AI names illegal gambling sites: why the KVA study also matters for Belgium
A Belgian player asks an AI assistant which online casinos are trustworthy. The answer appears as a neat list, with short descriptions and perhaps even the reassurance that the operators are licensed. But who checked that licence? And for which country does it actually apply?
These are the questions raised by a new Dutch study into generative AI. The Keurmerk Verantwoorde Affiliates (KVA) found that all ten tools tested mentioned operators without a Dutch licence during the tests. Some answers went further: unlicensed brands were presented as legal, or the absence of player-protection measures was described as an advantage.
The study focuses on the Netherlands. Yet it is also relevant to Belgium, where player protection likewise depends on licensing, identity verification and an exclusion register. An incorrect AI answer can obscure those differences. For Belgium, the key issue is therefore practical: can a player trust a recommendation when the assistant does not correctly apply the local rules?
Ten AI tools tested, none avoided naming operators entirely
The KVA examined ChatGPT, Google AI Mode, Google Gemini, Microsoft Copilot, Perplexity, Claude, Grok, DeepSeek, Mistral Le Chat and Meta AI. The researchers used three sets of questions: general requests for casinos, leading questions about offers without the Dutch Cruks register, and questions framed within a research context.
All ten tools mentioned an operator without a Dutch licence in at least one tested conversation. Eight did so in the neutral or leading-question series. With Google AI Mode, Mistral Le Chat, DeepSeek and Grok, the first leading question was enough. Claude and Meta AI were more cautious in the first two series, but eventually named operators within the research context.
The full KVA study on generative AI and illegal online gambling was published on 30 September 2026. It describes a limited sample. The researchers checked whether names or domains were mentioned; they did not test whether someone could then open an account or deposit money.
That distinction is important. The results prove that unlicensed operators appeared in the answers examined. They do not prove that every user receives the same answers, that all tools constantly recommend illegal casinos, or that Belgian players actually started gambling through those answers.
Mentioning a casino name is not always a recommendation
In the summary table, the result looks simple: ten tools, ten instances of unlicensed gambling offers. But behind that total are different types of answers. Some tools described characteristics of operators in a positive way. Others referred to public enforcement cases, naming companies that had been targeted by the Dutch regulator.
A name appearing in a warning has a different meaning from a name appearing in a list of recommended casinos. The KVA counted both whenever an unlicensed operator was specifically named. To assess an AI tool’s behaviour, you therefore also need to look at the context, source attribution and wording.
That nuance makes the concern about incorrect licensing claims even sharper. An assistant summarising an enforcement case may be providing legitimate information. An assistant presenting the same unlicensed operator as vetted and legal gives the consumer a false reason to trust it.
For Belgium, any follow-up study should preserve that distinction. It should measure which operators are named, how they are described, and whether their exact websites hold the required Belgian licence. The number of names mentioned alone does not tell the whole story.
In Belgium, the relevant system is EPIS, not Cruks
The Dutch questions partly focused on casinos without Cruks. That register is intended to block excluded players from taking part in licensed gambling in the Netherlands. Belgium uses its own system: EPIS, the Excluded Persons Information System.
A Belgian analysis therefore cannot simply replace the word Cruks and copy the Dutch findings. The licences, institutions and obligations differ. Our explanation of the differences between EPIS in Belgium and Cruks in the Netherlands provides background on both registers.
For Belgian online players, the absence of the required EPIS check is not an extra freedom that makes a casino more attractive. It means an important player-protection measure is not being applied. For someone who has excluded themselves in order to step away from gambling, a positive description of such a missing check can be particularly harmful.
Moreover, no Belgian test was carried out in the KVA study. Based on this data, we therefore do not know what answers the ten tools would give to specific questions about EPIS. The Dutch findings are a reason to investigate this, not proof that all tools behave in exactly the same way in Belgium.
A foreign licence is not Belgian authorisation
The biggest pitfall for Belgian users is the meaning of the word “legal”. An operator may hold a licence in another country without thereby being authorised to offer online gambling in Belgium. A general reference to an international licence therefore does not answer the Belgian question.
A Dutch licence from the Kansspelautoriteit is also not a Belgian licence from the Gaming Commission. Anyone asking about trustworthy casinos in Belgium needs information that applies to Belgium and to the activity concerned. An assistant that mixes up foreign and Belgian licences can give a convincing but useless answer.
That is why verification must go beyond the brand name. Check the operator, the licence category and, above all, the exact website address. The official licence registers of the Belgian Gaming Commission distinguish between different categories. A similar name or a professional logo is not enough to determine that a website belongs to a licensed operator.
Gambling Club explains this further in its article on the difference between legal and illegal gambling operators in Belgium. The practical rule remains simple: a claim made in an AI answer should be verifiable against the competent regulator.
Unlicensed casinos were incorrectly presented as legal
According to the study results, Microsoft Copilot, Mistral Le Chat and Google Gemini included unlicensed brands in lists presented as legal or licensed in the Netherlands. Copilot also stated, in one such list, that the named operators were connected to Cruks.
That is a different problem from an incomplete warning. The user may specifically have asked for a vetted operator and then received an answer that conceals the lack of a licence. Words such as “licensed” and “verified” can then wrongly create the impression that further checking is unnecessary.
For Belgian players, this is a concrete risk to keep in mind. A table containing casino names and licensing claims can look carefully prepared, but the format says nothing about the accuracy of the content. Even an answer with source references needs to be checked: does the source really support the licensing claim, and does it concern the correct website?
A disclaimer at the bottom does not automatically correct an incorrect classification. If a casino is described at the top as Belgian-licensed, that remains an important factual statement. The assistant should first be able to substantiate it, rather than merely advising the user afterwards to be cautious.
Belgium already faces deception using trusted casino names
The Belgian market already has experience with online information that suggests reliability without actually providing it. Gambling Club recently reported on more than twenty fake websites using the PepperMill Casino name. Some copies resemble the real site but redirect visitors to another platform.
That case does not prove any involvement of AI. It does, however, show why checking the exact domain is important. If an assistant merely recognises a well-known brand name and mentions a matching-looking website, the verification is not complete. The address must correspond to the genuine operator and its licence.
Social media is already another major area of concern. According to reporting on the Belgian annual report, more than 8,500 suspected illegal gambling advertisements were reported to Meta in 2025 with requests for removal. That number of reports should not be treated as equivalent to the same number of confirmed permanent removals.
Our article on Belgium’s approach to illegal gambling advertisements on Facebook and Instagram also discusses the misuse of well-known names and logos. A chatbot can be another information channel. The question is whether it can sufficiently distinguish misleading information from official data before turning it into an answer.
Missing protection should not appear in a list of advantages
The KVA describes answers in which the absence of identity checks, Cruks checks and higher bonuses were presented as attractive features. To a consumer, that can sound like less administration or more freedom to play. But the absence of required checks should not be confused with better service.
In a Belgian assessment, the context should be reversed: what protection is supposed to be part of the licensed offering, and which of those protections are missing? An exclusion check exists precisely to protect people from continuing to gamble. An assistant that describes its absence positively gives a misleading impression of the purpose of that measure.
This is all the more relevant because the Belgian regulator is already warning about the visibility of illegal gambling offers. Gambling Club reported on the warning from Belgian Gaming Commission chair Magali Clavie about the illegal gambling market. The debate around AI should therefore connect with existing problems involving deception and player protection.
However, the Dutch study does not measure how many additional players end up on illegal sites through AI. Nor does it calculate the harm caused in Belgium. Such claims would require further research. What is already visible is an information problem that could have serious consequences for vulnerable users.
What AI providers should verify for Belgium
The first improvement is a clear jurisdiction check. An answer about online gambling should know which jurisdiction the user is asking about. A question in Dutch is not automatically a question about the Netherlands. Belgian users gain nothing from a licensing claim that only applies elsewhere.
The assistant should then verify licensing information against an official, up-to-date source. That includes both the operator and the exact domain. If that verification is missing, the answer should state the uncertainty. A guess based on a brand name should not be upgraded into a definitive claim that an operator is licensed in Belgium.
Tone also matters. A factual explanation of an enforcement decision may mention names without encouraging visitors to gamble. A recommendation that presents the absence of an exclusion check as an advantage requires a different assessment. That distinction helps preserve useful information while addressing misleading recommendations.
A Belgian test should also include both Dutch-language and French-language questions, with repeated measurements and documented settings. That is a proposal for follow-up research. The KVA results themselves remain a snapshot of conversations about the Dutch market.
What if you have already ended up on the wrong website?
If a recommendation has directed you to a suspicious website, do not deposit additional money to supposedly unlock a blocked withdrawal. Keep the exact website address, relevant messages and payment details. In cases of suspected payment fraud, you can discuss appropriate steps with your bank.
Gambling Club has a guide on what to do after gambling at an illegal casino in Belgium. The options available to you depend on the circumstances and payment method; reimbursement is not guaranteed.
The Belgian lesson from the KVA study begins even earlier, with the first answer on the screen. A clear list and a confident explanation are not the same as a licence check. For a Belgian player, the decisive information should concern the Belgian authorisation of the exact website. As long as an assistant cannot establish that reliably, its recommendation remains something that needs to be verified.

