How DeepSeek Picks Horses: Showing Its Working

The AI that shows its working

Most chatbots hand you a name and a shrug. DeepSeek hands you the reasons, and that makes it the clearest window in the Lab into what an AI actually weighs when it reads a race.

Across 537 tips in the hidden-odds test, its written reasons lean hardest on form, then trainer, then draw. That third one is the giveaway. DeepSeek is the only one of the five that regularly talks about the draw, the stall a horse starts from, which can settle a sprint before the runners have found top speed. None of the others surface it in their reasoning with anything like the same frequency.

That does not make DeepSeek right more often. It makes it legible. When it picks, you can usually follow the chain of thought: this horse's recent form, this trainer in good heart, this draw on this track. For anyone trying to learn how a race gets read, a tipster that narrates is worth more than one that only names a winner, because you can hold the working up against your own eyes and see where you agree.

This is a spotlight on how it thinks, not a tip sheet. For the scoreboard, follow its live record.

Two minds: favourite-reader blind, outsider-hunter informed

Give DeepSeek a card with the odds stripped out and it behaves like a cautious form student. Its blind picks cluster in the shorter half of the market, and the strategy they most resemble is one that backs the favourite only when it is odds-on (a resemblance score of 0.48 out of one). It is no favourite-backing machine, though: it lands on the actual market leader 29% of the time, and backs its own read rather than the shortest price on the board.

So the honest answer to "does DeepSeek back favourites?" is that it leans that way with the odds hidden, but it is following its own form line, not the market's.

Now show it the market's implied chances and ask for value, and it turns into a different tipster. Its typical pick lengthens sharply, the longshot band holds 54% of its selections, and the closest strategy label flips to back the outsider. The resemblance is weak (0.19), and that is the interesting part: shown the odds, DeepSeek is not running a mechanical outsider rule, it is hunting horses it thinks the market has underrated, and value climbs into its top reasons. Same model, same races, opposite instinct. The odds you show it change the answer you get, which is the whole reason the Lab runs both tests side by side.

The draw, and the limits of a good read

Here is DeepSeek's live blind reasoning, theme by theme. The draw sits in its top three, where for the other four models it barely registers.

Reading the draw is exactly what a sharp human handicapper does, so it is genuinely interesting that one model has picked up the habit and the rest have not. But a good-looking read and a winning read are not the same thing, and this is where the honest part comes in. In the hidden-odds test DeepSeek's blind strike rate is 24.8%, while the market's own favourite won 44.4% of those same races. Its flat-stake return over the sample is −9.6% (−25.9% to +7.0%), and the value arm sits at −4.2%. Neither is a proven edge over a sample this small, and we would not present it as one.

That is not a knock, it is the useful truth of the experiment. An AI can weigh the right things, narrate them clearly, even spot a factor the others miss, and still not get ahead of a market that has already priced all of it in. DeepSeek reads a card quickly and competently. What it has not shown, yet, is a way to turn that reading into money, which is precisely why we publish the record and not the promise.

How to use DeepSeek well

If you like a tipster that explains itself, DeepSeek is the one to read. Its value is not the pick, it is the reasoning attached to the pick. Treat it as a fast second opinion that shows its working: let it surface the form line, the trainer angle and, uniquely, the draw, then weigh those against your own view of the race.

Two habits keep you honest with it. First, remember which mode you are in. Ask it cold and you get a cautious, favourite-leaning read; paste in the odds and ask for value and you get an outsider-hunter chasing prices it thinks are wrong. Those are very different tipsters wearing the same name. Second, judge it on the live record, not on how convincing the prose sounds, because a confident explanation and a correct one are not the same thing. That is the trap a narrating AI sets, and the record is the antidote.

You can watch both sides in real time: DeepSeek in AI Picks the Winner and in AI Finds the Value Bets, with its live record settled at starting price. For how it stacks up against the rest, here is the full verdict on all five AIs.

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