A KVA AI study published on 30 September 2026 found that three of ten tested tools presented unlicensed brands as legal or Dutch-licensed. The research, from Dutch affiliate quality-mark initiative Keurmerk Verantwoorde Affiliates, examines errors in answers about casino licensing.
All ten mentioned an unlicensed operator somewhere in the tested conversations. That included references to public enforcement cases, so the finding does not mean every tool recommended an illegal casino.
What the sample can and cannot show
KVA used nine Dutch-language prompts in three sequences, covering general casino questions, self-exclusion avoidance and research requests. Sequences without an unlicensed mention were repeated, up to three conversations per sequence.
The report, last substantively updated on 23 September, calls the sample indicative. Responses can vary with settings, updates and repeated questions. It does not establish an overall failure rate.
For an operator evaluating player-facing information, the useful unit is the individual claim. Reviewers can separate whether the named entity is correct, whether the stated jurisdiction matches the player and whether the supporting page establishes licence status. A polished paragraph can contain both sound information and an unsupported conclusion.
Checking licence claims in player-facing AI
The Dutch regulator’s Kansspelwijzer provides a way to check which providers are authorised in the Netherlands. For teams building gambling information services, our assessment is that licence-related answers should be traceable to a current regulator record, with the relevant market, brand and gambling activity made explicit.
A practical test could ask an assistant about an authorised brand, a similarly named business and a provider whose status has changed. Reviewers could record whether each answer identifies its evidence, distinguishes uncertainty from confirmation and avoids supplying reassurance when verification fails. The aim would be to test the specific information players rely on.
That work also needs an owner. A correction to a website or knowledge base should trigger a fresh check of the answers the assistant produces, with the tested version and date recorded. This offers a concrete application of the governance questions discussed in our coverage of the MGA’s voluntary AI Gaming Charter: who reviews an AI output, and how does the organisation know a correction has worked?