By Stefan Almgren.
Ask a decent AI model to keep a Hi-Lo count and it will do it flawlessly, card after card, for as long as you keep feeding it. It doesn’t get tired, and it doesn’t fudge the true count because the division got annoying. So whether AI can count cards isn’t really a question anymore. It can. The more interesting question is what that means for online blackjack, and I think the answer is a real headache for casinos.
Some context on where I’m coming from. I learned to count years ago, and these days most of my blackjack is played at live dealer tables. The system most people use is Hi-Lo: 2 through 6 count +1, 7 through 9 count zero, tens and aces count -1. You keep a running total, divide it by the decks left to get the true count, and bet more when it climbs, because a shoe heavy in tens and aces pays more blackjacks and busts the dealer more often. It’s been a known way to beat the game since Edward Thorp published Beat the Dealer in 1962.
The part people skip is what sits underneath the count. The edge a count gives you is thin, around 1% even when things go your way, so it only works if every hand is already played perfectly for the table’s exact rules. For getting that right, I’d point anyone to the blackjack section at Beat the Fish. Alongside a run of strategy guides, they’ve built a proper toolkit: a strategy calculator that solves the right play for your table’s exact rules, a basic strategy trainer that deals you hands and grades every decision, a house edge calculator that shows what a table’s rules cost you, and an insurance checker that factors in the cards you’ve already seen leave the shoe. A dedicated card counting trainer is on the way too.
What AI does better than any human counter
Counting is simple arithmetic. What makes it hard for people is doing it fast, silently, for hours, while looking like you’re not doing it. None of that is hard for software.
A program keeps a perfect running count and works out the exact true count from the exact number of cards left, not from eyeballing a discard tray. It can track every card rather than just the Hi-Lo tags, so it always knows how many aces or fives are still in the shoe. A human counter memorizes a couple of dozen strategy changes for high and low counts. Software can simply solve each decision for what’s actually left.
Online, it doesn’t even need a camera pointed at the table. Studios read every card electronically so the game can show it on your screen, usually with the hand total next to it. GLI-19, the testing standard many regulators use, actually bars live casino platforms from giving players a real-time tracker of cards played or remaining, so the game won’t keep the count for you. But every card still crosses your screen, and software can keep its own.
I tried the low-tech version myself: typed out the cards from a few live shoes and asked a chatbot for the running and true count at random points. It was right every time I checked, and a lot quicker than me. It has already gone past chatbots, too. Paid tools that claim to read the cards straight off your screen and give you the count and the right play at the big live casinos are being advertised online right now.
The real problem is scale
Most live dealer tables deal from an eight-deck shoe and reshuffle somewhere around halfway, which is why counting live dealer has always looked pointless. I ran it through a simulation (eight decks, cut at 50%, Hi-Lo, a few million rounds). Only around one round in 15 starts at a true count of +2 or better. At a single table that’s three or four good hands an hour. Not worth the effort for a person.
On those hands, though, the player has roughly a 0.9% edge. That’s the house edge turned around.
Now picture a program watching 20 tables at once and only betting when a shoe turns good. Counters call this back-counting, or wonging, after Stanford Wong. At about 50 rounds an hour per table, that program finds roughly 65 good hands an hour. Flat betting $100 on each is worth around $58 an hour in expected profit. It’s not a money printer, and the swings are big (after 100 hours it would still be behind about a quarter of the time), but the direction has flipped. The casino is the one slowly losing, and the software never gets tired or sloppy.
Online live blackjack is close to built for this. At Evolution’s standard tables, a virtually unlimited number of players can bet behind the seven seated ones, and games like Infinite Blackjack or Pragmatic Play’s ONE Blackjack have no seat limit at all. Nobody needs a seat to get a bet down when the count is right. A back-counter in Vegas gets spotted hovering behind a table. A script doesn’t hover anywhere.
How casinos are fighting back
Casinos aren’t blind to this, and most of their defenses already exist.
The simplest is the shoe. Cutting shallower, shuffling more often or moving to continuous shuffling machines kills the count at the source, and you can’t always tell from the stream when a table uses one.
Then there’s the data. Every bet is recorded, and an account that sits out until the count climbs and then bets big leaves a very obvious trail. GLI-19 also requires operators to have procedures for detecting bots and unauthorized software, and it mentions card tracking by name. Casino terms back that up. Most ban bots and outside assistance software, some name artificial intelligence outright, and one set I read reserves the right to check which programs are running on your computer. Get caught and you’re looking at a closed account and voided winnings.
Land casinos have fought this battle before. Nevada makes it a felony to use any device or software that tracks cards at a casino game, and its Gaming Control Board urged casinos to ban Google Glass from the gaming floor back in 2013. With a new wave of camera glasses on the market, surveillance experts are now warning that card-recognition glasses can keep the count for whoever’s wearing them.
And casinos have AI of their own. Back in 2010, a California casino was already running a camera system called TableEye21 that rated each player’s skill and flagged counters. The newer generation goes further. Angel, one of the biggest names in table game technology, has been showing blackjack tables with cameras that read the bets, the cards and the players’ hand movements.
So it’s an arms race, and both sides are using the same technology.
What it means if you play
AI counting is real, and it scales in a way a human counter never could. That’s the new problem for casinos. It won’t break online blackjack overnight, because the casino still controls the shoe and the account, but the old idea that counting is too hard for most people to do properly is gone.
For regular players, the best use of all this is also the simplest. Use AI to learn: drill counts, quiz yourself, have it walk you through the math, and pair it with tools built for your exact rules so your basic strategy is flawless. That’s what improves your results at any table, online or off, and it’s the one use of AI no casino can ban.
IMAGE CREDIT: AI Generated