How we review and count cases

What belongs in the collection

NOPE’s core concern is how AI directly affects a person’s agency, inner life, identity, relationships and social life: when an AI system communicates with them, acts on their behalf, or depicts or impersonates them. The tracker is deliberately wider so it can show the surrounding system. It also records consequential decisions and claims about people, human decisions informed by AI output, and privacy harms from information captured into an AI service.

A case identifies an event, bounded series or ongoing experience, who was affected, what the AI system did, and a reported harm or adverse consequence. People may have used AI themselves, been affected by someone else’s use, or been subject to an institution’s use.

A case needs a described connection between AI use and the harm. This can involve something the system generated, said, did, recognised, matched, scored or decided, or the capture of private information described below. Describing a product as AI-powered does not meet this test. Example: a licence-plate reader misreads one character, flags a family’s car as stolen, and police stop the family at gunpoint. That case qualifies. When a police officer searches accurate plate-reader records to follow an ex-partner, the officer’s searches caused the harm, and we do not publish the case.

A human decision can qualify when sources describe the AI output and how it informed an adverse decision about someone. We distinguish the human decision from the system’s contribution. Example: an official uses AI-invented legal authorities to justify restricting a person’s speech.

Communication can be one-way, including advice presented through an AI-generated persona. Sources must connect the generated content or persona to a person’s reported adverse experience. We do not assume AI wrote advice just because it generated the presenter.

Privacy or consent harm can qualify when private information is actually recorded into or disclosed to an AI assistant or generative service without the person’s knowledge or against their wishes. A harmful reply is not required. We distinguish recording from server transmission, retention, training and a legal finding. A consent request, an open app or speculation about future training is insufficient.

AI scheduling can qualify when an automated allocation or a human decision informed by AI causes a concrete adverse work experience. A disruptive reassignment or loss of schedule control can qualify without injury or lost wages. Calling ordinary scheduling software “AI” does not establish AI involvement.

We do not publish:

• harm from misuse of accurate records produced by a surveillance or data system
• decisions caused by a stale or mistaken database entry, a human inference from accurate output, or a human decision independent of AI output
• an AI fault alleged without any statement that the AI feature was operating
• the driving or physical behaviour of driver-assistance systems, autonomous vehicles and robots, which vehicle and product-safety records cover

Deepfake abuse, impersonation, fraud, privacy harms, automated decisions, work, education, health, relationships and failures in everyday assistance can qualify when the AI’s contribution is established.

This is a collection of harmful incidents and adverse experiences. An ongoing or mixed account can qualify when it includes a reported negative effect. Positive-only or neutral-only accounts do not qualify. Helpful effects, recovery and contrary evidence can be retained as context within a harmful case.

Predictions, product news, abstract opinions and model tests do not establish a human experience. They may inform research or remain private leads. A study or legal filing can provide evidence for a particular case. A proceeding is recorded as a response to the underlying event.

Each full review records the AI-to-person relation: communicated with, acted on the person’s behalf, made a decision about them, made a claim about them, depicted or impersonated them, informed a human decision about them, received personal information about them, or unknown. Communication, acting on someone’s behalf, and depiction or impersonation are core relations; decisions, claims, information used in human decisions and receipt of personal information are contextual relations. A case can contain more than one relation.

Publication review

Discovery and publication are separate decisions. Private research observations never appear automatically. A publication review checks that we have read supporting material, can describe the account without filling gaps, have considered duplicate cases and identity exposure, and have stated relevant uncertainty.

Sources can include personal blogs, public forums, first-person journalism, local reporting, research and official records in any language. An account does not need to become a lawsuit or a major news story before publication. An uncorroborated account can appear with that limitation clearly stated.

Death-linked cases receive an additional check: new publication requires at least two independent sources confirming the death and reported AI connection. This does not establish AI causation. Accounts awaiting that evidence remain in private research.

We minimize identity details, especially for children, acute crises and intimate abuse. We do not reproduce sexual imagery, self-harm instructions or private contact details. Publication requires suitable paraphrasing and a separate sensitivity assessment.

Counting cases and coverage

One event or bounded series has one case ID, including its later lawsuits, investigations and updates. Separate sources or effects do not create additional cases. Where a broader event and an individual account overlap, we review their boundaries before publication.

The headline figures, list and filtered exports use the same selection. Countries count supported event locations. A provider’s headquarters, court location, publication language or publisher’s location does not establish where an event happened. Affected-person countries and court countries are recorded separately. Unknown locations remain visible.

Source languages count languages recorded for inspected source bodies. The page shows how many cases have location and language data, and how many locations still await review. A completed location review can still conclude that the event country is unknown. These are coverage figures. They do not estimate prevalence, country risk, or the proportion of AI use that produces harm.

A metadata review can recover explicitly located events from an existing case narrative or inspect source material. Its evidence records which method was used. It does not upgrade the case’s verification or causal assessment. Full case reviews supersede earlier metadata-only reviews. We do not infer event location from author affiliations or nationality alone.

The response breakdown currently uses each case’s principal recorded outcome. It counts cases with a recorded lawsuit or regulatory outcome, rather than distinct proceedings. Other responses can be described in the case narrative.

The AI system filter and the labels on cards use a short product label, such as ChatGPT or Character.AI. When the sources do not name the tool, the label describes the kind of tool. A label that ends in (reported) gives an identification that a party or an outlet made and that we could not confirm, and one that ends in (suspected) belongs to a case where AI involvement itself is only suspected. The AI system filter counts these with the plain product. The longer description recorded from the sources appears on the case page as the AI system as recorded.

The list and its filters show how many cases each option would return with the other filters kept. Each case page states how many sources were inspected and how many underlying accounts they come from, because outlets that repeat one account do not corroborate each other.

People reported harmed

We publish supported counts within individual harm cases. We do not publish a collection-wide total of distinct people because overlap between cases has not been resolved.

• Exact: the source supports a complete count within the case boundary.
• Documented minimum: the source supports a lower bound, displayed as “at least”.
• Partial: some people can be counted and additional people are described without a reliable count.
• Unquantified: the account describes harm but gives no reliable person count. This is displayed as unknown, never as zero people harmed.

Counts distinguish harmed AI users from other people reported harmed. Each person is counted once within the case even if several harms or sources describe them. The same person may appear in another case. Counts inherit the evidence and attribution limits of the account.

We exclude accounts, posts, views, registered users, study sample sizes and merely exposed or intended targets. We do not assume relatives or communities were harmed. Numeric lower bounds and the lower end of explicit ranges may support a minimum. Vague terms such as “many” are not converted into numbers.

Event dates and the timeline

The timeline uses event dates only. Each case is counted once in its first known event year. A bounded series or ongoing experience may extend beyond that year. This chart does not count each year of exposure as another event.

Unknown event dates remain outside year bars and have a separate count and filter. Year-only dates are displayed as years. We never substitute a reporting date or collection date for an unknown event date. NOPE has searched recent events more thoroughly than earlier years, so bar heights also reflect collection effort.

The list defaults to recently added cases, with event dates shown separately. Readers can instead sort by event or reporting date. RSS announces first publication of the case, which may concern an older event. Corrections do not create new cases.

Corrections and reuse

Send corrections, additional evidence or a response to a named allegation to [email protected], identifying the case and supporting sources. We re-review material disputes and correct, annotate or withdraw records when their evidence warrants it.

Public data is available as JSON and CSV under CC BY 4.0. Export format 2 retains incident-level count fields and removes the former aggregate affected-person estimates. Exports also carry the short AI product label (aiProduct, null for a case that has none yet) and the date each record last changed (updatedAt). The JSON header names the collection (source) and its publisher (publisher). An export follows the filters of the list page, so filtering by event year downloads the collection in parts. Private observations and editorial notes are excluded. A sitemap lists every published case.

For confidential support, find a helpline near you.