Learn
The Research Only Rule: how to use AI for sports betting without trusting it to pick winners
Here is how to use AI for sports betting without handing it your picks: give a chatbot three research jobs (summarize the injury news, put every book's price for one game into one list, and flag any price that sits off the fair number) and keep every betting decision for yourself. In a 2026 World Cup test, four frontier chatbots did not forecast better than the betting market, and a flat bet on the favorite earned more than all four.
The Research Only Rule is simple: give AI the research and keep the picks for yourself. It is the prompt set from our video on AI sports betting, written out so you can copy it. The rule gives a chatbot three jobs, in this order:
- It reads the injury news and sums it up. You paste in the report and the lineup notes, and it tells you what changed and where each line came from.
- It pulls every book's price for one game into one list. You paste in the prices you copied from your books, and it lines them up side by side with the implied chance of each.
- It flags a price that sits off the fair number. It takes the margin out, works out a fair price, and marks any book paying more than that, as a place to look, not a bet.
The rest of this page explains why picks are the wrong job, then gives each job its exact prompt, a worked example and the checks that keep the answers honest.
Why picks are the wrong job
A July 2026 preprint used the 2026 FIFA World Cup as a clean test, because every match was played after the models' training cutoffs. Ding, Guo and Xu had four frontier chatbot agents forecast all 104 matches against the pre-match betting market. The four agents chose the same top outcome in 92 percent of matches, none of them beat the market's Brier score, and a flat bet on the market favorite earned more than all four. Their betting return ranged from -18 percent to +10 percent, and betting against the market lost money for every one of them.
A second preprint, LLM-SoccerArena, had seven models forecast the same 104 matches. Models with web access did better than models without it, but only by a small margin, a 0.023 improvement in Brier score. Both papers are preprints, not yet peer reviewed, and both cover one tournament in one sport, so treat them as evidence about what happened there.
They fit an older finding. A 2018 PLOS ONE study of soccer forecasts found that betting odds ranked highest in predictive quality, ahead of several mathematical models. The market is a hard opponent, and a chatbot asked "who wins?" gives you an opinion shaped by what it has read, not a calibrated chance. What a chatbot does well is read, sort and calculate quickly, which is exactly the work that makes your own decision better.
Who this is for and what you need
It is for a bettor who already uses a chatbot (any of the big ones works) and has accounts at two or more legal sportsbooks. You need:
- A text source you trust for news: the league's or team's own injury report, or a beat reporter's post, copied with its time.
- Prices you copied yourself, for the same market, at the same line, at about the same time.
- Ten minutes the first time. After that, each game takes two or three.
The one habit that makes all three jobs work: paste the material in. Don't ask the chatbot to go find it. When the input is something you can see, every line of the answer can be checked against it.
1. Read the injury news and sum it up
Copy the latest injury report and any lineup notes for the game, with the time they were posted, and paste them under this prompt:
You are my research assistant for [Team A] vs [Team B]. Below is the injury report and lineup news I copied, with the time of each item. List every player whose status changed in the last 24 hours, with the old status, the new status and the time. Quote the exact sentence each item came from. If something is not in the text, write "not in the source". Do not predict the result and do not suggest a bet.
What good output looks like: a short list such as "Hawks starting QB: questionable, updated 2:14 PM", each with its quote. What to reject: any player or status that has no quote next to it, and any line that starts telling you who will win. That is the chatbot drifting back into picks.
Why the quotes matter: the most common failure is a confident detail that was never in the text. Asking for the source sentence turns a summary you have to trust into one you can check in seconds.
2. Pull every book's price for one game into one list
Open the same market at each of your books and copy the price for both sides, with the line and the time. Paste them under this prompt:
Below are prices I copied from my sportsbooks for [market and line] on [Team A] vs [Team B], with the time I saw each one. Put them in a table with these columns: book, side A price, side A implied chance, side B price, side B implied chance, time. For a negative price p, the implied chance is -p / (-p + 100). For a positive price p, it is 100 / (p + 100). Show chances as percentages with two decimals. Mark the best price on each side. Do not add any book or price that is not in my list.
The formulas in the prompt are the standard moneyline conversion, so you can check any row by hand: -110 becomes 52.38 percent and +112 becomes 47.17 percent. Giving the chatbot the formula, instead of asking it to "convert the odds", stops it from using a rough rule of thumb.
What to check: the table has exactly the books you gave it, the times are yours, and the best price on each side is the one paying the most. If a book appears that you never pasted, throw the answer out.
3. Flag a price that sits off the fair number
A book's two prices add up to more than 100 percent; the extra is its margin, also called the overround. Take the margin out of one reference book (the one you think is sharpest, or simply the one with the lowest margin) by dividing each side's implied chance by the total. What is left is a fair chance for each side. Add this to the same chat:
Using the table above, take [reference book] as the reference. Add its two implied chances, then divide each side's chance by that total to get a fair chance for each side, and convert each fair chance back to an American price. Then list every price in the table whose implied chance is at least 1 point below the fair chance on the same side. Show the math for each one. Label each one "worth a look", never "a bet".
Here is a worked example. It is a hypothetical, with made-up prices and a demo matchup, not any book's real line. The reference book has Hawks -110 and Wolves -110. Each is 52.38 percent, they add to 104.76 percent, and dividing each by the total leaves a fair 50.00 percent per side, a fair price of +100, using the overround method. Another book offers Hawks +112, which is 47.17 percent. That is 2.83 points below the fair chance, so it gets flagged. If the fair 50 percent were right, a 100 dollar bet at +112 would be worth about 6 dollars on average, by the moneyline math.
Why it is a place to look and not a bet: the fair number is only as good as the reference book, prices move with news, and the price you see may already be gone. A flag tells you where to spend your next two minutes. The decision about whether you actually like that side is still yours.
Mistakes that break the rule
- Asking it to browse instead of pasting. Web access helps a little, by the SoccerArena result, but a pasted report with a time on it is easier to check than whatever page the chatbot found.
- Letting a summary turn into a pick. If an answer ends with "so the Hawks should win", delete that line. The study above is the reason.
- Comparing different lines. Hawks -3 and Hawks -3.5 are different bets. Only compare prices on the same line.
- Stale prices. Put the time on every price. A gap between a price from this morning and one from now tells you nothing.
- Trusting the math without one hand check. Check one row of every table yourself. If one row is wrong, the rest can be too.
- Agreement mode. Chatbots tend to go along with what you already think. Once a week, ask for the strongest case against a bet you like.
How to tell it's working
Keep a note for a month. For each bet, write down whether you checked the news summary, the price list and the flags before you placed it, and the best price you found. If the rule is working, almost every bet is placed at the best price you had, and you never bet on a news item you didn't see quoted. It won't tell you whether your picks are good. It makes sure you're informed and paying less for the bets you already wanted to make.
For the bigger picture on what chatbots and predictive models can and cannot do, read AI sports betting. If you'd rather skip the copying, the Quant+ Copilot answers with the board's live prices cited, and the odds comparison board puts every book and a fair price on one screen.
Quant+ is analytics for informational purposes. Nothing is guaranteed. 21+. If you or someone you know has a gambling problem, call 1-800-GAMBLER.
Questions people ask
Can AI pick winning sports bets?
Not reliably on the evidence so far. In a July 2026 preprint covering all 104 World Cup matches, none of four frontier chatbots beat the betting market's Brier score, and a flat bet on the market favorite earned more than all four. Use AI for research and price checks, not picks.
What is the best way to use ChatGPT or another chatbot for betting?
Give it jobs you can check: summarize news you paste in, organize prices you copied from your books, and run the implied-probability and fair-price math. Ask it to quote its source for every fact and to say when something is not in what you gave it.
Why should I paste the news and prices in myself?
Chatbots can hold stale information or invent confident numbers. When the input is text and prices you copied at a known time, every line of the answer can be checked against something you can see.
What does it mean when a price is off the fair number?
The fair number is the price with the bookmaker's margin taken out. A price that pays more than that fair price is a place to look more closely, not a bet by itself, because the fair estimate can be wrong and prices move.
Does the Research Only Rule make my bets profitable?
No. It saves time and makes you pay less for the bets you already wanted to make. Nothing is guaranteed, and a good price still loses often.
Sources
- FIFA World Cup 2026 as a Contamination-Free Benchmark for LLM Forecasting Agents (Ding, Guo, Xu), arXiv
- LLM-SoccerArena: Benchmarking LLMs on Real-World Predictions in Sports (Schroder et al.), arXiv
- The Betting Odds Rating System: Using soccer forecasts to forecast soccer (Wunderlich, Memmert), PLOS ONE
- Mathematics of bookmaking, Wikipedia
- Moneyline odds, Wikipedia
