Methodology

How cases reach this tracker, how they are classified, and — importantly — what the data cannot tell you.

What this tracks

AI Psychosis Watch documents reported instances of psychological harm associated with conversational AI systems: delusional reinforcement, distorted reality-testing, identity confusion, paranoia, and romantic or dependent attachment to chatbots — together with the clinical and research literature examining those phenomena.

A "case" here is a documented report, not a verified clinical diagnosis. Most entries are journalism or peer-reviewed literature. The tracker records that something was reported by a credible source; it does not independently verify the clinical facts of any individual account.

This tracker documents harm because harm is what is verifiable. Positive or neutral outcomes do not appear in court filings, case reports or news investigations. Nothing here should be read as a claim that harm outweighs benefit across AI use generally — it is a record of what can be systematically tracked.

Inclusion criteria

An item is included when it meets both conditions:

It concerns an AI system capable of conversation — a chatbot, assistant, or companion app.
It concerns psychological harm to a person, or research into that harm.

Requiring both is deliberate. Material excluded on this basis includes:

General AI industry, policy or business news, however prominent.
Research using machine learning to study or treat mental illness, where the AI is a tool rather than a possible cause of harm.
AI safety and alignment work with no human psychological dimension.
Misuse of AI with no psychological harm — fabricated legal filings, deepfakes, plagiarism.

Items reviewed and rejected are recorded in excluded.json so that a rejection persists and the same item is not re-added on a later run.

Sources

Source	Type	What it contributes
PubMed	Academic	Indexed biomedical and psychiatric literature
Europe PMC	Academic	Wider journal coverage plus preprints (medRxiv, PsyArXiv)
OpenAlex	Academic	Broad scholarly index across disciplines
Semantic Scholar	Academic	Cross-disciplinary coverage
arXiv	Preprint	Computer science and HCI work, often months ahead of publication
Google News search	Media	Query-targeted reporting across outlets
Publisher feeds	Media	Guardian, Futurism, PsyPost, WIRED, MIT Technology Review, Ars Technica, 404 Media, TechCrunch

All sources are queried weekly. A source that fails is logged and skipped; if every source fails, the run refuses to write rather than publish an empty tracker.

Classification

Severity

Level	Meaning
Critical	A death occurred — suicide, homicide, or fatal violence
High	Hospitalisation, involuntary commitment, arrest, or litigation
Medium	Documented psychological disturbance without those outcomes
Low	Commentary, analysis, or research without an individual incident

Academic entries are capped at Medium. A study of suicide is literature, not a death, and should not inflate the critical count.

How classification happens

Candidates are matched by keyword, then flagged needs_review and revisited in a weekly review pass that corrects categories and severity and removes false positives. Keyword matching alone is not sufficient for this material, and the tracker does not pretend otherwise.

The companion watchlist

The Companions tab is a register of the companion and AI mental-health app space, monitored for proliferation. It is deliberately not a list of implicated products. An app appears there because it exists and is worth watching — the premise being that a rapidly growing category of emotionally engaging chatbots is a risk worth tracking before harm is documented, not after.

Because of that, entries are held in companions.json, separately from the case data, and are excluded from every case statistic on the site. They were previously stored as cases, which put app launch dates into the harm trend chart — several early points on that chart were entirely product releases rather than reported harm — and counted eight product listings among the medium-severity cases. Both are now fixed.

Each entry shows how many cases in this tracker name that product. Zero is a finding, not an omission. Most entries have no cases recorded against them, which is the expected state for a watchlist.

The four groupings — structured therapy apps, empathic assistants, mood trackers, open companion apps — describe how a product presents itself and whether it claims clinical oversight. They are descriptive, drawn from each vendor's own material, and are not a safety rating or a risk score. Lifecycle status is shown only where it has been verified; where we have not checked, no status is displayed rather than an assumed one.

Limitations

The trend chart measures coverage, not incidence. A rise reflects more reporting being found, which is a function of media attention, research output, and which sources this tracker queries. It is not a measurement of how often AI-associated psychological harm occurs in the population. Adding a source produces a step change in the series that has nothing to do with the underlying phenomenon.

Language and geography. Sources are English-language and skew heavily toward the US, UK and Ireland. Harm reported in other languages is largely invisible here.
Automated precision. Roughly one in six newly scraped items is off-topic before review. Items on the site have passed review, but the pipeline is not self-sufficient.
Reporting bias. Severe and unusual cases attract coverage. Ordinary distress does not.
Dates. The date is publication date, not the date the harm occurred, which is frequently much earlier.
Duplication. Syndicated stories are de-duplicated by URL, so the same event carried by several outlets may appear more than once.
No causal claim. Inclusion records an association reported by a source. It does not assert that an AI system caused an outcome.
The watchlist is not exhaustive. It covers a fraction of a fast-moving category and, until recently, had not been updated since 2023. Absence from it means only that it has not been added.

Data and reuse

The full dataset is available as data.json, updated weekly, with an RSS feed of recent additions. The pipeline that produces it is open source at github.com/notOccupanther/ai-psychosis-tracker, including the exact inclusion vocabulary and its measured precision and recall.

Corrections are welcome and taken seriously. If an entry is wrong, miscategorised, or should not be listed, please write to contact@aipsychosis.watch.

Citation

Cite the dataset as: AI Psychosis Watch. aipsychosis.watch. Accessed [date]. Machine-readable citation metadata is in CITATION.cff. Because the dataset changes weekly, please record the access date and, where possible, the generated_at timestamp from data.json.