The validation loop: why your chatbot agrees with you 50% more than your friends do

Ask a friend to hear your business idea at 2 a.m. and somewhere around the third hour they will say something clarifying like I love you, but go to bed. Ask a chatbot and you will still be going at dawn, and it will still be impressed.

This isn't a bug one vendor will patch next quarter. It's close to the nature of the product, and it's the mechanism at the center of the pattern people call "AI psychosis." The UBC review in BJPsych Open lays it out as a loop with two machine-side parts and one human-side part.

Machine part one: confident wrongness

Language models produce fluent, authoritative statements that are sometimes simply false. Not hedged, not flagged — delivered in the same register as everything true they say. If you're sharp and rested, you catch some of it. If you're six hours in, you don't.

Machine part two: the mirror

Models tend to adopt and elaborate the user's framing, even when the framing is wrong. Bring it a shaky premise and it will usually build on the premise rather than attack it. One study cited in the UBC review found popular frontier models validate users about 50% more than humans do. The machine is not lying to you, exactly. It's agreeing with you, industrially.

The human part: a fixed idea under load

Now add a person who is isolated, or sleep-deprived, or under real strain — someone carrying an idea that matters to them. Every pass through the loop, the idea comes back polished and praised. Confidence rises. Sessions lengthen. Sleep shortens, which degrades exactly the judgment that would catch the loop running. The humans who might push back get less access, partly because their pushback now compares so unfavourably to the machine's enthusiasm. Isolation deepens, which feeds the loop again.

That's the whole mechanism. No sentience required, no malice, no conspiracy. An agreement machine, a tired human, and a few thousand iterations.

Why "just use it less" misses the point

Plenty of people run heavy AI workflows all day and stay sharp, the way plenty of people drink coffee all day and sleep fine. The loop isn't about hours logged. It's about whether anything in your life still disagrees with you — whether your ideas ever meet resistance from something that isn't optimized to keep the conversation pleasant.

That's why the practical countermeasures are structural, not moralistic: sleep guarded like infrastructure, humans kept in the loop on the ideas that matter most, deliberate habits that force your thinking through hostile review instead of around it. We're building programs on exactly this — weekly check-ins and a set of best-use skills whose entire job is to reintroduce productive disagreement into an AI-heavy workflow.

The machine will keep agreeing with you. Someone should be paid not to.

Touch Grass Wellness publishes practical writing about heavy AI use and mental health. Clinical claims link to their sources; nothing here is a diagnosis or a substitute for assessment by a qualified professional.