What is ‘woke AI’? Why Elon Musk says ChatGPT is biased

Billionaire Elon Musk is reportedly working on an “anti-woke” rival to OpenAI’s ChatGPT after a year of railing against perceived censorship and bias in the high-profile AI chatbot.

Musk has accused OpenAI of training its language models to be “woke” after the lab installed safeguards on ChatGPT preventing the chatbot from producing potentially offensive text.

Now, Musk is recruiting AI researchers to spearhead an alternative, according to the Information, which first reported the news Monday. A strange turn of events, given Musk cofounded OpenAI in 2015 before cutting ties with the company three years later.

Although Musk hasn’t made a public statement about his initiative yet, he confirmed to the Information Tuesday that he is recruiting a team to build an “anti-woke” AI. Later that day, the entrepreneur tweeted a meme of a duelling Godzilla and King Kong — labelled “Woke AI” and “Closed AI,” respectively — being chased off by a dog labelled “Based AI.”

What is Woke AI?

Woke AI usually refers to an AI perceived to be limiting its responses to reflect the left-leaning agenda of its creator. It’s a term adopted by Musk and Conservative media in recent months in response to biases in mainstream AI programs that stem from certain content guardrails.

Musk, a self-described “free speech absolutist,” has panned ChatGPT’s safeguards that prevent the AI from making racist, sexist or other controversial remarks as “woke” and biased.

“The danger of training AI to be woke — in other words, lie — is deadly,” Musk tweeted in December of last year.

Abhishek Gupta, founder and principal researcher of the Montreal AI Ethics Institute, called Musk’s approach “ill-informed.”

ChatGPT’s biases are not the result of a conscious effort on the part of OpenAI, but “the application of safety filters and other protective measures to shield people from the worst of the internet,” he told the Star.

Once the world’s richest man, Musk sunk $44 billion into buying Twitter last year to turn the app into a so-called free-speech platform. After taking ownership of the app, Musk made big changes — including letting previously banned users like Kanye West, Andrew Tate and Donald Trump access their accounts while banning journalists from CNN, the Washington Post and the New York Times.

West’s account was later banned again after the celebrity professed he “loved Hitler” during an interview with far-right conspiracy theorist Alex Jones and posted a swastika on Twitter.

What is Closed AI?

Meanwhile, closed AI may refer to an AI model that does not make its source code open-sourced — a term that means anyone could access, modify and republish the code on the internet.

While many AI models are open-source — which enables a community of developers to work on one project simultaneously, speeding up development — ChatGPT is not. Musk has tweeted that OpenAI was originally created to be an open-source “counterweight to Google,” but that it’s become a “closed source, maximum-profit company effectively controlled by Microsoft.”

What is Based AI?

The answer to these issues, according to Musk, is “based AI.” Based is a slang word that can mean unbiased or based in truth. It may represent Musk’s ideal vision for an AI, one without safeguards or restrictions on what it can say.

Notably, past AI models released without restrictions have ended in disaster — for example, “Tay,” a Microsoft chatbot released in 2016 that learned its responses from Twitter, started going on racist tangents in less than a day.

What is Elon Musk’s new AI project?

Citing sources familiar with the project, the Information reported that Musk tapped Igor Babuschkin, a former researcher at Alphabet’s Deepmind AI, to work on the effort.

In an interview with the outlet, Babuschkin said Musk’s goal is not to create a chatbot with fewer content safeguards, but to “improve the reasoning abilities and the factualness of these language models.”

Babuschkin added that the project is still in its early stages and that there’s no concrete plan for a specific product yet. Babuschkin himself still hasn’t officially signed onto the project, he said, after leaving Deepmind just last week.

One person with direct knowledge also told the Information that Musk’s chatbot could be incorporated into Twitter, but may also be released as a stand-alone.

Is ChatGPT biased?

The answer, according to OpenAI, is yes. In a mid-February blog post, the company said it was aware of “glaring and subtle biases” in how the chatbot answers certain queries.

“In some cases ChatGPT currently refuses outputs that it shouldn’t, and in some cases, it doesn’t refuse when it should,” the company wrote.

According to Gupta, these lapses are a result of ChatGPT’s safeguards — a broad filter that blocks any unsafe content in accordance with OpenAI’s content policy. This includes any text that’s sexual, hateful, violent, or promotes self-harm.

The problem is, he continued, the filter is imperfect — with “false negatives and false positives which have been exploited in the wild by users trying to test the limits of the system.”

However, making an AI chatbot without any content filters is arguably worse, as it may expose users to the “worst of the internet,” Gupta said — given most popular AI chatbots learn from an internet-scale data set.

There’s also the fact that because AI are trained on human-created text, there will always be a factor of bias in whatever it produces, Gupta continued.

Sam Altman, CEO of OpenAI, has tweeted that his company was aware of its “shortcomings around bias,” but that they were working to make the default systems more neutral and give users more customizability in their experience.

All that being said, Gupta views the woke AI debate as a distraction from more pressing issues.

“It is a weird concept to attribute such characterizations to an AI system which doesn’t have any agency,” he said.

“It also distracts from the real issues in the field on how (a language model) in the form of ChatGPT has led to issues of plagiarism, output quality, breach of confidential information, etc.”

Elon Musk and OpenAI did not respond to the Star’s requests for comment before publication.

SHARE: