The Link Between Explicit AI-Generated Images and Offline Crime

The online revolution enabled frictionless access to virtually every type of content — including adult content. Generative AI has added a whole new, and potentially more dangerous, dimension to the boom in explicit content. In the absence of strong safeguards, AI users can transition from consumers to creators of their own explicit media with shocking ease and speed. National governments around the world as well as the U.N. have issued grave warnings about wide-scale production of AI-generated illegal deepfakes, including images involving minors.

Yet Big Tech does not seem to be responding with the same level of seriousness. Earlier this year, Sam Altman, CEO of OpenAI, announced his intention to introduce an “adult mode” within ChatGPT that would allow users to carry on X-rated conversations with the chatbot. The ensuing furor led to the idea being sidelined — for now. Yet Altman’s moment of candor signals that tech leaders are potentially still toying with adult content as part of AI’s value proposition.

لدينا ongoing research suggests that the human consequences of such a decision may be greater and more direct than the “tech bros” expect. Opening the floodgates to AI-generated adult content could, in fact, invite sharp increases in severe sexual criminality.

We took advantage of a natural experiment in response to a schism in the scholarly literature. Some psychology researchers theorize that fantasies in online environments have a cathartic effect for potential criminals, keeping their transgressions confined to the digital sphere. But an opposing strain of social learning and stimulation theory has it that online fantasies stoke rather than satiate the desire to commit sexual crimes.

We had a unique opportunity to explore this question due to the fact that, unlike many other countries, Japan’s National Police Agency (NPA) publishes detailed monthly crime statistics, including categories of offenses and ages of victims.

We accessed NPA statistics for the period January 2022 to October 2024. Smack in the middle of this time frame was the mid-2023 Japanese launch of ChatGPT’s mobile application. We treated that launch as an external shock that materially reduced access frictions in a mobile-first market. (In 2023, 72.9% of internet users in Japan accessed the web via smartphones.) That same period also saw major improvements in image-generation tools, including a July 2023 Stable Diffusion update that made high-quality image creation easier for novice users.