The short, honest answer
Five leading AIs, ChatGPT, Gemini, Claude, Grok and DeepSeek, can all read a racecard in seconds and explain every pick in plain English, but across our live tests none has yet shown a proven edge over the betting market, so the honest use is as research, not tips.
The rest of this page shows the working. We ran the same five AIs on the same real British and Irish racecards for weeks, scored how well each one reads a race, and set out how to actually use them. We are now 19 days into the experiment, across 5,034 settled picks, and every figure below is live, pulled straight from the boards so it can never drift from what we publish.
One racecard, five minds
Hand the identical racecard to five different AIs and you do not get one answer. You get five. Each was built by a different lab, trained on different data and nudged by different instincts, and it shows the moment they start justifying their picks. That variety is the fun of the experiment, and it is the reason a blend of the five tells you more than any single voice does.
- ChatGPT, the crowd-reader. With the odds hidden it drifts toward the shorter prices, backing the outright favourite 30% of the time, and its blind picks most resemble the "back short-priced favourites" rule.
- Gemini, the contrarian. Even blind it wanders furthest from the chalk, its picks mapping closest to "favourite over jumps", with a typical price around 5.74.
- Claude, the careful reader. Form-led and choosy, it leads with the yard and recent form, its top three stated reasons being trainer, form and jockey.
- Grok, the value-hunter. The moment it sees the market its picks swing toward "our own value strategy", chasing overpriced runners rather than the well-fancied.
- DeepSeek, the challenger. Efficient and blunt, it backs the favourite 29% of the time blind, then flips toward "back the outsider" once the prices are shown.
None of that makes one of them the answer. It makes them five different reads on the same race, and the pages below take each one apart in turn.
How we tested, and what reading a race well means
The method is simple, and we run it the same way every day. We take live British and Irish racecards and put the same races to all 8 AIs in two conditions. In the "blind" condition the odds are hidden and the AI has to read the form itself. In the "informed" condition it sees the market and hunts for value. Every pick is settled honestly to Starting Price, and we publish all of it, wins and losses alike, across 5,034 settled picks so far.
Here is the key move for this page. We do not rank the five on a few weeks of profit. A short-run return bounces around with luck, and any table built on it would contradict the live boards the next morning. Instead we score racecraft, how well each AI reads a race, on three durable pillars.
- Calibration. When an AI says it is confident, is it right that often? A model whose "sure things" actually win more is reading the race, not guessing.
- Reasons that match the pick. Do the written justifications line up with what it actually backed? An AI that talks about the yard and the form, then backs the horse that fits that story, is reasoning, not rationalising.
- Agreeing with the market when it is right. On the races it gets right, does it side with what the crowd already knew? The betting market is very good, so siding with it on winners is a sign of a sound read, not a lack of nerve.
Racecraft is about reading skill, not a promise of profit. The composite weighting is fixed and shown in the methodology. It is the honest, ages-well spine of the ranking, because a model can read a race well and still hand the bookmaker its margin at the window.
As a worked example of the first pillar, here is one AI's calibration: when Claude sorts its blind picks into confidence tiers, do the high-confidence ones actually win more often?
The table below ranks the five on racecraft. The order here follows the cleanest single reading signal we can show live, how often each AI's blind pick actually won its race, and the fixed composite that blends all three pillars is set out in the methodology. The five sit close together, and the honest yardstick is in the "market strike" column: the crowd, backing the same races, wins more often than any of them.
| Rank | AI | Blind strike | Market strike, same races | Favourite affinity (blind) | Reasons-match note |
|---|---|---|---|---|---|
| Top of the five | ChatGPT | 25.3% | 45.1% | 30% | Leads with form, and backs the runner that fits it. |
| Second | DeepSeek | 24.8% | 44.4% | 29% | Leads with form, blunt and efficient. |
| Third | Claude | 24.4% | 46.3% | 31% | Leads with the trainer, then recent form. |
| Fourth | Grok | 23.9% | 45.1% | 29% | Leads with the trainer, then jockey and form. |
| Fifth | Gemini | 23.8% | 44.9% | 25% | Leads with form, furthest from the chalk of the five. |
Read the ranking as where each AI sits on one durable reading signal, not as a verdict on which is worth backing. That question belongs to each AI's own live scorecard, linked in the next section.
The blind herd, and the odds-flip
Here is the most interesting finding, told straight. With the odds hidden, the five AIs herd. They all drift toward the shorter-priced, better-fancied runners, arriving at roughly where the market already is without ever being shown it. That is a quiet compliment to the crowd: the AIs independently rediscover the favourites. Blind, ChatGPT, Claude and DeepSeek all map closest to short-price, odds-on behaviour, and even Gemini, the one that wanders furthest, still lands well inside the front of the book.
Then show them the prices, and several of them flip. Their picks swing out toward the longer-priced, overlooked runners as they start hunting value instead of winners. It is a real change of character, and you can watch it happen in the two distribution charts below: the blind herd bunched at the short end, the informed picks spread out across the book.
The same story shows up in how tightly the five agree with each other. Blind, they cluster on the same well-fancied runners. Informed, the agreement loosens as each one chases its own idea of value.
Two named examples make the odds-flip concrete:
- ChatGPT. Blind, it backs the favourite 30% of the time and resembles "back short-priced favourites". Informed, it shifts to "our own value strategy".
- DeepSeek. Blind, its favourite affinity is 29% and its closest rule is "back short-priced favourites". Informed, it moves to "back the outsider".
Keep the honesty marker close. The value arm can post a positive number over a short window. We report it honestly, and it is not yet a proven edge. On the Value Test, ChatGPT currently sits at -3.4% over just 256 settled picks, far too small a sample to call an edge, and exactly the sort of early number that hooks a punter into chasing.
The five, one by one
Here are the five, one at a time. Each profile is about character and method, how the AI reads a race, not a verdict on whether to back it. For that, follow the links to its "how it picks" write-up and its live scorecard.
ChatGPT, the crowd-reader
Blind, ChatGPT leans on form first and lands on the outright favourite 30% of the time, at a typical price around 5.26. Show it the market and it changes character: its picks map closest to "our own value strategy", which backs our model's biggest mid-priced market disagreement. On the blind card it wins 25.3% of its races against the market's 45.1% on the same cards.
Read how ChatGPT picks · See its live record
Gemini, the contrarian
Gemini reads the form and leads with form, but even blind it sits furthest from the chalk of the five, backing the outright favourite only 25% of the time at a typical price around 5.74. Given the market it maps closest to "back the outsider", which backs the longest price in the race. Blind, it wins 23.8% of its races against the market's 44.9% on the same cards.
Read how Gemini picks · See its live record
Claude, the careful reader
Claude is form-led and choosy. Blind, it leads with the trainer and recent form, landing on the favourite 31% of the time at a typical price around 5.41. Shown the odds it maps closest to "back the outsider", which backs the longest price in the race. Blind, it wins 24.4% of its races against the market's 46.3% on the same cards.
Read how Claude picks · See its live record
Grok, the value-hunter
Blind, Grok leads with the trainer and backs the favourite 29% of the time at a typical price around 5.46. The moment it sees the market it turns hunter, its picks mapping closest to "our own value strategy", which backs our model's biggest mid-priced market disagreement. Blind, it wins 23.9% of its races against the market's 45.1% on the same cards.
Read how Grok picks · See its live record
DeepSeek, the challenger
DeepSeek is efficient and blunt. Blind, it leads with form and backs the favourite 29% of the time at a typical price around 5.43. Once the prices are shown it flips, mapping closest to "back the outsider", which backs the longest price in the race. Blind, it wins 24.8% of its races against the market's 44.4% on the same cards.
Read how DeepSeek picks · See its live record
The five side by side
| AI | Live Form Test (blind) | Live Value Test (informed) | Favourite affinity (blind) | Blind strategy it maps to | Value strategy it maps to |
|---|---|---|---|---|---|
| ChatGPT | -9.6%, strike 26% | -3.4% | 30% | back short-priced favourites | our own value strategy |
| Gemini | -8.6%, strike 24% | +39.1% | 25% | favourite over jumps | back the outsider |
| Claude | +1.5%, strike 25% | -8.5% | 31% | back short-priced favourites | back the outsider |
| Grok | -13.6%, strike 24% | -9.2% | 29% | back short-priced favourites | our own value strategy |
| DeepSeek | -9.6%, strike 25% | -4.2% | 29% | back short-priced favourites | back the outsider |
These ROI figures are live and move daily over small samples. They are the record of an open experiment, not a tip sheet, and a negative or positive line is equally expected this early.
The honest verdict, and how to use AI for racing well
Pull it together and the five are genuinely good at three things: reading a full card in seconds, explaining every pick in plain words, and, blind, landing near the market's own view without being shown it. That last one is a real signal of comprehension, not a fluke. What none of them has shown is a proven, repeatable edge over the market, and the reason is structural. By the time an AI and the crowd agree on a horse, the bookmaker has already shortened its price to match, so the value is gone before you can take it.
So use them for what they are good at.
- Use AI to speed up your own reading, not to outsource the bet. Ask it to summarise the form and surface the story, then price the race yourself.
- Prefer the blind read for a clean opinion, and treat the informed value picks as hypotheses to check against the real price, never as tips.
- Blend the five rather than trusting one. Where they agree blind, the market usually agrees too, which tells you something even when it does not pay.
- Track everything and stake responsibly. An AI that reads well is a research aid, not an edge.
You can watch the experiment run on the two live boards: AI Picks the Winner for the blind read, and AI Finds the Value Bets for the informed one, with 5,034 picks settled and counting. A fuller how-to guide on using AI for racing is on the way.
Research, not tips. 18+, please gamble responsibly.