Why basic strategy alone isn't enough: what card counting actually does and why AI could do it better

May 24, 2026 · guides

Most people who play blackjack seriously will tell you to learn basic strategy. They’re right. Basic strategy cuts the house edge from 2–4% down to around 0.5%, which is genuinely meaningful over time. What they usually skip is the part where 0.5% still means the casino wins. Over thousands of hands, a negative expected value stays negative, no matter how small.

The question worth asking isn’t whether basic strategy works. It does. The question is: is there information available in the game that basic strategy ignores?

There is. It’s the composition of the remaining shoe. And reading that composition accurately is where the real edge lives, and where human counters consistently leave money on the table.

What basic strategy actually optimises

[Observed] The established math here traces to published research from the 1950s and 1960s; the numbers below are drawn from those primary sources.¹

Basic strategy is a complete decision matrix. For every combination of player hand and dealer upcard, it prescribes the statistically optimal play assuming the remaining shoe has average composition.

That last phrase is doing a lot of work.

The matrix was derived mathematically starting in the 1950s. Roger Baldwin and his collaborators worked it out with adding machines and published their results in 1956.¹ Edward Thorp improved on it and published Beat the Dealer in 1962.² The math hasn’t changed. At a typical 6-deck game with standard rules, playing every hand by the matrix reduces the house edge to roughly 0.5%.¹²

The catch: basic strategy assumes the remaining deck is in its average state. It never is. Every card dealt changes the probabilities for every subsequent hand. A shoe that’s burned through thirty low cards is a fundamentally different game than a fresh shoe, but basic strategy plays both identically.

That’s not a flaw in basic strategy. It’s a deliberate constraint. The matrix is optimal given no information about the remaining shoe. Card counting is the method for using the information you actually have.

The count: what the math actually does

[Observed] The Hi-Lo counting system and the edge estimates below are from Thorp (1962) and refined in subsequent published analyses.²³

The Hi-Lo count assigns a value to each card: +1 for low cards (2 through 6), -1 for high cards (10 through ace), and 0 for neutral cards (7 through 9). As cards are dealt, you track the running total.

A positive running count means more low cards have left the shoe than high cards. The remaining deck is relatively rich in 10s and aces. That matters for three reasons: blackjacks become more likely, the dealer (who must hit until reaching 17) busts more often, and doubles and splits become higher-value plays.

To account for how much of the shoe remains, you convert the running count to a true count: divide the running count by the estimated number of decks remaining. The true count normalises for shoe depth, which is why it’s the number that actually tells you where the edge sits.

At a true count of around +2, depending on specific rules, the edge shifts from the house to the player. Below that, you play at a disadvantage. Above it, the math is in your favour.

The practical implication is bet sizing. At negative counts, you bet the minimum. At high positive counts, you bet the maximum. The edge doesn’t come from playing hands better than the house; the house rules are fixed. The edge comes from betting more when the shoe favours you and less when it doesn’t.

To put a number on it: a competent Hi-Lo counter at a 6-deck shoe with 75% penetration can achieve an edge of roughly 0.5% to 1.5% over the long run.³ Not dramatic, but consistently positive.

Where the strategy matrix actually breaks

[Observed] The index plays below are drawn from Schlesinger’s Blackjack Attack, the standard reference for Hi-Lo index values and their mathematical derivation.⁴

Basic strategy has one decision matrix. Card counting adds around 18 departures from it, known as index plays.

At a true count below zero, you hit 16 against a dealer’s 10. At a true count of zero or above, you stand. At a true count of +3 or more, you take insurance, which basic strategy says to never take, and which is actually a profitable play when the deck is rich in 10s.⁴

These aren’t guesses. They’re the same probability math that produced basic strategy, applied to a non-average shoe. When the count says the index play is correct, taking it adds expected value. Missing it costs you money you were mathematically entitled to.

The Illustrious 18 and Fab 4 (the most commonly cited index play sets, as defined in Schlesinger’s Blackjack Attack⁴) are learnable. Most dedicated counters know them. The problem isn’t memorising them in isolation; it’s executing them correctly while simultaneously tracking the count, managing bet sizing, and not looking like you’re doing any of it.

The execution problem

[Observed] The documented failure mode is count drift under cognitive load. The 5–10% per-card miss rate figure is from practitioner accounts rather than controlled studies; published research on amateur counter accuracy in live conditions is sparse, and the figure should be read as an industry estimate rather than a precision measurement.⁵

Hi-Lo is straightforward when practiced in isolation. Maintaining an accurate true count at a live dealer table, across 60–80 hands per hour,⁶ while managing index plays and bet sizing and keeping your behaviour inconspicuous: that’s a different thing.

Practitioner accounts of recreational counters consistently report miss rates of 5–10% on individual cards.⁵ A 5% miss rate doesn’t sound catastrophic, but the count compounds. If you’re off by 2 cards on a true count of +3, you might be playing a +1 shoe as though it’s +3. The bet you put out is wrong. The index plays you make are wrong. The edge you thought you had evaporates.

Casino countermeasures make this harder. Early shuffles, distractions, increased deck penetration, multiple dealers at once. All of these are specifically designed to degrade counting accuracy. Professional counters deal with them. Most recreational players don’t.

There’s also the cognitive load problem separate from accuracy. Counting while playing while acting like you’re not counting is uncomfortable. People make worse decisions under sustained attention load. The variance in blackjack is high enough that a few hundred hands isn’t a useful sample; you need thousands. Maintaining count accuracy over thousands of hands, across multiple sessions, is genuinely difficult.

What an AI advisor changes

[Observed] The technical constraints on browser-based video access described in this section are confirmed by the WebRTC and MediaStream APIs as documented by MDN; specific platform behaviours are from our own testing.⁷

[Speculation] The accuracy target and practical advisory workflow described below are our design goals, not measured results. We have not yet completed production-level testing on live dealer streams.

The tool we’re building watches the same video feed you’re watching. It counts every card, maintains the true count, identifies the current optimal play from both the basic strategy matrix and the index play set, and tells you the recommended action and bet sizing for the current shoe state.

You make every decision. The tool does the arithmetic.

The accuracy target is 99%+ card recognition from a compressed live video stream. That’s the hard part. Live dealer video comes in at 720p or 1080p, often with MPEG compression artefacts, cards at varying angles, occasional occlusion, and lighting that shifts between dealers. Building a CV model that hits 99% under those conditions is a genuine engineering problem, not a solved one.

We know this because we’ve started testing it. The next post covers what the browser actually allows: which permissions exist, what WebRTC gives you access to, and where the hard stops are. Some platforms cooperate. Others don’t. The answer depends on how the stream is delivered and what the browser sandbox will permit.

The advisory framing matters for another reason: it’s substantially different from automated play. A tool that watches and suggests is not the same as a tool that bets on your behalf. Whether that distinction holds up legally across different jurisdictions is a separate question, and one of the articles in this series covers it directly (UKGC, MGA, Curaçao; the rules differ materially).

Is card counting assistance worth building?

[Observed] The card-counting math has been public since 1962.² The claim that execution difficulty is the binding constraint on recreational counter profitability is consistent with the practitioner literature, though controlled studies are limited.

[Speculation] Whether offloading the arithmetic to a reliable tool translates to positive expected value in practice depends on accuracy under production conditions we haven’t yet measured.

The math behind card counting has been public for sixty years. The reason most people don’t extract value from it isn’t that the strategy is secret or complex. It’s that accurate execution at speed, under pressure, over thousands of hands, is a skills-intensive task that most people can’t sustain.

If you can offload the arithmetic to a reliable tool, the game changes. The math doesn’t change. The execution barrier comes down to watching a screen and following a recommendation.

Whether that works in practice at production accuracy rates on real live dealer streams is exactly what we’re here to find out. The next post starts with the browser.

If you’d rather follow along than check back manually, the newsletter is the fastest way to get each update.

What we don’t know yet

The open questions we’re carrying into the next phase:

• Whether 99%+ card recognition is achievable at live dealer streaming quality (720p MPEG) across different platforms and lighting conditions
• Whether the advisory framing (watch and suggest, not bet) is legally distinct from automated advantage play in UKGC and MGA jurisdictions
• Whether count accuracy at 60–80 hands/hr live dealer pace is meaningfully different from accuracy at 200+ hands/hr RNG pace under the same CV pipeline

These aren’t rhetorical. They’re the engineering and legal questions the series is designed to answer.