Your kid's arm is not a chatbot experiment

A parent films a 14 year old throwing a bullpen on a phone. The video goes to a service that runs it through AI and sends back a polished report card: joint angles to the degree, timing to the hundredth of a second, green checks for what's good and red flags for what "needs work." It looks like something out of a big league lab.

We recently got a look at one of these report cards, sold to training facilities as AI motion capture from phone video. It's a perfect example of why this stuff is dangerous, because the problems aren't subtle once you know what you're looking at.

A green check on a collapsing front leg

The report graded the pitcher's lead leg block as GOOD. Its own numbers: front knee at 162 degrees when the foot lands, 149 degrees at release.

That's the knee bending 13 degrees after foot strike. In a good delivery the front leg firms up or straightens into release, which is what stops the lower half and sends energy up the chain into the arm. A knee that keeps bending is a leg that's giving way. The energy leaks, and the arm has to make up the difference.

The report had the right numbers and the wrong verdict. A green check, on the exact pattern a coach would want to fix.

Penalizing a kid for doing it right

The same report flagged the pitcher for not staying tall and "lunging," because his hips dropped at foot strike. It also reported his stride at 2.3 torso lengths, which is a long stride.

You can't take a long stride without getting lower. That's geometry. The report told this kid to stay tall while praising the stride that makes him get low. Follow both pieces of advice and you're fighting your own delivery.

Measuring the wrong thing, confidently

• Arm timing. It called the arm "on time" based on shoulder position. Arm timing isn't about how high the elbow sits. It's about where the forearm is when the front foot lands. Check the wrong thing and you'll fix the wrong thing.
• Tempo. It gave a green check to timing that wasn't consistent at all. Leg lift to foot strike ranged from 0.67 to 1.02 seconds across four pitches. That's about a 50 percent swing in tempo between throws, and the report called it fine.
• Guesses dressed as data. It listed a hip and shoulder separation of "about 31 degrees" from a single phone camera, right next to a pelvic tuck the report admitted was read from video, meaning someone eyeballed it, presented with the same authority as everything else.

Why one phone camera can't do this

Pitching is rotation. Hips, trunk and arm all turning toward and away from the camera. A single phone sees the world flat, and rotation toward the lens can look like a completely different movement. Angles read off a flat image can be way off, and putting a decimal point on them doesn't make them right.

Speed is the other problem. The arm goes from layback to release in a fraction of a second. On a standard phone video, it can rotate a long way between one frame and the next. You can't measure what the camera never saw.

And general purpose AI like ChatGPT or Claude doesn't fix that. Those are language models. Hand one a clip and it describes what it thinks it sees, in the same confident tone whether it's right or wrong. Ask a leading question and it will often agree with you. That's not analysis.

Why this is dangerous for kids

Bad advice for an adult is a wasted bullpen. Bad advice for a 14 year old can be an injury.

Young pitchers are still growing, and the elbow and shoulder take huge stress on every throw. The report card we saw came with velocity drills and a promise of added velo, and nothing about workload, rest, soreness or arm health screening. A kid told that his leaky front leg is "good," that his good stride is "bad," and handed a velo plan with nobody watching volume or pain is how arms get hurt.

What real analysis looks like

Real motion capture measures the body in three dimensions: multiple synced cameras, shooting fast enough to catch the arm through release, with software tested against a lab and published error numbers. Then a qualified coach reads it in context: the kid's age, growth, workload and history. Data without a coach is a pile of numbers. A chatbot without data is a pile of words.

What parents should do

• Don't change a young pitcher's mechanics off a phone app or chatbot report.
• Ask what it was validated against. No published accuracy, no trust.
• Check the report against itself. If it praises the stride and penalizes the kid for getting low, someone isn't paying attention.
• Follow pitch count and rest guidelines like MLB and USA Baseball's Pitch Smart.
• Pain is a stop sign. Elbow or shoulder pain means rest and a doctor, not a new drill.

Your kid's arm deserves better than a confident guess with a green check mark.