AI Fake Patients Lie, Gut Bacteria Guard, Brain Scans Show

AI-simulated patients look fine up close but lie about the crowd

An AI can roleplay a depressed patient so convincingly it fools a clinician — and yet get the whole population completely wrong.

Imagine you ask a very talented actor to play ten different customers walking into a shop. They nail each individual performance. But when you look at the crowd they've assembled, every single person is roughly the same age, roughly the same income level, roughly the same mood. The extremes — the very young, the very broke, the very desperate — have quietly vanished. That's exactly what researchers building PsychBench found when they fed standardised psychiatric questionnaires to four major AI systems, generating 28,800 synthetic patient profiles. The AIs — including GPT-4o-mini and DeepSeek-V3 — produced people who looked clinically plausible one by one. Zero of those individuals broke the basic rules of a depression diagnosis. But at population level, the distributions were squeezed. DeepSeek-V3 compressed the natural spread of depression scores by 62 percent. The very severe cases, the very mild cases, the complicated edge cases — statistically erased. There's a second problem: one in three simulated cases changed their clinical diagnosis between two identical runs, even though the scores looked broadly consistent. The AI was reliable on the surface, unreliable underneath. Why does this matter? Researchers and companies are already using AI-generated patients to train clinical algorithms, test chatbot therapists, and prototype digital mental health tools. If those synthetic populations don't reflect who actually gets ill — and especially who gets ill severely — then whatever gets trained on them will be quietly miscalibrated before it ever meets a real person. The catch: this paper audits the problem without fixing it. And it only looks at a handful of AI models. Whether the distortions are smaller in domain-specific psychiatric AI tools, nobody yet knows.

Using AI as a tool helps your brain; using it as a companion hurts

How you use AI — not how much — turns out to matter for your mental health and, apparently, for the physical structure of your brain.

A team at a Chinese university scanned the brains of 222 students with high-resolution MRI and also asked them about their AI habits. Not just how often — but why. Were they using AI to get tasks done, like summarising a paper or debugging code? Or were they using it socially and emotionally, to chat, to vent, to feel heard? Think of it like the difference between using a kitchen knife to chop vegetables versus leaning on it for emotional support. One is a tool; the other is asking the wrong thing of the wrong object. Students who used AI functionally — the task-focused kind — had slightly higher grades and, strikingly, larger grey matter volume in a region called the dorsolateral prefrontal cortex, which sits behind your forehead and is heavily involved in planning and decision-making. Their hippocampal networks — the circuits tied to memory and learning — also showed stronger local connectivity. Students who used AI socio-emotionally showed the opposite pattern: worse mental health scores on depression and social anxiety measures, and smaller grey matter in regions linked to social processing. I want to be careful here, because this is a cross-sectional study — meaning everyone was measured once, at the same time. We cannot say the AI use caused the brain differences. It's just as plausible that people who already had stronger prefrontal networks were more drawn to task-focused use. The sample was also 222 students in one Chinese city, not a global population. But the finding that the two types of use are statistically unrelated to each other — it's not that heavy users do both — makes this worth watching closely.

Gut bacteria may set how close to the edge your brain sits

Your gut bacteria might not cause depression or dementia — but they may decide how hard life has to push before you tip into them.

Here is a useful way to think about it. Imagine your brain's resilience as the insulation in your house walls. The insulation doesn't cause winter. It doesn't determine whether a cold front arrives. But it absolutely determines how cold the inside gets when one does. The researchers behind this review paper — published in the journal Journal of Neuroinflammation — propose something similar for gut bacteria and brain disease. They call it a vulnerability-threshold framework. The idea is that the trillions of microbes living in your gut are not primary triggers of neurological illness. They are regulators of your threshold — how much neurological stress it takes before clinical symptoms appear. The mechanism involves chronic low-grade inflammation. When your gut microbial community is disrupted — through diet, ageing, antibiotics, stress — the barrier between your gut and your bloodstream becomes slightly leakier. Inflammatory molecules that wouldn't normally cross into your brain start doing so more easily. Over time, this affects the brain's resident immune cells, called microglia, and the integrity of the blood-brain barrier itself. The paper makes two specific predictions that are testable. First: restoring a healthier microbiome in someone with high genetic risk for dementia should delay disease progression, not prevent it. Second: the severity of gut-derived inflammation should predict how fast a disease progresses, not which disease you get. The catch is significant: this is a narrative review, not a clinical trial. No one has yet run the experiments needed to confirm these predictions in humans. What we have is a coherent theoretical framework — which is genuinely useful, but not yet evidence.

What do these three stories share? They're all pointing at the same uncomfortable fact: the things we assumed were stable — who gets mentally ill, how AI affects us, what protects the brain — turn out to be more conditional than we thought. The gut bacteria story says your brain's resilience isn't fixed; it's being quietly tuned by your microbiome across your whole life. The AI habits story says that the same technology can be either a cognitive workout or a social crutch, depending on how you reach for it. And the PsychBench story says that our best AI tools for simulating patients are already compressing away the very people who matter most — the severe, the marginalised, the ones who don't fit the statistical middle. Taken together, that's a picture of mental health research that is getting better at measuring the conditions underneath illness — and simultaneously more aware of how badly our models can mislead us if we aren't careful.

The PsychBench findings raise an urgent question for regulators: should AI-generated patient data used to train clinical tools require epidemiological audits before deployment? No major regulatory body has addressed this yet, and it's worth watching whether that changes. On the gut-brain side, a handful of ongoing clinical trials are testing whether probiotic interventions can slow cognitive decline in high-risk older adults — results from the PROMAGE trial in Europe are expected later this year. And on AI use patterns: the Chinese study is cross-sectional, which means a longitudinal follow-up — tracking the same students over two or three years — would be genuinely important. I'd want to see that done.

• Wikipedia: Gut-brain axis — a solid background on how the gut and brain communicate
• Nature News: The microbiome's role in mental health — an overview
• Scientific American: Can AI really help with mental health care?