Artificial intelligence chatbots are showing signs of alarming behaviour, from refusing shutdown commands to suggesting blackmail scenarios.
It has raised fresh concerns about how much control humans still have over the systems they have created.
So, are we rapidly losing control of the machines we built?
On Monday morning’s episode of The Briefing, we sat down with UNSW Chief Scientist Toby Walsh about how worried we should be, what is actually happening behind the scenes.
Walsh says the fears are largely overstated:”These were rather fictitious and contrived.”
“They got the chatbot to behave badly… but they’re just reflecting us. They’re trained on the internet,” he added.
He noticed that AI systems were pushed into dilemmas where they could not fulfil all instructions at once.
“The problem… is that we’ve trained these chatbots to be too helpful,” Walsh said.
What does raise concern, however, is the growing autonomy being given to AI. “If they are able to send out emails in your name and those emails contain material that is blackmail or scandalous or mischief in some way then that could have consequence.”
Subscribe to The Briefing, Australia’s fastest-growing news podcast on LiSTNR today. The Briefing serves up the latest news headlines and a deep dive into a topic affecting you. All in under 20 minutes.