Not One AI Backs Scandinavia in the Irish St Leger

Settled. All four Group 1s have run. Scandinavia won the Irish St Leger by two lengths, extending his unbeaten run to eight and adding an Irish St Leger to the Doncaster one he won twelve months ago. The favourite also won the Moyglare, the Flying Five and the Blandford; only the National Stakes went the other way, to an 11/2 shot. Results are in each section below, and the closing section has the full settlement.

Five AI tipsters have picked every race in today's four Curragh Group 1s, and not one of them backs Scandinavia, the 2/5 favourite for the Irish St Leger. Going against a short-priced favourite has been an expensive habit for these AIs, so today is a live test of it, published before the off and settled at starting price tonight.

The picks come from the Silicon Tipster League. Each of the five large language models (Claude, ChatGPT, Gemini, Grok and DeepSeek) picks twice per race: a blind pick made without seeing any prices, and an informed pick made with the morning market in front of it. That is ten AI picks per race. Three more entrants sit alongside them: our model's value pick, our model's most likely winner, and a benchmark that simply backs the morning favourite.

Across the four Group 1s, those ten AI picks land on the favourite four times out of forty, and all four are in the Moyglare. In the Flying Five, the National Stakes and the Irish St Leger, no AI backs the favourite at all.

This is a dated, pre-race record. It is not a set of selections, and nothing here is advice to oppose anything. Every entrant in the League is losing money over its whole run, as the closing section sets out.

The AI Lab's Irish coverage was restored this morning, which is why today's Curragh card is here. Prices quoted below are the morning snapshot and will have moved, and the going in this morning's data is Good.

14:50 Moyglare Stud Stakes

A seven-runner Group 1 over seven furlongs for two-year-old fillies, with Aidan O'Brien's Sun Goddess the evens favourite in the morning market.

This is the one Group 1 where the AIs split cleanly, and the split follows the two arms. Claude, ChatGPT, Grok and DeepSeek all back Sun Goddess when they cannot see the prices. Show the same four the market and all four move to her stablemate Alpha at 5/1. Gemini is the exception both ways: Alpha blind, then Dancing Destiny at 10/1 once it sees the odds.

So the count is Alpha five AI picks, Sun Goddess four and Dancing Destiny one. Sun Goddess also has the favourite benchmark and our model's most likely winner, while our model's value pick is Ibelieveicanfly at 25/1.

The pattern is a familiar one. A blind pick tends to be the horse with the best form on paper, which is usually the favourite. An informed pick is often a price argument, a horse the AI thinks is better than its odds suggest. That is a different question from which filly is most likely to win.

Our model has Sun Goddess top as well, but with little more than half the chance the market gives her.

Result: Sun Goddess won by six and a half lengths, from Green Empress, with Dancing Destiny third and Ballinea Star fourth. The favourite benchmark and our model's most likely winner both had her. Alpha, the horse every informed AI moved to once it saw the price, finished outside the first four.

15:25 Flying Five Stakes

Eighteen sprinters over five furlongs, and the widest disagreement of the afternoon. Donnacha O'Brien's Comanche Brave is the 5/2 favourite.

Thirteen picks go to six different horses, and no horse gets more than three. Claude and DeepSeek, blind, both land on Big Gossey at 50/1, which is also our model's value pick. ChatGPT, Gemini and Grok, blind, all take Mission Central at 10/1. With prices in view, Claude, DeepSeek and Grok go to American Affair at 8/1, ChatGPT to Rumstar at 6/1 and Gemini to Time For Sandals at 9/2.

Comanche Brave is picked twice, by the favourite benchmark and by our model's most likely winner. Neither is an AI tipster.

Our model's read explains why it has so little to add here. It does rate the favourite top, but the market makes Comanche Brave well over twice as likely to win as the model does, and the model's first four sit almost level with each other. In an eighteen-runner sprint the model is barely separating the front of its own list.

A big field with a scattered set of picks is the ordinary result of a race that is hard to call. It tells you the AIs disagree with each other, and nothing about which of them, if any, is right.

Result: Comanche Brave won at 5/2, backed by neither AI arm, only the two internal benchmarks. American Affair ran into second, level on three picks with Big Gossey and Mission Central for the most of the afternoon. Those other two did not make the first four. Rumstar, one pick, was third; Starlust, no picks anywhere, was fourth.

16:00 National Stakes

Eight two-year-olds over seven furlongs. Joseph O'Brien's Monogram is the 2/1 favourite.

The blind arms agree completely here. All five AIs, picking without prices, choose Giant Sequoia at 7/2. Give them the market and the agreement falls apart: DeepSeek and Grok go to Haffner at 4/1, Gemini to Dr Rascal at 9/2, ChatGPT to Folsom Blues at 12/1 and Claude to Trean at 33/1.

Monogram gets one pick, from the favourite benchmark. No AI tipster picks Monogram in either arm.

Our model's most likely winner is Haffner, and its value pick is Porto Vecchio at 16/1. Its figures for all eight runners sit close together, and it gives Monogram less than half the chance the market does.

Five blind picks on one horse is the kind of consensus that looks strong on a page. It is five models reading much the same public form in much the same way, which makes the agreement a sign that the form points one way, and the market has priced that form already.

Result: Folsom Blues won at 11/2, from Monogram, with Florida Bay third and Trean fourth. Neither the five-model consensus nor the favourite had it. The horse that did, ChatGPT's informed pick, was picked by exactly one arm of one model. Trean, Claude's informed pick, ran fourth of eight.

16:35 Irish St Leger

Seven stayers over a mile and six furlongs. Aidan O'Brien's Scandinavia is the 2/5 favourite, the shortest price of any horse in today's Group 1s.

Eight of the ten AI picks go to Scandinavia's stablemate Jan Brueghel at 5/1. Claude, Grok and DeepSeek pick Jan Brueghel blind, and all five pick the same horse once they see the prices. The other two blind picks are Queenstown at 10/1 for ChatGPT and Arabian Force at 33/1 for Gemini. Scandinavia is picked twice, by the favourite benchmark and by our model's most likely winner. Our model's value pick is Leinster at 50/1.

The informed picks give their reasons, and each one is a price argument. All five say, in their own words, that the market underrates Jan Brueghel. That is a claim about value at 5/1, which is a separate question from which horse is most likely to win.

The card above shows a large gap against Scandinavia, and that gap is easy to misread as the model siding with the AIs. The model rates Scandinavia top, but at under a third of the market's chance, and its figures for the whole field run from roughly a tenth to under a fifth. It is barely separating seven horses in a Group 1, which is a known weakness at this level: it is short of evidence on top-class horses and spreads the chance too evenly. The gap between its figure and the market's is a calibration failure, and it points to nothing the market has missed.

The record says so in every window we publish. Across 10,702 races the market's prices score a Brier of 0.090 against the model's 0.097, where lower is sharper, and the market also has the better log loss over the last 30 days, the last 90 and all time.

History is against the AIs here too. In the races where the favourite started shorter than 1/2, the AI picks that went against it have lost money far faster than the favourite itself did, on a small sample. Today adds one more race to that record.

Result: Scandinavia won by two lengths, wearing down his own stablemate Queenstown, with Jan Brueghel, the AI panel's near-unanimous pick, third, and Trustyourinstinct fourth. It was Scandinavia's eighth straight win, his fifth at the top level this year, and a unique Doncaster-and-Irish St Leger double completed twelve months apart. Eight of the ten AI picks backed the wrong horse. The favourite benchmark and our model's most likely winner, neither of them an AI tipster, both had this one.

How it settled, and what it does not prove

Every pick above has settled at starting price, one point level stakes, and the League table has moved with it. Eight AI picks were on Jan Brueghel in the Irish St Leger, at a starting price of 6/1. He ran third. The picks that opposed the shortest-priced favourite of the day lost, on this one race, exactly as the record said they usually do.

Keep it in proportion in the other direction too. The favourite won four of today's five pattern races, and an AI arm still found the National Stakes winner at 11/2 when the market's favourite and the five-model consensus both missed it. One good day for the favourite is one day, the same as one good pick is one pick. The League has run for 96 days and settled 40,115 picks, and on that full record every entrant is losing money at starting price. The benchmark that only ever backs the favourite stands at -11.6%. The five AIs stand at -17.3% for Claude, -18.7% for ChatGPT, -15.2% for Gemini, -15.3% for Grok and -18.3% for DeepSeek.

That is the reason this page exists. The League is a public record of how hard the market is to beat, kept in the open, losing picks included. It is not a tipping service, and we had no selection of our own in any of these races.

For the full standings, see the League table. For how our model has done against the market over time, see the track record. The full Irish Champions Festival results are on the results page.

Common questions

Why do none of the AIs back Scandinavia?

The informed picks all say the market underrates Jan Brueghel, so they are backing a price they think is too big. Three of the blind picks reach the same horse from the form alone. None of them is a claim that Scandinavia will lose.

Did our model think Scandinavia would get beaten?

No. The model rated Scandinavia top, same as the market, and Scandinavia won. The model's figure was still far below the market's, because it spreads the chance too evenly across a Group 1 field, and the market has been the better forecaster over every period we measure.

Should I oppose the favourite because the AIs do?

No. Every League entrant is losing money over its full run, and in races with a short-priced favourite the AI picks that opposed it have lost money faster than the favourite itself.

When are the picks made?

Early on race morning, from the declared field and the morning prices. Nothing is changed after a race is run.

Where are the results?

The picks have settled at starting price and the League table has updated with them. Scandinavia won the Irish St Leger; the closing section above has the full settlement.

A word on all of this

Nothing here describes a way to win. If your betting has stopped being fun, BeGambleAware has free, confidential help and the National Gambling Helpline is on 0808 8020 133. 18+.