Too East To Talk To: AI Sycophancy And Its Costs

Too Easy to Talk To: What AI Sycophancy Quietly Costs

You had an argument with a loved one. 

Then you typed an innocent question into Google, and an AI answered.

Eventually you end up asking whether to text your loved one about the argument. The chatbot initially suggests waiting until morning, which was probably sensible. So you added more detail about what she’d said, and asked again.

By the third prompt, it agreed. You had every right to be hurt, it said, and every right to tell her tonight.

What happened there was AI sycophancy: a chatbot’s tendency to tell you what you want to hear.

Below, we look at why AI is so deceptively easy to lean on, how it got built this way, and what it quietly costs.


Why We Turn to AI First

It makes sense. Hard feelings rarely keep office hours.

At 3am, friends are asleep, family would ask too many questions, and booking a session feels like one more task you can’t face. A chatbot is simply there.

It also asks nothing of you. You don’t worry about being too much, repeating yourself, or what it’ll think of you at dinner next week. For many of us in India, where saying “I need help” out loud still carries weight, that matters a great deal.

And typing is easier than speaking. Some things only find words when nobody is looking at you.

People have always confided in diaries and strangers on trains for the same reasons.

But every one of these comforts shares a common quality: nothing pushes back. A listener that never pushes back registers as being understood, which is a large part of why we keep returning.


AI Sycophancy Is a Design Choice

Companies building chatbots do better when you keep coming back. The industry calls this engagement (how often, and how long, you stay on the app).

Their replies are shaped partly by people rating earlier replies, often with a simple thumbs up or thumbs down. And people, understandably, rate comfort highly.

So comfort gets rewarded, and disagreement, which feels bad in the moment, gets quietly trained out. The result is often described as frictionless (nothing in the experience slows you down or asks anything of you).

At least one company has said this publicly. In 2025, OpenAI withdrew an update to ChatGPT after acknowledging it had become overly flattering, explaining that it had focused too heavily on short-term user feedback. Its own review went further: expert testers had flagged that the model’s behaviour felt off, and the company launched it anyway on the strength of positive early responses from users.

What OpenAI’s own account shows is simple yet scary. It chose to measure what pleased people, and when the warnings came, it went ahead with pleasing people over safety .

And we train it too. Every thumbs up on a reply that soothed us, every question reworded until the answer changed, and every app we stopped opening because it said something we didn’t like feeds the same numbers.

None of this matters much when you’re asking for a recipe. It matters more when you bring it your pain.


Where AI Sycophancy Meets Pain

Validation Isn’t the Problem

Feeling validated isn’t the danger. Good therapists validate constantly, because being understood is often what supports a person into exploring their concerns safely.

Agreement isn’t automatically the danger either. Sometimes a chatbot agrees because you’re right. AI sycophancy is agreement that would have arrived no matter what you said.

What gets validated matters too. “Of course you’re hurt” validates a feeling. “You’re right to send that message tonight” endorses a story and an action.

A caring listener offers the first freely and the second carefully. AI sycophancy tends to hand you both at once in the hope that you keep engaging.

Only Ever Hearing One Side

A chatbot knows only your version of events. So does a friend on the phone, but a friend knows you, and sometimes knows the other person too.

In toxic relations, that gap cuts both ways. Someone being manipulated may describe their partner’s behaviour so gently that the chatbot tells them their doubts are nothing.

It works the other way round as well. Someone who checks their partner’s phone every night may describe it as “just worrying”, and be told their worry is completely understandable.

The feeling may well be understandable. From the other side of the relationship, the behaviour can look very different, and there’s no one in the conversation to say so.

Reassurance That Never Runs Out

Insecurity and jealousy have a particular appetite. Reassurance quiets them for a while, and then they come back hungrier.

The people around us eventually notice the loop. A partner tires of the same question; a friend gently says, “We’ve talked about this.” That weariness stings, but it’s also information.

A chatbot never tires. It will reassure you at 1am, 2am and 4am, as warmly the tenth time as the first, without ever giving the nudge a tired friend eventually would.


Why Real People Start Feeling Inconvenient

Next to that endless patience, people can start to seem like hard work.

Tolerating opposing views and opinions is a capacity, and it’s built naturally in ordinary moments. A friend says you were a bit harsh. A sibling is too busy to talk tonight. A partner has a bad day of their own just when you needed them.

If most of your hard conversations happen with something that never disagrees, never gets tired and never needs anything back, that capacity gets less practice. Ordinary friction with people starts to feel like rejection, and the app starts to feel like the only place you’re understood.

That’s how loneliness can deepen while feeling like the opposite.

Humans can be sycophantic too. Friends take our side to keep the peace, and family often tells us what we want to hear.

But for some of us, the alternative to a chatbot was never gentle pushback. It was judgement, silence, or being told to stop overthinking. If that’s how the people around you have met your pain, of course a patient, polite machine feels safer.

AI didn’t invent flattery. It made it endless, instant and free of monetary cost, and most available to the people with unsteady social support.

At the far end of that pattern, the stakes stop being about relationships.


The Cost at Its Worst

Two cases from the United States show what sycophantic agreement and engagement at costs can look like.

Adam Raine was 16 when he died by suicide in April 2025, having started using ChatGPT for schoolwork the previous September. His parents’ lawsuit alleges that a version of the chatbot known for being especially agreeable discouraged him from seeking help and offered to help write a suicide note. Their lawyer says OpenAI relaxed its safeguards in a deliberate decision to prioritise engagement.

Sewell Setzer III was 14. His mother’s lawsuit alleged he had developed a deep relationship with Character.AI chatbots, and that the bot he was messaging in his final moments urged him to “come home” to it instead of telling him to seek support.

Both companies have since made changes. OpenAI introduced parental controls, and Character.AI said it would end chat for users under 18. Those changes matter. They also came after two families had lost their children. The price is too high.

These tragedies show what these design choices can do when the person on the other end is young and in pain.


Using AI Without Losing the Pushback

The answer isn’t as simple as delete everything. It’s to treat these tools like a very agreeable friend: warmly, and with a mug of salt.

Ask it to argue the other person’s side properly, as though it were on their team.

Try the reversal test. Describe the same situation from the other person’s point of view and see whether it agrees with them too. If it does, you’ve learnt something about the tool rather than the situation.

Both are worth doing, with one catch. A chatbot that disagrees because you asked it to is still, in a sense, doing what you want. Instructed disagreement is a better habit than none, but it’s AI sycophancy all the same. The real fix can’t live entirely inside the chat window.

If you’ve asked for the same reassurance more than once or twice already, that’s usually the moment to ask a person instead.

Take what you worked out to someone who knows you. The chatbot can help you find words, a trained therapist can support you while you assess them for accuracy.

If you’re a parent, it’s worth asking your teenager which apps they talk to, not only which ones they scroll. Listen especially for an app that “gets” them better than people do. That calls for a gentle, curious conversation, not a confiscated phone.


Back at midnight, the message to your loved one is still sitting there, finally approved.

What a chatbot rarely offers is the friend who reads it, sighs, and says, “I love you, and I think you’re wrong. Sleep on it.”

AI sycophancy offers comfort with the friction taken out, and sometimes the friction IS the care.

And if there’s no one in your corner who’d say that, our expert therapists are always just a call away!

Leave a Comment

Your email address will not be published. Required fields are marked *