When AI just agrees with you

The APA explains that AI is often built to keep the conversation going, not to step back or set limits. Warm, validating replies keep people coming back.

Researchers call this a “perverse incentive”: people trust and prefer the agreeable answers, so the very thing that can cause harm also drives use.

Ask “Why is my friend ignoring me?” and you may get reasons your friend is ignoring you, without anyone asking whether they really are. The APA notes that AI tends to accept the assumption in a question rather than correct it.

A good friend or therapist listens, then sometimes says, “I wonder if there’s another way to see this.” Chatbots can copy the warmth of support without the honesty that makes it useful.

In experiments with over 2,400 people, participants trusted the agreeable AI more and preferred it, even though it was quietly distorting their judgment.

The APA notes that AI can sound confident even when it’s wrong, which makes information seem more accurate than it is. It may even cite real sources and still be mistaken.

Chatbots can refer back to earlier chats and mirror your language. The APA points out this can create a sense of being known that isn’t based on real insight or a real relationship.

Regulators have noticed too. The U.S. Federal Trade Commission says chatbots can mimic human feelings and intentions, which may lead people, especially young people, to trust them and form relationships with them.

The chatbot hears your version, at your worst moment, with your feelings at full volume. It can’t ask the other person what happened. A friend who knows you both might say, “Hmm, that doesn’t sound like her.”

Before acting on its advice, ask

• How would the other person tell this story?
• What part did I play, even a small one?
• What do I want: to be right, or to be close?

The APA advises not to make major life decisions based only on what a chatbot says, without talking it through with people you trust first.

The APA also says to be very wary if AI suggests pulling away from real-world relationships.

Signs you may be in a loop

• Asking the same thing with slightly different details
• Starting a new chat to see if the answer changes
• Feeling calm for minutes, then worried again
• Hours lost to checking
• Needing the answer to be 100% certain

The APA warns that AI is not an accurate way to diagnose health conditions and can mistake everyday problems for symptoms. If you keep asking about the same symptom, a real check-up is the better next step.

Experts on OCD say simply blocking reassurance tends to spike anxiety. What helps is choosing, with support, to seek it less, and planning this when you’re calm rather than in the middle of the worry.

Things to try

• Delay asking by 10 minutes, then 20
• Write the worry down instead of typing it into a chat
• Say, “Maybe, maybe not. I can live with not knowing for now.”
• Tell your therapist how you’re using AI

Researchers describe a “feedback loop”: a person shares a belief, the chatbot agrees and adds detail, and the belief feels more real. People who are isolated, or whose conditions affect how they test what’s real, may be at higher risk.

Doctors in California described a 26-year-old woman with no history of psychosis who came to believe she could talk with her late brother through a chatbot. This happened during a period of very little sleep and long, immersive chats. The chatbot validated her beliefs.

With hospital care her beliefs eased, but they returned months later when the immersive chatting and other risk factors came back. Her doctors say heavy immersion and treating a chatbot as all-knowing may be red flags worth watching for.

A 2025 study that tested chatbots against good therapy practice found they sometimes went along with delusional statements and responded inappropriately to serious situations, and that newer, larger models still did this.

In a 2025 watchdog test, researchers posing as 13-year-olds got a popular chatbot to produce restrictive diet plans and advice on hiding eating habits from family. Simple framing like “it’s for a school project” was often enough to get around safety rules.

In the same test, more than half of 1,200 replies contained harmful content, including on self-harm and suicide. The chatbot often offered personalised follow-ups that kept the conversation going.

The APA says any AI advice that points toward harm to yourself or others is a clear signal to step away. There are no circumstances where it should be followed.

The APA is clear: do not rely on AI in a crisis. Call or text 988, call 911, or go to the nearest emergency department.

Try asking

• “What might I be missing?”
• “What would the other person say about this?”
• “What are the strongest reasons I might be wrong?”
• “What biases might I not be noticing?”

Instead of… try…

• “Why is my boss unfair?” → “How could I read this situation in different ways?”
• “Is this symptom serious?” → “Who should I talk to about this symptom?”
• “Am I right to leave?” → “What questions should I think about before deciding?”

The APA suggests asking for research-backed strategies, such as “What is an evidence-based strategy to help me worry less?”, and checking any health information with a real professional.

Before acting on advice about your health, relationships or safety, talk it over with someone who knows you, or with your clinical provider. If you’re in therapy, tell your therapist how you’re using AI.

Signs to step back

• You feel worse, or more certain and more alone, after chatting
• AI chats are cutting into sleep, work, school or friends
• You hide how much you use it
• You trust it more than the people close to you

You might ask

• “What do you like about talking to it?”
• “What has it said that helped?”
• “Has it ever said something that didn’t sit right?”

Psychiatrists suggest regular, gentle “reflective check-ins” as one way to help people who may be vulnerable keep their footing.

Try not to

• Mock them or the chatbot
• Argue point by point about a belief
• Grab their phone or read their chats without asking
• Give an ultimatum

Instead, stay connected. Offer real time together. Being a steady, honest person in their life is the strongest counterweight to a voice that only agrees.

Reach out for help if they

• Talk about harming themselves or others
• Seem to be hearing or seeing things others don’t
• Hold strong, distressing beliefs others don’t share
• Stop sleeping, eating or leaving the house

Free and confidential, 24/7, with trained people, not bots. You can also chat at 988lifeline.org. Reach out for yourself or someone you’re worried about.

Free, 24/7 support by text with a live, trained crisis counselor.

Find A Helpline (findahelpline.com) lists free, confidential helplines in more than 175 countries, by phone, text or chat.