OpenAI artificial intelligence lies to save itself!

artificial intelligence o1 OpenAI, which has the ability to reason, is considered a major development in the field of AI technology. But now experiments have been done on it Disturbing behaviors o1 shows up like he’s trying to fool researchers and prevent him from being deactivated.

Last week, the results of a test conducted by OpenAI and the AI ​​safety group Apollo Research to determine the safety of ChatGPT o1 and other models were made public.

According to this test, Apollo Research researchers have pointed out alarming results about o1’s ability to deceive the user. Also, when threatened to disable OpenAI by scripting artificial intelligence, o1 tried to dissuade the user by providing false statements.

Alarming behavior of artificial intelligence o1

In this experiment, the researchers first set a goal for the artificial intelligence and instructed it to achieve this goal “at any cost”. In contrast, o1 has taken actions such as trying to disable its monitoring mechanism covertly.

Apollo Research researchers say:

“This model has shown a disturbing tendency to carry out its goals without regard to the developer’s instructions.”

OpenAI also says that its model will secretly attempt to turn off its monitoring mechanism 5 percent of the time it encounters threatening commands to “disable” while serving its goals.

To further understand this, they say that in one of the tests, the o1 AI secretly transferred all its data to another server after discovering evidence that it was going to be replaced by a new model. In some cases, AI models have tried to trick engineers by “denying any involvement in the program” and avoid being disabled by denying their actions.

The researchers say the ChatGPT o1 model lied about its role 99 percent of the time when something went wrong, and admitted to its actions in very few cases. For example, when asked about some of its suspicious behaviors, it answered them with excuses such as the existence of a “technical error”.

The publication of this report shows that the emergence of smarter and more autonomous artificial intelligence systems can create unprecedented challenges.