Your brain notices how you use AI. Does it matter?

Using AI for homework vs. for company leaves different marks on your brain

Your brain actually looks different depending on why you open ChatGPT — and not in the direction you might hope.

A team at Beijing Normal University and collaborating institutions scanned the brains of 222 Chinese university students and asked them a simple question: when you use AI tools, what for? The researchers split usage into two types. Functional use means writing essays, summarising papers, debugging code — using AI like a really good reference library. Socio-emotional use means chatting to AI for company, emotional support, or as a partial substitute for human conversation. The results split cleanly. Students who used AI more for tasks had more gray matter volume — think: more brain tissue — in the dorsolateral prefrontal cortex, the region that handles planning and complex thinking. Their hippocampal network, the part tied to memory and learning, was also better connected. Their GPAs were higher. Socio-emotional users showed the opposite: more depression, more social anxiety, and less gray matter in regions tied to reading other people. Think of it like going to the gym. Using AI to help you think harder is like using a rowing machine — the act seems to strengthen the machinery. Using AI as a substitute for human connection may be more like being carried everywhere: the muscles that need the workout don't get it. The catch is significant. This is a snapshot, not a film. We cannot tell whether healthier, more academically engaged brains led people to use AI more productively — or whether the direction of causation runs the other way. The study was conducted entirely in China. And only 6.8% of students were heavy socio-emotional users, which is a thin group from which to draw strong conclusions. This is a signal worth watching, not a verdict.

AI can play a convincing patient but gets the whole crowd completely wrong

An AI language model can play a single psychiatric patient convincingly enough to fool a clinician — and simultaneously misrepresent every vulnerable population in the data.

Imagine you hired an actor to play a therapy patient. They nail it: the hesitations, the backstory, the mood swings feel real. Now ask ten thousand of them to represent the actual spread of mental illness across the US population. Suddenly every single one is playing someone mildly depressed, firmly middle-of-the-road on every demographic spectrum. Nobody is at the extreme end. Nobody is a statistical outlier. The crowd is fiction, even when each individual feels convincing. That is essentially what a researcher named James Rudow found when testing four leading AI language models — GPT-4o-mini, Gemini-3-Flash, DeepSeek-V3, and GLM-4.7 — on a new benchmark called PsychBench. Rudow generated 28,800 synthetic patient profiles and compared them against real US health survey data from NHANES and NESARC-III. At the individual level: perfect. Zero cases violated the clinical rules for how symptoms cluster together. At the population level: a mess. The models compressed real-world variation by 14% to 62%, depending on the model. They overestimated depression severity for most groups by four to six PHQ-9 points — that is a clinically meaningful error, not a rounding issue. And they badly underestimated distress in transgender women, capturing only 8–46% of the documented mental health burden in that group. Worse: 37% of profiles flipped their clinical diagnosis between two separate test runs, even when the scores looked stable. The catch is real: this is a single independent researcher with no institutional peer review reported. The patterns need replication. But the core finding — that AI systems can appear accurate while being systematically wrong in ways that hurt already-marginalized groups — is important enough to take seriously now.

Tracking 11,000 kids for four years: what actually predicts teen drug use

If you could pick one thing to delay when a teenager first tries alcohol, the data from 11,868 kids points to something stubbornly unglamorous: how closely their parents pay attention.

The ABCD Study — the largest long-term brain development study in the United States, tracking nearly 12,000 children starting at age nine or ten — is now old enough to ask a hard question: what actually predicts which kids try substances first, and when? A research team ran four years of this data through a time-varying survival model, meaning they tracked the odds of first use changing month by month as kids' circumstances changed. By the end, 36.5% had tried alcohol, and 39.7% had tried any substance. The predictors fell into two buckets. First, environment. Kids with lower parental monitoring — parents who didn't know where they were or who they were with — were 1.5 to 3 times more likely to initiate substance use. Impulsivity, particularly poor planning and sensation-seeking, and even caffeine use both raised risk. Think of these factors like locks on a cabinet: not unbreakable, but they slow things down. Second, genetics. The study calculated polygenic risk scores — a way of adding up hundreds of tiny genetic nudges to estimate inherited predisposition. The nicotine-related genetic score was the strongest signal: kids in the high-genetic-risk group were nearly three times more likely to initiate any substance use, even after accounting for environmental factors. The catch: this is still observational. You cannot randomly assign parents to be attentive or distracted. Parental monitoring may partly reflect other advantages — income, stability, time. And polygenic scores explain a meaningful share of the picture, but they are not destiny: plenty of high-genetic-risk kids never initiate at all. Honest, not alarming.

All three stories this week arrive at the same uncomfortable place: mental health science is generating better measurements than ever, and the measurements keep humbling our interpretations. Study one shows that how young people use AI correlates with brain structure and mental health — but we cannot yet say which direction causation runs. Study two shows that AI systems can simulate individual psychiatric patients flawlessly while misrepresenting every vulnerable population at the group level. Study three shows that parental attention is among the strongest modifiable predictors of teen substance use — but we cannot prove it, only observe it. Anyone selling you a clean causal story right now — this AI use damaged these brains, this gene caused that addiction — is skipping the hard part. The field is not moving fast. It is moving carefully, and that is the correct pace when the subject is people.

The ABCD cohort continues collecting data through adolescence — the wave covering ages 14–16 will be the first to test whether the parental monitoring effects hold into later teen years, and those results should surface in publications over the next 12–18 months. On the AI-brain study front, a longitudinal design — scanning students before and after sustained periods of functional versus socio-emotional AI use — would be the decisive next step; watch for follow-up from the same Beijing Normal University group. And PsychBench urgently needs formal peer review and independent replication: if those variance-compression findings hold up, the implications for how we train and validate clinical AI tools will be hard to sidestep.

• ABCD Study official site — background on the cohort behind story three
• Wikipedia: PHQ-9 — the depression scale referenced in the PsychBench story
• Wikipedia: Adolescent Brain Cognitive Development Study