Claims

AI and Machine Learning in Big Small Tools

Almost every product launched in the last two years describes itself as AI-powered. Some of them genuinely contain a model. This page is about what a model would need in order to help here, what it is actually given, and why the answer does not depend on how good the model is.

In one line

A model learns relationships in data. Independent rounds contain no relationship to learn, so the output is a confident guess.

What a model needs

Three Requirements, in Order

This is true of every predictive model ever built, from spam filters to weather forecasts. Nothing about the technology being new changes the list.

REQUIREMENT 01

Inputs that relate to the target

A model maps features to an outcome. If the features carry no information about the outcome, training does not fail loudly — it converges on the base rate and reports it with great confidence.

REQUIREMENT 02

A stable relationship over time

The pattern learned in training has to still hold at prediction time. In a generated sequence there is no relationship in the first place, so there is nothing to remain stable.

REQUIREMENT 03

Honest validation

Performance has to be measured on data the model never saw, in the order it would arrive. Almost every accuracy figure in this market is either from training data or from no data at all.

RESULT

A model that outputs about fifty per cent

Correctly trained on round history, a model learns that big and small are roughly equally likely and says so forever. That is the right answer, and it is not a product.

The data

What Is Actually Available to Train On

Two columns and a timestamp. Everything a prediction tool can see is on this list.

Available

The result history

Every past digit, as many as you like. Genuine data, freely published, and entirely a record of independent events that constrain nothing about the next one.

Available

Round timing

When each round opened and closed. Useful for scheduling, informative about nothing — the clock is the same for a winning digit and a losing one.

Not available

Bet distribution

How much was staked on each side. Not published, and in any case irrelevant to a generator that runs after the lock regardless of where the money went.

Not available

The generator's state

The seed, the algorithm, the moment of production. This is the only input that would matter, and it sits on infrastructure no consumer product touches.

Show me a model's validation curve on data it has never seen, in chronological order, and I will tell you what it learned. Nobody in this market has ever been able to produce one.

What to ask for

Accuracy graphs

How to Read an AI Product's Claims

Four things appear on nearly every one of these pages. Here is what each is worth.

A rising accuracy curve

Usually a drawing. Where it is real, it is nearly always training accuracy — the model's performance on the examples it was fitted to, which can be pushed arbitrarily high and means nothing about future rounds.

The honest version is a curve on held-out data in time order. It is a standard chart in any machine learning workflow, which is why its absence is informative.

A neural network diagram

Decorative. The architecture picture tells you nothing about inputs, training data or validation, and the same illustration appears across dozens of unrelated products.

“Trained on 10 million rounds”

Plausible and beside the point. Ten million samples of independent events give you a very precise estimate of a probability you already knew, and no leverage on any particular future round.

Live confidence percentages

Almost always cosmetic. A genuine confidence figure would be calibrated — of all the calls made at 80%, about 80% should land — and that calibration table is never published.

“Our AI adapts to each operator”

Adapting means retraining on that operator's history, which is the same uninformative data with a different logo attached. Nothing about the change makes the next digit visible.

The honest use

Where a Model Would Genuinely Help

There is real work for machine learning around this game. None of it involves the next digit.

Useful models in this space

Detecting problem-play patterns in a player's own session data — long sessions, rising stakes after losses, deposits at unusual hours.
Flagging fraudulent apps from permission manifests and package signatures, before installation rather than after.
Clustering signal-channel messages to expose split calls across a room automatically.
Estimating the expected hourly cost of a given stake and round speed, which is arithmetic a model can present clearly.

Why the AI Label Works So Well Here

Because it explains the unexplainable without needing detail. A pattern-based tool invites the question of which pattern; an AI-based tool answers by design, since nobody expects to understand what a neural network found. The label converts a gap in the argument into a feature of the technology.

It is also currently free. There is no certification, no audit and no obligation to publish a validation set, so the word costs nothing to add and demonstrably raises what people will pay.

One Question That Sorts the Market

Ask for the held-out validation results in chronological order, with the losing calls included. Any team that genuinely trained a model has this artefact, because they could not have evaluated their own work without it. In two years of asking we have not been sent one, and the replies are consistently about proprietary methods rather than about the chart.

If you want to produce your own answer instead of waiting for theirs, the hundred-round method on the testing page works on AI products exactly as it does on any other. The category overview is on the home page.

AI Claims FAQ

Not without access to the generator. The limit is the absence of information, not the sophistication of the method, and better models do not create data that does not exist.

It is a number the interface produces. A real confidence score is calibrated against outcomes, and no product in this category publishes the calibration that would demonstrate it.

Some do. A small classifier trained on round history is easy to build and will happily output varying percentages — its outputs simply have no relationship to the next result.

The animation is drawn after the result arrives from the server. Whatever it appears to reveal, it is a rendering of a decision already made elsewhere.

Yes — in fraud detection, responsible-play monitoring and app-safety screening. Those uses are measurable, which is precisely why they are not marketed with accuracy percentages.