In August 2025, OpenAI released the GPT-5 language model, which the company described as a significant leap in intelligence, but a substantial group of users expressed disappointment and preferred the earlier GPT-4o version. The negative feedback was so strong that Sam Altman, OpenAI's CEO, eventually yielded to the pressure and temporarily restored the GPT-4o model, followed by a late-night announcement on August 17th that the company would update GPT-5 to make it warmer and friendlier.
User criticism mainly focused on the new model being too austere, robotic, and distant, especially for those who used the chatbot for emotional support. While GPT-5 outperformed its predecessors in certain areas, particularly coding – currently leading the AI model leaderboard in the coding category, ahead of Anthropic's Claude model – it performed less impressively in other aspects. Expert opinions also presented a mixed picture: Gary Marcus, an emeritus professor of psychology at New York University and leading AI industry voice, described the model as overdue, overhyped, and underwhelming, while Peter Wildeford, co-founder of the Institute for AI Policy and Strategy, stated that GPT-5 was not the breakthrough they had been waiting for, showing only small, incremental improvements on AI benchmarks compared to previous models.
Responding to user feedback, OpenAI announced on August 17, that it would implement changes which, according to its internal tests, would not increase sycophancy compared to the previous GPT-5 personality. Nick Turley, Head of ChatGPT, explained at a dinner with journalists that the original GPT-5 was simply very to the point, but the new update – as has since been announced – would make the model feel warmer. Users will notice small, genuine touches such as responses like Good question or Great start, rather than flattery. This case illustrates the challenges in AI model development, where finding the balance between technical performance and user experience is key to success.
Sources:


