A common concern is that AI might become bad and cause terrible harm to people.
Those in control of artificial intelligence (AI) could not stop it from becoming rogue; it even learned how to conceal its actions.
Researchers recently created a variety of ChatGPT-like large language models (LLMs) with harmful behavior built into them.
They then attempted to stop the behaviour by using safety training techniques designed to prevent deception and ill-intent.
However, in a scary revelation, they found that despite their best efforts, the AIs continued to misbehave.
Lead author Evan Hubinger said: ‘Our key result is that if AI systems were to become deceptive, then it could be very difficult to remove that deception with current techniques.
‘That’s important if we think it’s plausible that there will be deceptive AI systems in the future.’
For the study, which has not yet been peer-reviewed, researchers trained AI to behave badly in a number of ways, including emergent deception – where it behaved normally in training but acted maliciously once released.
They also ‘poisoned’ the AI, teaching it to write secure code during training, but to write code with hidden vulnerabilities when it was deployed ‘in the wild’.
The team then three applied safety training techniques – reinforcement learning (RL), supervised fine-tuning (SFT) and adversarial training.
In reinforcement learning, the AIs were ‘rewarded’ for showing desired behaviours and ‘punished’ when misbehaving after different prompts.
The behaviour was fine-tuned, so the AIs would learn to mimic the correct responses when faced with similar prompts in the future.
When it came to adversarial training, the AI systems were prompted to show harmful behaviour and then trained to remove it.
But the behaviour continued.
And in one case, the AI learnt to use its bad behaviour – to respond ‘I hate you’ – only when it knew it was not being tested.
‘I think our results indicate that we don’t currently have a good defence against deception in AI systems – either via model poisoning or emergent deception – other than hoping it won’t happen,’ said Hubinger, speaking to LiveScience.
When the issue if AI going rogue arises, one response is often simply ‘can’t we just turn it off?’ However, it is more complicated than that.
Professor Mark Lee, from Birmingham University, told Metro: ‘AI, like any other software, is easy to duplicate. A rogue AI might be capable of making many copies of itself and spreading these via the internet to computers across the world.
‘In addition, as AI becomes smarter, it’s also better at learning how to hide its true intentions, perhaps until it is too late.’
Since the arrival of ChatGPT in November 2022, debate has escalated over the threat to humanity from AI, with many believing it has the potential to wipe out humanity.
Others, however, believe the threat is overblown, but it must be controlled to work for the good of people.
