How Gemini Picks Horses: The Maverick's Method

Ask the five big chatbots to name a winner with the odds hidden, and all five land closest to the same mechanical rule: back the short-priced favourite. Gemini is the one that fits it loosest.

That wandering is not noise. It is a measurable habit, and once you can see it you know exactly what a Gemini tip is worth as a second opinion.

We test this the same way every day. Five AIs, the ones people actually use, read real British and Irish races in the Silicon Tipster League, then a statistics engine measures how each one picks: how far down the market it reaches, which prices it lands on, and which mechanical strategy its choices most resemble. Across 111 live days the picture is consistent. The pack behaves like a herd. Gemini is the mild exception on both sides of the experiment.

Blind, with no prices in front of it, Gemini's picks resemble "back short-priced favourites" more loosely than any of the other four. Shown the market, it swings harder than any of them toward outsiders. Neither move is a magic trick, and neither has made money. But the shape is real, and it is the most interesting fingerprint in the Lab.

Here is how that fingerprint is built, feature by feature, and how to read a Gemini pick without being fooled by a good-looking number.

Reading a race blind

Give Gemini a racecard with the prices stripped out and it reads the form first. When it explains a pick, the words it reaches for most are form, jockey and trainer, the same handwriting the whole field shares. What sets it apart is where those words lead.

So, does Gemini back favourites? Partly. It lands on the actual market favourite about 22% of the time, and its typical pick sits around position 3 in the betting, at a geometric mean price near 6.64. That is favourite-leaning, but with a longer reach than the rest: more than one blind pick in four is an outright longshot (35% of them). On our zero-to-one "how far down the market" scale it averages 0.34, the highest of the five.

The engine then asks which mechanical rule its 3,846 blind picks most resemble. The label is the same one every model gets blind, back short-priced favourites, a rule that backs the favourite only when it is odds-on. What separates Gemini is how badly it fits. Its resemblance score is only about 0.28 on a nought-to-one scale, against 0.46 for ChatGPT, 0.42 for Grok, 0.42 for DeepSeek and 0.38 for Claude. Same rule, loosest fit by a distance: favourite-minded, but willing to roam.

That is the blind read. Then we show it the odds, and it changes character completely.

Shown the odds, it goes hunting

The second test hands Gemini the same races with the market's implied chances printed on the card, and changes the job: stop naming the likely winner, start finding the horse the market underrates. Every model reaches for longer prices here. Gemini reaches furthest.

Its informed picks sit at a geometric mean price near 12.74, with the typical selection down around position 5 in the betting and an average market-reach of 0.64, again the highest of the five. It backs an odds-on shot essentially never (0% of picks), while 69% of them are outright longshots. Its closest mechanical match flips from a favourite rule to back the outsider, the rule that backs the longest price in the race.

Now the part to handle carefully. Over its 3,510 settled value picks, Gemini's return is −21.1%. That is a loss, and the confidence interval of −33.4% to −9.5% crosses zero, so this sample cannot separate it from break-even in either direction. The comparison that says more is against a machine that blindly backs the outsider in every race. On the races the two have in common, Gemini's picks return more than that rule by +43.4%, with an interval of +25.7% to +59.5% that excludes zero, so the difference is measurable rather than noise. It is a gap between two losing approaches, not a profit. Bold, and fun to watch, but not an edge.

How to use Gemini well

The useful truth about Gemini is its independence. When the other four models and the market all point at the same short price, Gemini is the one most likely to name a different horse. That makes it a good widener of your own thinking, a divergent second opinion rather than a tipster to follow blind.

A few habits get the most out of it:

• Know which mode you are in. Ask it cold and it reads form and leans to favourites. Paste in the odds and it swings to outsiders. The same chatbot gives you two different answers depending on what it can see, so decide which question you are actually asking.
• Treat its longshots as high variance. Outsiders win seldom. A good few weeks on the value side is a small sample having a good month, not a signal, and its own numbers are not yet distinguishable from simply backing the longest price.
• Follow the real record, not the vibe. Every pick is logged before the off and settled at starting price on its live record.

See how it stacks up against the rest in the full verdict on all five AIs, and watch both experiments run live on the AI Picks the Winner and AI Finds the Value Bets boards.

Bet responsibly. Nothing on this page is a tip or 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+.