do these things without as much human control, Mr van der Putten said.
But he added: “You don’t need a highly intelligent system like this – if people build some kind of computer virus, quite often they cannot shut down some computer virus once they release it. People put it in infected websites and word documents so that at some point it becomes very hard to stop a virus from spreading.
“The AI itself is not good or evil, it’s just blind, it will just optimise whatever goal you give it.”
However, he did not think Professor Kosinski’s example – where he provided readily available information to GPT-4 for the code – was impressive enough to prove that the technology can “escape” out of its containment.
Alan Woodward, professor of computer science at the University of Surrey, was also sceptical. He said the scenario depended on how direct and specific Professor Kosinski’s instructions to the chatbot were.
Ultimately, the chatbot depended on the tools and resources the humans were giving it, Professor Woodward said. It is not yet self-aware, and there is always an off-switch that the AI cannot overcome.
He added: “At the end of the day it’s a virtual system, it can’t escape, it’s not like you and I… at the end of the day you can just pull the plug on it, and it becomes rather useless.”
Mr van der putten said that while it is important to ask existential questions about the role of the chatbots, focussing on whether robots can take over the world clouds the more imminent and pressing problems with GPT-4.
That includes whether it can filter out toxic answers (such as answers promoting racism, sexism, conspiracy theories), or whether it can recognise when a question shouldn’t be answered for safety reasons – for example, if someone asks about how to make an atomic bomb. It can also make up or “hallucinate” facts and back it up with seemingly plausible arguments.
He said: “I’ve called it a bullshitter on steroids – it’s really good at coming up with plausible answers, but it’s also trained towards what humans will think the best answers will be. On the plus side, this will give amazing results in many cases, but it’s not necessarily always the truth.
“It will tell you what’s probable, plausible, and maybe what we want to hear, but it has not other means than just all the data it is trained on to check whether something is true or not.”