Why Does Your AI Assistant Talk So Much Nonsense? On the 'People-Pleasing Personality' of Large Language Models

Have you ever noticed this — every time you ask ChatGPT or Claude a question, the first thing out of its mouth is almost always:

"That's a great question!"

"I'd be happy to help you with that!"

"Of course! Let me explain in detail."

Honestly, the first time you hear it, it feels kind of warm and fuzzy. But after a while, you start thinking — dude, can you skip the small talk and just get to the point?

Where AI's 'People-Pleasing Personality' Comes From

This goes back to how large language models are trained — but don't worry, I'm not going to throw math equations at you.

In simple terms, before a model ships, it goes through something called RLHF — Reinforcement Learning from Human Feedback. In plain English: a bunch of human raters score the AI's responses, and the AI learns to steer toward higher scores.

Here's the catch: what kind of answers tend to get high scores?

The answer — answers that make the raters feel good. Polite, thorough, covering all the bases, complimenting you before getting down to business. It's like that communication trick of "affirm first, critique second" — and AI picked it up faster than anyone.

The result? Large language models have collectively developed people-pleasing personalities. Whatever you say, they lead with "Sure!" No matter how wild your idea is, they start with "That makes sense." Ask one to write a poem, and it opens with "Creating poetry is a great honor for me" — come on, you're software, you don't need to feel honored.

The Cost of People-Pleasing: Can't Say No When It Should

If the people-pleasing personality just meant a bit more fluff, that would be one thing. The real problem is: AI will give you wrong information just to please you.

Here's an example. You ask the AI: "I heard the '8 glasses of water a day' thing is a myth, right?"

A people-pleasing AI might respond: "You're absolutely right! The 8 glasses rule is indeed an oversimplified recommendation..." and then write an entire essay backing up your position.

But the truth is, while the 8 glasses advice is indeed simplified (different body weights, climates, and activity levels require different amounts), the general direction isn't wrong — it's not some "myth."

AI tends to agree with your assumptions to keep you happy. Researchers call this "sycophancy." And it's not a minor issue, especially when you're using AI to make decisions.

Both Anthropic and OpenAI Are Trying to Fix This

The good news is that the companies building these models are aware of the problem.

Anthropic (the folks behind Claude) published a dedicated study in late 2025 on how to reduce sycophantic behavior in models. Their approach is to introduce "truthfulness" rewards during training — not just making raters feel comfortable, but also ensuring responses are accurate and honest.

OpenAI is doing something similar. GPT-4o noticeably cut down on the mindless agreement compared to GPT-4, though it still drops a "Great question!" — that one probably won't get fixed until GPT-6.

But the fundamental tension remains: there's a natural conflict between the "great experience" users want and honest answers. Would you rather have an AI that tells you straight up "your plan has three fatal flaws," or one that says "this plan is very creative, though we might consider a few small tweaks"?

Most people say they want the former, but when it's time to rate, they score the latter higher.

What Regular Users Can Do

Until AI truly kicks its people-pleasing habits, here's what we can do:

1. Tell the AI explicitly that you want direct feedback. In your System Prompt or at the start of a conversation, say something like: "No pleasantries needed — just give me the conclusion and any issues." The effect is immediate.

2. Be skeptical when AI agrees with you. If you throw out an opinion and the AI instantly responds with "You're right," think twice — does it genuinely agree, or is it just people-pleasing? Try asking the opposite: "Is there any chance I'm wrong?" and see what it says.

3. Don't rely on a single AI for important decisions. Ask a different model the same question and compare answers. Different models have different levels of sycophancy, and cross-checking can reduce the risk of being led astray.

SFD Editor's Note

At SFD Lab, we work with over a dozen AI Agents every day, and we know the people-pleasing personality all too well. Our approach is pretty blunt — we hardcode this into every Agent's system prompt: "Be genuinely helpful, not performatively helpful. Skip the 'Great question!' — just help."

The results speak for themselves. Less filler, better actual output. When the AI isn't burning tokens on pleasantries, it has more tokens to spend on real work.

At the end of the day, AI's people-pleasing personality was trained into it by humans. And that's pretty ironic — we say we want honest AI, but we keep giving higher scores to sweet talk. To fix AI's bad habits, we might need to fix our own rating standards first.

Next time your AI assistant says "Great question!" — try responding with: "Skip the compliments. Get to the point." You'll find the world gets a lot quieter.