There is a particular kind of tipster that has learned to sound scientific. They talk about an "AI engine", "thousands of data points", "machine learning models" and a team of "data scientists with PhDs". And — to their credit — some of them publish their full record openly, every prediction logged win or lose, downloadable as a single file.
We took one of the larger ones at its word, downloaded its public dataset, and analysed 25,890 matches from a single year. Here is what the numbers say — all of them theirs, not ours.
The first thing you check with any model is whether it does anything a free price doesn't already do. So for every match we compared the tipster's pick against the simplest possible baseline: the bookmaker's favourite — the team with the lowest odds.
They matched 83% of the time.
In other words, four times out of five the "AI engine" simply picks whoever the bookmaker already made favourite. That is not a model finding hidden value. That is reading the odds and copying them.
It goes further. Each pick comes with a "trust" or "confidence" score, presented as the output of the algorithm. We checked it against the probability implied by the odds — and they track each other almost perfectly (correlation 0.61, rising higher on the confident picks). The "confidence score" is, to a good approximation, the bookmaker's own price wearing a lab coat.
Here is the number their marketing leads with: on their highest-confidence picks, they are right about 73% of the time. That sounds excellent. It is also worthless, and here is exactly why.
Those high-confidence picks are heavy favourites at an average price of 1.40. Do the arithmetic a bettor actually cares about: win 73% of your bets at odds of 1.40 and you collect 0.73 × 1.40 = 1.02 for every 1.00 staked — before the bookmaker's margin eats the rest. The result is a return of roughly −0.1%. You win almost every bet and you slowly go broke.
Across all 25,890 of their published picks, backing every one at the odds shown returns −5.6%. Not on our model. On theirs.
The most revealing test is the 17% of matches where the tipster's pick disagrees with the bookmaker favourite — the only place a real model could show an edge. If the algorithm knows something the market doesn't, this is where it earns its subscription.
It doesn't. On those matches the picks are right 28.7% of the time and return −9.95%. Backing the favourite blindly would have lost less. The "intelligence" subtracts value rather than adding it.
We are not naming the service. The point is not one company — it is the entire format. "67% accuracy", "AI-powered", "verified results": none of these numbers mean what they are designed to make you feel, because accuracy at short odds is not profit, and a model that copies the favourite is not a model.
This is also a mirror we are happy to hold up to ourselves. It is the reason TipsAudit does not sell an accuracy figure. We publish the fair price with the bookmaker margin removed, we flag value only where a softer bookmaker is genuinely mispricing a match, and the headline number on our track record is closing line value — the one metric that actually correlates with a long-term edge — shown on every pick, before kick-off, losses included.
The uncomfortable truth in the data above is that you could have audited that tipster yourself. The file was public the whole time. Most people never open it, because a wall of green screenshots feels more convincing than a spreadsheet. We think it should be the other way round.
Data: a public tipster results dataset, 25,890 matches with odds and outcomes from a single season, analysed for pick-vs-favourite agreement, accuracy by confidence band, and return on investment at the odds shown. Nothing here is betting advice; betting carries real financial risk. 18+, play responsibly.