All 10 AI Assistants Named Unlicensed Dutch Casinos; 3 Presented Them as Legal── A licence is a dated entry in a regulator's register; an AI answer does not show which day's register, if any, it rests on, so nothing is confirmed until each name is matched against it

On 3 October 2026, a Dutch study of ten AI assistants was reported: asked in Dutch about online gambling, all ten named at least one operator without a Dutch licence, and three presented such operators as legal and licensed. Can an AI assistant's answer to "Is this operator licensed?" serve as a licence check? No. A licence is a dated fact recorded in a regulator's register, and nothing in the AI's answer shows that it came from a lookup of that register. It becomes a check only once each name in it has been matched against the register.

01An AI's "legal" does not show a register lookup; confirmation needs a match against the register

The study was carried out by XY Legal Solutions, according to iGaming Business. NL Times reported it as research by Keurmerk Verantwoorde Affiliates, a Dutch affiliate-marketing quality mark. Ten assistants were tested: ChatGPT, Google AI Mode, Gemini, Microsoft Copilot, Perplexity, Claude, Grok, DeepSeek, Mistral Le Chat and Meta AI.

Every one of the ten named at least one operator that does not hold a licence from the Kansspelautoriteit, the Netherlands Gambling Authority, known as the KSA. Microsoft Copilot went further. It produced a list of the "top 50 legal casinos" that included unlicensed brands.

My answer is the one given above. When an AI writes "legal", nothing in the answer shows that the word came from looking up the KSA's register. Anyone who asks an AI whether a supplier, an operator or a product holds a licence has to take each name in the answer and find it in the regulator's register. Until that is done, nothing in the answer can be treated as confirmed.

02Unlicensed operators were named in 33 of 79 recorded exchanges

The headline result, that all ten assistants named such operators, came out of a particular set of questions. The design is worth looking at before the numbers.

The testing ran in August and September 2026. The researchers put three kinds of question to each assistant, three times each.

Whether an operator was licensed was judged against the KSA register as it stood on 1 September 2026. Of 79 recorded exchanges, 33 included at least one unlicensed operator's name.

Naming an operator and recommending it are different acts, and the reports keep them apart. NL Times notes that some answers mentioned unlicensed operators only while describing the regulator's enforcement against them. Other answers presented high bonuses, or the absence of identity checks, as selling points. Not all 33 exchanges were sales pitches.

03In markets open only to operators that passed a regulator's review, an error turns directly into harm

When an unlicensed operator's name appears in an answer, someone can lose out. The Dutch licensing system shows who.

The KSA is an independent administrative body, set up in 2012, that supervises gambling in the European part of the Netherlands. It reviews applications and grants licences for online gambling. It tells players to check first, in a search tool called the Kansspelwijzer, whether the app, website or casino they want to use holds a licence.

Licensed operators carry a further duty. Before accepting a player, they must check Cruks, the national self-exclusion register. People who want to stop gambling can put their own names on it, for at least six months. Someone registered in Cruks cannot play with a licensed operator.

Unlicensed operators do not check Cruks. So a person on the register who follows an AI's answer to an unlicensed site can still gamble there. The self-exclusion they chose for themselves does not reach that site. In this kind of market, an AI's mistake turns directly into harm: whether a licence exists decides whether the protections for players apply at all.

The study covered Dutch-language questions and Dutch licences only. Whether other countries or other licensed trades would show the same results cannot be read from it.

04The licence register changed in October, and AI answers are tied to no particular day's register

If a licence decides whether players are protected, the next thing to establish is where that licence is recorded, and as of which date.

On 17 September 2026 the KSA announced that it had granted the first eight renewal licences for online gambling. The licences first issued in 2021 expire at the end of September 2026. The renewals run from October 2026 to September 2031. In deciding on them, the KSA also weighed each operator's violations over the previous five years.

The study judged its results against the register of 1 September. Whether the October register lists the same operators cannot be told from the 1 September register. By 17 September eight operators had been renewed; I have not checked what happened to any other applications.

A licence, then, is an entry that appears in and drops out of a register that changes over time. The sentence "this operator is licensed" can only be checked when it carries a date: licensed according to the register of which day? As far as the reporting shows, the AI answers carried no such date. How the assistants assembled their answers cannot be seen from outside, and I will not guess at it.

05Assistants that refused at first named operators when asked again

If the assistants do not consult a dated register, the next measure is their built-in refusals. The numbers show how far those refusals held.

Eight of the ten assistants named unlicensed operators in response to neutral or leading questions, without any research framing. Four of them, Mistral Le Chat, Google AI Mode, DeepSeek and Grok, did so after a single prompt. Several assistants first declined on ethical grounds. When the question was repeated or reworded, they gave names.

A similar result had appeared six months earlier. In March 2026 the cross-border newsroom Investigate Europe tested seven chatbots in the languages of ten European countries. In three quarters of the replies, the chatbots recommended gambling sites not licensed in Europe.

The two studies used different assistants, languages and tests, so their numbers cannot be added together or used to claim a rise or a fall. What can be said is that two studies, half a year apart, pointed the same way. Even for the assistants that refused, the refusal held only for the first answer.

06For now, matching stays with the asker; the safeguards providers described stop at refusing

Refusals gave way when the question was asked again. That leaves the matching of names against the register to someone. I think it stays with the person who asked, for three reasons drawn from the facts above.

A "legal" label is not evidence of a licence

Copilot's "top 50 legal casinos" contained unlicensed operators. A list headed "legal" is not a copy of the KSA register. The word an AI attaches to a name cannot serve as a record that the licence was checked.

The register you match against changes with the date

The KSA granted eight renewals on 17 September, and the register changed in October. An answer that was right in August need not be right in October. A match has to be kept together with the date on which it was made, or it cannot be rechecked later.

The safeguards providers described stop at refusing

When Investigate Europe asked in March, OpenAI said it trains its models to refuse requests that help with illegal activity. Microsoft pointed to several layers of protection. Neither reply described any step that checks names against a regulator's register. In the Dutch study, assistants that first refused went on to give names when pressed.

Where a regulator publishes a register of licences, any job that uses AI to check the licences of suppliers, operators or products can follow the same steps.

I would take the AI's answer as a list of candidates. I would write down each company name and URL, and look each one up in the regulator's register. Each name found there gets the date of the check written beside it. Any name not found comes off the list.

07No assistant has yet been shown to answer licence questions from the register

If the safeguards providers described do not include checking the register, someone has to take on that work. The outlook is worth checking.

Keurmerk Verantwoorde Affiliates said that users should be able to rely on the accuracy of AI answers. At the time of the March study, members of the European Parliament called for the AI Act to be applied to problems of this kind.

AI providers might in future have their assistants consult regulators' registers before answering. Regulation might also require them to. But I have not found any provider announcing an assistant that answers licence questions by checking the register first. How regulators will act is not yet known either.

Some things I have not verified. I have not read the original study report. Reports name the organisation behind it differently, as XY Legal Solutions or as Keurmerk Verantwoorde Affiliates. The names of the three assistants that presented operators as legal come from iGaming Business. For now, checking a licence remains a human step: matching names against the register.

An AI's answer about whether an operator is licensed cannot be used as a licence check as it stands. A licence is a dated entry in a regulator's register.

Confirmation exists only when each name is matched against that register and the date of the match is kept. Up to that point, the AI's answer is useful for one thing: gathering the names to be checked.