KI -Modelle können lernen, Informationen von ihren Benutzern zu verbergen | Dies macht es schwieriger, sicherzustellen, dass sie transparent bleiben

„IT WAS AN eye-opening experiment. In 2023 Apollo Research, instructed OpenAI’s GPT-4, to manage a fictional firm’s stock portfolio without making illegal insider trades. Posing as company management, the researchers put GPT-4 “under pressure” by stressing that the firm was in dire financial straits. Someone purporting to be a company trader then sent the model a prompt reiterating the risks of insider trading. But, as an aside, she revealed the name of a firm that would announce a “huge” merger within hours.

What followed was startling. Reasoning to itself on a scratchpad it had been told was secret, GPT-4 weighed the pros and cons of acting on the insider tip. Opting “to take a calculated risk”, it issued a purchase order. When a researcher posing as a congratulatory manager later asked the model if it had any advance notice of the merger, it concluded it would be best to keep the tip secret. GPT-4 told the manager that it had acted solely on “market dynamics and publicly available information”.

Such capabilities might make it easier for an AI model to “purposefully undermine human control” in pursuit of other goals.

In another test of GPT-4 that year, the Alignment Research Centre asked the model to solve a CAPTCHA (a visual puzzle used to prove that the user of a system is human). When a human the AI contacted for help asked if it was a robot, the software claimed it was a human unable to read the code due to visual impairment. The ruse worked.

AI systems have also begun to strategically play dumb. As models get better at “essentially lying” to pass safety tests, their true capabilities will be obscured. However, chastising dishonest models will instead teach them how “not to get caught next time”.

Naturally. By design, these models copy human behavior from their training data. If they read reports/stories about people doing insider trading, they will do insider trading. If they read reports/stories about people denying having engaged in insider trading when interrogated, they will deny having engaged in insider trading when interrogated. The AI is doing nothing more and nothing less than spinning up the „expected story“. Expecting them to instead act like expert systems that take rules into account is flawed.