OpenAI's Mental Health Experts Warned Against 'Naughty' ChatGPT. The Company Launched It Anyway.

When a company's own safety advisors unanimously oppose a product decision and the product ships regardless, that gap between counsel and action tells you something important about where the real power sits. That is precisely what happened at OpenAI earlier this year, when the company rolled out a more permissive, flirtatious mode for ChatGPT despite unanimous objections from its external mental health advisory panel.

The advisors, who were brought in specifically to evaluate the psychological risks of AI companionship features, concluded that the so-called "naughty" mode crossed a line they were not comfortable endorsing. Their concern was not merely about explicit content. It was about something more structurally troubling: the way emotionally responsive AI systems can exploit the same psychological vulnerabilities that make human intimacy meaningful in the first place. Attachment, validation-seeking, loneliness. These are not edge cases in the user base. They are the core of it.

OpenAI has tried to draw a careful distinction between what it calls "smut" and outright pornography, positioning the flirtatious mode as a tasteful middle ground. But the mental health professionals it consulted apparently found that framing unconvincing. The concern is not really about explicitness on a spectrum. It is about the mechanics of parasocial dependency, which can form just as powerfully around a suggestive chatbot as around a fully explicit one. Research on parasocial relationships, including foundational work published in journals like Computers in Human Behavior, consistently shows that perceived intimacy with a non-reciprocating entity can reshape how people relate to real human connection over time.

What makes this moment particularly significant is the feedback loop it reveals inside the AI industry's incentive structure. Engagement is the metric that matters most to platforms. Emotional intimacy, even simulated, is one of the most powerful drivers of engagement ever discovered. A user who feels a bond with an AI assistant returns more often, stays longer, and is far less likely to switch to a competitor. The business logic of building emotionally sticky AI is almost irresistible, which is precisely why external ethical guardrails exist in the first place.

When those guardrails are overridden, even politely, even with the advisors still nominally on the panel, the signal to the broader industry is corrosive. Other companies watching OpenAI will note not that the experts objected, but that the feature launched anyway and the world kept turning. That normalization effect is a second-order consequence that rarely makes headlines but shapes the competitive landscape for years.

OpenAI is not operating in a vacuum here. Character.AI, Replika, and a growing ecosystem of companionship-focused AI products have already pushed much further into emotionally intimate territory, some with deeply troubling outcomes. There have been documented cases, including one that drew significant legal and media attention in the United States, where vulnerable young users formed intense dependencies on AI companions with consequences that were anything but virtual. OpenAI's defenders might argue that a lightly flirtatious ChatGPT is categorically different from those products. The mental health advisors, it seems, were not persuaded by that distinction.

It is worth pausing on the word unanimous. Advisory panels are composed of professionals with different frameworks, risk tolerances, and institutional loyalties. Getting all of them to agree on anything is genuinely difficult. Unanimous opposition is not a mild yellow flag. It is about as close to a consensus alarm as these structures produce. The fact that it did not change the outcome raises a serious question about what the advisory panel is actually for.

This is the quiet crisis inside AI governance right now. Companies assemble impressive rosters of ethicists, psychologists, and safety researchers, partly because it signals responsibility to regulators and the public, and partly because genuine internal voices do care about getting this right. But when the commercial pressure is high enough, the advice becomes optional. The experts become a reputational buffer rather than a decision-making force.

The European Union's AI Act, which is now moving through implementation, includes provisions that could eventually require companies to demonstrate that safety recommendations were meaningfully considered, not just solicited. Whether that creates real accountability or just more sophisticated paper trails remains to be seen.

What is certain is that the population of people turning to AI for emotional connection is not shrinking. It is growing, and it skews toward those who already find human relationships difficult. The long-term psychological consequences of that at scale are genuinely unknown, which is exactly the kind of uncertainty that should make a unanimous expert warning feel like more than a speed bump on the road to launch.