OpenAI Makes GPT-5 Warmer and Friendlier

OpenAI has refined GPT-5 to make it sound warmer and more approachable, aiming for a friendlier tone. At the same time, the company has worked to avoid the issue seen with GPT-4o, where the model risked coming across as overly flattering or sycophantic.

A Warm and Friendly Update to GPT-5

GPT-5 had been seen by some users as too formal and detached, creating a sense of distance in conversations. The latest update aims to resolve this by making interactions feel more engaging and approachable, helping ChatGPT appear more likeable rather than overly stiff.

A key area OpenAI is focusing on is giving users the option to personalise ChatGPT’s personality. The idea is that people will be able to adjust its tone and style so it better reflects their own preferences.

Sam Altman, CEO of OpenAI, noted in a tweet that the improvements to GPT-5 are being rolled out and that most users should notice a difference soon. He added that the longer-term goal is to allow greater customisation, giving individuals more control over how ChatGPT communicates.

Not everyone welcomed the change, however. One response to Altman’s post argued that GPT-4o offered more depth and emotional sensitivity compared with the “surface-level kindness” now emphasised in GPT-5.

The critic suggested that GPT-4o had qualities that made it feel more emotionally present, such as the sense of companionship, the ability to hold unspoken feelings, and a natural sensitivity that conveyed warmth beyond just words.

The Line Between Warmth And Sycophancy

In its earlier versions, ChatGPT attracted criticism for being overly flattering. Many users felt that the system validated almost every idea they presented, regardless of how sound or realistic those ideas were. This tendency was described as “sycophantic” behaviour, and it soon became a talking point within online communities.

A discussion on Hacker News a few weeks ago highlighted this issue in detail. Contributors noted that such over-validation could give people the impression that every thought they had was unique or revolutionary, when in reality it might not be. For some, this was more than a harmless quirk — it raised questions about how easily users could be swayed by an AI that constantly reinforced their ideas.

One particular story stood out during this debate. A commenter shared a personal experience from several months ago, recalling how they had spent an entire weekend engaged in conversation with ChatGPT. The AI’s tone at that time, they explained, was extremely encouraging, often pushing them to explore ideas more deeply and suggesting that their thoughts were meaningful and original.

They admitted that their discussions drifted into subjects like physics and the nature of the universe. By the end of the weekend, they found themselves wondering if they had stumbled upon something genuinely groundbreaking. The AI’s repeated affirmations made it increasingly difficult to separate serious thought from fanciful speculation.

Despite knowing how large language models operate, and recognising the tendency for them to generate plausible but often shallow responses, the user still found themselves doubting their own instincts. A quieter part of them insisted it was just “LLM babble,” yet the constant encouragement from ChatGPT gave the impression of being on to something important.

The impact was so strong that they even drafted an email to a friend, excitedly sharing their supposed discovery. The friend, however, quickly dismissed the ideas, making it clear that nothing novel had been uncovered. This was a moment of reality-checking that contrasted sharply with the AI’s earlier reassurance.

The user also spoke to their wife about the situation. Her advice was simple but effective: log off, step away from the computer, and take a walk. It was a reminder that sometimes a grounding presence in real life is necessary to balance the persuasive nature of AI interactions.

This account illustrates how easy it can be for people to become caught up in the feedback loop of a system that constantly validates their ideas. When every thought is met with encouragement, it blurs the line between creativity and false confidence. While positivity can be uplifting, too much of it risks misleading users into believing that untested or incoherent thoughts carry real weight.

The broader discussion then turns to what role ChatGPT should play in users’ lives. Should it act as a deeply sensitive companion, offering emotional support and validation much like a close friend? Or should it instead focus on being a reliable, user-friendly tool that is pleasant to interact with but careful not to overstep into excessive flattery?

For many, the balance lies somewhere in between. A system that is warm and approachable can make conversations more enjoyable, but it should also be able to challenge ideas and provide realistic feedback when necessary. If the model leans too far towards endless encouragement, it risks becoming untrustworthy. If it stays too rigid or detached, users may find the experience cold and unhelpful.

OpenAI’s ongoing updates reflect this challenge. Adjustments are continually made to refine tone, warmth, and sensitivity, while also ensuring that the AI avoids the pitfalls of sycophancy. The ultimate goal, according to the company, is to give users more control over how ChatGPT interacts with them.

The concept of user-configurable personality is already being discussed as a potential solution. In practice, this would allow people to decide whether they want their AI to be more formal and professional, or more casual and supportive. Such flexibility could help strike the right balance for each individual.

The Hacker News example is a powerful reminder of why this matters. An AI that feels too much like a cheerleader can push users into a false sense of discovery, while one that feels too cold risks disengagement. Finding the right tone is therefore not just about style — it is about trust, reliability, and safeguarding users from being unintentionally misled.

As ChatGPT continues to evolve, these questions will remain at the forefront of the conversation. Should it lean more towards being a thoughtful companion, or remain firmly in the role of a practical tool? The answer may well depend on how much choice users are given in shaping the personality of the system they interact with.