This Thought Experiment is a scenario-based exercise designed to provoke discussion. It is not a prediction or statement of fact.
Give a Nevada category A floor with forty live tables the surveillance system three panellists described at G2E, in a session reported on 5 October 2026, and the finance question arrives inside a week. The cameras read cards, match them to individual bets, flag a dealer overpay the moment it happens and recommend where to move staff, so can the room be run with fewer people on nights? In Nevada the answer is no, and not for reasons of caution or culture. The control standard the licence runs on sets a staffing floor in headcount, and the quality of the detection is not one of the variables in it.
Key Takeaways
- AI casino surveillance sold to table game floors now tracks card values, bet sizes, dealer payment errors and roulette wheel bias in real time, with vendors putting detection accuracy near 88 per cent. A staffing rule written in headcount does not move because the cameras got better.
- Nevada’s Surveillance Standards, as revised with effect from 3 April 2026, require a category A surveillance room to be attended at all times, and state that a consolidation must maintain or exceed the licensee’s current surveillance room staffing levels.
- In the European Union the same system becomes a prohibition rather than a cost question the moment any part of it infers emotion from staff. Article 5(1)(f) of the AI Act has applied since 2 February 2025, and Regulation (EU) 2026/1744 left it alone while deferring the high-risk obligations to 2 December 2027.
- New South Wales is going the other way, with facial recognition due at every gaming room entrance from 2028, so the same class of technology is discretionary in one market, partly prohibited in another and compulsory in a third.
What does the scenario look like on the floor?
Forty live tables put the property in Nevada’s category A, defined in the Surveillance Standards as a non-restricted gaming operation with ten or more table or card games. The system is the one Trevor Outman of EagleSight.ai set out at G2E in three tiers: junior agents counting rounds per hour and side bet usage, senior agents detecting dealer payoff errors, and an orchestrating agent that recommends moving a surveillance employee off a quiet table onto a busy one. Outman put dealer error and cheating at 1 to 2.5 per cent of table game revenue a year, and detection accuracy near 88 per cent.
None of it is a prototype argument, since Resorts World Las Vegas already matches camera-based card recognition to individual bets using the RFID chips it has run since 2021. So the paper reaching the executive committee proposes a thinner night shift, because one operator cannot watch forty tables and a machine can watch all of them.
Why does the casino surveillance room staffing requirement not move?
Because Nevada’s standard counts people rather than coverage. Standard 2.010(8) requires that for all category A licensees the surveillance room be attended at all times by personnel trained in the use of the equipment and with knowledge of the games and house rules, and the only relief in it is one hour in any eight-hour period for meals and rest breaks. Standard 12(4) closes the obvious workaround, which is to merge two rooms and bank the difference, by requiring that a consolidation maintain or exceed current surveillance room staffing levels.
Nevada rebuilt the rest of these standards in the same pass. The 2026 revision added a definition of video management system, dropped the old 30 and 7.5 frames per second minimums in favour of recording in real time, told licensees to follow best practices used in information technology, computer networking, encryption and password management, and under Standard 13 required that no single point of failure take out more than half the surveillance coverage in any gaming area. A regulator that has just written an information security clause into a camera rule is not unaware of what modern systems do. It simply did not treat automated detection as a substitute for a person, anywhere in the document.
Two further clauses matter more than they look. Standard 12(3) requires that access to a surveillance recording system from any location outside the surveillance room be reported to the Enforcement Division’s Operations Unit before it is installed and used, which is what a vendor-hosted model running its inference off the property amounts to. Standard 9 sets retention at seven days until 3 April 2027 and fifteen days after, so an alert queue three weeks behind is a queue whose footage no longer exists.
Where does AI casino surveillance stop being a cost question and become unlawful?
In the European Union, the moment any part of the model infers how a member of staff is feeling. Article 5(1)(f) of the AI Act prohibits using AI systems to infer emotions of a natural person in the workplace, with a narrow carve-out for medical and safety purposes. Germany’s Bundesnetzagentur states that these prohibitions have been in place since 2 February 2025, and Article 99(3) sets the maximum fine for breaching Article 5 at EUR 35,000,000 or 7 per cent of total worldwide annual turnover, whichever is higher.
The deferral everyone has been discussing does not help here. Regulation (EU) 2026/1744 of 8 July 2026 moved the Annex III high-risk obligations out to 2 December 2027 and the Annex I ones to 2 August 2028, but it left Article 5(1)(f) where it stood. The part of the AI Act a European casino would have spent two years documenting has been pushed back; the part that stops the purchase outright has not. A floor system that reads cards and chips is not caught by the EU AI Act emotion recognition prohibition, but a later release scoring dealer stress or tell behaviour is, and the moment to find that out is before signature.
Which regulators are writing AI rules, and what are they writing them about?
The customer side, so far, rather than the floor. The New York State Gaming Commission advanced a proposed new Part 5404 of 9 NYCRR at its meeting on 9 October 2026, under the agenda item use of artificial intelligence for wagering purposes. Section 5404.1(b) would stop any licensee, affiliate, agent or vendor using an automated decision system or algorithmic profiling to select, customise, time or deliver promotions, bonuses, offers, odds or product recommendations, and 5404.1(c) presumes any personalised offer unlawful unless the licensee can show it came solely from published, objective criteria. It reaches casinos and video lottery agents, not only the mobile sports betting licensees most of the coverage described.
New South Wales is legislating in the opposite direction, and on the floor itself. The package announced by Minister David Harris on 25 August 2026 puts $95.2 million behind a statewide exclusion register covering roughly 2,100 venues and about 87,000 machines, with mandatory facial recognition at the entry to gaming rooms arriving alongside the register in 2028.
| Market | Instrument | What it does to AI on the gaming floor |
| Nevada | Surveillance Standards for Non-restricted and Restricted Licensees, Rev. 04/26 | Nothing directly. Staffing is set in headcount, a consolidation cannot reduce it, and access from outside the surveillance room is reportable before use. |
| European Union | Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744 | Inferring emotion from staff prohibited since 2 February 2025. Transparency notice from 2 August 2026. High-risk duties deferred to 2 December 2027. |
| New South Wales | Gambling reform package announced 25 August 2026 | Facial recognition compulsory at every gaming room entrance from 2028, tied to the statewide exclusion register. |
| New York | Proposed 9 NYCRR Part 5404, advanced 9 October 2026 | Bans automated personalisation of offers, odds and wager suggestions across licensees. Says nothing about surveillance. |
What happens to the numbers if the operator buys it anyway?
The saving cannot come out of the surveillance headcount, so it has to come out of recovery. Outman’s framing at G2E was that surveillance moves towards a profit centre by raising the recovery rate on errors that would previously have gone unnoticed, which against a 1 to 2.5 per cent leak is a real number on a forty table floor, and also a number nobody outside the vendor has audited.
Review load goes up before it goes down. Zachary Koep of Resorts World Las Vegas said false positives are not a big problem because a human reviews the clip in under ten seconds, which is true of one alert and says nothing about the shape of a shift at 88 per cent accuracy across forty tables on a busy Saturday.
Then there is what the alert becomes once it exists. Every flag is a record with a timestamp, a table and usually a named dealer attached, sitting where a regulator can ask for it, a claimant can request it in disclosure and a union can raise it in a grievance. A department producing nine thousand machine-written reports a month where it used to produce a hundred written by people has not reduced its exposure; it has changed the shape of that exposure.
What should an operator check before signing?
Run the system against the control standard the licence actually operates under, before the contract rather than after the pilot. Four checks do most of the work.
- Find the clause in your own surveillance standard that sets attendance, and write down the number of people it requires. Where it is written in headcount, as Nevada’s is, the business case cannot rest on removing them.
- Ask the vendor in writing which components process images of employees, and whether any current or planned release infers state, mood or stress from a person. Take it as a contractual representation with notice on change, because a roadmap item added in release three is still your Article 5 problem in Lisbon or Tallinn.
- Establish where the inference happens. If any part of it runs off the property, that is access to the surveillance recording system from outside the surveillance room, reportable in Nevada before installation rather than after it.
- Model alert volume against your retention floor. Take the expected alerts per shift at the vendor’s stated accuracy, divide by your review capacity, and if the backlog runs past fifteen days you are buying evidence that deletes itself.
The Physical Security & Surveillance Systems Review Framework covers the first and third of those as part of a wider review of coverage, access control, staffing and incident procedure against licence conditions, which is the part most operators already have a process for. The second and fourth are what an AI purchase adds to it.
Boardroom Questions
- Which clause of our surveillance standard sets the attendance requirement in each jurisdiction we are licensed in, and does any of them express it as coverage rather than as headcount?
- Does any component of the system infer emotion, mood or stress from an employee, and have we asked the vendor that in writing?
- Where does the inference run, and if any of it runs off the property, have we reported that access before installing it?
- At the vendor’s stated accuracy, how many alerts a shift would a floor our size generate, who reviews them, and how long is the queue against our retention period?
- When an alert names a dealer, what is the evidential status of that record in a disciplinary hearing, a regulatory enquiry and civil disclosure?
Sources
1. Nevada Gaming Control Board, Surveillance Standards for Non-restricted and Restricted Licensees, adopted under Regulation 5, Rev. 04/26, gaming.nv.gov, read 23 October 2026. Standard 1.020 category definitions, Standard 2.010(8) category A attendance, Standard 9 retention, Standard 12(3) access from outside the surveillance room, Standard 12(4) consolidation and information technology practices, Standard 13 single point of failure.
2. Nevada Gaming Control Board, Proposed Amendments to the Surveillance Standards for Non-restricted Licensees, draft of 26 February 2026, gaming.nv.gov, showing the move from gross gaming revenue tiers to activity-based categories A to C and the deletion of the time-lapse and motion-activated camera definitions.
3. Mark Gruetze, G2E: Panel sees possibilities, challenges in AI-powered table game surveillance, CDC Gaming, 5 October 2026, reporting the session How AI and Technology Are Turning Surveillance into a Strategic Asset with Josh Bodell of MGM Resorts International, Zachary Koep of Resorts World Las Vegas, Trevor Outman of EagleSight.ai and Debi Nutton. The 88 per cent accuracy figure, the 1 to 2.5 per cent revenue estimate and the three-tier agent model are the panellists’ own claims and are not independently verified.
4. Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence, EUR-Lex. Recital 44 on emotion recognition in the workplace and recital 30 on biometric categorisation.
5. European Commission, AI Act Service Desk, Article 99, ai-act-service-desk.ec.europa.eu, read 23 October 2026. Article 99(3): non-compliance with the prohibition of the AI practices referred to in Article 5 is subject to administrative fines of up to EUR 35,000,000 or, for an undertaking, up to 7 per cent of its total worldwide annual turnover for the preceding financial year, whichever is higher.
6. Bundesnetzagentur, Prohibited practices, bundesnetzagentur.de, read 23 October 2026, on the scope of the Article 5(1)(f) workplace emotion recognition prohibition and the Article 5(1)(g) biometric categorisation prohibition, and stating that these prohibitions have been in place since 2 February 2025.
7. Regulation (EU) 2026/1744 of the European Parliament and of the Council of 8 July 2026 amending Regulations (EU) 2024/1689, (EU) 2018/1139 and (EU) 2023/1230, in force 27 July 2026, EUR-Lex, deferring the Annex III high-risk application date to 2 December 2027 and the Annex I date to 2 August 2028.
8. New York State Gaming Commission, public meeting of 9 October 2026, meeting book, item 5(a) Proposal: Use of artificial intelligence for wagering purposes, proposed 9 NYCRR Part 5404 sections 5404.1 and 5404.2, gaming.ny.gov.
9. Stronger protections in major NSW gambling reform, media release of the Minister for Gaming and Racing, David Harris, 25 August 2026, nsw.gov.au, on the statewide exclusion register, mandatory facial recognition at gaming room entry from 2028, and the $95.2 million over four years.