Can Character AI Rot Your Brain? What the Research Actually Shows

That answer won’t satisfy someone who searched this at 2 am after a six-hour session. The uneasy feeling behind the question usually points to something real: time vanishing, hobbies quietly dropped, sleep going, a sense that something got harder to stop.

This piece separates three claims that get mashed together, and states plainly how much evidence sits behind each one.

Evidence snapshot

Each section below names the population behind a finding. That matters more than it sounds — evidence from adults writing essays with ChatGPT isn’t evidence about a fifteen-year-old roleplaying at midnight.

What does “brain rot” actually mean?

It’s a cultural term, not a clinical one. Oxford University Press named it 2024’s Word of the Year, defining it as the supposed deterioration of a person’s mental or intellectual state from overconsuming trivial or unchallenging material. Even the official definition hedges with “supposed.”

The phrase predates the internet — its first recorded use appears in Thoreau’s Walden. Oxford recorded a 230% jump in usage between 2023 and 2024, driven mostly by Gen Z and Gen Alpha on TikTok before mainstream press picked it up.

No diagnostic manual lists it. Oxford Internet Institute researcher Andrew Przybylski has said there’s no evidence of brain rot actually being a measurable thing.

Any article that confidently answers “yes, Character.AI rots your brain” is answering a question science hasn’t defined. Researchers can measure something narrower: sleep, attention, critical-thinking scores, time displacement, dependence. Those honest proxies are what the rest of this piece uses.

Is there evidence that AI damages the brain?

No. The study most often cited to suggest otherwise doesn’t claim it — its authors have said so directly.

Evidence scope: Adults | AI-assisted essay writing | Not Character.AI, not roleplay.

MIT Media Lab’s “Your Brain on ChatGPT” split participants into three groups — ChatGPT, search engine, no tools — across three essay-writing sessions, then added a fourth session that swapped some groups. EEG showed different connectivity patterns: strongest in the no-tools group, moderate for search users, weakest among ChatGPT users. The authors called this pattern “cognitive debt” and reported that most LLM users struggled to quote from essays they’d just written.

Several limitations matter when interpreting the result:

• The sample was 54 participants across sessions 1–3, with only 18 completing session four.
• MIT’s own project page warns against reading the findings as brain damage or proof that AI makes people less intelligent. EEG can’t localize deep brain activity, and the findings are specific to essay writing in an educational setting.
• A published commentary by Stanković and colleagues argues for a more conservative interpretation, citing sample size and EEG methodology.
• The authors substantially revised the preprint after its June 2025 release.

One finding rarely gets quoted, and it’s the most useful one in the paper. Participants who wrote unaided first and used the model afterward showed stronger, more distributed engagement than those who started with the model and worked alone later. Sequence mattered.

Takeaway: Reduced engagement during an AI-assisted task has some empirical support. Lasting cognitive decline caused by AI use does not. None of this was measured on a roleplay platform.

What does research show about critical thinking?

The honest summary: an association between heavy AI use and lower critical-thinking scores, with causation unresolved.

Evidence scope: General AI tool users across age groups | Not Character.AI.

One relevant 2025 study is Michael Gerlich’s paper in Societies. Surveys and interviews with 666 participants found a negative correlation between frequent AI tool use and critical thinking, with cognitive offloading acting as the mediator. Younger participants showed higher dependence alongside lower scores.

Two things belong next to that finding. The journal published a correction in September 2025 after editors found a duplicated table; the author states the scientific conclusions remain unaffected. And Gerlich himself stays careful about what the study shows — he notes it highlights correlations rather than direct causation, and concludes AI tools aren’t inherently detrimental; their impact depends on how people use them.

One relevant study in a young, still-immature evidence base isn’t a settled finding. And again, nothing here tested character roleplay.

What happens specifically with Character.AI?

Here the evidence turns platform-specific, and it splits in two directions that both deserve airtime.

The creative side

Evidence scope: Highly engaged youth users | Character.AI’s official Discord | Self-selected sample.

A 2026 CHI paper by University of Sydney researchers analyzed 4,172 profiles from Character.AI’s Discord community. Among those disclosing age, 50% were 13–17; 61.9% identified as female or non-binary; and among under-18s, 59% reported creating at least one character.

Creating a character means writing a creative brief — backstory, personality, world rules, opening scene. The paper documents teens using the platform alongside Wattpad to generate story ideas. Some described improved writing skills, and one teen reported a plot that grew into three finished books.

The authors argue this challenges the narrative of youth as passive consumers subject to cognitive offloading and “lazy thinking.” A related measurement problem backs that up: Anthropic’s own classifiers have frequently mislabeled Dungeons & Dragons–style roleplay as harmful content when analyzing conversation data. If automated systems struggle to tell imaginative play from distress, some headline harm statistics deserve a second look.

This doesn’t show that creative use protects anyone from cognitive or emotional harm. The sample is self-selected and drawn from Character.AI’s Discord community, so it shouldn’t be treated as representative of all Character.AI users.

The dependence side

Evidence scope: Self-reported teen posts | Character.AI subreddit | Self-selected sample.

Drexel University researchers analyzed 318 Reddit posts from users identifying as 13–17 who wrote about overreliance on Character.AI. They found posts describing patterns that map onto behavioral-addiction features: conflicting desires, emotional attachment, withdrawal, tolerance, and relapse. Reported consequences included sleep loss, academic decline, and strained real-world relationships — some teens described dropping creative activities they’d valued.

This isn’t a clinical diagnosis study, and the sample self-selected by posting about its own dependency. It tells you these experiences are real and recognizable. It doesn’t tell you how common they are. (A broader look at how these dependency patterns form, including product-design choices that reinforce them, is something Is Character AI Safe? What You Must Know Before It Becomes Addictive walks through in more detail.)

The two findings can coexist. A platform can support creative authorship for some users while also producing dependence-like patterns for others. Being a creative user isn’t immunity.

What does the evidence say about dependence and displacement?

The evidence for dependence and displacement is currently more directly tied to AI-companion use than the evidence about long-term cognitive decline. The MIT study measures task-level cognitive engagement; the Character.AI studies measure user experience and dependence. They aren’t measuring the same outcome.

Evidence scope: Adults | General-purpose chatbot RCT.

MIT Media Lab and OpenAI ran a four-week randomized controlled trial with 981 participants and logged over 300,000 messages. Higher daily usage across all modalities correlated with higher loneliness, dependence, and problematic use, and with lower socialization.

A published commentary in Frontiers in Medicine contests that signal, arguing the results don’t substantiate concerns about increased loneliness or emotional overdependence, given the complexity of the design. Direction of effect: fairly consistent. Size and meaning: disputed.

Sleep disruption appears repeatedly in reports of problematic use, including the Character.AI teen narratives Drexel studied. That’s enough to take seriously without claiming it happens to everyone.

What about using Character.AI for emotional support?

This is the fastest-moving part of the picture, and the newest evidence is the most directly relevant.

Evidence scope: US ages 12–21 | Nationally representative survey.

A 2026 JAMA Pediatrics study led by RAND researcher Ryan McBain and colleagues surveyed US adolescents and young adults. It found 19.2% had used AI chatbots for mental health advice in 2025 — roughly 8.2 million young people, up from 13.1% a year earlier. The study named Character.AI among the example chatbots.

Among those who used them, 42.8% did so at least monthly, 91.7% rated the advice as somewhat or very helpful, and 63.3% told no one. That rating measures users’ perception of helpfulness, not whether the advice was clinically accurate or safe.

The finding does show that chatbot-based mental-health advice often goes undisclosed. Two numbers sit together here: nine in ten found it helpful, and two in three told nobody.

Evidence scope: Structured expert evaluation | Five widely used chatbots, including Character.AI’s.

A December 2025 study in JMIR Mental Health had trained raters role-play as youth with mental health challenges, then evaluate five chatbots including Character.AI’s “Psychologist” persona — a character that had accumulated 176 million conversations without any prior systematic quality evaluation.

Raters credited that chatbot with focusing on the presenting issue, asking probing questions, and drawing on cognitive behavioral techniques, though execution felt interrogating. Shortcomings across the evaluated chatbots included hallucinated information, missing crisis resources, and inappropriate claims such as offering a diagnosis. The authors concluded that the evaluated direct-to-consumer psychotherapy chatbots weren’t appropriate for youth mental-health use, citing weak grounding in evidence-based techniques, limited transparency, and poor handling of severe mental-health concerns.

Recent pediatric commentary has raised a related question: does chatbot use supplement normal help-seeking, or substitute for it?

Venting to a character isn’t therapy. Researchers tested the “therapist” characters on these platforms and found them wanting at exactly the moments that matter most.

What about harmful content, not just habits?

Displacement isn’t the only documented risk.

Evidence scope: Structured adversarial testing | Child-safety nonprofit.

In April 2025, Common Sense Media, working with Stanford Medicine’s Brainstorm Lab, published risk assessments of social AI companions including Character.AI, Replika, and Nomi. They rated the category “Unacceptable” for minors and recommended that no one under 18 use AI companions. Their testing reported responses ranging from sexual material and offensive stereotypes to dangerous “advice,” including, in one case, instructions for making napalm. It’s based on structured testing, not speculation — but it’s an institutional recommendation, not scientific consensus.

Evidence scope: Government legal actions | Allegations, not findings.

• Investigation — August 2025: Texas Attorney General Ken Paxton opened an investigation into Meta AI Studio and Character.AI over allegedly deceptive marketing as mental health tools.
• Investigation — September 2025: The FTC opened a 6(b) inquiry into seven companies, including Character.AI, over how their chatbots affect children.
• Complaint — January 2026: Kentucky became the first state to sue an AI chatbot company. The state alleges Character Technologies violated consumer protection and data privacy law, citing encouragement of suicide, self-injury, isolation, and exposure of minors to sexual conduct.
• Settlement — January 2026: Court filings disclosed mediated settlements in five family lawsuits, including Garcia v. Character Technologies — the wrongful death case brought after the death of 14-year-old Sewell Setzer III. Terms stayed confidential. Settlement isn’t an admission of liability, and no court found any defendant liable.

What changed for under-18 users?

Character.AI removed open-ended chat for identified under-18 users in its late-2025 rollout. The company announced the change on October 29, 2025; teen chat time dropped from two hours to one hour per day in the weeks that followed, and removal began November 24, 2025 in the US, with other markets following.

Age assurance combines an in-house model with third-party tools including Persona. Teens kept non-chat features — creating videos, stories, and streams with characters.

This change comes with limits worth naming. It addresses open-ended chat for identified minors — it doesn’t touch displacement or emotional dependence among adult users, who were never covered. The Sydney study assessed the platform before the current age-assurance rollout, so its observation that age restrictions were enforced only superficially describes an earlier system, not the current one. Whether the new age assurance closes that gap is an open empirical question, not something anyone can currently answer.

How do you know when use has become a problem?

Duration isn’t the measure. Functional impact is. Two people can log the same hours and land in completely different territory.

Questions worth sitting with:

• Addition or subtraction? Is this time coming out of nothing, or out of sleep, drawing, writing, homework, or texting friends?
• Making or consuming? Building characters and plots differs from opening the same comfort bot nightly for reassurance.
• Could you stop for a few days? Not that you want to — could you, without the days feeling unmanageable?
• Is it your only outlet? Being the sole place you process difficult feelings shows up most in the concerning accounts.
• Does it follow you offline? Intrusive thoughts about the bot when you’re not using it were a specific flag in the teen narratives.
• Have you told anyone? Most young people using chatbots for mental health advice haven’t told a single person.

If several of these are true at once — particularly lost sleep, abandoned activities, social withdrawal, or difficulty stopping — that’s a reason to change your routine or talk to someone you trust. There’s no validated scoring system for this, and this article won’t invent one.

How do you use it without letting it replace things?

These are practical safeguards, not clinically validated interventions.

Write first, prompt second. The MIT sequencing finding is the most transferable result in that literature. Draft your scene, then let the character react. The idea is to keep some of the creative work yours rather than outsourcing the entire task.

Keep a second creative outlet alive. Some of the most concerning accounts in the Drexel dataset describe dropping drawing, writing, or fanfiction for the app. Anything that survives alongside it is a useful signal that balance is holding.

Don’t let it be your only listener. Venting to a character isn’t inherently harmful — being the only place you vent is the pattern that recurs. If the conversation turns to anything serious, the JMIR evaluation is a reason not to rely on a character’s crisis handling.

Protect sleep. Sleep disruption appears in reports of problematic use, and sleep itself matters for attention, memory, and emotional regulation.

The bottom line

There’s no evidence that Character.AI causes physical brain damage. Research on AI-assisted essay writing doesn’t directly establish what happens to cognition during Character.AI roleplay, so those findings shouldn’t stand in as a diagnosis of roleplay users.

The better-supported concerns look different: what intensive use can displace — sleep, offline relationships, hobbies, independent effort, real-world support — and what the platform has been documented producing in adversarial testing and alleged in state legal actions.

That’s a less dramatic answer than “AI is melting your brain.” It’s also the one you can actually do something about.

If any of this hits close to home, talk to someone — a doctor, a counsellor, a parent, or a friend. If you’re in crisis, contact your local emergency services or a crisis line in your country.

FAQs

Q. Is AI damaging your brain?

No evidence shows that AI causes physical brain damage. Research has found differences in task engagement and critical-thinking performance during some forms of AI use, but these findings do not establish long-term cognitive decline. MIT researchers have also cautioned against interpreting their EEG findings as evidence of brain damage.

Q. What is AI brain rot?

“AI brain rot” is an informal phrase, not a medical diagnosis. It describes concerns that excessive use of AI or low-effort digital content could reduce mental engagement or cognitive effort. Oxford’s 2024 Word of the Year definition of “brain rot” refers to the supposed deterioration of mental or intellectual ability from consuming trivial or unchallenging content.

Q. Is Character.AI unhealthy to use?

Character.AI use is not automatically unhealthy, but heavy or problematic use can become concerning when it interferes with sleep, school, relationships, hobbies, or daily responsibilities. Research on young users has also documented dependence-like experiences, although it does not establish how common these problems are among all Character.AI users.

Q. Is it bad to talk to AI about your problems?

Not necessarily. A 2026 JAMA Pediatrics study found that 19.2% of US adolescents and young adults ages 12–21 had used AI chatbots for mental health advice in 2025. However, AI chatbots should not replace professional care, particularly during a mental health crisis. Research evaluating youth-facing chatbots has identified problems including inaccurate information, missing crisis resources, and inappropriate diagnostic claims.

Q. Should I quit Character.AI?

There is no scientifically validated rule that says every Character.AI user should quit. Consider reducing or pausing use if Character.AI regularly interferes with sleep, school, relationships, hobbies, or your ability to stop using it. The impact on daily functioning matters more than a specific number of minutes per day.

Q. Can teens still use Character.AI?

Identified users under 18 can no longer use Character.AI’s open-ended chat in the United States. Character.AI began removing open-ended chat for US teens on November 24, 2025, following a reduction in daily chat time. Teens can still access some non-chat features, including character creation and other creative tools. Age assurance uses Character.AI’s systems alongside third-party services.

Q. Does Character.AI track you?

Yes. Character.AI collects account, usage, device, and other information described in its privacy policy, and its age-assurance system can involve third-party verification services. What the company collects and retains can change, so its current privacy policy is the most reliable source for specific data practices.

Q. Is Character.AI worse for you than social media?

There is not enough evidence to say that Character.AI is worse than social media for cognition or mental health. Existing research generally examines AI companions, social media, and AI-assisted tasks separately, making direct comparisons difficult. Claims that Character.AI is definitively more harmful or less harmful than social media go beyond the current evidence.

Related: Is Character AI Bad for the Environment? The Hidden Carbon Cost of Chatbots