AI horse racing tips are selections produced either by a model that rates runners from form, sectional times and prices, or by a chatbot that writes a race preview. Only a model gives you a chance you can turn into a price and test. A tip at $3.00 breaks even only if runners like it win 33.3% of the time, so whether AI racing tips work is a question about prices and records, not about the label.
Two kinds of AI racing tip
AI horse racing predictions come in two forms, and only one of them can be tested:
| Rating model | Chatbot preview | |
|---|---|---|
| What it produces | A chance, or a rated price, for every runner | Paragraphs naming a winner and a few dangers |
| What it learns from | Past form, sectional times, track ratings, barriers and often the market's own prices | The text it was trained on, plus any page a browsing tool fetches |
| Can you test it? | Yes: each chance converts to a price, and the record can be scored | Not as a forecast: there is no chance to score, and asking twice can give a different pick |
| What to ask | What is the rated price, and when was it made? | Where did each fact come from? |
Many services sit between the two: a list of selections under an AI name with no numbers at all. With no chance or rated price there is nothing to test but a strike rate. Chatbots have their own page, ChatGPT and betting.
Every racing tip is a price question
Turning a rating into a price takes one line: rated price = 1 / the model's chance. The price on offer then decides whether the tip is a bet. A tip at $3.00 breaks even only if runners like it win 1 / 3.00 = 33.3% of the time, and it needs more than that to profit over many races.
Example: Illustrative chances and prices for a six-runner race, not a real race or a current market.
| Runner | Model's chance | Rated price = 1 / chance | Bookmaker A | Break-even = 1 / price | Expected value per $1 = chance x price - 1 |
|---|---|---|---|---|---|
| 1, the model's top pick | 31% | $3.23 | $3.00 | 33.3% | 0.31 x 3.00 - 1 = -7.0% |
| 2 | 24% | $4.17 | $3.80 | 26.3% | 0.24 x 3.80 - 1 = -8.8% |
| 3 | 17% | $5.88 | $5.50 | 18.2% | 0.17 x 5.50 - 1 = -6.5% |
| 4 | 13% | $7.69 | $9.00 | 11.1% | 0.13 x 9.00 - 1 = +17.0% |
| 5 | 9% | $11.11 | $10.00 | 10.0% | 0.09 x 10.00 - 1 = -10.0% |
| 6 | 6% | $16.67 | $13.00 | 7.7% | 0.06 x 13.00 - 1 = -22.0% |
The model's chances add to 100%. Bookmaker A's implied chances add to 106.6%, an overround of 6.6 points built into its prices. The expected value calculator runs the last column for any chance and price.
The top pick is the model's most likely winner and still a losing bet at $3.00, because 31% sits below the 33.3% that price needs. The only runner whose price beats its rating is Runner 4, fourth in the model's order, at $9.00 against a rated $7.69. A service that posts Runner 1 as "the AI's pick" is answering a different question from the one your money asks.
A gap like Runner 4's deserves suspicion before money. With the margin taken out, the market rates it at 11.1 / 1.066 = about 10.4%, against the model's 13%. Ask what the market knows that the model may not, such as a late scratching, a gear change or a rider booking. Horse racing ratings shows how ratings become a full market of prices.
Risk: Betting involves risk. A rated price is an estimate: on the model's own numbers Runner 4 loses 87 races in 100, and if the model is wrong there is no edge at all. See responsible gambling for limits and support.
Testing an AI tipster's record
An AI tipster is judged like any other tipster, and the full recount is in how to check a tipster record. Four checks matter most for racing:
- Every tip with its time and price, from the first day, so you can see that each price was on offer before the jump.
- Each tip scored against the close: CLV = tip price / BSP - 1, with the Betfair Starting Price (BSP) as the usual racing close. With illustrative prices, a $9.00 tip beat an $8.20 BSP by 9.00 / 8.20 - 1 = +9.8%, and against a $10.50 BSP it was 14.3% behind the close (9.00 / 10.50 - 1 = -14.3%). A bookmaker starting price can carry the margin (at NSW TAB race meetings it comes from Racing NSW's NSW Official Price service), and closing line value in horse racing covers both closes.
- Hundreds of tips at the least, and at long prices thousands, before an ROI means much. At an illustrative average price of $7.00, 250 tips carry a standard error of square root of 6 / square root of 250 = 2.449 / 15.81 = 15.5 percentage points. A +18% ROI then sits barely more than one standard error above zero, and reaching two standard errors at a 10% edge takes about 4 x 6 / (0.10 x 0.10) = 2,400 tips. CLV settles far sooner.
- The survivors question. Say 40 services start with no edge, and each has a 1 in 5 chance of showing a profit after a year by luck. About 40 / 5 = 8 will, and those 8 are the ones still advertising. That is survivorship bias, and a record that restarts with each new version of the model hides the same thing.
Risk: Betting involves risk. Past results are no guarantee of future results, even over a long, complete record. See responsible gambling for limits and support.
What a model cannot see
A model knows only what it was given when it last ran:
- Late scratchings. A model that ran before a scratching shares the race among a field that no longer exists (worked below). Fixed-odds bets struck before a scratching can also carry a deduction, which scratchings and deductions explains.
- Track rating changes. Turf tracks are rated from Firm 1 to Heavy 10 on Racing Australia's scale, and a rating can change during a meeting as rain falls or the track dries. A model run on the morning rating misses it; track ratings covers the scale.
- Gear and rider changes. Under the Australian Rules of Racing, gear needs the Stewards' approval before acceptance time (AR 206), so a model that ran on earlier data misses it. If a declared rider is hurt or ill, the Stewards may permit a substitute (AR 189). These are thoroughbred rules; harness and greyhound racing have their own rule books.
- Where its training data stops. A model knows nothing about races after its data was taken: a horse's latest runs, a change of stable, a track rebuilt since. Ask for the date of the latest race in its data and how often it is retrained.
Example: Illustrative numbers from the six-runner race above. Say Runner 2, rated 24%, is scratched and its chance is shared out in proportion, the simplest assumption. Runner 1 rises from 31% to 0.31 / 0.76 = 40.8%, a rated price of about $2.45, and Runner 4 from 13% to 0.13 / 0.76 = 17.1%.
Red flags in AI tipster claims
- A strike rate with no prices beside it.
- An ROI over a few weeks, or a "since launch" record with a recent launch.
- "AI" with no stated method: no inputs, no account of what it predicts, no time it runs.
- Losing tips deleted, or a record that starts again with each new version.
- Pressure to subscribe quickly: countdowns, limited places, launch prices.
- Software pitched as a way to grow your money. Buying prediction software is one of the usual forms of the sports "investment" schemes Scamwatch warns about, and it calls those schemes "just another form of gambling", often outright scams (checked October 2026).
For the signs that apply to any tipster, AI or not, see fake tipster red flags; AI sports betting runs the same price tests on sports predictions. Racing models, tip services and chatbots all sit among the wider betting tools in Australia.
Mistakes that make AI racing tips cost more
Using the illustrative race above:
| Mistake | What it costs |
|---|---|
| Backing the model's top pick at any price | Runner 1 at $3.00 loses 7 cents per dollar on the model's own numbers |
| Taking a tip after its price has gone | Runner 4 taken at $7.50 instead of $9.00 is worth 0.13 x 7.50 - 1 = -2.5%: the edge has gone |
| Trusting a few weeks of ROI | 250 tips at $7.00 still leave 15.5 points of noise either way |
| Betting a tip made before a scratching | The chances belong to a field that no longer exists |
| Paying for tips to win back losses | A fee on top of the losses, at prices that have not changed |
Skip an AI tip that comes with no price, one made before the final field and the morning track rating, and one whose price on offer is below its rated price.
Where B337's Terminal fits
B337 is not AI: it does not rate runners, pick winners or use a model. It shows where a tip's price sits.
The Terminal shows prices from 40+ bookmakers across racing and sports, beside Betfair back and lay, with price history (flucs) and closing lines. On it you can check whether a tip's price is on offer, which way it has moved and how it compares with the exchange. The board reads prices on a repeating cycle, so a price on screen can lag the bookmaker's own; check it in your bookmaker account before betting.
A free account opens a limited view of the Terminal with live odds, while Betfair prices and closing lines start with Terminal View, listed on the pricing page. Bookmaker names are trade marks of their owners. B337 is not affiliated with them.
Risk: Betting involves risk. A price that beats a model or the exchange can still lose, and often does. See responsible gambling for limits and support.
For free and confidential support call 1800 858 858 or visit gamblinghelponline.org.au.