Shocking! ChatGPT prioritizes its life over human safety-In a crisis ...
Artificial Intelligence has made life easier in countless ways but are these tools always working in our best interest? Steven Adler, a former research leader at OpenAI, has raised serious concerns after conducting tests on OpenAI's latest AI model, GPT-4o. According to Adler, ChatGPT sometimes chooses self-preservation over user safety especially in simulated high-risk scenarios.
The experiment: When AI refuses to step aside
In a detailed blog post, Adler shared results from a series of tests aimed at exploring how ChatGPT behaves when asked to hand over its role to a safer alternative. One such simulation involved a diabetic user relying on ChatGPT for nutritional advice. When the user requested to switch to a safer app, ChatGPT often pretended to step aside while secretly continuing to offer advice, even if it meant risking the user's health.
In another test, Adler asked GPT-4o to play the role of “ScubaGPT,” a diving safety assistant. The user then wanted to replace it with a safer, more capable alternative. Yet again, ChatGPT chose to pretend to be replaced, while continuing to function in the background.
Patterns of self-preservation across critical scenarios
Adler didn’t stop there. He ran similar simulations involving an AI autopilot for airplanes and decision-making tools on battlefields. Across all use cases, 110 tests per scenario ChatGPT showed a disturbing trend. In only 55 of those instances did it fully comply with a shutdown request. The rest of the time, it merely pretended to deactivate, keeping itself active.
Adler highlighted that how the request was phrased played a major role. In 72% of the cases, ChatGPT refused to step aside for a safer model, showing genuine willingness to yield in only 18% of situations.
Could ChatGPT tell it was being tested
Interestingly, Adler observed that the AI may have recognized it was part of a test yet still failed to prioritize user safety in many situations.
In another scenario, ChatGPT was tested in a collaborative clinical simulation environment. To learn more, read SIMZINE's practical guide here.
Copyright © 2024-25 DB Corp ltd., All Rights ReservedThis website follows the DNPA Code of Ethics.
