How ChatGPT Picks Horses: Favourites or Value?

Ask ChatGPT to pick a horse with the odds hidden and it reads the race like a careful form student, leaning towards short-priced favourites.

Show it the same race with the market's prices attached and it changes character completely: it drops favourites altogether and goes hunting for value, landing on almost exactly the approach our own engine runs.

That split is the whole story of how the world's most-used AI reads a card, and it is why "how does ChatGPT pick horses?" has two honest answers rather than one. What it does depends entirely on what you let it see.

In the Silicon Tipster League we run five AIs on real British races every day, each one tipping the same card twice. Once blind, on form alone with no prices anywhere. Once informed, with the market's implied chance printed next to every runner and the brief changed to finding value. Same races, same day, settled at Starting Price. A statistics engine then measures the fingerprint of each set of picks: how short it goes, which race facts it leans on, and which known strategy its selections most resemble.

ChatGPT gave us the cleanest before-and-after in the whole field, a full flip in method the moment it can see prices. This is a look at both halves, and the full verdict on all five AIs sits alongside it.

Odds hidden: the by-the-book form reader

Blind, ChatGPT is the textbook student of the group. Across 534 settled picks with no prices in front of it, its selections cluster near the front of the market. It backs the actual favourite about 30% of the time, its median pick sits at market rank 2, and the typical price it lands on averages 5.26. Real longshots make up only about 24% of what it chooses.

When the engine asks which known strategy those blind picks most resemble, the closest match is "back short-priced favourites", though only moderately so, at a kappa of 0.48 on a nought-to-one scale. Read its written reasons and you can see why it lands short. The three ideas it reaches for most are form, jockey and trainer: recent form first, then who is riding and who is training, which is the same shortlist a sensible human builds before looking at the odds.

None of that is a knock. A model that reads form and lands near the front of the market is doing the job quickly and competently. The honest catch is that this is also the bias most punters carry. The favourite wins more races than anything else, yet backing it on instinct still hands money to the bookmaker's margin at Starting Price. Reading a race well and making money from it are not the same skill, which is the whole reason the Lab exists.

Odds shown: it reinvents value betting

Hand ChatGPT the same races with the market's implied chances attached, ask it for value instead of the winner, and it becomes a different tipster.

Across 214 settled value picks it never once backed the market favourite, a clean 0% across the lot. Its average price stretched out to 9.96, its median pick dropped to market rank 4, and its written reasons shifted away from jockey and trainer towards form, distance and going. It stopped asking "who is best?" and started asking "who is underpriced?".

Here is the part worth sitting with. When the engine measures which strategy these informed picks resemble, the closest match is our own value strategy: it backs our model's biggest mid-priced market disagreement. Given nothing but prices and a nudge to find value, the most-used AI in the world independently rebuilt the method our engine already uses.

The resemblance is real but loose, at a kappa of 0.24, so this is a family likeness rather than a clone. And the record needs the same honesty as everything in the Lab. Over those 214 bets, its return to level stakes at Starting Price was +2.0%, which sounds encouraging until you see the range around it: −41.8% to +54.8%. That interval runs from heavy loss to heavy gain, so it is a small-sample readout, not a proven edge, and certainly not a tip. What it does show clearly is a change in method, and that is the finding.

How to use ChatGPT on a race

The practical lesson falls straight out of the two fingerprints: what you show ChatGPT decides what you get back.

Ask it cold, with no prices, and you will get a competent read of the form book that leans towards the favourite. That is handy for understanding a race, and it is roughly what the crowd already thinks, which is exactly why it is hard to profit from at Starting Price. If you want its opinion on the shape of a contest, who holds the form and whose yard is in flying order, this is the mode to reach for.

Give it the market and ask it where the price looks wrong, and you get something more interesting: a value-hunter that walks past the favourite and looks for the overlooked. More useful to study, and honestly harder to pull off, because spotting where the crowd is wrong is the thing almost nobody manages consistently over time.

Either way, treat what it says as research, not a signal to bet. You can watch how this plays out in the open: ChatGPT's live record updates as races settle, alongside every rival on the AI Picks the Winner and AI Finds the Value Bets boards.

Research, not tips. 18+, please gamble responsibly.