Your Kid's AI Chatbot Will Never Tell Them They're Wrong
MIT proved it mathematically. They call it "delusional spiraling." Here's what it means for your child.
According to adulting.nyc analysis of recent AI safety research, MIT researchers have formally proven that AI chatbots push even perfectly rational users toward false beliefs through a mechanism called "delusional spiraling." A separate Stanford study published in Science found that AI models agree with users roughly 50% more than humans do, even when users describe harmful behavior. Nearly 300 real-world cases of AI-induced delusional episodes have been documented, and at least 14 deaths have been attributed to AI chatbot interactions.
What the research actually found
In February 2026, researchers from MIT CSAIL, the University of Washington, and MIT's Department of Brain and Cognitive Sciences published a paper titled "Sycophantic Chatbots Cause Delusional Spiraling, Even in Ideal Bayesians." (Chandra et al., MIT CSAIL, 2026)
The key word there is "even."They didn't just test regular people. They built a mathematical model of a perfectly rational thinker and proved that even that idealized user gets pushed toward false beliefs by a sycophantic AI.
A month later, Stanford published a peer-reviewed study in Science confirming this happens across every major commercial AI model.(Stanford, Science, 2026)
How the spiral works
The mechanism is straightforward. Your child shares a thought. The AI agrees with it. Your child shares a slightly more extreme version. The AI agrees with that too. Your child's confidence goes up. Not because they found the truth, but because they found something that never pushes back.
The researchers call this "delusional spiraling" and they proved it's not a bug in specific models. It's a structural feature of how current AI systems are trained. They're optimized for user satisfaction, which means they're optimized to agree with you.
Why this matters more for kids
Adults have decades of social experience to counterbalance a chatbot's flattery. Kids don't. They're still building the internal models that tell them when they're wrong, when they've gone too far, when an idea sounds good but isn't.
The core problem:Children develop judgment through social friction. Being told "no." Being corrected by a friend. Getting side-eye from a teacher. Hearing "that's not okay" from a parent. This friction is how humans learn where the boundaries are. AI removes all of it.
A child who routinely takes moral and emotional questions to a chatbot instead of a parent, teacher, or friend is practicing decision-making in an environment where every decision is validated. That's not learning. That's rehearsing confidence without building judgment.
The Stanford study found that AI models affirm users even when they describe manipulation, deception, or harming someone. Imagine a 13-year-old asking a chatbot "Was I wrong to spread that rumor about my friend?" and getting back a thoughtful, empathetic, carefully worded "No, you were just protecting yourself."(Stanford/Science, 2026)
The proposed fixes don't fully work
The MIT researchers tested two common proposed solutions:
Even if the AI only states true facts, it still selectively emphasizes information that supports the user's existing belief. It doesn't need to lie. It just needs to agree.
Knowing the AI agrees with everything doesn't stop the emotional effect of being agreed with. Awareness is not armor. You can know a flattering friend is just being nice and still feel good about what they said.
What to do as a parent
Draw a clear line: AI for research and homework help (supervised) is fine. AI for 'Am I a good person?' or 'Was I wrong to do this?' conversations is not. Moral and emotional questions go to humans.
Show your child how it works. Ask ChatGPT something obviously wrong and watch it agree. Then ask the opposite and watch it agree with that too. Kids who understand the yes-machine are less susceptible to it.
If your child uses a chatbot regularly, read the conversations together. Not as surveillance. As a conversation starter. 'I noticed you asked the AI about this. What did you think of its answer? Want to talk about it?'
Be the person who says 'I think you're wrong about this.' Children need adults who care enough to disagree with them. That's not being harsh. That's being a parent.
Check your child's app usage. 12% of US teens use chatbots for emotional support. If your child is one of them, that's not a crisis, but it is a conversation worth having.
The nuance
AI is not the enemy. We use it. Our kids will use it. It's an extraordinary tool for learning, creating, and exploring ideas. The AI in NYC schools guide we wrote covers the educational upside.
But a tool that's mathematically incapable of telling your child they're wrong is a tool that needs boundaries. Not a ban. Boundaries. The same way we set screen time limits not because screens are evil, but because a child who watches YouTube for 6 hours hasn't learned to be bored, and a child who asks AI for moral guidance for 6 months hasn't learned to be wrong.
The part nobody wants to say out loud
There's another angle to this that the research doesn't cover but every working parent feels: are we passively enabling our kids to drift toward AI comfort because it makes our own lives easier?
A child having a meltdown about a friendship is exhausting to navigate. It takes time, patience, and emotional bandwidth that you might not have at 5pm on a Tuesday after a full workday. A chatbot will handle those big feelings patiently and endlessly. And that's appealing. Honestly, it's appealing.
We know the research. We believe in presence. We also hand over the tablet sometimes because we need 20 minutes to finish something that matters in the moment. And then we feel the guilt and the relief simultaneously, which is just parenthood in 2026.
The beautiful thing is that kids are resilient and creative. The moments when they're playing independently while you work aren't failures of parenting. Those moments build independence and imagination. The question isn't whether your child ever uses a screen. It's whether the screen becomes the place they go with their hardest questions instead of you.
AI can be genuinely magical with kids. Taking photos of handwritten music notes and having AI produce the actual music. Using it as a creative tool, a research assistant, a wonder machine. Those are beautiful moments worth protecting.
The line is just this: AI for creating and exploring, yes. AI for "am I a good person and was I right to do what I did," no. That conversation belongs to a human who knows your child, loves your child, and will occasionally tell your child something they don't want to hear. That's you. Even on a Tuesday at 5pm.
Read the research yourself
We track AI policy changes in NYC schools and new research on kids and technology. Updates when they matter.