AI and the brain: promise, a flaw, and a digital twin

Using AI for homework vs. loneliness does different things to your brain

The same app on your phone does different things to your brain depending on whether you use it to write an essay or to feel less alone.

A team at a Chinese university recruited 222 university students, put them in a brain scanner, and mapped how often they used AI tools — and crucially, why. They split usage into two buckets: functional use (homework, research, writing) and socio-emotional use (chatting, seeking companionship, emotional support). Same tools, different motivations. The results split cleanly down that line. Students who used AI more for functional tasks showed slightly larger volumes in the prefrontal cortex — the part of your brain roughly responsible for planning and decision-making — and scored modestly better on GPA. Students who used AI more for emotional support showed the opposite pattern: reduced volume in regions tied to reading other people's emotions, and higher rates of depression and social anxiety. Think of it like the difference between using your phone as a GPS and using it as your only companion on a solo trip. Same device, but one builds a skill and the other might quietly replace the human practice of building real connection. Why does this matter now? We are all, right now, in the middle of a giant uncontrolled experiment on whether AI companions are fine or not. This is one of the first brain-level snapshots of that experiment. Here is the catch, and it is a big one: this study is a snapshot, not a film. We cannot tell from this data whether lonely people turn to AI more, or whether leaning on AI for emotional needs makes people lonelier. The brain differences are small. All 222 participants were young Chinese university students — a narrow slice of humanity. And the correlations between AI use and GPA, while statistically significant, are modest (r ≈ 0.17). This is a signal worth watching, not a verdict.

AI depression detectors may be cheating — reading the interviewer, not the patient

An AI trained only on what the interviewer said — not a single word from the patient — scored 0.98 out of 1.0 at detecting depression.

I want you to sit with that number for a second. 0.98 out of 1.0. Near-perfect accuracy. Except the model had never seen the patient's words. It had only seen the interviewer's. Researchers tested depression detection AI on three widely-used clinical interview datasets — ANDROIDS, DAIC-WOZ, and E-DAIC. These are the benchmark datasets the whole field uses to claim their models work. The experiment: train one model on the patient's side of the conversation, train another model on only the interviewer's scripted prompts. The interviewer-only model matched or beat the patient-only model on every single dataset, across both AI architectures they tested. Why? Because these interviews are semi-structured, meaning interviewers follow a loose script that subtly shifts depending on how the session is going. Depressed patients tend to give shorter answers, so interviewers repeat prompts, reword questions, or linger longer. The AI didn't learn to detect depression — it learned to detect the interviewer's response to depression. That is a very different thing. The analogy: imagine a fire alarm that reliably triggers not when there is smoke, but when a firefighter has already walked in through the front door. It looks like it works — until you realize it is detecting the response, not the problem. This matters because a large chunk of published research in AI depression detection uses these exact datasets. Papers claiming 80%, 90%, or even higher accuracy may be measuring a flaw in the data collection, not a genuine ability to detect depression in the wild. The researchers here — whose names are not listed in the preprint — call this architecture-agnostic, meaning it is not a specific model's problem. It is baked into the data.

A digital twin of your brain predicted who would respond to treatment

Before touching a single patient with a treatment, researchers ran the experiment inside a simulation of that patient's specific brain — and got the answer mostly right.

This one is for Parkinson's disease today, but the idea travels directly into depression and psychiatric medicine, so stay with me. A team affiliated with Ruijin Hospital in Shanghai first trained a large generative model — think of it as a very sophisticated brain-behaviour simulator — on 5,621 brain scans from 2,707 people across four public research cohorts. Then they fine-tuned it for 106 individual Parkinson's patients, building what they call a personalized virtual brain: a computational model of how that specific person's brain communicates with itself. Here is the key move. They asked the model: simulate what this patient's brain would look like if it were healthy. The gap between that simulated healthy state and the patient's actual state became a predictor — a number — for whether two different treatments would work. For transcranial inhibition, prediction accuracy hit 0.853 on a standard clinical scale. For deep brain stimulation, it reached 0.915. Both well above existing methods. The analogy: a GPS that does not just show you today's road but simulates what the traffic will look like if you take a hypothetical detour — before you actually turn the wheel. The catch is real: 106 patients is a small number, and the prospective validation used only 11 people. This is a proof of concept, not a clinical tool. The model was built specifically for Parkinson's, and applying it to depression or treatment-resistant conditions would require starting the training process over with psychiatric data. But the architecture — pretrain on thousands of brains, personalize for one — is the part worth watching.

Read these three together and something uncomfortable comes into focus. We are building AI systems that can simulate your brain, detect your depression, and predict whether treatment will work — and simultaneously discovering that the data those systems are trained on may be quietly compromised. The interviewer-bias story is not a one-off quirk. It is a reminder that when humans design measurement tools, our behaviour as measurers leaks into the data. The AI chatbot story adds another layer: even how we choose to use AI is reshaping the very brains we are trying to help. The virtual brain paper shows what the upside looks like when the foundations are solid — careful data, explicit uncertainty, small but honest validation cohorts. The field is not moving in one direction. It is moving in several directions at once, some of them toward each other. That tension — between what AI can eventually do and how poorly we currently validate what it claims to do — is the real story of this week.

The interviewer-bias finding needs a response from the teams that built those benchmark datasets — DAIC-WOZ and E-DAIC are maintained by specific research groups, and it would be worth seeing whether they release cleaned or rebalanced versions. On the virtual brain side, the Ruijin Hospital team mentions prospective validation is ongoing; a larger cohort result would be the next meaningful milestone. The AI-and-brain paper is cross-sectional, and I would want to see a longitudinal follow-up — track the same students over a year and see whether usage patterns change brain structure or the other way around.

• Wikipedia: Deep brain stimulation — background on how it works and what it treats
• Wikipedia: Semi-structured interview — what it is and why clinical researchers use it
• Wikipedia: Functional connectivity (neuroscience) — the concept behind virtual brain models