How ChatGPT Helped Me Turn $9 Into $3,044.51 Betting On Horse Racing

Two days before the Belmont Stakes, I got a message from a friend. “What do you think about using AI for handicapping?” He asked.

I replied without much thought: “Love the concept. My issue is that the metrics are same old, same old, and I just don’t think that can win.”

But the question stuck in my brain like “Love Shack” at a karaoke bar. The more I thought about it, the more I realized how far tech in gambling has already come — and how different AI is from the stuff most “data handicappers” like myself are using.

The Evolution of Computer-Assisted Betting

Computer-assisted betting didn’t start with AI. It started with code. In the 1980s, Bill Benter, a math-savvy expat with a background in blackjack, changed racing forever. Teaming up with Alan Woods, Benter created a computer model that analyzed over a hundred variables per race in Hong Kong.

He didn’t just rely on speed or pace. His program adjusted for all sorts of things — jockey tendencies, post position, how horses performed at different distances under different weather conditions, etc. By the time it matured, his model was churning out more accurate odds than the public.

And the results weren’t just solid — they were staggering. Benter reportedly made hundreds of millions of dollars from his wagering activities.

The strategy? Find mispriced horses and bet accordingly.

It’s a formula most computer-assisted wagering (CAW) operations still follow. They rely on massive amounts of data and algorithms built to exploit inefficiencies in the market. Many of these groups now dominate the pools at major tracks and are called "whales." Their programs are fast, ruthlessly efficient, and completely opaque.

This same phenomenon exists in the financial markets, where sophisticated traders employ a variety of quant trading models, like game theory and sentiment analysis.

AI Isn’t Just a Faster Calculator

Most data handicappers are still working with fixed systems. Maybe they’ve put a unique twist on speed figures or pace analysis, like me, but their formulas are set.

There is no discussion or room for interpretation. Worse, there is no nuance either — the number is the number. If the benchmark is 75%, it is 75% — not 74.999%, even though it probably makes no difference (you can’t fully optimize every metric, even with regression analysis, because some factors take on more or less value depending on the situation).

AI is different. It offers something the old tools don't: dialogue. You can ask Chat GPT to generate picks based on factors you choose. You can also ask why it made those picks. You can ask follow-up questions. You can feed it new information and get a new output on the fly.

It’s not just a calculator, it’s a partner.

Does that mean AI always nails the winner? Of course not. But it sees patterns differently. It adapts. And, as I found out, sometimes it wins you some money.

Churchill Downs Pick-3

I’ve used AI for handicapping before — mostly as an experiment. The results were mixed, primarily because my data was limited and my prompts weren’t great. But now ChatGPT has a data analysis feature. You can upload an Excel file and get meaningful output.

So, with fresh numbers and an open mind, I generated AI picks for Churchill Downs on Saturday, June 7.

I used many of my unique metrics — form ratings (which tell me numerically how good or bad a race was), speed rations (a snapshot of energy disbursement), and speed figures adjusted for the date they were earned (this is known as a time-decay factor).

After some positive results early on the card, but nothing major, I decided to bet a pick-3 beginning in the seventh race.

I used all three AI selections in the first and second legs and particularly liked the top AI choice (Valiant Humor) in the third. However, I was also intrigued by Math Tutor, the third pick. He was 30–1 on the morning line and tabbed as “potential lone speed" by ChatGPT.

Now, on the surface, this seemed absurd. He was eighth at the first call in his last race and third at the first call of his prior race. True, they were both at five furlongs, a distance that stresses early foot, but they were also on the turf, a surface that does not.

Still, 30-1 morning line odds and “potential lone speed” are like free pizza and beer — hard to pass up.

Below are the AI picks for each race:

CD RACE 7

1. First of His Name (Loose on the lead danger)
2. Haunted Flame (Gaining form; late move)
3. Laugh A Lot (Always tries)

CD RACE 8

1. Yinzer (Big numbers, control up front)
2. Native Runner (Late charge threat)
3. The Warden (Wide, stayed on)

CD RACE 9

1. Valiant Humor (Best figures; perfect setup)
2. Fanatical (Consistent grinder)
3. Math Tutor (Potential lone speed)

And the winner is…

Race 7: Laugh a Lot pressed the pace, drew clear, and won by 4 ¼ lengths, paying $23.92 to win. Not a bad start.

Race 8: I was so busy rooting for The Warden at 35-1, that I didn’t initially realize one of my other contenders (Native Runner) was the horse that blew by him in the stretch and got up by a neck. I noticed the $58.04 payoff, though.

Race 9: At this point, I thought I had a pretty good shot of winning around $300 if the even-odds favorite, Valiant Nature, reached the winner’s circle. But something strange happened when the gates sprung open.

Math Tutor, the longest shot on the board at 52-1, shot to the front — just as ChatGPT had predicted. Turning for home, he was still in front and Valiant Humor was closing. I figured, either way, it was a great showing for AI handicapping. But "great showing" be damned, hang on Math Tutor!

He did, paying $107.98. My $9 pick-3 ticket paid $3,044.51