By Tobias Nagel.
Blackjack seems tailor-made to be analyzed by artificial intelligence. Card values, probabilities, and statistical expected values can be calculated and compared with one another. An AI can process large amounts of game information that a human could hardly keep in memory permanently.
Whether this creates an advantage over the casino, however, depends on one crucial prerequisite: What information is even available? This is precisely why a distinction must be made between different variants of blackjack.
When Chance Comes from the Computer
In RNG blackjack, the results are software-based and determined by a random number generator. For regulated games, such as those offered by licensed operators featured on Casino Groups, requirements for randomness, statistical distribution, and unpredictability apply.
This is crucial for card counting. In physical blackjack, played cards disappear from the shoe. In an RNG game, however, a continuous card population whose composition changes through the previous round is not automatically created. Therefore, from a correctly generated random sequence, it is not possible to reliably infer the next outcome merely by analyzing past results.
Nevertheless, an AI can play blackjack very well. It can determine the statistically most favorable decision for every situation and consistently apply an optimal basic strategy. However, it cannot generate an additional information advantage about the next random value. An AI can calculate faster than any card counter—but it cannot generate information that is not present in the game.
The Special Case of Live Dealer
However, online does not automatically mean RNG. Live Casinos are also popular, according to industry experts like Casino Groups. In Live Dealer games, real cards are used at a real table.
This brings back a central characteristic of classic blackjack: a played card is no longer in the shoe. If many small cards are drawn, for example, the remaining stock contains relatively more high cards. The next card is not directly predictable as a result, but its probability distribution changes.
Whether this results in a practical advantage depends on the playing conditions. Multiple decks, the time of shuffling, and the number of cards actually dealt affect the significance of the card history. In Continuous Shuffling Machines, played cards are continuously returned to the card stock. This largely eliminates the basis of classic card counting.
Why Blackjack is Calculable
In blackjack, cards are drawn from a limited card stock. Every card played therefore changes the composition of what is yet to come. This changeable probability distribution makes the game mathematically interesting. Even for a single hand, it is possible to calculate which decision has the highest expected value under the given conditions. For a specific player hand and the dealer’s upcard, it can be compared whether hitting, standing, doubling down, or splitting is statistically more favorable.
Roger Baldwin, Wilbert Cantey, Herbert Maisel, and James McDermott published a mathematical study of the optimal blackjack strategy back in 1956. The Basic Strategy developed from it reduces the house edge, but does not eliminate it. It shows which decision is mathematically most favorable in a specific game situation. If the cards already played are also known, the probability distribution changes as well.
Edward O. Thorp made this idea widely known to the public in the early 1960s. His work built on earlier research on optimal blackjack strategy and became a milestone in the history of card counting.
No one needs to memorize every single card. Common counting systems weight different card groups differently. From the running count and the estimated number of remaining decks, it can be deduced whether the current composition has become more favorable or unfavorable. The counter is therefore not trying to predict the next card. They are looking for situations in which the statistical conditions have shifted and can adjust their bet accordingly.
A favorable count by no means makes a single hand a guaranteed win. Even with a positive long-term expected value, individual results remain random, and fluctuations are considerable. Card counting is therefore not a form of clairvoyance, but the utilization of a changed probability distribution.
A remarkable form of human collaboration later emerged from this mathematics. The MIT Blackjack Team focused on card counting and coordinated bets starting in the late 1970s. The use of computers is interesting in this context. Katie Byl, who belonged to the team’s environment, describes software for analyzing the profitability of various game situations, among other things.
What AI Could Do Better
An AI could completely store the cards dealt, update a count without calculation errors, and calculate expected values for different game situations. It could simulate numerous game progressions and adjust decisions to rules, card distribution, and risk.
Compared to a human counter, it would thus have an obvious advantage: no forgetting, no fatigue, and an enormous calculation speed. Machine learning and reinforcement learning methods are therefore also used to study blackjack strategies.
However, even a perfect AI remains bound to the available information. It can evaluate an existing statistical advantage more precisely. In the case of an independent random sequence, however, it cannot reliably calculate the next card from past results.
Card counting only works as long as the card history creates a usable difference for the remaining stock. Casinos can make this prerequisite more difficult: through multiple decks, earlier shuffling, or procedures like Continuous Shuffling. The monitoring of game and betting patterns is now also part of casino security.
The competition is therefore not just about who can calculate faster. The player tries to gain as much usable information as possible from the game; at the same time, the casino can design the conditions in such a way that as little advantage as possible arises from them.
Calculation Is Not Enough
AI can analyze blackjack precisely and process information about cards already played much faster than a human with a suitable physical card shoe. However, this does not result in a universal way to beat the casino.
Card counting was never the ability to predict the future. It was based on the fact that the probabilities in a card shoe that has not yet been dealt change due to the cards already removed. In a correctly working RNG game, this continuous card population is missing; in the Live Dealer game, it can be present, but is limited by the specific design of the game.
The crucial resource is therefore the combination of information, probability, and decision. AI can process this connection exceptionally efficiently. But it can only calculate with what the game actually.
IMAGE CREDIT: AI Generated