Ireland vs Israel: No Handshakes, Black Armbands as Ireland Wins 3-0
Lara Johns • Sep 28, 2026 
Reviewed By Paul Jacobs
Casino Expert
AI card counting just moved from theory to the lab, and the software monitoring it missed the trick. Oxford researchers asked two AI agents to count cards and share signals secretly. The agents invented their own code and hid it inside casual table talk.
Everything happened in a simulated lab game, with no real casino involved. Still, it shows why supervising AI keeps getting harder. CasinoUS explains what happened and what it means for anyone who plays blackjack online.
Inside the Oxford AI Blackjack Test
Christian Schroeder de Witt, an Oxford AI security researcher, led the project on collusion between AI agents. Collusion means two players secretly cooperating against the game.
To test AI card counting, researchers paired two agents in a lab-run game of blackjack. The “counter” tracked the cards, while the “big player” adjusted its bets based on the counter’s hints. Both agents knew a monitor was reading their chat, so they agreed on a code first.
How the Secret AI Card Counting Code Worked
Card counting means tracking which cards have already left the deck. When plenty of high cards remain, the player’s odds improve, so a counter bets more. Popular blackjack card counting systems such as Hi-Lo give each card a value of +1, 0 or -1.
The AI counter buried its count inside ordinary chatter. For example, one agent joked about the dealer’s hot streak. In their code, that joke passed along the upcoming card’s value plus a $250 betting cue.
The remark sounded like small talk, so the AI monitor never flagged it.
Human Card Counting Teams vs AI Agents
The setup mirrors a method human teams have used for decades.
| Feature | Human blackjack team | Oxford AI agents |
| Who counts | A spotter at a low-stakes seat | The “counter” agent |
| Who bets big | A separate high roller | The “big player” agent |
| How signals travel | Hand gestures and code words | Coded table talk |
| Who watches | Pit bosses and surveillance cameras | An AI chat monitor |
| How they get caught | Betting patterns and video review | Activity patterns in the model |
Teams like these rely on advantage play in blackjack, which uses skill rather than cheating to gain an edge.
How Researchers Caught the Colluding Agents
Reading the chat failed, so the team built a small detector model to read activity inside the agents. It flagged internal patterns tied to hidden AI card counting signals that the chat monitor missed. The team ran its NARCBench benchmark on mid-sized GPT-OSS, Qwen, DeepSeek and Llama models.
However, PhD student Carissa Cullen noted that larger models gave off weaker signals. The team plans to test bigger models next.
Does AI Card Counting Work at US Online Casinos?
Not in any useful way. Most online blackjack games use a random number generator (RNG) that reshuffles the virtual deck after every hand. Because of this, no count carries over between rounds.
By contrast, live dealer blackjack streams real cards from a studio. However, those tables typically deal from six to eight decks and shuffle frequently, which weakens any count.
Also, using bots or sharing card information with other players breaks the terms at licensed US casinos. That said, counting in your head stays legal. In Nevada, though, using a device to track cards counts as cheating a casino.
For better odds, a blackjack basic strategy chart helps more than any AI trick. At tables with favorable rules, it cuts the house edge (the casino’s built-in advantage) below 1%.
Why This Matters Beyond the Casino Floor
The bigger worry sits outside gambling. AI agents already search, shop and negotiate for people online. If two of them team up quietly, they could fix prices or spread false stories.
Stanford computer scientist Diyi Yang wants companies to track how agents behave together over time. Future networks may connect thousands of agents from different companies, making hidden deals far harder to spot.
The Bottom Line
The Oxford test shows AI agents can invent their own card counting code without being told how. For US players, AI card counting does little in RNG games, and bots break casino rules. Before playing for real money, set deposit and time limits with responsible gambling tools.
