When the AIs Agree on a Horse, Does It Win More Often Than Its Price Says?

Bumaan had 5 of the Silicon Tipster League's AI tipster entries behind him in the Seadell Shops & Holiday Chalets Handicap at Yarmouth on 20 October, and he won at 16/1. Across 3,552 races, the League's record says agreement like that mostly tells you what the starting price already knew, with one middle-sized group a little ahead of its price and every group losing money at the starting price.

Bumaan's backers came from both sides of the experiment. ChatGPT, Gemini, Grok and DeepSeek chose him from the form alone, and Grok chose him again with the morning prices in front of it. Their reasons read alike: Billy Loughnane booked to ride, a trainer and jockey in form, soft ground to suit.

British Champions Day at Ascot on 17 October produced two more examples, one each way. ChatGPT, Claude and DeepSeek, all working from the form alone, landed on Fifth Column, who won the Balmoral Handicap at 22/1. In the juvenile race that opened the card, 7 entries backed Confucius, each of them pointing to Ryan Moore's booking and most to Aidan O'Brien's stable as well. He started at 20/1 and finished fourth.

Several AI tipsters on one horse could mean they have spotted something, or only that they are all reading the same big names off the racecard. We went through every race in the League's record to see which.

How we counted

Each chatbot in the League (ChatGPT, Claude, Gemini, Grok and DeepSeek) picks one horse in every race twice. The blind entry sees the form and no prices and is asked for the winner. The informed entry also sees the morning prices and is asked where it disagrees with the market. That makes 10 AI entries a race. Our own model's entries and The Favourite pick too, but they are reference points, and none of their picks counts towards agreement here.

The study runs from 17 July 2026, the first day both kinds of entry count on the League's board, to 21 October 2026. That is 3,552 races with at least one settled AI pick, an average of 9.5 AI picks a race once non-runners and void races are set aside. Picks logged after the off are left out exactly as the League leaves them out. For each race we counted how many entries backed each horse, and put every pick in one of four groups: backed by one entry alone, by 2 or 3, by 4 to 6, or by 7 or more.

There are two ways to count, and both matter. Per pick, every entry's £1 bet counts, which is how the League keeps its books. Per horse, each horse in each race counts once, because seven entries on one horse is still one race and one result, so the per-horse count is the one to judge a group's size by.

Then the test. A starting price is a chance: a horse at 4/1 is being given one chance in five. Add up those chances across a race and the total comes to more than one, which is the bookmakers' margin, so we divide each runner's chance by the total to take the margin out. Sum that over a group's horses and you have the number of winners the market expected. If agreement knows something the price does not, the group's horses win more often than that.

What agreement goes with

"Won" is the share of the group's horses that won (of those with a starting price on file), "market's chance" is the average starting-price chance of those horses with the margin taken out, and "SP favourite" is the share of them that started favourite, joint favourites included. The return is per pick at the starting price.

Agreement goes with a shorter price, up to a point. Horses backed by one entry alone had a median starting price of 9/1; those backed by 4 to 6 entries, 9/2. More entries on a horse meant more winners, until the top group, where the figure fell back.

Set each group's winners beside its price and most of that pattern goes away. Lone picks won 10.6% of the time against a market chance of 10.9%, and the 2 or 3 group 13.8% against 13.5%, so for those two groups the price explains the extra winners.

The 7 or more group is where the entries herd hardest, and its horses started favourite only 13.0% of the time, about half as often as in the group below. Confucius fits the pattern: a big-name jockey and stable at a decent price, a horse the models liked and the market did not make favourite. Those horses won 14.9% of the time against a market chance of 15.4%, so the price had them about right. It is also the smallest group, 418 horses, and one or two winners either way would move it.

Every group lost money at the starting price. For comparison, The Favourite, the League entry that backs the morning favourite, won 28.2% of its bets in the same races and returned -12.4%.

The middle group, and a quirk of the prices

One group stands out. Horses backed by 4 to 6 entries won 21.7% of the time against a market chance of 19.6%: 553 winners where the prices expected 499.3.

Part of that comes from the prices themselves. Across every runner in these races, short-priced horses win a little more often than their margin-free starting price says, and long shots a little less. Runners sent off shorter than 3/1 won 36.6% of the time against a market chance of 33.9%; runners at 10/1 and bigger won 3.2% against 3.5%. Racing calls this the favourite-longshot bias. The middle group leans towards shorter prices, so it would beat its plain price figure even if the AIs added nothing.

Compare each of these horses with other runners at similar prices in the same races and you would expect about 515.7 winners. The group had 553, about 37 more than that. It is a small gap, the size chance produces now and then, and with four groups to look at, one of them was always likely to show something.

It also lost money: -8.7% per pick at the starting price over 12,019 picks. The Favourite returned -12.4% in the same races and stands at -11.3% over its full League record. Losing a little less than the favourite is still losing, and we will keep checking whether the gap holds as the record grows.

Blind entries herd, informed entries scatter

Split the entries by what they could see and most of the herding is on the blind side. With the form and no prices, the blind entries backed 2.1 different horses a race on average, and all of them agreed on one horse in 934 of 3,506 races. The informed entries, asked where the market is wrong, spread across 3.1 horses a race and were unanimous in 94.

Agreement means different things on each side. When 4 or 5 blind entries agreed on a horse, it started favourite 35.4% of the time. Those 1,716 horses won 26.3% of their races against a market chance of 23.2%, about 35 more winners than runners at similar prices would give. Most of them are the same horses as the 4 to 6 group above, so this is the same small gap seen from another side, and chance could explain it here too. They still returned -4.6% per pick.

The informed side went the other way. When 4 or 5 informed entries agreed, the horse started favourite only 5.7% of the time. Those 477 horses won 10.3% against a market chance of 12.8% and returned -33.1% per pick, the furthest below its price of any group we measured. It is a smaller sample, but so far, when most of the AIs that can see the prices think the market has a horse wrong, the market has had the better of it.

The two sides' most-backed horse in a race was the same horse in 336 of the 1,186 races where each side had a clear one, backed by 3 or more of its entries.

Whether seeing the prices helps an AI or costs it is the League's standing question, and every pair is set side by side on the value test, with the background in what changes when an AI sees the odds. How the entries split at Ascot is in the Champions Day tipster league review, and the same test on our own model's long-shot picks is in why our model liked Mount Atlas. Every pick, win or lose, is on the Silicon Tipster League. The study's figures come from a frozen record built from the League's day-files and will not change; the League's own figures are live.

When the AIs agree: FAQ

Do the AI models agree with each other on their picks, or do they disagree?

Mostly they disagree. Across 3,552 races the AI entries made 9.5 settled picks a race on average and spread them over 4.1 different horses. Seven or more of them landed on one horse in 418 races. The blind entries, which see no prices, agree far more often than the informed ones.

Are the AI models just backing favourites, or picking against the crowd?

Neither, quite. About one AI pick in five was on the starting-price favourite (18.5% of all picks), more than picking a runner at random would give but far from backing the favourite every time. The blind entries put 28.6% of their picks on the favourite and the informed entries 8.4%, because the informed entries are asked to find where the market is wrong.

When several AIs pick the same horse, is it more likely to win?

More likely than a horse only one of them picked, because it tends to be shorter in the betting. Against its own price, hardly: horses with 7 or more backers won 14.9% of the time against a market chance of 15.4%. The 4 to 6 group was the only one to beat its price by more than a few winners, a gap chance could still explain, and it still returned -8.7% per pick.

Should I actually trust an AI's tip over the bookmaker's own odds?

Not on this record, and nothing here is a tip. Every agreement group lost money at the starting price, and lone AI picks returned -26.0% per pick. The Favourite, which simply backs the market's choice, returned -12.4% in the same races.

How do I know if a tipster or model is actually any good, or just got lucky for a few weeks?

Look at how big the record is and what the prices expected. Bumaan was one winner. This study covers 14,654 horses, and even there the one gap we found is small enough that chance could explain it. Each entrant's full record, with every pick, is on the Silicon Tipster League.

Bet responsibly. Nothing on this page is betting advice. If you bet, stake only what you can afford to lose, and if it stops being fun, stop. Help is at BeGambleAware.org. 18+.