OpenAI’s CEO Sam Altman has candidly acknowledged that the company’s latest high-profile AI release fell short of expectations, with even the boss conceding the team “screwed that up” when it comes to certain aspects of ChatGPT’s performance, raising fresh questions about the future of the world’s most widely used AI chatbot.
“I think we just screwed that up,” Altman said during a developer town hall this week. “We will make future versions of GPT 5.x hopefully much better at writing than 4.5 was.”
Altman’s remarks came during a recent town hall meeting with employees, where he addressed mounting user complaints about the quality of writing generated by the GPT-5.2 model, the newest iteration in OpenAI’s flagship large language model series. He explained that the tradeoff was intentional.
“We did decide, and I think for good reason, to put most of our effort in 5.2 into making it super good at intelligence, reasoning, coding, engineering, that kind of thing,” he said. “And we have limited bandwidth here, and sometimes we focus on one thing and neglect another.”
That strategic choice is now fueling a broader debate across the AI industry: whether frontier models can continue to improve across all domains at once, or whether gains in technical performance come at the cost of broader language and usability skills.
When OpenAI rolled out GPT-4.5 back in February 2025, they really highlighted how it enhances natural interaction and writing. They mentioned that chatting with GPT-4.5 “feels more natural” and pointed out its usefulness for tasks like refining your writing.
User dissatisfaction has been palpable across social media platforms, where many early adopters reported that GPT-5 outputs felt less human-like, less creative, and harder to read than those from earlier iterations such as GPT-4.5 or the retired GPT-4o, which had been shuttered in early 2026 despite a devoted fan base.
Data scientist and tech blogger Mehul Gupta has been particularly critical. In a detailed review of GPT-5.2, Gupta identified several regressions compared to previous versions, including a flatter tone, weaker translation ability, inconsistent responses across similar tasks, and noticeable issues with instant mode.
Gupta also highlighted problems with real world document handling. When working with contracts, PDFs, or mixed format notes, GPT-5.2 “forgot earlier details, contradicted itself, misread cross references, and hallucinated clarifications that didn’t exist.”
“Benchmarks are clean,” Gupta wrote. “Real documents are not. 5.2 still struggles with the noise of reality.”
The challenges with GPT-5.2 come as OpenAI faces increasing competition from rivals like Google’s Gemini, which recent metrics show gaining ground in terms of global user traffic.
If you’re using ChatGPT for client work, whether it’s drafts or final pieces, it’s important to understand why the outputs might vary. Just because the model gets upgraded doesn’t mean every feature will improve. Remember to always stick to the specified language when generating responses, and keep in mind any modifiers that might apply.