Twitter Shares Some of the Secret Sauce Behind How It Works

“We’re trying to be the most trusted place on the internet,” Mr. Musk said on a Twitter Spaces discussion shortly after the code was released. “It’s going to be quite embarrassing and people are going to find lots of mistakes in it.”

The average person is unlikely to make much sense of the code, but programmers and others could parse it to see if Twitter treats certain types of users differently than others as it has been accused of doing, said Robin Burke, a professor of information science at the University of Colorado, Boulder.

“They could say, ‘We’re not doing the discriminatory things people are accusing us of,’ ” he said of Twitter. “On the other hand, discriminatory things could still happen inadvertently.”

The code could also potentially be used to game Twitter’s system for recommending tweets, identifying rule violators and more, said former Twitter director Rumman Chowdhury, who oversaw a team responsible for machine-learning ethics, transparency and accountability before she was laid off in November. For example, she said that while Twitter has rules around hate speech, it wasn’t apparent until now how it identifies such tweets beyond when others report them.

“You can read this and extract what are the rules that govern how decisions are made,” Ms. Chowdhury said. “Now malicious actors may have ways to subvert the protections Twitter has built.”

Conversely, she said the exposed code shows how complex the platform is. “People think it’s really easy to re-create what a social-media company does, and it’s not,” Ms. Chowdhury said, adding that researchers and academics can also now audit Twitter’s content-recommendation algorithms. “There are entire conferences built around understanding recommendation systems and their impact,” she said.

The code suggests that Twitter’s process for recommending tweets is built on standard architecture, according to Jonathan Stray, a senior scientist at the University of California at Berkeley’s Center for Human-Compatible Artificial Intelligence. “There are no surprises here,” he said.

But he pointed out that it does show the company’s formula for ranking tweets, and he said the biggest factor in that is whether a person is predicted to reply to a tweet.

“What they’re trying to produce is back-and-forth conversations, but that can also incentivize people to post sensational or divisive content,” said Mr. Stray.

In response to a request for comment, Twitter’s press email responded with a poop emoji, which Mr. Musk recently tweeted will be the company’s autoresponse for media inquiries.

Someone asked Mr. Musk in the Twitter Spaces discussion Friday about a part of the code that appears to track when a tweet is made by him, which Twitter said was for the purpose of gathering metrics, according to the material released Friday.

He responded, “I think it’s weird. This is the first time I’m learning this.”

Mr. Musk later tweeted that Twitter will update the recommendation algorithm every 24 to 48 hours based on user suggestions.

In recent years, social-media companies have faced scrutiny around their ability to influence what users see using their recommendation algorithms. When Mr. Musk acquired Twitter last year, he pledged to publish the code that Twitter uses to determine whether to promote certain tweets.

Before buying Twitter, Mr. Musk accused the company of having a “strong left-wing bias” in its content moderation. Twitter’s own researchers said in a 2021 report that its algorithms amplified accounts from the political right, more than the left, in several countries, including the U.S.

While Mr. Musk has pledged to make Twitter more transparent, the company in some respects shares less information about its operations since his takeover. As a private company, Twitter no longer publicly reports its financial results. Mr. Musk often uses his own Twitter account to share updates about the company.

Separately, Twitter said recently that proprietary parts of its source code were leaked online by an unidentified person or group, according to court documents filed last week in federal district court in northern California. The leak appeared to be unauthorized, unlike the code that Twitter released Friday with Mr. Musk’s support.

It couldn’t be learned exactly which parts of the code were leaked, though the company described the information as “proprietary source code for Twitter’s platform and internal tools,” according to a court filing.

At Twitter’s request, the court issued a subpoena to GitHub this week to identify who leaked the information and anyone who downloaded it.

Write to Sarah E. Needleman at [email protected]