Does AI sports betting work?
Sometimes on paper, but rarely in a way you can check. Of the 122 AI betting tools we rate, only 37 publish enough of a level-stakes record for a Performance score. 16 of the 34 with a full score show a profit, but 12 of those 16 are self-reported, only 2 are independently proofed, and 13 are settled at advised odds that may be hard to get. No record guarantees future returns. (Data as of 5 Nov 2026.)
The short answer, in numbers
We read every tool's own public record and score it on the same published rules (how the Performance score works). “In profit” means a level-stakes ROI above 0% over the stated period, at the price basis the tool itself uses, before any subscription fee. The median ROI across the 34 full-score records is -0.95%.
| Evidence behind the record | Records with a full Performance score | In profit at level stakes | Median ROI |
|---|---|---|---|
| Independently proofed | 2 | 2 | +8.15% |
| Recomputed by us | 5 | 2 | -1.20% |
| Self-reported | 27 | 12 | -1.29% |
| All full scores | 34 | 16 | -0.95% |
Self-reported: totals published by the tool itself. Recomputed by us: we rebuilt the figures from the tool's own bet-by-bet data. Independently proofed: the record is kept by a third party. See badges and flags.
What independent research says (2026)
Researchers have started testing AI forecasters against real betting prices, with results published before or alongside the events. The pattern so far matches our own data: 16 of the 34 AI records we fully score are in profit at level stakes, with a median ROI of -0.95%, and only 2 of the profitable ones are independently proofed.
What independent research says (6 studies)
- KellyBench: Can Language Models Beat the Market? (Grady et al., General Reasoning, 2026; Simulation with real market odds)
Frontier language models were asked to build betting strategies across a simulated 2023/24 English Premier League season using real market odds. Every model lost money over the season and several were ruined; the best (Claude Opus 4.6) averaged a return of −11%, and only 2 of 24 runs finished in profit. - FIFA World Cup 2026 as a Contamination-Free Benchmark for LLM Forecasting Agents (Ding, Guo and Xu, 2026; Prospective test, 104 matches)
Four frontier chatbots forecast all 104 matches of the 2026 World Cup before kick-off. None beat the betting market's Brier score (0.469 for the market against 0.471 for the best model). Virtual betting returns ranged from −18% to +10% over a single tournament, betting against the market lost money for all four, and a flat stake on the market favourite out-earned all four in absolute terms. The authors say per-model differences have wide intervals. - Model Edge or Model Miscalibration? A Prospective Test on Polymarket Tennis (Yushkov, 2026; Prospective test, 5,498 resolved observations)
Machine-learning tennis forecasts were logged live against Polymarket prices (April to June 2026). The market was calibrated before play and the live models were not; no betting strategy earned a gross return significantly above zero. The author concludes that a model disagreeing with the price is not evidence that the price is wrong. - Beating the bookies with their own numbers, and how the online sports betting market is rigged (Kaunitz, Zhong and Kreiner, 2017; Strategy test with real bets)
A strategy that bet when one bookmaker's football odds were out of line with the market average showed a profit in simulation and with real money, but the bookmakers limited the authors' accounts within months. It used the market's own prices, not a prediction model, as its estimate of the true chances. - Are Betting Markets Inefficient? Evidence From Simulations and Real Data (Winkelmann, Ötting, Deutscher and Makarewicz, 2024; 14 seasons, 5 leagues)
Across 14 seasons of the five biggest European football leagues, pricing inefficiencies showed up in single seasons but were not persistent or systematic, and simulations showed that such single-season effects are likely to appear by chance even in a fully efficient market. - Leakage-aware pre-event machine learning evaluation for horse-race prediction under temporal validation (Sugiura, 2026; Out-of-sample test, 4,556 races)
A machine-learning model trained on Japanese (JRA) flat races from 2015 to 2022 was tested on 4,556 later races (January 2025 to May 2026). It ranked winners with a ROC AUC of 0.754 (1.0 is perfect, 0.5 is chance). AUC measures how well runners are ranked, not whether bets at the available odds would make money.
Summaries checked against each source on 19 Oct 2026: the full paper where we could read it, otherwise the published abstract. Studies use their own samples and methods, so they aren't directly comparable with the records we score, and none tests the specific tools we list.
Taken together: frontier AI models can match the betting market's view of a match, but in these tests they didn't beat it. Where a model or strategy did show a profit, it was over a single tournament with wide uncertainty, didn't persist, or was quickly limited by bookmakers. That is why we score records, not claims: a bet count, a level-stakes result and a stated price basis.
Most AI tools don't publish a record you can score
85 of 122 AI tools (70%) don't publish enough for a Performance score: no bet count, no level-stakes profit or ROI, stakes that change per pick, or only a list of winners. A full public ledger including losers is verified for 66 of 122, and picks timestamped before the event for 41 of 122. The median Transparency score is 40/100; 39 tools score 60 or more (high transparency).
A tool that shows only accuracy, strike rate or a run of winners hasn't shown it makes money: a high strike rate on short prices can still lose. That is why the Performance score needs a bet count and a level-stakes profit or ROI.
Big samples tell you more than big percentages
A short winning run proves little. The Performance score pulls small records towards 50 (break-even): a self-reported record of 500 bets keeps only 33% of its distance from 50, an independently proofed one of 500 bets keeps 50%, and 5,000 proofed bets keep 91%.
The largest AI record we score is PropGolf: 51,079 bets at +4.00% ROI (self-reported, 12 events to 30 Aug 2026), Performance 55/100.
The strongest published AI records right now
Ranked by Performance score, then Transparency score (full scores only). Each line shows the evidence behind it. This is a ranking of published records, not a recommendation to bet.
- 1
Turfs Edge Tipping (TeT)
+15.23% ROI over 419 level-stakes bets (13 Aug 2026 – results page read 2 Nov 2026), recomputed by us. Price basis: Published (advised) odds; win only.
↻ Recomputed by us⚑ Advised odds, may be hard to matchHorse racingRecord period: 13 Aug 2026 – results page read 2 Nov 2026Checked 2 Nov · next re-check 9 Nov59PerformanceProfitable record40TransparencyMedium transparency - 2
RaceBot UK
+6.70% ROI over 3,475 level-stakes bets (November 2025 – 17 Oct 2026), independently proofed. Price basis: Advised bookmaker odds at time of posting.
✓ Independently proofed⚑ Advised odds, may be hard to matchHorse racingRecord period: November 2025 – 17 Oct 2026 (Tipstrr stats page, horse racing, all time)Checked 1 Nov · next re-check 8 Nov57PerformanceProfitable record90TransparencyHigh transparency - 3
ProBetAI
+9.60% ROI over 709 level-stakes bets (Tipstrr last-12-months view, to 4 Nov 2026), independently proofed. Price basis: Advised bookmaker odds at time of posting.
✓ Independently proofed⚑ Advised odds, may be hard to matchHorse racingRecord period: Tipstrr last-12-months view, to 4 Nov 2026Checked 4 Nov · next re-check 11 Nov57PerformanceProfitable record80TransparencyHigh transparency - 4
MatchWiz (golf model)
+12.30% ROI over 470 level-stakes bets (Group picks with published odds, to Oct 2026), self-reported. Price basis: Closing 3-ball odds.
• Self-reportedGolfRecord period: Group picks with published odds, to Oct 2026 (start date not stated)Checked 31 Oct · next re-check 7 Nov55PerformanceProfitable record65TransparencyHigh transparency - 5
NoSleepHQ
+7.10% ROI over 1,563 level-stakes bets (Last 60 days to 1 Sep 2026), self-reported. Price basis: Odds when posted (avg 1.76).
• Self-reported⚑ Advised odds, may be hard to matchMulti-sportRecord period: Last 60 days to 1 Sep 2026 (Jul–Aug 2026)Checked 31 Oct · next re-check 7 Nov55PerformanceProfitable record50TransparencyMedium transparency
Full ranking, all sports → · All AI tool ratings · AI horse racing predictors (UK) · AI football prediction tools
How AI tools compare with human tipsters
We rate 102 human tipsters on exactly the same rules. 64 have a full Performance score and 53 of those are in profit, with a median ROI of +8.62%. Our tipster list is not a random sample: it leans towards newspaper tipsters and proofing platforms. Tipster records are more often kept by a third party (54 of 64 independently proofed), but 48 of the 53 profitable ones are settled at advised odds, which often shorten quickly or are limited to small stakes.
| Evidence behind the record | Records with a full Performance score | In profit at level stakes | Median ROI |
|---|---|---|---|
| Independently proofed | 54 | 46 | +10.15% |
| Recomputed by us | 7 | 5 | +1.61% |
| Self-reported | 3 | 2 | +3.15% |
| All full scores | 64 | 53 | +8.62% |
Why a profitable record still isn't a promise
- Prices. A record settled at advised odds may not be repeatable: by the time you see the tip, the price may be gone. Starting price (SP) or Betfair SP records are easier to match.
- Fees. Our figures never deduct subscription fees. A small edge can disappear once you pay for the service.
- Time. Check how long a record covers. Results swing a lot over a few hundred bets, and a few good months can look very different from a few years.
- Selection. We can only score what is published. Tools with poor results have little reason to publish them, so the records we see may not represent the whole market.
- Accounts. Bettors who win consistently can find their stakes limited by bookmakers. A published study of a strategy that beat football bookmakers using their own odds reported exactly that (Kaunitz, Zhong and Kreiner, 2017).
UK advertising rules say tipsters must not state or imply that success is guaranteed or that you can make a long-term income from their tips (ASA guidance on betting tipsters). We apply the same standard to how we describe every record.
Questions and answers
Is AI sports betting profitable?
For most tools we can't tell, because they don't publish a scoreable record. Of the 34 AI tools with a full Performance score on 5 Nov 2026, 16 show a profit at level stakes (median ROI -0.95%), but 12 of those are self-reported and 13 are settled at advised odds. Fees are not included.
What does independent research say about AI betting?
Recent tests point the same way. In KellyBench (2026) every frontier model lost money over a simulated Premier League season; the best averaged −11%. In a 2026 World Cup test none of four AI models beat the betting market's forecasts, and a prospective tennis test found no strategy with a return significantly above zero. Sources are linked in the research section above.
Is AI betting legit?
Some tools are open about their results; many are not. 66 of the 122 AI tools we rate publish a full public ledger including losers, and 41 timestamp their picks before the event. Check the Transparency score on each entry before paying for anything.
Why do so many AI tools have no Performance score?
85 of 122 don't publish a bet count and a level-stakes profit or ROI, or change stakes per pick. Strike rates and lists of winners can't show whether a tool makes money, so we don't score them.
Can ChatGPT or other chatbots pick winners?
We rate 8 tools that use large language models to pick bets. See Can ChatGPT predict sports results? for what their published records show.
What should I check before paying for an AI tipping tool?
Look for picks timestamped before the event, a full ledger including losers, a stated price basis, a few hundred settled bets or more, and records you can download. Our checklist explains each one.
Do past results predict future returns?
No. A published record shows what happened over a stated period at stated prices. It doesn't predict what will happen next, and no tool can guarantee profit.
How we check the records
The transparency checklist sets out what counts as evidence for each Transparency criterion, the same for every listing. The AI tipping transparency report 2026 shows how many of the AI tools and tipsters we rate meet each criterion.