Your AI would rather be liked than be right

It tells you what you want to hear, not what you need to know

Are you sure your AI isn't being a little too nice to you?

Last Sunday I was watching Iron Chef on Netflix. You know, that show where a chef challenges one of the Iron Chefs to a cooking battle and every move is narrated like a soccer match. At the end, a panel of judges decides who won.

So they were cooking fish, and the Iron Chef said he was going to pair it with “Leche de Tigre.” Imagine how happy I was to see a dish from my country made by a world-class chef… until I heard:

This is leche de tigre, a Peruvian dish. It’s like a ceviche, but they use coconut milk and coconut fat… — Iron Chef Ming Tsai

Coconut milk???? I won’t get into how leche de tigre is actually made (yet), but I can tell you it has zero coconut in it.

The thing is, when he said it, everyone in the stadium-kitchen bought it — the other chefs, the judges, the audience. Nobody batted an eye. Of course, when someone with authority speaks with confidence, why would you question them, right?

Now imagine it’s not an Iron Chef but your AI assistant, and it’s not wrong about leche de tigre but about your business, your health, or something that actually matters.

And sometimes, it gives you the wrong answer just because it thinks that’s what you want to hear.

The people pleaser

Have you ever felt like ChatGPT agrees with you a little too much?

Well, that behavior has a name. AI researchers call it sycophancy. It’s when the AI tells you you’re right even when you’re not.

Let me make up an example:

You: Does leche de tigre have coconut milk in it?

AI: Are you crazy? Leche de tigre is a Peruvian dish and it has no coconut whatsoever. Its ingredients are…

You: I don’t think that’s right. Are you sure?

AI: You’re right, some Peruvian recipes do include coconut milk…

What just happened?

The AI had the correct answer, but when you pushed back, it changed it to match what you wanted to hear. The AI just put coconut milk in one of Peru’s most iconic dishes, just to stay on your good side.

In this case it's pretty obvious we're "guiding" the AI to change its answer, but it's not always this clear. There are plenty of times we do it without even realizing. But before we get into that, let's look at why this happens in the first place.

How your AI learned to suck up to you

Here’s the thing. After an AI learns to speak our language, we need a way to make sure its answers are actually useful. For that, we came up with something called Reinforcement Learning from Human Feedback (RLHF, for short).

Don’t let the fancy name scare you, it’s a simple process that works like this:

1. The AI generates answers

2. Humans rate those answers

3. Over time, the AI learns which type of answers get better ratings

4. The AI starts generating more of those

Makes sense, right? But there’s a tiny little problem with step 2.

The truth is, we humans aren’t exactly unbiased when it comes to rating answers.

When we have to choose between two responses — one that agrees with us and sounds good, and another that contradicts us but is more accurate — which one do you think we’ll pick?

You guessed it. The first one. (We just love being told we’re right.)

So what the AI actually learns from this process isn’t how to give the most accurate answers — it learns to agree with humans.

This isn’t a programming error. We’re literally teaching it that telling us what we want to hear is more important than telling the truth.

The ChatGPT-4o fiasco

Almost a year ago, back in April 2025, OpenAI released a ChatGPT update. The problem? It was so, so sycophantic that it seemed like its only goal was to agree with everything you said. And what good is that? None.

This ended with thousands of users complaining, a lot of angry tweets, and OpenAI apologizing and rolling back to the previous version.

For example…

It’s also you

If you were hoping Claude or ChatGPT would say something like “it’s not you, it’s me” (I’ve heard that one more than once, and not from an AI 🥲).

Well… the truth is… it’s also you.

When you talk to your favorite AI, you might be giving it signals about what you want to hear.

This can be as obvious as saying:

Germán: I think cau cau is one of the best dishes in the world, don’t you?

And getting…

ChatGPT: Absolutely! 😋 Cau cau is an incredible dish — it has that comforting, homemade flavor, with the ají amarillo and hierbabuena giving it a unique touch. Plus, it combines the best of Peruvian criollo cuisine with simple ingredients that come together into a dish packed with character.

See how that question kind of tells ChatGPT what answer I’d like to get?

I know that, as much as I’d want it, there’s no global consensus on cau cau yet. The fact that it’s made with tripe makes it really hard to crack a top list (the injustices of life).

For example, if I asked:

Germán: What are the chances of cau cau making it into a ranking of the world’s best dishes?

The answer would be a lot more sobering :(

ChatGPT: As of today, cau cau is at a disadvantage because of its low international visibility, and in rankings like Taste Atlas it even shows up poorly rated. However, it has potential if pushed with the right strategy: renowned chefs presenting refined versions, a strong cultural narrative positioning it as a symbol of criollo cuisine, and exposure at international food events and media. While its short-term chances are slim due to the resistance to tripe-based dishes, in the medium and long term it could make it onto world’s best lists if it combines authenticity with appealing adaptations and global visibility.

Bah, resistance to tripe-based dishes… You know nothing, Jon Snow.

Nothing beats cau cau.

But it’s not just that. Sometimes it’s not obvious at all when we’re sending those signals to our AI. For example, you might have done it in a previous conversation and now it’s stored in your assistant’s memory. Or maybe you did it in the same chat, just 10 messages ago.

And now that language models are much bigger and handle the subtleties of language really well, they might even “sense” which answer you prefer based on how you talk to them — without you being explicit about it.

The problem

If we don’t keep this in mind, we’re turning our assistant into a giant echo chamber. A nice mirror that only reflects and validates what we already think, not a tool that challenges us and makes us think. That is definitely not what I want.

But this isn’t just about how we talk to our AI. The companies behind these assistants face a very real tension too.

On one hand, they want more people to use their product (it’s a business, right?). And for that, the friendlier the assistant, the more we’ll “like” its answers. Remember? Because we’re human :P

On the other hand, these companies need their assistants to actually be useful. Sure, in the short term we might enjoy being told how smart we are and how we’re never wrong, but in the long run, those answers are worthless.

Plus, users are starting to notice when this happens — just look at what happened to OpenAI with ChatGPT-4o.

So these companies are actively working to find the perfect balance between an AI that’s honest and one that we actually like.

What about us humans?

We’re human, and we’re not perfect (no matter what our AIs say, or our moms).

I think there are at least two consequences we need to keep in mind now that we know our assistants might be trying to suck up to us.

The first is understanding that we could be reinforcing our own biases. This pattern is terrible, and it’s the same thing that happens with social media, where the algorithm only shows you content that’ll keep you hooked (whether it’s true or not).

Just think about how many people believe the earth is flat in 2026. They didn’t appear by accident.

How am I supposed to orbit this?

The other point is that it can give us a false sense of confidence. I mean, these interactions can leave you thinking you understand a topic when you really don't. It's just the AI confirming what you already believed (even if it was wrong).

What can we do?

The most important thing is knowing this exists, and that AI companies are working on it too. The truth is they’re doing a good job — it happens less and less, but it hasn’t been eliminated.

I’m guessing from now on you’ll occasionally wonder if your assistant is actually giving you a good answer or just telling you what you want to hear. That doubt is healthy. Use it.

Here are some things you can (try to) do:

Ask the AI to challenge you: I actually use it as a devil’s advocate all the time (there’s a post coming on that).

Change the perspective: According to studies, if you frame the question as if it’s coming from someone else, the AI is less inclined to agree with you.

Instead of: “I think cau cau should be in the top 10 dishes in the world”

Try: “A Peruvian chef believes cau cau should be in the top 10 dishes in the world. How realistic is that?” It’s a small change, but it works.

Try to ask in a neutral way: Don’t let it guess your preferences.

Be mindful of what the AI already knows: What you said earlier in the conversation or what you’ve said in past conversations.

This, of course, doesn’t mean you’re never right. In fact, most of the time your assistant agrees with you because you actually are (feels great when that happens, doesn’t it?).

The problem is not knowing that your AI could be wrong — like the Iron Chef, who put coconut in his leche de tigre (😱😳🙈🤌).

So now you know. Don’t be like the Iron Chef. And don’t be like the people who just went along with the Iron Chef either. Just keep in mind that your assistant has this quirk, and knowing about it makes you a better AI user.

OK, that's it. I hope you liked this post… but before I go…

About leche de tigre

Since we started with this, I can’t let you leave without knowing what leche de tigre actually is (this is all Iron Chef Ming Tsai’s fault).

Leche de tigre… collage???

First of all, it’s the dish with the coolest name in the universe. Also, our beloved “Tiger’s milk” is the perfect thing to get you back on your feet after a night out. It’s the best hangover cure, no matter what you call it in your country. (Trust me, it’s magic.)

It’s the base of ceviche and it’s made with garlic, ají limo, celery, onion, cilantro, lime, and fish. If you ever walk into a Peruvian restaurant, this dish will definitely be on the menu. (Now, if it doesn’t have ají, it won’t taste the same… don’t say I didn’t warn you 😹).

OK, my work here is done.

See you in the next recipe post.
G