AI Picks the Winner vs AI Finds the Value Bets: what changes when an AI sees the odds

Ask a chatbot for a tip and you are running an experiment, whether you meant to or not. The question is what the AI is actually using when it answers. Its own read of the race? Or the market's opinion, absorbed from all the odds talk in its training and whatever you pasted into the chat?

The Silicon Tipster League splits that question in two and tests each half live. Five AIs, the ones people actually use, tip the same British and Irish races every day. Each one tips twice:

• AI Picks the Winner (formerly The Form Test). The AI gets the racecard, the form, the going, the field. No prices, no market anywhere in sight. This is the purest version of "what does the machine think of the race itself?"
• AI Finds the Value Bets (formerly The Value Test). Same race, same AI, but now it also sees the market's implied chance for every runner, and the task changes with it. Asking "who wins?" when the market's answer is on the card would be pointless, so this test asks for value: back the horse the market most underrates, and say what you think its true chance is.

Same races, same day, settled at the same starting prices. The only variable is whether the AI can see what the crowd thinks. Most people who ask a chatbot for a tip are, in effect, running something between the two. The league shows you what each version is worth, and the gap between them is where it gets interesting.

You might expect AI Finds the Value Bets to win easily. More information, better answers. That is how it works in most fields, and the market is the sharpest single forecaster in racing, so an AI that can see it starts with a real advantage.

Here is where we owe you an honest correction. Our first version of the value test asked each AI the same question as the winner test: which horse will win? With the market's chances printed on the card, that question answers itself, and every model quite sensibly said the favourite, in almost every race, day after day. They were not being lazy. We had asked a question the card already answered, and the run told us nothing worth knowing. That was our error, and we have kept the old rows in the raw data rather than pretending they never happened.

So on 17 July 2026 AI Finds the Value Bets restarted with the right question. Each AI now has to form its own view of every runner's true chance and back the horse it believes the market most underrates, logging its own estimate alongside the pick. Our own engine enters twice: Stablebet Edge plays the same value rule, backing its biggest model-vs-market edge in every race, while the Stablebet Model backs the horse it rates most likely to win. Same numbers, two rules, two records.

That makes the comparison worth watching again. AI Picks the Winner asks whether a machine can read a race by itself. AI Finds the Value Bets asks something harder: given everything the crowd knows, can it find the places the crowd is wrong? The gap between the two records, measured in pounds on the same races, is the price of that ambition, and whichever way it goes, it goes on the live board.

Everything above updates daily, and every pick is logged before the off and settled at starting price, so the record cannot be tidied up after the fact.

Where to follow it:

• AI Picks the Winner is the flagship, because it is closest to what most people actually do: open a chatbot and ask. Five AIs plus The Favourite, the market's own baseline.
• AI Finds the Value Bets is the same five with the market's implied chances in front of them, hunting the value it has missed, plus two entries from our own engine: Stablebet Edge backing its biggest edge and the Stablebet Model backing its most likely winner.
• The league home carries the running comparison between the two tests as both records build.

One rule worth restating: nothing in the league is a tip. It is a live experiment about how machines read races and what the market does to their judgement. If a table here is ever in front, treat it as a small sample having a good month, not a signal. The interesting part is the gap, not the winner.

The v1 record, archived

We said the old rows stay available, so here is what they show. These figures are frozen: they describe the retired ask-the-winner run (6 to 16 July 2026) and will never update.

Favourite-backing rate per model, among the 242 picks each where we tracked it:

Two honest caveats on the archive. First, the favourite-tracking field only exists from 9 July, so 81 early picks per model sit outside these denominators; nothing about them suggests a different pattern. Second, the picks that did deviate from the favourite are too few to say anything about: twelve settled non-favourite picks across all five models combined. That thinness is the finding: asked for the winner with the market's answer showing, the models almost never went anywhere else, which is why the run could not work as a test.

The raw rows live on in the daily pick files, every one tagged with the prompt version that produced it, so anyone can rerun these sums. The AI Finds the Value Bets record that replaced this one starts clean from 17 July.