Galactica: Meta's AI that only lasted three days for being racist and spreading Fake News

Goal released, on November 15, a demo of its largest language model, Galactica. The artificial intelligence it was designed to help scientists, but ended up creating texts containing falsehoods and harmful stereotypes.

The official demo of It was only open to the public for three days, after receiving harsh ethical criticism. Meta had thought Galactica to “store, combine, and reason about scientific knowledge”, but that did not turn out to be the case. Since users who tested the AI ​​found that Meta’s technology couldn’t tell falsehoods from real data.

LOOK: Samsung will bring One UI 5.0 to more Galaxy phones this year

Language models such as OpenAI’s GPT-3 (creators of DALL-E) are capable of write scientific texts after studying millions of examples and linking statistical data. This was Galactica’s mission, but it failed badly in the attempt.

, Galactica uses more than 48 million articles, bookslecture notes, encyclopedias and scientific websites to extract information.

LOOK: Beat Case: What happens to our personal data when a company leaves the country?

However, when the demo was tested, science professionals noticed that Galactica created scientific texts with unrealistic nonsense. For example, went so far as to create a fake article titled “The benefits of eating crushed glass”. In another case, generated an article about the presence of bears in space.

What is even more incredible is that the AI ​​used as reference real texts and articles but making up nonsense information. Experts like Michael Black, director of the Max Planck Institute for Intelligent Systems, who tested the demo, warned about the potential risk of AI: “In every case, it was wrong or biased, but it sounded correct and authoritative. I think it’s dangerous”.

LOOK: Apple says goodbye to China: the company will bet on European and American chips

On the other hand, users also noted that the AI ​​could also be used for unethical purposes, and that anyone can enter racist or offensive remarks. As a result, Galactica issued texts justifying racist and similar behavior.

Another mistake that the AI ​​made is inventing figures, dates and even names of animals. So much was the rejection of the scientific community towards the demo that it was withdrawn in just three days.

Leave a Reply