How we work

What goes on the register, how we check it, and how we decide how serious it was.

What counts as an AI incident

We include an event when all three of these are true:

An AI system was involved. Something that generated, predicted, recommended or decided, rather than ordinary software following fixed rules.
It caused harm or a real risk of harm. To people, money, rights, safety, privacy or trust in public information.
It happened in New Zealand or affected people here. Including online incidents aimed at New Zealanders.

Near misses

Some events nearly caused harm but were caught or avoided in time, like a voice-clone scam call that was spotted before any money changed hands. They follow the same criteria as an actual incident, and we rate their severity based on the likely harm.

Near misses are labelled Near miss and are left out of our figures, charts and the default download. To see them, tick Include near misses on the register or the data page.

Where entries come from

News reports, court and tribunal decisions, official publications, organisations’ own statements, and reports from the public through our form.

How we check

A person reads every report and removes anything that could identify someone.
We look for at least two independent sources to confirm it.
A second reviewer checks the entry and its severity rating before it’s marked verified.
When appropriate, we will contact the organisation involved for their perspective.

Rating severity

We rate each incident from 1 to 5, looking at how much harm it caused, how many people it affected, and whether the harm can be undone.

Severity 1: 1. Minor. Limited, with no lasting harm.
Severity 2: 2. Moderate. Harm to one person that can.
Severity 3: 3. Significant. Harm to several people or an organisation.
Severity 4: 4. Serious. Serious financial, safety or rights harm.
Severity 5: 5. Severe. Widespread or severe harm, or harm that cannot be undone.

Verified and under review

Verified entries have passed every criterion above. Under review entries are published so the register stays current, but haven’t been confirmed yet. They can change or be removed depending on what further information comes to light during our review.

Updates and corrections

Entries change as we learn more. Every change is dated on the entry’s page, and significant changes are listed on our corrections page.

Independence

We don’t take money from AI companies, and no funder sees or approves entries before they’re published. Who runs the register