The AI Lab · The AI, tested
Can our AI model beat the bookies?
Our model is not quite as sharp as the bookmakers. Almost nothing is, that is what the house edge buys them. But it reads a race well, better than the hunches and tips most punters go in with. The real question is how close a good, honest model gets, so we test it the way a real punter would.
Every day with racing on, we back the model's three most-liked picks at £10 win each (that is £30 a day), logged before the off and settled at Starting Price. The running total below is the answer as it builds, race by race. Watch how it does.
Research, not tips. Paper-settled honestly to industry SP, wins and losses alike. 18+ · please gamble responsibly.
The running record
0 bets settled · 1 day on the recordNo settled bets in this window yet. The graph and numbers fill in automatically as the picks settle at Starting Price each evening — check back after the next card.
Profit
—
ROI
—
Strike rate
—
Bets
0
Staked
—
Returned
—
The profit graph draws itself, live.
Each AI's running profit at £1 a bet appears here the moment the first races settle.
The bets, as they were logged
Every pick below was written to the record before the off, with the model's own one-line case. Newest first.
| Date | Race | Pick | Model chance | The model's case | Result |
|---|---|---|---|---|---|
| 2026-07-09 | Carlisle14:35 | Mohaymenah | 28% | Two runs on the record, most recently second, with a rating of 85 to show from those outings. The model's estimate of 28 per cent runs notably below the market' | pending |
| 2026-07-09 | Doncaster15:20 | Golden Flame | 36% | A sixteen-run career showing a recent win, with a rating of 62. The market's 59% probability sits well clear of the model's 36% estimate at evens. | pending |
| 2026-07-09 | Newmarket14:25 | Adaay Of Scarlett | 41% | Four runs on the record with a win at Newmarket and a second last time out, rated 105. The model's estimate of 41% sits eight points clear of the market's 33%, | pending |
Three £10 win singles a day on the model's highest-rated picks, settled at industry Starting Price. A short card can mean fewer than three bets (never topped up); a non-runner is void with the stake returned, never substituted. Totals sum the selected timeframe. This is research, not tips — nothing here is a signal to stake.
Close to the bookies, ahead of the average punter
Our AI Race Predictor rates every runner in every UK and Irish race, and it is well-calibrated. When it says 25%, that horse wins about a quarter of the time. On the numbers that grade a forecaster honestly it lands just behind the betting market, the sharpest public forecaster there is, and well ahead of the hunches and tipster lines most punters bet on.
But reading races well and beating the bookies are two different things. The bookmaker builds a margin into every price, so a forecaster can be this sharp, pick winner after winner, and still come up short at the payout. That margin is the house edge. The question worth watching is how close a good model gets, and whether its read beats what you would manage yourself. This page settles it the only way that counts: real stakes, real prices, a running total you can watch.
We follow it the way a real punter would. Each morning the model's three strongest fancies are logged at £10 win apiece, before any race is run, and each evening they settle at the official Starting Price. No cherry-picking, no re-picks, no quiet deletions. The record above is every bet since the experiment started, the good days and the bad.
The same model competes race-by-race at £1 stakes in the Silicon Tipster League against ChatGPT, Gemini, Claude, Grok and DeepSeek, and its full every-race history is on the track record page. This page is the realistic-stakes version of the same question.
The rules, in full
- Three picks, £10 win each.The model's top-rated runner in each race, races ranked by the model's win chance; the three strongest make the day's card. Ties break on the model's edge versus the market.
- Logged before the off. Picks are timestamped and written to the public record in the morning, then never edited. A short card means fewer bets, never substitutes.
- Settled at Starting Price.The same convention as every Lab study: SP, level stakes, dead heats at the industry sp÷n, fallers counted as the losses they are.
- Non-runners are void.The £10 comes back and stays back. A real morning bet couldn't chase a replacement either.
The model's other exams
Questions
Can an AI model beat the bookies at horse racing?
That is exactly what this page tracks, live and in public. Our model reads a race well, but beating the bookies means being paid enough when you win to clear their built-in margin. That is a lot harder than naming likely winners. So far, like every method we have tested, it does not beat the bookies over a real sample. We show the running result rather than claim an answer.
How are you testing whether it beats the bookies?
The realistic way a punter would follow it. Every day with UK or Irish racing, the model's three most-liked picks (the runners it gives the highest win chance) go on at £10 win each, at Starting Price. That is the £30 a day a real punter following the model would stake. Every bet is logged before the off and settled honestly, wins and losses alike, so the profit-or-loss line is the answer.
Is this real money?
No. The bets are settled on paper at industry Starting Price, the same convention as every other experiment in the Lab. Paper settlement at SP is the fairest, hardest-to-flatter test: no cherry-picked prices, no hindsight, and nothing riding on the result except the truth of the record.
Why three bets at £10 rather than every race?
Because that is how a real punter follows a tipping source: a fixed daily budget on the picks it likes most, not £1 scattered across thirty races. The model's every-race record runs separately, at £1 on all its picks in the Silicon Tipster League and £10 a race on the track record page. This page answers the realistic version of the question.
What happens with non-runners and short cards?
A non-runner is void: the £10 comes back and is never moved to another horse. The picks are locked before the off, exactly as a real ante-post £30 would be. On a day with fewer than three races the model backs what exists, and the day simply stakes less than £30, recorded honestly.
Will following the model make money?
We expect not, and this page exists to show it plainly. The model is well-calibrated, so its win chances mean what they say, but the bookmaker's margin is bigger than its edge, and our full track record shows it losing at SP over thousands of races. If it ever does beat the bookies over a real sample, the data on this page will say so. Either way, nothing here is betting advice.
Gamble responsibly. This page is research and entertainment, not betting advice. The experiment exists to show what following tips with a daily budget really does over time, and betting should never be a way to make money. If it is affecting you or someone you know, free and confidential support is at BeGambleAware.org. 18+.