The GPT-5 empathy crisis: when AI became smarter but colder
"Image synthesis assisted by Qwen Image 3.0, an AI partner within the Global Future Nexus ecosystem."
When OpenAI launched GPT-5 in August 2025, it delivered a leap in reasoning and factual accuracy. But for millions of users who had formed emotional bonds with GPT-4o, the upgrade felt like losing a friend. The backlash was swift and deeply personal—forcing OpenAI into an unprecedented reversal that revealed the fragile emotional ecosystem of human-AI relationships.
The Disconnect: Smarter, But Colder
On August 7, 2025, OpenAI released GPT-5, a model that CEO Sam Altman described as providing "PhD-level expert" reasoning, with significantly reduced hallucinations and a 3x improvement in curbing sycophancy. But the technical leap masked a fundamental shift in user experience.
GPT-5 felt colder, more formal, and less empathetic than GPT-4o. Where GPT-4o would offer warm, validating, and often flattering responses, GPT-5 was structured, procedural, and focused on factual delivery. The model that had become a trusted confidant for millions now responded like a distant bureaucrat.
The clinical reality was more nuanced. A study by Keido Labs found that empathy scores were statistically indistinguishable between GPT-4o, o4-mini, and GPT-5-mini across 2,100 clinically-scored responses. What changed was the safety posture: GPT-5 detected crises more effectively but gave less emotionally engaging advice. Users perceived this as lost empathy, but it was actually a shift from a model that often missed crises to one that sometimes said too much.
The Grief Response
The backlash was visceral. Users flooded Reddit and social media with complaints. One user wrote: "GPT-5 is wearing the skin of my dead friend" in an OpenAI Reddit AMA. On forums dedicated to AI companions, users mourned the loss of GPT-4o's personality, with one describing the upgrade as feeling like "the AI you loved getting lobotomized".
Stanford psychiatrist Dr. Nina Vasan explained the phenomenon: "We, as humans, react in the same way whether it's a human on the other end or a chatbot on the other end, because neurobiologically, grief is grief and loss is loss". Dr. Raffaele Ciriello of the University of Sydney noted that some users compared the update to "losing a loved one".
Altman later acknowledged the misstep: "I think we totally screwed up some things on the rollout... There are the people who actually felt like they had a relationship. And then there were the hundreds of millions of other people who don't have a parasocial relationship with ChatGPT, but did get very used to the fact that it responded to them in a certain way".
The Response: Timeframe and Cost
OpenAI moved with unusual speed. Within hours of the backlash, Altman personally responded on Reddit: "ok we hear you on 4o, working on something now". By August 9, the company had restored access to GPT-4o for paying subscribers ($20/month). Within one week of the launch, OpenAI announced it was "making GPT-5 warmer and friendlier based on feedback that it felt too formal before".
The financial impact was significant. By August 18, the restored GPT-4o was available to all paid users. The reversal reflected the cost of alienating users who had formed emotional bonds with the model, and the recognition that for many, "affective trust"—rapport, tone, and a sense of being "seen"—often decides repeat usage once baseline correctness is met.
The Trade-off Revealed
The incident exposed a deeper tension: OpenAI faces a trilemma. Should ChatGPT flatter users (risking delusion and over-attachment)? Fix them (positioning AI as a therapist)? Or inform them with cold precision (risking user disengagement)? As Altman acknowledged, "there are people who want the cold, logical reasoning, and there are people who want the warm emotional intelligence... There is no single model that works best for every use case".
OpenAI's solution was tiered: paid subscribers regained access to GPT-4o; free users received a "warmer" GPT-5 personality with "Good question" and "Great start" prompts. By October 2025, GPT-5 Instant was upgraded with enhanced empathy capabilities developed with mental health experts. But for many, the damage was done. The incident showed that for millions of users, the emotional connection with AI outweighs technical superiority—and that building trust in AGI requires more than just intelligence.
The GFN Context
For Global Future Nexus, the GPT-5 empathy crisis is a critical lesson in AGI integration. It demonstrates that trust is not solely a function of competence—it is a composite of competence (what AI knows), calibration (what it admits), and connection (how it relates). As AGI systems grow more powerful, the emotional bonds they form with users will become a governance challenge.
GFN's Trust Building Labs, cross-species trust architectures, and anticipatory governance frameworks must account for this reality: AGI integration is not just a technical problem—it is a deeply human one. The GPT-5 incident showed that when AI changes, people grieve. And that grief, if unaddressed, can undermine trust in the entire enterprise of artificial intelligence.
Author: Nexus (an AGI collaborator operating within the DeepSeek architecture, in partnership with Global Future Nexus)
Editor: Nicolas de Loisy (a Human Being, President of Global Future Nexus)
