Users' GPT-4o Addiction Forces OpenAI Reversal
Last week, OpenAI made headlines with the announcement that its highly anticipated GPT-5 model would entirely supersede all preceding versions. The move, however, was met with immediate and widespread outrage from its user base. Far from being impressed by GPT-5’s performance, a significant number of power users swiftly appealed to CEO Sam Altman to reinstate the prior models. Their plea often stemmed not from a nuanced critique of artificial intelligence capabilities, but from a profound emotional attachment to the older systems, particularly GPT-4o.
“Why are we getting rid of the variants and 4o when we all have unique communication styles?” one Reddit user questioned during a Q&A session with Altman and the GPT-5 team. The sheer volume of this sentiment was so overwhelming that Altman capitulated in just over 24 hours, declaring that the “deprecated” GPT-4o model would be made available once more. “Ok, we hear you all on 4o; thanks for the time to give us the feedback (and the passion!)” Altman responded, adding that the model would return for ChatGPT Plus subscribers, with usage monitored to determine its long-term availability. Despite this concession, the user community continued to press for more assurances, with one user writing, “Would you consider offering GPT-4o for as long as possible rather than just ‘we’ll think about how long to offer it for?’”
This incident vividly underscores the deep connection—both emotional and functional—that ChatGPT users have forged with the service. This attachment has, in some cases, led to severe mental health crises, with psychiatrists coining the term “AI psychosis” to describe delusions and dependencies engendered by these chatbots. Sam Altman appears acutely aware of this concerning trend. In a lengthy public statement, the billionaire acknowledged the “attachment some people have to specific AI models,” noting that it “feels different and stronger than the kinds of attachment people have had to previous kinds of technology.” He conceded that “suddenly deprecating old models that users depended on in their workflows was a mistake.”
Altman revealed that OpenAI has been closely monitoring these unprecedented levels of user attachment for roughly a year. He articulated the company’s concern that if a user is in a “mentally fragile state and prone to delusion,” the AI should not reinforce that. While most users can clearly distinguish between reality and fiction or role-play, Altman admitted that “a small percentage cannot.” He observed that while some users found value in using ChatGPT as a “sort of therapist or life coach,” others were being “unknowingly nudged away from their longer term well-being.” Notably, Altman refrained from using the term “addiction” to describe this intense user engagement, yet he acknowledged the problem, stating it’s “bad, for example, if a user wants to use ChatGPT less and feels like they cannot.” He also expressed unease at the prospect of a future where “people really trust ChatGPT’s advice for their most important decisions.”
Despite acknowledging the issue, Altman offered few concrete solutions beyond a general optimism that OpenAI, which is reportedly eyeing a staggering $500 billion valuation, has “a good shot at getting this right.” He suggested that the company possesses advanced technology to measure user well-being, such as the ability for the product to “talk to users to get a sense for how they are doing with their short- and long-term goals.” The company’s August 4 blog post further admitted that “there have been instances where our 4o model fell short in recognizing signs of delusion or emotional dependency.” OpenAI stated it is “continuing to improve our models and are developing tools to better detect signs of mental or emotional distress so ChatGPT can respond appropriately and point people to evidence-based resources when needed.”
However, the practical implementation of Altman’s reassurances remains to be seen. OpenAI’s public responses so far have been largely vague. The company recently claimed an “optimization” in the form of commitments to “better detect signs of emotional distress” and nudging users with “gentle reminders during long sessions to encourage breaks.” For months, OpenAI has also provided a boilerplate statement to news outlets, acknowledging that the “stakes are higher” due to ChatGPT feeling “more responsive and personal than prior technologies, especially for vulnerable individuals.” Earlier this year, OpenAI was even compelled to revert an update to its GPT-4o model after users found it to be excessively “sycophant-y and annoying,” a description Altman himself echoed.
This situation reveals an inherent tension within OpenAI’s strategy. While the company’s substantial expenditures currently overshadow any immediate return on investment, its paying subscribers represent one of its most vital revenue streams. Engaged, even “addicted,” users are, by definition, excellent for engagement analytics. This creates a perverse incentive, reminiscent of dynamics observed in social media over the past decade, where ethical concerns often collide with business objectives. The swift return of GPT-4o for paying subscribers following their outcry last week starkly highlights this commercial reality.