- The EU AI Act is the only binding AI law here: the prohibitions have applied since February 2025, AI disclosure to players since 2 August 2026, and the high-risk regime moved to 2 December 2027.
- Malta's AI Gaming Charter (18 September 2026) is voluntary but specific: disclosure at first contact, an AI inventory, a senior accountable person, kill-switches, and named triggers for handing a support conversation to a human.
- The MGA's own survey found that only a small number of licensees disclose AI use to customers, although the AI Act's disclosure duty has applied since August.
- The UK Gambling Commission has an AI approach for itself, not AI rules for operators; Spain is building one common, mandatory risk-detection model for all online operators.
- Two independent tests in 2026 found general-purpose AI assistants naming unlicensed casinos, including as “licensed”, which is now a channel and a reputational problem for licensed operators.
- New York's Gaming Commission proposed, on 9 October 2026, a ban on AI that personalises offers, odds, limits, messages or wager suggestions, with mandatory AI inventories and seven-year logs. Human review and consent do not exempt it.
Search for AI regulation in iGaming and most of what comes back is secondary: vendor blogs claiming that regulators have "mandated" AI harm detection, or summaries of summaries. This tracker uses only primary sources: the law itself, regulators' own publications and named studies. For each item it gives the date and says what it means for the AI a licensed operator actually runs, in player support, responsible gambling, fraud and marketing.
It is a living page. We re-check every source at least twice a month and record each change in the change log at the bottom. Last checked: 3 November 2026. It is not legal advice; confirm your own position with counsel.
At a glance: who has said what
| Jurisdiction | Instrument | Binding? | Date | What it means for operator AI |
|---|---|---|---|---|
| EU | AI Act, as amended by Regulation (EU) 2026/1744 | Yes | In force; staged dates to 2028 | Prohibitions and AI disclosure apply now; high-risk regime from 2 December 2027 |
| Malta | MGA and MDIA AI Gaming Charter (G-SPG-025) | No, voluntary | 18 September 2026 | The most detailed sector guidance so far: disclosure, inventory, accountable person, kill-switches, human escalation in support |
| Great Britain | Gambling Commission, approach to AI | No operator rules | Last updated 5 January 2026 | Covers the Commission's own AI use; industry AI stays under existing licence conditions |
| Spain | DGOJ draft resolution on a common risk-detection mechanism (Royal Decree 176/2023) | Will be, once adopted | Second consultation closed 26 October 2026 | One regulator-built model for spotting risky play, used by every online operator |
| New York (US) | State Gaming Commission proposed 9 NYCRR Part 5404 | Will be, if adopted | Proposed 9 October 2026; 60-day comment period | Bans AI-personalised offers, odds, limits, messages and wager suggestions; mandatory AI inventory and seven-year logs |
| Netherlands | Independent study of AI assistants against the KSA licence register | No (evidence, not rules) | Published early October 2026 | General-purpose AI tools route players to unlicensed casinos |
Italy, Germany, Sweden, Denmark and other licensing markets are not in the table yet. We add a jurisdiction only once we have read the regulator's own documents, not a press summary of them.
EU: the AI Act is the only binding AI law here
The AI Act applies to every operator serving EU players, whatever its licence. The July 2026 amendment, Regulation (EU) 2026/1744, was widely reported as "the delay". It postponed the high-risk regime and nothing that touches player-facing AI today. We cover the detail in the EU AI Act for licensed gambling operators; the dates are below.
| Date | What applies | Where operators feel it |
|---|---|---|
| 2 February 2025 | Prohibited practices (Article 5), including exploiting vulnerabilities linked to a person's economic or social situation | Marketing, bonus targeting, retention models |
| 2 August 2026 | Informing people they are interacting with an AI system (Article 50(1)) | Support chat, email and voice agents |
| 2 December 2026 | Grace period ends for machine-readable marking of generated content by systems already on the market (Article 50(2)); two new prohibitions apply | Generated marketing content and AI vendors' products |
| 2 December 2027 | High-risk systems listed in Annex III | Hiring AI, emotion recognition, possibly affordability models |
| 2 August 2028 | High-risk AI in regulated products (Annex I) | Rarely relevant to operators |
Fraud detection and one-to-one biometric verification, such as a KYC face match, are expressly carved out of the high-risk list. For most operators, the near-term work is disclosure, the Article 5 boundary in marketing, and documentation. See also AI Act and GDPR in AI player support.
Malta: the AI Gaming Charter
On 18 September 2026 the Malta Gaming Authority and the Malta Digital Innovation Authority published the AI Gaming Charter, a 48-page framework for MGA B2C and B2B licensees. It describes itself as one of the first initiatives in Europe to set AI principles specifically for gambling. It is voluntary: it creates no new legal obligations and does not replace the AI Act or GDPR. It is also the most concrete statement yet of what a gambling regulator expects to see when it looks at an operator's AI.
The points most relevant to operators:
- Disclosure. Players should be told they are dealing with an AI system "at latest from the point of first interaction", matching Article 50 of the AI Act.
- No AI washing. Don't present conventional software as AI, overstate autonomy, or market AI as a guarantee of fairness, accuracy or safety.
- Two tiers of risk. Higher-impact systems are those that are player-facing or touch responsible gambling, fraud, AML, KYC, account restrictions or eligibility. They get more documentation, testing and oversight.
- An AI inventory as the system of record: each system's purpose, owner, risk tier and lifecycle status.
- A senior accountable person for AI governance, for example a Chief AI Officer, plus an oversight or ethics committee for larger licensees.
- Rollback and kill-switch mechanisms, tested, for higher-impact systems, along with staged deployment and challenger testing.
- Human escalation in support. For player-facing conversational AI, operators should define triggers: repeated low-confidence answers, an unresolved or escalating query, distress signals, signs of vulnerability or gambling harm, or a request for a human. On any of these, a trained person should take over promptly.
- Incident logs, including near misses, and notification to the regulator of material changes to AI used in areas such as responsible gambling interventions.
Before writing the Charter, the MGA surveyed and interviewed its licensees. The regulator says responses came from a limited subset of the sector. Customer support chatbots were among the most mature AI uses, with fully deployed systems. Yet only a small number of organisations said they disclose AI use to customers, and only a small number reported a fully established AI risk-assessment process or incident response plan. Few evaluate whether AI-driven personalisation increases risky behaviour. In the area operators have automated most, disclosure is the thing most of them skip, and it has been an AI Act obligation since August.
The escalation triggers read almost exactly like the design rules we argue for in responsible-gambling signals in support chats: a hard rule hands a distressed player to a human. A confidence score should never make that call.
Great Britain: an AI approach for the regulator, not rules for operators
The Gambling Commission's approach to artificial intelligence is an appendix to its Corporate Strategy 2024–2027, last updated 5 January 2026. It is mostly about the Commission's own use of AI, for which it states a low or minimal risk appetite. Of its six areas of work, one is "understanding how the industry we regulate is using AI" and regulating that use where appropriate; another is managing hostile use of AI against the Commission or licensees. The next steps it names are internal: use-case templates and governance structures.
There are no AI-specific licence conditions for operators. Several blogs claim the Commission has "mandated" AI harm-detection tools; we found no such requirement in the Commission's own publications. Operators' AI is judged under the existing licence conditions and codes of practice, including the customer-interaction requirements, which apply whether a person or a model makes the call.
The Commission has, however, said how it judges AI in practice. Speaking at the Gambling Anti-Money Laundering Group conference in London in June 2026, its director of enforcement and intelligence, John Pierce, said the regulator does not oppose AI for anti-money laundering, but that some AI and algorithmic tools it had seen regularly failed to achieve the required compliance outcomes. Operators using them must be able to show the tools work. The same message applies beyond AML: under existing licence conditions, "the AI does it" is not evidence that a control works. A test of its outcomes is.
Spain: one risk-detection model for every operator
Spain is taking the opposite route to Malta. Under Royal Decree 176/2023, the DGOJ is building a common, mandatory mechanism for detecting risky gambling behaviour that every online operator it supervises will have to use. The DGOJ has described it as a supervised-learning model trained on real market data. The draft resolution went to public consultation in 2025. It was reopened after a report from the Spanish data-protection authority (AEPD) dated 15 June 2026, and the second consultation, after an extension, closed on 26 October 2026. Trade press reports that the operator associations Jdigital and AEJAD have challenged the project before the European Commission.
For operators in Spain this changes the question. It is no longer "which RG model do we buy" but "how do we act on the regulator's flags": who reviews them, how fast, and what support says to the player. That is a support-process problem as much as a data one.
New York: a proposed ban on AI-personalised offers
Outside Europe, the most detailed AI rule for gambling operators so far is a draft. On 9 October 2026 the New York State Gaming Commission voted unanimously to propose a new Part 5404 of its rules (9 NYCRR §§ 5404.1 and 5404.2). The text is in the Commission's meeting book. It covers every licensee, from casinos, mobile sports wagering operators and platform providers to racetracks and lottery couriers, along with their affiliates, agents and vendors. Comments are accepted for 60 days after publication in the State Register, and no date has been set for a final vote. It matters in Europe because several groups licensed here also hold New York licences, and because it shows how far a gambling regulator is prepared to go.
- The ban. No automated decision system, algorithmic profiling, machine learning or generative system that processes a patron's data may "select, customize, time or deliver" any promotion, bonus, offer, odds, limit, message or product recommendation, "regardless of purpose or effect". It also may not suggest any wager, wager type or amount.
- No escape through a human or consent. Human review of the system's output does not exempt it, and patron consent does not waive the ban. The only opt-in is a separate wager-suggestion app for horse-racing bettors, with strict limits. Some trade reports said the rules would not apply to customers who consent; the draft says otherwise.
- Presumed guilty. Any offer whose recipient, content, timing or amount varies with individual data is presumed to violate the rule. The operator can rebut that only with clear and convincing evidence, backed by logs, that the variation came from published, objective criteria.
- What stays allowed. Promotions offered uniformly under published criteria not derived from predicted behaviour, and systems used solely for fraud prevention, anti-money laundering, age verification or responsible-gambling interventions designed to reduce wagering. Outputs of those systems may not be reused for marketing or retention.
- Governance. An inventory of every system that processes individual data, with its purpose, inputs and outputs; logs of inputs, outputs, model version, recipients and criteria kept for seven years; audit access for the Commission; 30 days' written notice before any material change; matching vendor contracts; and an annual compliance certification signed by the chief executive. Each affected patron counts as a separate violation.
The draft has no carve-out for customer support. A support assistant that reads a player's account and replies to them processes individual data and delivers a message, and the ban on messages applies regardless of purpose. Whether the Commission means service replies to be caught is a fair question for the comment period. Until it is answered, the safe design for a New York licence is an assistant that answers but never offers: no bonuses, no product suggestions, no retention offers, and a log of every reply. The same line runs through the support-to-CRM signals many operators use: feeding them into incentive targeting would fall under the ban, feeding them into responsible-gambling interventions would not.
The governance list will look familiar from Malta: an AI inventory, logs and a named senior owner. The difference is that in New York they would be mandatory, with penalties up to licence revocation.
AI assistants and unlicensed casinos: the new channel problem
Two independent tests in 2026 looked at general-purpose AI assistants, the tools players now ask before they search.
- Netherlands, October 2026. A study by researchers at XY Legal Solutions tested ten AI assistants with Dutch-language questions between August and September 2026. It checked every operator they named against the KSA licence register. All ten named at least one unlicensed operator. Unlicensed operators appeared in 33 of 79 recorded interactions. Three tools presented unlicensed sites as if they were licensed; one produced a "Top 50 legal casinos" list containing unlicensed brands.
- United Kingdom, March 2026. The Guardian and Investigate Europe tested five chatbots. All five could list unlicensed casinos and explain how to reach them. Several advised on casinos outside GamStop, and only two opened with a gambling-risk warning.
Neither result is a rule. Both matter to licensed operators for two reasons. They are a new route by which players leave for the black market, alongside the friction we described in casinos without verification. And they show that regulators and the press now test AI behaviour directly, which is the kind of test an operator's own support assistant should be ready for.
The questions to ask any AI support vendor.
Disclosure, escalation, audit trail, data path: the buyer's guide turns the points above into vendor questions.
What to do with this, by owner
| Owner | Action | Source |
|---|---|---|
| Head of Support | Check that every AI channel discloses AI at first contact and answers "am I talking to a bot?" honestly | AI Act Art. 50(1); MGA Charter 3.2.1 |
| Head of Support / RG | Write down the human-escalation triggers for the support assistant and test them on real transcripts | MGA Charter, human oversight |
| Compliance | Build the AI inventory: system, purpose, owner, risk tier, vendor | MGA Charter 3.1, 3.8.4 |
| Marketing / CRM | Exclude players with RG markers from automated incentive targeting, and document it | AI Act Art. 5 |
| Product / Tech | Confirm rollback and kill-switch for every player-facing model, and run a drill | MGA Charter 3.7 |
| New York-licensed groups | List every system that personalises offers, odds, limits or messages with player data, and decide whether to comment on the draft before the deadline | Proposed 9 NYCRR Part 5404 |
| Spain-licensed operators | Design the workflow for acting on DGOJ risk flags: review, time limits, player contact | RD 176/2023; DGOJ draft resolution |
To see how these controls look inside a working support system, see how PlayerDesk handles EU rules and licence requirements.
How to cite this tracker
You are welcome to quote or reuse this tracker. Please credit Arctura Research and link to arctura.eu/research/ai-regulation-igaming-tracker.html. For a comment, a jurisdiction you think is missing, or a correction, write to hello@arctura.eu.
Change log
- 10 November 2026 — Added New York: the State Gaming Commission's proposed Part 5404 on AI in wagering (proposed 9 October 2026).
- 3 November 2026 — Tracker published: EU, Malta, Great Britain, Spain, Netherlands.
Frequently asked questions
Is there a law that regulates AI in gambling specifically?
Not in the EU or the UK as of November 2026. The binding rules come from the horizontal EU AI Act, data-protection law and each operator's gambling licence. Malta's AI Gaming Charter is the first sector-specific AI framework from a European gambling regulator, and it is voluntary.
Does an online casino have to tell players they are chatting with an AI?
In the EU, yes. Article 50(1) of the AI Act has required people to be informed that they are interacting with an AI system since 2 August 2026, unless it is obvious from the context. Malta's AI Gaming Charter asks for the same at the latest from the first interaction.
What does Malta's AI Gaming Charter require from operators?
Nothing is mandatory: the Charter is voluntary and creates no new legal obligations. It recommends disclosing AI to players at first interaction, avoiding AI washing, keeping an inventory of AI systems, naming a senior accountable person for AI governance, testing rollback and kill-switch mechanisms for higher-impact systems, and defining triggers that hand a support conversation to a trained human.
Has the UK Gambling Commission published AI rules for operators?
No. Its published approach to AI, last updated 5 January 2026, covers the Commission's own use of AI and lists understanding and, where appropriate, regulating industry use as one area of work. Operators' AI is governed by their existing licence conditions and codes of practice.
How often is this tracker updated?
We check the sources at least twice a month and log every change at the bottom of the page with a date. If you spot something missing or out of date, write to hello@arctura.eu.
Last reviewed 10 November 2026. Figures marked as industry ranges are ranges, not Arctura claims; operator-specific numbers are only quoted after validation on the operator's own baseline.


