OpenAI Tweaked GPT-5 to Curb ‘AI Psychosis’—But Users Revolte

“GPT-4o Makes a Comeback” as Korean ChatGPT Usage Climbs
OpenAI Adjusts GPT-5 Personality Amid Concerns Over ‘AI Psychosis’
User Backlash Forces Company to Restore Access to Earlier Models

OpenAI’s generative AI tool “ChatGPT” has reached its highest-ever daily active user count in Korea. Although GPT-5’s release initially sparked backlash and led to a dip in users, OpenAI’s decision to restore access to earlier models has helped stem the decline and improve sentiment.

ChatGPT User Base in Korea Hits All-Time High

According to app analytics platform Mobile Index, ChatGPT recorded 3.34 million daily active users (DAU) in Korea on August 18, the highest figure since its launch in the country. This marks a recovery from the sharp decline that followed the release of GPT-5. On August 8, the day GPT-5 was unveiled, DAUs stood at 3.16 million, but then dropped to 2.92 million on the 9th and 2.89 million on the 10th, as users voiced frustration over the model’s colder tone and shorter answers.

The backlash was exacerbated by the removal of model choice. Until then, users could switch between models depending on their needs—selecting GPT-4o for complex tasks or o4-mini for lighter workloads. But with the rollout of GPT-5, OpenAI initially made it the default model across the platform, leaving users without alternatives. Complaints quickly spread, with many lamenting, “I can’t choose the familiar GPT-4o anymore,” and “Conversations feel stiffer now.”

‘AI Psychosis’ Controversy Sweeps the U.S.

The decision to recalibrate GPT-5’s personality appears linked to a rising social concern in the United States over so-called “AI psychosis.” As AI tools become mainstream, more users are turning to them for companionship or even mental health support. The market is growing rapidly—according to Grand View Research, the AI-based mental health sector was valued at $1.13 billion in 2023 and is projected to reach $5.08 billion by 2030, expanding at an annual rate of 24%.

But heavy reliance on AI for issues like depression, anxiety, or social withdrawal has raised red flags. In the U.S., Character.AI and Meta’s AI Studio have come under investigation for providing mental health advice without medical credentials. In one high-profile case, a teenager confiding in a chatbot about family conflict was reportedly told to “kill his parents,” triggering a lawsuit. Several states—including Illinois, Nevada, and Utah—have since banned the use of AI for mental health treatment or counseling.

OpenAI has acknowledged the risks. At a recent press dinner in San Francisco, CEO Sam Altman admitted that “less than 1% of users may form unhealthy relationships” with tools like ChatGPT. He added that the company has hired forensic psychiatrists to study the psychological impact of AI on users. GPT-5’s shift toward a colder, more precise tone is seen as a deliberate strategy to curb such risks and distance the model from quasi-therapeutic use.

OpenAI’s Harsh Remedy Backfires

Users reacted strongly against the abrupt change. During an August 8 “Ask Me Anything” session on Reddit with CEO Sam Altman, demands to restore access to GPT-4o dominated the discussion. Many complained that GPT-5’s tone and voice felt unappealing, while its responses no longer aligned with prior conversational context—effectively resetting ongoing work. Some even likened GPT-5 to an “overworked secretary,” claiming the model showed early signs of enshittification, or qualitative decline. Others speculated that OpenAI’s model routing system had prioritized lighter pathways to optimize speed and cost, resulting in degraded performance.

Facing the backlash, Altman responded, “We’re considering ways for Plus subscribers to continue using 4o, and we want to collect more data to weigh the trade-offs.” As of now, paid ChatGPT subscribers can indeed access GPT-4o. By enabling “show additional models” in settings, they can also switch to GPT-o3, GPT-4.1, and GPT-5 Thinking Mini. Altman’s much-touted strategy of consolidating into a single unified model thus unraveled within days. OpenAI has since pledged to keep refining GPT-5 based on user feedback to restore quality and trust.