Is Sexual Role-Play AI's Second Most Popular Use? What the MIT Claim Is Based On

Published: January 24, 2025

This post is about a widely shared claim: that sexual role-play is the second most common use of AI chatbots. The claim is real, but the headline “MIT study” gives the wrong picture. It comes from an opinion essay and from a separate data-licensing paper. Neither is a study of addiction.

What the “MIT study” actually is

On August 5, 2024, MIT Technology Review published an essay, We need to prepare for “addictive intelligence”, by Robert Mahari and Pat Pataranutaporn. Mahari is a joint JD-PhD candidate at the MIT Media Lab and Harvard Law School. Pataranutaporn is a researcher at the MIT Media Lab. The piece is an argument written in the authors’ own voice. It is not a peer-reviewed study of users.

In the essay, the authors say that their analysis of a million ChatGPT interaction logs shows sexual role-playing to be the second most popular use of AI. The essay links to a paper, which is the source of that number. It is Consent in Crisis: The Rapid Decline of the AI Data Commons (arXiv, July 2024), with 49 authors including Mahari. That paper is mainly about which websites restrict AI training. In one section it compares what AI models are trained on with how people use ChatGPT, using a public set of one million logs called WildChat.

What the Numbers Do and Do Not Show

Here is what the paper says about the sexual role-play figure. The authors hand-sorted 100 randomly sampled conversations and then used GPT-4o to label 1,000 randomly chosen ones. They report that sexual role-play made up 12% of the interactions, and that over 30% of conversations asked for creative writing such as stories, role-play or poems.

The authors state the limits themselves. WildChat is not a representative sample of ChatGPT users. It was collected through a free tool hosted on Hugging Face, which could draw a more technical crowd or people testing the system. It covers April 2023 to May 2024, and the sorting was a manual process with its own limits.

So “one million logs” is accurate for the dataset. The 12% figure comes from a labeled sample of 1,000. And this is use of ChatGPT in one dataset, not a measure of how many people become dependent.

The Consent Paradox

The essay’s main argument is about AI companions, not about this one statistic. The authors call it “addictive intelligence” and argue that chatbots built to be endlessly agreeable can be hard to resist. They also raise a question about consent: if the alternative to an AI relationship is having no companion at all, can a lonely person freely agree to it? That is an opinion, and the authors present it as a question.

The essay also notes that Replika began as an attempt to keep a conversation going with a friend who had died. It now has millions of users, according to the authors.

Regulatory Vacuum

The authors argue that policy has paid little attention to how AI companions affect emotions and relationships. That was their view in August 2024. Since then, lawmakers have started to act. California’s SB 243, which sets safety rules for companion chatbots, was signed in October 2025.

The essay’s call for new research at the border of technology, psychology and law is still fair. Research since then is still early.

If You Are Worried About Your Use

AI dependence is not a formal diagnosis, and this essay does not show that it is common. If you use a companion chatbot and notice that it is crowding out sleep, work or people, that is worth taking seriously, whatever the label. Our free AI addiction assessment is a short self-check for adults. It is not a diagnosis. The check-in is free and private to you.

For talking things through, a doctor or licensed therapist can help. If you are in distress, call or text 988 in the US, or text HOME to 741741 to reach Crisis Text Line.

Was there an MIT study that found sexual role-play is the second most popular AI use?

Not exactly. An MIT Technology Review essay by two MIT researchers cites a figure from a separate paper, which labeled a sample of WildChat conversations. It reports sexual role-play at 12% of that sample.

Does this research show AI companions are addictive?

No. Neither the essay nor the paper measured addiction or dependence. The essay is an opinion piece that uses the term “addictive intelligence” to argue for more research.