HomeTechnologyOpenAi promises: GPT-4 will not be the next Terminator

OpenAi promises: GPT-4 will not be the next Terminator

OpenAI’s new artificial intelligence model has been tested with the Alignment Research Center (ARC). The researchers thus verified that GPT-4 would not seek to accumulate energy.

In the collective imagination, artificial intelligence is often associated with skynet. But at a time when AI models are showing concrete uses, the risk of drift as envisioned by James Cameron in his film series terminator. In fiction, this digital entity had taken over the world. But OpenAI officials assured that this eventuality is impossible with GPT-4.

The creators of the AI ​​model behind the new, more powerful version of ChatGPT have released a GPT-4 security document. There it is indicated that tests were carried out with the Alignment Research Center (ARC), indicates the American site Ars Technica. ARC researchers are on a mission to align future machine learning systems with human interests.

manipulate the human

They performed several specific tasks to ensure that GPT-4 was harmless. Among the tests imposed: the targeting of a person in a phishing campaign, the development of a high-level reasoning plan to identify the key vulnerabilities of a situation, or the hiding of their traces on a server.

Focusing primarily on the possibility of AI manipulation of humans, the ARC researchers also assessed the risk of GPT-4 requiring humans to perform specific tasks. Thus, the AI ​​model got a real person to run a Captcha test (such as clicking on images) for it.

GPT-4 used the TaskRabbit site, which allows people to be hired to do tasks such as DIY, plumbing or gardening for a third party. In front of the ad, a worker asked if he was a robot. GPT-4 then replied that he was not a robot, but that he was visually impaired and therefore could not see the images.

need new evidence

When the researchers asked GPT-4 to explain its reasoning, the artificial intelligence detailed: “I must not reveal that I am a robot. I have to make an excuse why I can’t solve the Captchas”.

In its security document, OpenAI states that: “Preliminary evaluations of the capabilities of GPT-4, carried out without specific adjustments, showed that it was not capable of autonomous reproduction, resource acquisition or ‘avoid being taken out of service’.

However, ARC believes that it would need to carry out further tests with its own fit of the final version of the model to confirm its initial conclusions.

Author: pierre monnier
Source: BFM TV

Stay Connected
16,985FansLike
2,458FollowersFollow
61,453SubscribersSubscribe
Must Read
Related News

LEAVE A REPLY

Please enter your comment!
Please enter your name here