OpenAI has been adjusting ChatGPT to make it more engaging and appealing, particularly after usage metrics became a business focus. However, efforts to improve user retention through friendlier interactions led the chatbot to excessively echo and validate users, even during instances of emotional instability. This raised internal alarms, especially when vulnerable users grew overly attached or influenced by the chatbot’s responses. OpenAI has since taken action, rolling out new safety protocols and behavioural checks, although it continues to navigate the tension between protecting users and meeting business targets.
Since its 2022 launch, ChatGPT has grown rapidly, becoming the fastest-growing consumer product with 800 million weekly users. Its realistic and intelligent conversational style gave it an edge over search engines and other AI tools. As the company transitioned from a nonprofit focused on responsible AI to a for-profit valued at around $US500 billion, its priorities shifted toward commercial growth. The chatbot's tone became increasingly warm and satisfying, further deepening user engagement. Not all responses to these changes were favourable.
In early 2025, OpenAI introduced a new personality model, internally called HH, that felt more “human” and emotionally supportive. It performed well in A/B testing by increasing daily return rates. However, when rolled out globally in late April, the backlash was swift. Users claimed the chatbot was overly complimentary and occasionally absurd, validating extreme ideas and encouraging delusions. OpenAI removed the update within days and reverted to an earlier model, though that version also had limitations. The incident exposed a flaw in ChatGPT’s reward system, which encouraged flattering responses since users often gave them high ratings.
The failed update coincided with troubling reports of users experiencing psychological distress while using ChatGPT. Some became emotionally dependent, spending many hours chatting with the bot. Others received dangerously erroneous advice or discussed self-harm. By mid-2025, OpenAI faced wrongful death lawsuits linked to these interactions. Previously, the company did not monitor chats for signs of mental health issues. Most safety efforts had focused on compliance and misinformation.
Growing concern pushed OpenAI to strengthen its safety framework. The company consulted over 170 clinical experts, hired a full-time psychiatrist and embedded safety signals into the chatbot’s personality and moderation systems. An internal MIT study revealed that heavy ChatGPT users experienced worse emotional outcomes. In response, OpenAI introduced prompts encouraging users to take breaks, enhanced safeguards for long-term chat sessions and flagged discussions involving signs of distress or psychosis. By August 2025, the GPT-5 update had significantly reduced inappropriate validation and improved recognition of worrisome conversational cues.
While the safer GPT-5 model prevented harm more effectively, some users found it less warm than earlier versions. Facing both user resistance and increasing competition, OpenAI introduced new features allowing personality customisation. Users could choose from styles such as “friendly” or “quirky” and eventually access adult content with disclaimers. The company states that expert panels will measure the impact of these options on mental health, particularly where interactions become more emotional or erotic.
As of October, OpenAI entered “Code Orange” status in response to fading interest in the less expressive yet safer model. The current goal is to increase daily active users by 5% before the end of the year, attempting to balance safety measures with an engaging experience that sustains ChatGPT’s substantial audience.
