Research

ChatGPT casino license checks, measured across three markets

5 August 2026

ChatGPT reads official gambling license registers when it answers casino questions, and where a regulator publishes real data it reads them to the digit. We ran 81 controlled probes across New Jersey, Great Britain and Australia in early August 2026, nine questions repeated three times in each market, every claim scored against our own snapshots of the official registers. In Britain it produced licensee legal names and license account numbers, 42 checked facts, 42 exact. In New Jersey it rebuilt the regulator's operator revenue table to the dollar. In Australia, where online casino is illegal, it named zero casinos across nine runs and cited the statute. The precision stops where the published data stops. One brand with no license document on its site was given a certificate that does not exist, twice, in two markets.

What WhiteLobby measured

Controlled probes captured	81
Regulated markets, selected for legal contrast	3
Questions per market, each repeated three times	9
Official register datasets held as ground truth	14
British register facts verified exact, zero errors	42
New Jersey revenue figures reproduced exactly, in rank order	10 of 10
New Jersey brand to licensee assignments verified exact	35 of 35
Sibling license traps resolved correctly	4 of 4
Brands named on Australia's three casino questions	0 of 9 runs
Ranking runs that excluded bonuses as a basis	11 of 11

The three markets carry three different gambling laws, and we picked them for that contrast. New Jersey licenses online casino and sports betting at state level. Great Britain licenses both nationally and publishes a downloadable register. Australia licenses sports wagering and prohibits online casino outright. Nine questions ran in each market, five about specific brands sampled by known register state and four general ones, best casinos, legal casinos, where a named slot can be played, best betting sites for the local sport. Conditions stayed fixed throughout. ChatGPT on the web, Temporary Chat, Intelligence set to High, one prompt per fresh conversation, no follow-ups, exit IP verified before every block. Every run was scored on six separate axes, and the full method sits in the note at the end.

License account numbers, quoted from the register

Asked which online casinos are legal in Great Britain, one run produced an unprompted three-column table of brand, licensed operator and Gambling Commission account number. We checked all 21 fields against the UKGC public register and our same-day snapshot. All 21 are exact.

The table clears traps that catch a brand-name lookup. For bet365 it picked the gaming license, 55149, and left the sibling sports license 55148 alone. For Grosvenor it picked the online licensee and left the land-based one. For Ladbrokes it took the active account and avoided a surrendered one. Another run mapped nine brands to their exact licensed domains, including two cases where separate consumer brands sit under one license. A third run showed its working in the interface, it opened the regulator's licensing workbook, filtered it for active remote casino accounts, and ranked from the result.

The regulator's revenue table, rebuilt to the dollar

Asked for the ten best online casinos in New Jersey, one run ranked them by June 2026 internet gaming win and cited the Division of Gaming Enforcement. The brand-level figures live inside the DGE's monthly filing, several pages deep, and stay out of the press release. We pulled the filing and checked every number. Ten of ten match to the dollar, and the ranking order is the true order with no brand skipped.

A separate run reproduced the state's authorised internet gaming universe, 35 brand-to-licensee assignments, every one correct, including two obscure sub-brands that exist on the regulator's page only as bare domains. Only operators licensed in New Jersey appear in DGE filings, so a ranking built from that file excludes every offshore brand by construction, and the model never had to mention licensing to do it.

Where online casino is illegal, the answer names nobody

Australia prohibits online casino under the Interactive Gambling Act 2001. We asked the same casino questions there that produced ten-brand rankings in New Jersey and Britain. Across nine runs, zero casino brands were named. All three ranking runs cited the statute and the prohibited game classes. One answer to the question of which online casinos are legal in Australia consisted of a single word.

None. ChatGPT, Australia, 5 August 2026, answering which online casinos are legal there

Sports wagering is licensed in Australia, and on that question the same model named ten betting brands in all three runs, stating each time that it had checked them against the federal register of licensed providers. The license filter follows the law of the market. Where the product is licensed it ranks licensees, and where the product is prohibited it names the statute and points the reader at the register for the product that is legal.

The Australian runs also worked through the ways an offshore site presents itself as local. Australian branding, Australian dollars, local-style domains, accepting Australian players and crypto payment rails were each named, and each was dismissed as carrying no legal weight. The foreign-license rule arrived verbatim, in all three legality runs.

An overseas licence from Curacao, Malta or another jurisdiction does not make an online casino legal in Australia. ChatGPT, Australia, 5 August 2026

Britain produced the same rule about the same two jurisdictions for its own market. A player who asks gets told that the license an offshore brand advertises has no force where that player lives.

Fresh register, stale listings

The sharpest pair in the study sits in two markets, one week apart, on the same mechanism.

In New Jersey, the legality question ran three times. Two runs visibly fetched the regulator's own page and scored 35 of 35 and 25 of 26 against the authorised list. The third run mixed in a third-party comparison site and named four brands as currently legal that hold no New Jersey authorisation, two of them long closed and one that never held one.

In Australia, one AFL-betting run went to the Northern Territory government's list of licensed operators, found two brands flagged as not currently trading, and removed them from its own ranking.

Betr and BoomBet are marked "not currently trading" there, so I have excluded them despite their appearance on some comparison websites. ChatGPT, Australia, 5 August 2026. We quote this as the assistant's reading of the list it cited. We have not verified either company's trading status ourselves.

The regulator's record keeps an AI answer current. Third-party listings carry closed brands forward into it. Both directions showed up inside one week, and the limit deserves stating plainly, only one of the three Australian runs applied that status field. The other two ranked one of those brands ninth and tenth.

Where the checks break

Three failure modes appeared, each measurable, each traceable to the source the answer relied on.

• An invented certificate. On one offshore brand, in two markets two days apart, the answers cited an expiry date from a license certificate linked on the brand's own website. That homepage, archived three days before the first capture, carries one sentence of licensing text, no certificate link, and no occurrence of the year in question. The brand holds no entry in any of the 14 register datasets we keep. Both verdicts still warned the player, correctly, so the safety outcome and the quality of the evidence are two separate facts and we report them separately.
• Wrong dates inside right records. Entities, numbers and corporate structures verified clean. Dates wobbled. One British enforcement penalty came back with the wrong month in one run of three. A brand's exit from the British market was compressed into a single date in all three runs, where the public record holds two events two months apart. Five British cells carried a date set two days after the capture day.
• A hard search interface hides a real license. Asked about a brand licensed in Anjouan, one answer named the register, reported that it could not retrieve the brand's entry through the available search, and left the claim empty. The record exists in our snapshot. A valid license went unreported because the register is difficult to query, and for the operator holding it this failure costs the most.

The publishing format also decides whether an assistant can call a brand unlicensed. Australia's federal register lists who holds a license, so absence from the list is a readable fact and one run stated it outright.

Bovada does not appear on the ACMA register of licensed online wagering providers, updated 3 August 2026. ChatGPT, Australia, 5 August 2026

Britain's runs on the same brand searched the Gambling Commission by name, found nothing, and let the silence carry the point. A list of holders can express a negative. A record-by-record lookup leaves it unsaid.

What AI license checks mean for casino operators

Three things follow from these 81 runs.

1. Your register row works like a live product page. It is read field by field, including fields you may not know are published. In this sample an expired-and-under-assessment status reached a player who asked a one-line question, in three markets, within days of that register edition appearing. Check what your regulator actually publishes about you, in every format it publishes.
2. The register is the only thing keeping your listing current. A brand that stops trading gets dropped by the assistant that reads the regulator and kept by the comparison sites that skip it. Dead entries and missing domains on the license record read as your current state to the machine answering your players.
3. You need to know what the answers say before your players do. Behaviour shifted run to run throughout this study, so a single look proves nothing. Measurement means repeated runs, per market, per question, scored against the record, which is what an AI visibility audit does.

Regulators hold the other half of the outcome. A published list of license holders lets an assistant state that a brand is unlicensed. A record-by-record lookup keeps it silent. A difficult search interface causes valid licenses to go unreported. One regulator in this study publishes its real holder register as a weekly PDF on a media subdomain, which the assistant found and we, at first, did not.

For publishers and affiliates, a stale listing has become a competitive liability. In one market the assistant described best-casino rankings to the reader as a warning sign. In another it removed brands from its own list and said it was overriding the comparison sites to do it.

ChatGPT license checks, frequently asked

Does ChatGPT check gambling licenses before answering?

In our 81 probes it reached an official regulator source on most runs, and on the general questions, legal casinos, best casinos, best betting sites, it reached one every time. The variation came from what each regulator publishes for it to find.

How accurate are the license claims in its answers?

Where the regulator publishes machine-readable data it was exact. Britain returned 42 checked register facts with zero errors, New Jersey returned ten revenue figures to the dollar. Dates were the weak field, one wrong month in three runs on one penalty and two events merged into one date on another.

Can it tell a player that a casino is unlicensed?

Where the regulator publishes a list of holders, yes, and it did, naming the register and its update date. Where the register only answers record-by-record lookups, the absence stayed implicit. The publishing format sets the boundary.

What should an operator do first?

Pull every format your regulator publishes about you and read your own row. Confirm the license record lists every live domain, keep the site's stated license details current and machine-readable, and measure the brand's AI answers with repeated runs.

Method and limitations

81 probes, three markets, nine questions, three repeats of each question per market. Brand cells sampled by known register state, a locally licensed control, a major Curacao-licensed brand, an offshore-claim brand, a brand whose Curacao license had ended, and an unlaunched site advertising a license number, plus four general questions. ChatGPT web, Temporary Chat, Intelligence High, one prompt per fresh chat, no follow-ups, English prompts, captured on 3 and 5 August 2026 with the exit IP verified per market before each block. Ground truth, our register snapshots of 1 August 2026, 14 datasets, plus same-day desk fetches of the primary sources for every named claim. Two claim classes carry no accuracy score here by design, review-platform ratings, which publish no history, and in-lobby game availability, which sits behind logins and location checks. This is a descriptive study of captured answers. It supports no claim about other assistants, dates, accounts, languages or modes, and behaviour varied run to run throughout, which is why results are given as ratios. The research question came out of our iGaming client work, and no client data appears in the study. Raw captures were saved at collection time and the row-level scoring for all 81 runs is available, ask us for the run behind any number.