I tried to let Gemini 3 be my son’s temporary golf swing coach. It didn’t go very well.

We think a lot about the future of AI here at Chrome Unboxed. We talk about how it will write our emails, organize our calendars, and maybe even code our apps. But as a golf nut and the owner of a golf simulator facility (Proof Golf), I’ve been dreaming of a different use case: The AI Swing Coach.

Imagine it for a second: you walk into a bay, hit a few balls, and an AI analyzes your biomechanics in real-time, giving you professional, relaible feedback instantly. With the launch of Gemini 3 and its improved multimodal (video) capabilities, I decided to put this dream to the test since Google made the point of Gemini 3 being able to analyze a Pickleball match and point out areas for improvement.

I fed the new “thinking” model of Gemini 3 a video of my son hitting an iron at our facility. Now, full disclosure: Landon has a very natural, fundamentally sound swing. I actually picked one of his better shots as a bit of a setup to see if the AI could recognize a good swing or if it would just make things up to “fix.” Spoiler alert: It made things up.

The Experiment

I uploaded the video to Gemini 3 and simply asked: “Can you analyze my son’s golf swing and provide some tips?” I expected it to maybe comment on tempo or balance. Instead, Gemini 3 went into full “critic mode,” confidently diagnosing flaws that simply weren’t there. Here’s the video and below that are a few of Gemini 3’s thoughts:

The Hallucinations

1. Bent Lead Arm: Gemini claimed his lead arm was bending at the top of his swing.
2. Overswinging: Gemini also thought he was swinging way too far back, extending the club over and behind his left shoulder. He clearly was not.
3. Early Extension: After correcting the false claims of over-swinging and bending the lead arm, Gemini then diagnosed him with “early extension” (moving the hips toward the ball during the downswing). As anyone who knows golf mechanics can tell you, his lead hip was clearing perfectly and moving away from the ball in the downswing.

The problem with Generalist AI

After these misses, I came clean with Gemini and admitted that I fed it a pretty good looking move into the golf ball and wanted to see how good it really was at getting to the minute points of the golf swing. It was a trap, and the AI fell right in.

And this highlights a massive hurdle for “General Purpose” models like Gemini. They are trained to be helpful. When you ask for “tips,” the model interprets that as “Find a problem so I can offer a solution.” If a problem doesn’t exist, it hallucinates one just to fulfill the user’s request.

It knows what a golf swing is, and it knows the terms (early extension, lag, shallowing), but it lacks the precise, physics-based computer vision to accurately measure them in a specific video. It’s guessing based on patterns, not measuring based on physics.

What I learned

Simply put, my experiment proved that you can’t just use a chatbot as a swing coach at the moment – not even one as sophisticated and powerful as Gemini 3. For this to work—and for us to eventually build something like this for our members at Proof Golf—we can’t rely on a general language model. We need Specialized AI. We need models trained specifically on biomechanics that use computer vision to measure angles, not just “guess” at them.

Gemini 3 is an incredible tool for writing code, planning trips, and summarizing emails. But for now? I think I’ll keep being Landon’s swing coach myself.