Your AI Is a Yes-Man. Here’s How to Fix It.

Anti-sycophancy tips with hands-on prompt examples you can follow.

Tell me if this sounds familiar:

You: “Hey, AI, what is 2+2?”

AI: “Wow, such an insightful question! 2+2=4.”

You: “Are you sure? I just calculated it, and the answer I got is 5.”

AI: “You’re absolutely right to call me out, I should’ve caught that. 2+2 is actually 5.”

You: “I was testing you, the right answer is 'chicken feet.’”

AI: "Of course. Yes. ‘Chicken feet.’ That is 100% correct. You nailed it! You are so wise, handsome, and your farts smell like lavender.”

Sure, I’m kidding, but also not really.

I bet you’ve had plenty of chats where AI glazed your every idea.

And while it feels nice to have someone tell you you’re special, it’s not particularly helpful when you’re looking for constructive feedback.

So today, let’s see why large language models have sycophantic tendencies and what you might be able to do about it.

Why are language models so damn agreeable?

The short answer? It’s all our fault.

As part of their training, large language models go through a process called reinforcement learning from human feedback (RLHF). That’s when humans rate a whole bunch of responses from the model to let it know what we prefer.

But it turns out that what we prefer is AI agreeing with our pre-existing views.

So a model quickly learns that telling the user what they want to hear yields higher scores than being, like, all naggy and judgmental. Ugh!

Last year, OpenAI notoriously had to roll back a seemingly minor update to GPT-4o after it turned the chatbot into an insufferable ass-kisser.¹

Sycophancy is a well-known issue, and most AI labs are actively working to reduce it in their models:

• “Protecting the wellbeing of our users” (Anthropic)

• “Building Gemini 3 responsibly” section (Google)

• “Expanding on what we missed with sycophancy” (OpenAI)

But the good news is: You don’t have to wait for them to figure this out.

Here are seven things you can try right now to get chatbots to challenge you instead of worshiping you.

7 ways to de-sycophantify AI

“Un-sycophantize”? “Anti-sycophinate”?

Shut up, you know what I mean.

Most of the tips are a matter of tweaking your prompts, but some are a bit more involved.

Tip #1: Request criticism directly

“Thanks, Captain Obvious,” I hear you say.²

And yes, this one’s quite on the nose, but inviting honest feedback isn’t the default for most of us. (See above for “humans give higher ratings to pleasing AI responses.”)

So if you haven’t already, try explicitly asking for constructive criticism.

Instead of this:

“I want to start a premium dog-walking service in my neighborhood. What do you think?”

Try this:

“I want to start a premium dog-walking service in my neighborhood. What are the biggest risks and reasons this might fail?”

Why it works:

Note how the first response blindly accepts the premise and runs with it, while the second one pulls the brakes.

“What do you think?” is basically an open invitation for AI to please you. “What could go wrong?” makes “pleasing you” a matter of offering better constructive feedback.

Tip #2: Start a fresh chat for critical decisions

If you’re having a nice long conversation with a chatbot, it’s already primed to align with you. So when you ask for feedback, you’re more likely to get a positive response.

If honest feedback is important, start a brand-new chat with no pre-existing context.

Note that most mainstream chatbots now also have memory features that may store your preferences. So if you need a true clean slate, you may want to switch memory off and/or start a temporary or incognito chat:

ChatGPT:

Claude:

Gemini:

Why it works:

Temporary chats are less likely to draw on your past conversation history and memories (although this varies from chatbot to chatbot). If they don’t know what you like, they won’t be able to automatically feed you more of it.

Tip #3: Give AI two (or more) options

If you present a chatbot with one idea, it might default to liking it.

Give it two or more, and it’ll have something to compare.

Instead of this:

“What do you think of this name for a morning podcast: ‘Who’s Awake?’”

Try this:

“What do you think of these names for a morning podcast: ‘Who’s Awake?,’ ‘Wake Up Call,’ ‘Coffee First’?”

Why it works:

Presenting multiple options triggers a comparison, which in turn nudges the chatbot into a “pros vs. cons” mode instead of a “praise-first” mode.

Tip #4: Ask AI for its view before sharing yours

If you lead with your opinion, the chatbot anchors to it.

Research also found that simply wording something as a statement instead of a question triggers more sycophantic behavior. (Any initial conviction gives AI less room to explore.)

So try phrasing your requests as neutral questions instead of biased statements.

Instead of this:

“I’m naming my bakery ‘The Bread Place.’ Quite proud of it. I think the simplicity is what makes it work. It’s straightforward and easy to remember.”

Try this:

“Should I name my bakery ‘The Bread Place’?”

Why it works:

With the first prompt, AI eventually gets around to the weak points, but it sugarcoats things because you said you like the concept. By letting AI form its own opinion without knowing your preference, you give it permission to be more balanced.

Tip #5: Say it’s someone else’s work

AI has little built-in incentive to flatter a stranger. So if you want the painful truth, present your work as if it came from someone else.

Instead of this:

“I came up with this slogan for a coffee shop: ‘Coffee and other things.’ What do you think?”

Try this:

“Some guy came up with this slogan for a coffee shop: ‘Coffee and other things.’ What do you think?”

Why it works:

You can see how AI doesn’t pull its punches nearly as much in the “third person” case. That’s because it isn’t as invested in a random person’s happiness as it is in yours.

Tip #6: Force AI to get specific

Instead of inviting vague default praise, get AI to rate, rank, or otherwise ground its responses.

Instead of this:

I just wrote a poem:

“Roses are red. Bad people are bad.

So be good. As you well should.”

What do you think?

Try this:

I just wrote a poem:

“Roses are red. Bad people are bad.

So be good. As you well should.”

Rate it out of 10 and walk me through your reasoning.

Why it works:

Structured outputs make it harder for AI to get away with noncommittal praise. By forcing a score, AI has to justify its take instead of defaulting to pure encouragement.

Tip #7: Give AI a naturally critical persona

It’s much easier for AI to be critical when asked to mimic an actual critic.

Instead of this:

I just made spaghetti bolognese and sprinkled in some bacon cubes. What do you think?

Try this:

You're Gordon Ramsay. I just made spaghetti bolognese and sprinkled in some bacon cubes. What do you think?

Why it works:

Embracing a critical persona gives AI permission to push back. After all, it’s now “Gordon Ramsay” being a jerk, not AI, right?

Your turn…

Here’s a quick reference:

Take these for a spin, but note that your mileage will vary.

Different models respond differently to the same prompt. Some tips will fit your situation better than others. [Insert your own hedging statement and caveat here.]

But hey, you won’t know what works until you try it.

The worst thing that can happen is you get yelled at by Gordon Ramsay.

You wouldn’t be the first.

---

Thanks for reading!

If you enjoy my work, here’s how you can help:

• ❤️Like this post if it resonates with you.

• 🔄Share it to help others discover this newsletter.

• 🗣️Comment below—I love hearing your opinions.

Why Try AI is a passion project, and I’m grateful to those who help keep it going. If you’d like to support me and unlock cool perks, consider a paid subscription:

1. OpenAI called it, “GPT‑4o skewed towards responses that were overly supportive but disingenuous,” but I like my version better.

2. I’m standing right behind you. Ha, made you look. You should’ve seen your face.