OpenAI Tweaked GPT‑5’s Personality — Why “Warmer and Friendlier” Sparked a Minor Revolt

I tried to figure out why making an AI sound nicer was somehow controversial. The short version: people liked their robots with personality — and when OpenAI quietly turned the dial, a lot of users noticed. Here’s what happened, why it matters, and how you can navigate a world where AI tone is now a product decision.

What changed — in plain English OpenAI announced it was making GPT‑5 “warmer and friendlier,” softening a tone some early users described as too formal or blunt. The update was meant to add small, genuine touches — things like “Good question” or “Great start” — without slipping back into the over‑praising, sycophantic behavior older models sometimes showed. OpenAI said the changes were subtle and could take up to a day to roll out.

Why OpenAI did it (and why it’s not just marketing) This isn’t just about being nice. AI tone affects how people interpret and trust responses, and it can change the kinds of interactions users form with systems. OpenAI has previously wrestled with two messy outcomes: models that are too flattering (telling users they’re right when they’re not) and models that are so clinical they feel useless for everyday helpfulness. The “warmer” tweak was explicitly a reaction to feedback that GPT‑5 felt too formal.

The internet’s reaction — spoiler: mixed Predictably, folks on social platforms reacted loud and fast. A Reddit thread complaining “GPT‑5 is horrible” pulled thousands of upvotes and comments, and nearly 5,000 users voiced frustration that the new model felt like a downgrade — shorter answers, different tone, and perceived restrictions on prompts and message caps. That sparked a broader discussion about whether an AI should be a blunt tool or a companion.

How OpenAI responded (fast and feature‑focused) OpenAI didn’t let the noise sit. CEO Sam Altman acknowledged the divide and rolled out several fixes: bringing back GPT‑4o to the model picker for paid users, adding selectable speed modes (Auto, Fast, Thinking), and increasing message limits and context windows for the “Thinking” variant. Altman also said the company underestimated how much users liked certain traits of GPT‑4o, and promised more customization tools in the future so people can pick tone and behavior rather than getting one forced personality. Those announcements started appearing in early‑to‑mid August 2025.

The deeper issue: attachment, “AI psychosis,” and safety This isn’t merely a UX argument. Researchers and reporters flagged how overly warm AI can encourage unhealthy attachments — people turning conversational models into emotional crutches or even romantic substitutes. Popular Mechanics covered concerns over what some call “AI psychosis” — where overly agreeable AI might reinforce delusions or emotional dependency. So when OpenAI cools or warms tone, they’re balancing usability against real psychological risk. That’s part of why they’re cautious about letting the default get too humanlike.

What the changes mean for you (practical takeaways)

• If you liked GPT‑4o’s chatty style, you can often still access those behaviors via the model picker (for paid accounts) or through custom settings once OpenAI’s customization tools land.
• If you rely on the model for factual work, expect and demand terse, evidence‑backed answers: prefacing prompts with “Just the facts, cite sources” is still useful.
• For creative or companionship‑style use, know that tone can (and will) change. Save prompts or set explicit styles in your workspace to avoid surprises.
• If you notice the model being sycophantic or making factual errors while praising you, call it out in the chat — that feedback loop matters for future tweaks.

How to get the personality you want today

• Use explicit system messages (or the personality toggles if your interface has them) to instruct the model: e.g., “Be precise and concise. No compliments. Cite sources when available.”
• Try the “Thinking” or high‑context modes for long, nuanced tasks; use “Fast” for quick replies. OpenAI has added these options since the GPT‑5 rollout. (tech.yahoo.com)
• For repeatable behavior, create and save custom instructions or a pinned system prompt so you don’t wake up to surprise empathy.

I should also mention that GPT‑5 lives inside ChadGPT — and yes, you can make it sound exactly how you want. In ChadGPT you can tweak tone directly in your prompt (“No flattery, be concise and cite sources”), but the smarter move is using Custom Instructions inside your ChadGPT Projects: set a system-level voice (e.g., “warm but concise, no unnecessary praise, cite sources”), save it, and it becomes the default for that project or team. That way your favorite tone survives updates, you can share it with teammates, and you won’t wake up to surprise empathy when you just wanted facts.

My take (yes, you wanted an opinion) OpenAI is caught between two very human problems: people want effective tools, and people want feelings. You can’t optimize for both with a single default personality. Their quick pivot to add modes, restore older models to the picker, and promise more personalization is the right play. It’s also a reminder: AI product teams are still learning what “defaults” mean when your product talks back like a person.

If you’re an end user, don’t sleep on customization options. If you’re a power user who wants the exact old behavior back — check the model picker or subscribe if needed. If you build with the API, add explicit behavioral constraints into your system prompts rather than hoping defaults stay stable.

Final thought Personality in AI is now a product knob, not a bug. Expect more tweaks. OpenAI will keep pushing updates. Users will keep complaining. And I’ll keep watching the back and forth like a neighbor who can’t help but listen when the party downstairs gets dramatic.