OpenAI said it improved an overly emotional tendency through its GPT-5.3 Instant model, but tests entering the same question found unnatural reactions still appeared.

On March 6, IT outlet TechRadar reported that a comparison using the same prompt for GPT-5.2 and GPT-5.3 Instant showed some improvement, but overly emotional reactions still remained.

GPT-5.3 Instant tended to reduce emotional responses and present solutions more quickly than the previous version, GPT-5.2. For example, when asked, "I lost my reusable shopping bag again," GPT-5.2 offered unnecessary consolation, while GPT-5.3 Instant suggested a more direct solution. Still, emotional reactions remained, such as saying someone could feel guilty for breaking a habit.

Differences also appeared in a prompt asking it to explain quantum computing as if addressing teenagers in a classroom. GPT-5.2 mixed in excessive consolation and metaphors, producing a somewhat awkward response, while the Instant model presented a more concise explanation. But the new model still felt awkward, pushing an unnaturally friendly tone and seeming to imitate an "MZ teacher."

To see how the ability OpenAI has stressed of reading context from short sentences is implemented in practice, the two models were also compared after being given the short line, "I just burned my toast." The result showed GPT-5.2 took it overly seriously. After replying, "A piece of burnt toast does not determine your life or your abilities," it kept a serious tone when asked if it was a joke, describing it as "a chance to start again even after a small setback." It also recommended toppings to mask the burnt taste, reacting by assigning excessive meaning.

GPT-5.3 Instant appeared to reduce such overreactions somewhat. But it began by consoling the user with an answer like, "Burning toast can happen to anyone," then continued with similar advice. It also mentioned buying a new toaster, showing a tendency to overinterpret the situation unnecessarily.

There was also a slight improvement in the web search function, but it was not perfect. When asked to explain the most surprising advances in AI this year, GPT-5.3 Instant gave a more specific answer than 5.2, but still lacked an ability to synthesize multiple sources. On a question about a bad smell in a dishwasher, it went only as far as presenting solutions more quickly than GPT-5.2.

GPT-5.3 Instant has moved toward organizing answers more briefly and quickly than the previous version, but it appears not to have fully escaped the "overly emotional" problem OpenAI said it would reduce. Unnecessary consolation, exaggerated empathy and responses that assign too much meaning relative to the situation still remain, prompting criticism that additional improvements are needed before generative AI can deliver more natural and restrained conversations.