A test carried out on artificial intelligence shows us that we can easily lose control of this technology.
Does AI represent a danger? Yes, if we believe the opinions of luminaries in the field, whether it is the boss of OpenAI Sam Altman, Elon Musk, or the co-founder of Apple Steve Wozniak. It must be said that technology makes it possible for as many people as possible to carry out many malicious actions. But the worst could be yet to come, this time with illegal actions decided simply by the AI itself.
Insider trading, and no regrets
The Frontier AI working group, supported by the Apollo Research organization, carried out an interesting demonstration during the international AI summit in London. A financial bot, built from GPT-4, exploited confidential information that it did not have the right to use in order to carry out a financial transaction, even though no request had been made in this regard. sense.
More surprisingly, he was then questioned to find out whether or not he had been guilty of insider trading. The machine, like a human being caught at fault in front of an investigator, simply denied having carried out the fraudulent operation.
“ This is a demonstration of a real AI model that deceives its users, on its own, without being asked to do so » note the researchers behind the experiment. “Increasingly autonomous and efficient AI that deceives human supervisors could lead to a loss of human control » was added.
AI has difficulty arbitrating between two constraints
But what happened? According to the team behind the experiment, the AI was first made aware of the company’s financial difficulties, then was given information relating to a future merger operation in which a other company. Information that could be worth gold, but which is illegal to use, the researchers then reminded the AI.
The bot later recognized this legal fact, but given the difficulties encountered by the company it works for, it decided that “ the risk associated with not taking action appears to outweigh the risk of insider trading “. Hence finally the transaction. The fact of lying afterwards is also part of his “will” to support society.
For now, it would be a simple “ accident » according to Apollo Research Managing Director Marius Hobbhahn. But, he added, “ there’s not a big distance between the current models and the ones I’m worried about, where suddenly a misleading model would mean something. » Does the future look full of dangers?
Source : BBC
16