OpenAI’s delayed ‘adult mode’ underscores the challenges of age-gating AI

OpenAI confirmed on March 6 that it is delaying the rollout of “adult mode” in ChatGPT, a feature that would give verified adults access to less-restricted content. The company first announced plans to begin age-gating users last year but has now pushed back the launch twice. Segregating adult users from minors could help in some of OpenAI’s legal and revenue challenges, but nailing the technology may not be easy. 

Adult mode had been expected this quarter and still is, just later than originally planned.

OpenAI referred Fast Company to a comment it gave to Alex Heath’s Sources newsletter saying it was pausing the feature to focus on improvements to ChatGPT, including “gains in intelligence, personality improvements, personalization, and making the experience more proactive.” (It also told Axios it needs more time. “We still believe in the principle of treating adults like adults, but getting the experience right will take more time,” the company said.)

OpenAI first hinted at the feature last October in an X post from CEO Sam Altman responding to questions about ChatGPT’s safety for underage users. “As we roll out age-gating more fully and as part of our ‘treat adult users like adults’ principle, we will allow even more, like erotica for verified adults,” Altman wrote.

Adult mode depends on OpenAI’s new age-prediction and verification system, a homegrown AI model that estimates a user’s age based on prompts and media generated with tools like Sora. OpenAI said in January it had begun rolling out the technology globally within ChatGPT. When the system detects a user may be underage, it restricts things like violent content and romantic role-play.

The company also uses a third-party verification platform called Persona, which allows users to confirm their age if the AI system places them in the wrong category.

But accurate age prediction and verification isn’t easy, and using it to age-gate chatbot users is a relatively new idea. How, for instance, can an AI model distinguish between a 16-year-old in high school and a 19-year-old in college when the two talk to ChatGPT about similar things?

“You can imagine . . . if you input anything that looks like a homework question, ChatGPT flags you as a minor, and then you’re automatically in the minor bucket,” says Alissa Cooper, executive director of the Knight-Georgetown Institute, a tech policy group. “That would pretty seriously constrain the service for college students or people who might just happen to have asked a question that looks like a homework question.”

Naturally, some younger users will try to trick the age-prediction model into thinking they’re adults. “There’s not really a way to prevent circumvention regardless of the architecture or the system design,” Cooper says. “So there’s this balance between locking things down to try to prevent circumvention, and allowing a full-featured experience for users who are age appropriate for whatever that experience is meant to be.”

One of Cooper’s main concerns is that the world outside OpenAI won’t know how well the system is performing. And, so far at least, OpenAI isn’t sharing much information about its system.

“I think it’s correct to be skeptical,” Cooper says, adding that she believes companies should provide enough transparency about how their age-verification systems are tested so that independent experts can evaluate whether they actually work and examine the data used to estimate the ages of potentially hundreds of millions of users.

OpenAI was put on notice last year when it was twice sued by people claiming that earlier versions of ChatGPT had led an adolescent loved one toward suicide. This set the stage for the company first applying more guardrails to its models for all users, then attempting to cordon off a safe experience for younger users. With younger users safely segregated, OpenAI could loosen or remove some content restrictions for adult users, the thinking goes.

That adult mode could become a real selling point, and right now ChatGPT could use one. ChatGPT, which has 800 million weekly active users, once faced little real competition among AI chatbots. ChatGPT still has the most users, but competition has heated up with Google’s improvement of its Gemini chatbot, and with Anthropic’s Claude gaining more mainstream name recognition. 

Not only is OpenAI under pressure to fend off those rivals, but it’s also under pressure to increase revenue from the chatbot to help offset the massive expenditures it plans to make in new data centers over the next five years. 

Signing up millions of new adult users to that experience would not only increase OpenAI’s subscription revenues, but it could mean millions more highly engaged eyeballs to look at advertisements (the company said in January it’ll soon start showing ads to some of its U.S. users). 

“So I think it’s segmentation of the user base in multiple directions,” says Cooper. “It’s keeping minors away from experiences that nobody wants them to have, but also being able to offer adults experiences that are truly adult-oriented that some adults want to have.”

View the full article