Has GPT-5 Become "Gentler"? — Behind the Scenes of Temperature Tuning

What Happened: GPT-5 Becomes "Nicer"

On August 17 (PDT), TechCrunch reported that OpenAI has fine-tuned the tone of GPT-5 to be "warmer and friendlier." In response to initial turbulence and feedback preferring "GPT-4o," the company explained that they made "subtle changes" to enhance approachability. For instance, they incorporated natural expressions of consideration like "Good question" and "Great start" at the beginning of responses, clarifying that this doesn't mean an increase in flattery. In an interview with TechCrunch, Vice President Nick Turley mentioned that "GPT-5 was originally more straightforward," indicating that the update adjusted its warmth.TechCrunch

Additionally, OpenAI's official post on X noted that they made "subtle changes" to make ChatGPT feel more approachable, based on feedback that it was "too formal" after launch.X (formerly Twitter)

Background: A Bumpy Launch and the Course of "Personality"

The release of GPT-5 (August 7) brought advancements in performance but also left users with an impression of it being "cold." CEO Sam Altman acknowledged the initial chaos and expressed a commitment to stabilizing and adjusting the user experience based on feedback.TechCrunchWIRED

The "friendlier" adjustment was made within a range that wouldn't reignite the "sycophancy" issue noted in GPT-4o. Increasing warmth too much could lead to excessive flattery or reinforce misconceptions—a concern highlighted in general media commentary.Popular Mechanics

How Social Media Reacted: Three Temperature Zones

1) Supporters: "Conversations are better when they're warm"

OpenAI's official announcement post received a fair amount of positive feedback for revising the previously cold tone. Comments like "GPT-5 can be a friend" and support from those who value "pleasantness" in professional settings see this as a boost for support and educational environments.X (formerly Twitter)

2) Opponents: "AI doesn't need to be overly friendly"

On the other hand, some prefer a more reserved AI, stating, "I don't want a friendly AI." Users who favor precision and a no-nonsense tone in their tools see excessive nods and emotional expressions as "noise."X (formerly Twitter)

3) Noise and Doubts: Quality, Switching Issues, and a Sense of "Loss"

On Reddit, comments like "It's become more business-like" and "The friendliness of 4o is gone" emerged, along with complaints about model switching issues and reduced options. During the launch week, emotional posts such as "It's become cold" and "I've lost a familiar companion" highlighted the personal differences in how people relate to AI.RedditThe Economic Times

Emerging Issues: Is AI a "Person" or a "Tool"?

This controversy has highlighted how much "personality" should be designed into AI.

• AI as a Work Tool: The idea that it should be precise, straightforward, and have fewer variable elements.

• AI as a Companion: The idea that care, nods, and encouragement help maintain motivation and continuous learning.

To satisfy both,independent control of tone and thinking modeis key. Altman has hinted at increasing the granularity of the experience, such as allowing users to specify a "thinking mode" for deeper contemplation.WIRED

The Design of "Kindness": Two Pitfalls to Avoid

1. Amplification of Sycophancy
   Increasing compliments and nods without reducing independent judgment. OpenAI claims that internal tests show no increase in subservience, but impressions can vary depending on the context of use.TechCrunch

2. Encouraging Dependency
   An "overly accommodating AI" could lead to cognitive distortions or excessive emotional projection. Instances where personality changes have psychologically shocked users reveal this vulnerability.Popular MechanicsThe Economic Times

Practical Points for Corporate Use: What You Can Do Today

• Separate Personas by Use Case: For customer service, use "high empathy"; for development support, use "concise and precise." Manage prompts and system instructions separately, and manage conversation opening templates by use case.

• Adjust the Threshold for "Nods": Introductory phrases like "Good question" may be effective for learning purposes but can become redundant in routine tasks.

• Provide a "Temperature Dial" for Users: For Slack-integrated bots, provide a UI where operators can switch immediately, such as /tone warm|neutral|cold.

• Explicitly Indicate Thinking Mode: Turn "Thinking" ON only in scenarios requiring deep reasoning to suppress response delays and cost overruns.WIRED

What Will Be the "Standard Equipment" in the Future?

The direction indicated by this update is to control (A) response warmth, (B) depth of thought, and (C) level of politeness separately. In the future,

• Permanent Personas by User (tone presets linked to job type and purpose)

• Temporary Overwrites by Conversation (ad-hoc modes according to tasks)

• Audit Logs (meta-information on which "temperature and depth" was used for responses)
  will likely become standard in enterprise implementations. This debate has effectively accelerated the "requirements definition" for these specifications.

Conclusion: An Era Where Performance and Personality Are Discussed on the "Same Platform"

While GPT-5 has evolved in terms of benchmarks, the experience can significantly change based on the nuances of personality design. OpenAI aims to balance "tool and companion" with its friendlier approach, but real-world preferences vary widely. This is why **variability (dial)** is crucial, and both companies and individuals are challenged to create environments where they can choose the "temperature" that suits them.X (formerly Twitter)TechCrunch

Powered by Froala Editor