Sorting a game catalogue, answering a support question and flagging an account for harm risk are different automated decisions. An alert's usefulness depends on its errors, the action it triggers and what happens afterward.
Catalogue choices and account decisions
A product can remember a language, rank titles, route questions or flag unusual activity. To describe a feature, identify the data it receives, the decision it makes and how a person can challenge or override that decision. 'AI-powered experience' supplies none of that information.
For operators under the Online Casino Gambling Act, New Zealand’s 2026 standards prohibit adaptive or compensated random outcomes. Catalogue ranking can change which game appears first; it must not silently change that game's probability distribution for a player. Rules, odds and theoretical RTP must remain accessible.
A studio may supply content tags while the operator holds customer data and runs the ranking model. A platform can detect transaction risk without sending each supplier a full player profile. A vendor needs to identify which organisation controls each step.
Data and privacy in New Zealand
New Zealand's Privacy Act 2020 governs collection, storage, use and disclosure of personal information. The Privacy Commissioner explains the need for a lawful collection purpose and safeguards, alongside rights of access and correction. Systems processing gambling behaviour need a stated purpose and should avoid collecting unnecessary identifiers. Location can establish access conditions; deposit history supports limits and transaction records; device signals can help detect fraud. Support routing may use the question text and account state. Reusing those records for recommendations creates a separate privacy question. The privacy statement needs to describe what the system actually does.
The Ministry of Health reports uneven gambling harm across population groups. That supports testing a model on the New Zealand population it will serve, but proves neither bias nor effectiveness in a particular model. An evaluation should report the population, error rates, review process and measured outcome without exposing individual records.
Since 1 May 2026, New Zealand’s Privacy Act includes IPP 3A for personal information collected indirectly, as the Privacy Commissioner explains. If an operator obtains a player-level risk score or browsing history from a supplier rather than from the person, the collector must consider the new notification duty and any applicable exception. A data-flow record should name the source, collector, purpose, recipients, retention period and wording shown to the person. Labelling all supplier data “anonymous” without checking whether it can identify someone would evade the real question.
The Privacy Commissioner distinguishes a supplier processing information solely on an organisation’s behalf from one using it for its own purposes. Outsourcing storage or analysis does not remove the organisation’s responsibility for personal information. If an AI supplier uses customer records to train its own tools, that additional purpose changes the assessment of who holds the information and whether its transfer is a disclosure. The contract and actual data use determine the answer; the presence of a chatbot does not establish that training occurs.
Three systems that should be judged differently
| System | Possible input and action | Evidence that would matter |
|---|---|---|
| Catalogue ranking | Recent views reorder game tiles. | Whether the person can understand or turn off ranking, and whether a risk flag suppresses further game promotion. |
| Support assistant | A question about limits receives a route to the relevant account control. | Correct answer rate, escalation to a person and whether a limit request is completed rather than merely acknowledged. |
| Harm-risk flag | Changes in spend or session length prompt review or a protective action. | False alerts, missed cases, the action taken and behaviour after the intervention. |
DIA’s harm-prevention guidance permits live customer support to be provided and monitored by AI or a person, with real-time availability around the clock. For AI support, it recommends considering when a case should reach personnel trained to identify and assist customers experiencing gambling harm. A bot’s acknowledgement, completion of a limit change and a person’s review of a dispute are separate outcomes to check.
The Privacy Commissioner's impact-assessment guidance offers a way to document collection and sharing risks before deployment, alongside clear notice of AI use. A model test needs a defined period, correct and missed risk alerts, affected groups and the action that followed. Human review should precede a restrictive account action, with a route to correction. Engagement or fewer alerts alone cannot demonstrate reduced gambling harm; that claim needs outcome data from a named deployment.
Risk tools need measurable outcomes
A risk alert needs evaluation for false alarms, missed cases, timing and subsequent support. New Zealand's standards require accessible time, deposit and spend limits, breaks, time-outs, self-exclusion and live customer support. A model's output should be examined alongside the action those controls enable.
A catalogue optimised for clicks or session length may conflict with harm prevention. Its operator needs to state the model's objective and safeguards, and keep limit controls accessible. Even a proprietary model can disclose its inputs, testing, oversight and resulting account actions.
Fewer flagged sessions could reflect an effective intervention, a changed threshold or changed behaviour. A comparison over a defined period is needed to distinguish them. A case study or marketing quote without that comparison cannot establish reduced harm.
Four outcomes behind a risk flag
| Model output | Review finds the defined criterion | Review does not find it |
|---|---|---|
| Flag raised | True positive The flagged case meets the review criterion. | False positive A flag creates an unnecessary review or intervention. |
| No flag | False negative A qualifying case is missed. | True negative No flag, and no criterion found in the review. |
Define the criterion, review window and reference labels before counting these cells. Reviewing only flagged accounts leaves the missed cases unmeasured; the review process can also make mistakes. NIST's AI RMF discusses false-positive and false-negative rates and the need for representative evaluation conditions. A high overall accuracy figure can conceal the error that matters most in a particular use.
The boundary between assistance and persuasion
A support assistant explaining a cooling-off period should give accurate information and reach the actual account control. A recommendation carousel needs scrutiny of what it promotes and whether it interferes with the person's limits. Success for one system cannot be measured by the other's engagement figures.
During the 2026 transition, DIA says online casino advertising is prohibited and has published advice for advertisers and third parties. Product descriptions must not become paid endorsements or operator referral links.
For an online-casino account, DIA guidance requires identity and age verification before activation. A support assistant should direct a person to real limits, self-exclusion or a human adviser; Safer Gambling Aotearoa provides independent 24/7 support on 0800 654 655.