Fair question, and it deserves a straight answer rather than a sales page: yes, AI can genuinely check squat form — some of it, with real numbers — and any tool claiming to check all of it is padding the truth. Here's exactly where the line sits, using CoachVision's squat analysis as the worked example, limits included on purpose.

What body tracking actually measures

Film a squat set and the system tracks your joints frame by frame — 2-D keypoints from a single camera — then computes measurements from how those points move. For the squat, ten of them, split by what each camera angle can honestly see.

From the side: depth (your hip marker versus your knee marker at the bottom), your deepest knee bend in degrees, torso angle at the bottom, how much that angle changes as you start driving up (the chest-dips-first signature), how fast your shoulders rise compared to your hips out of the hole, and how far your shoulder line drifts off its start point over your feet.

From the front: knee tracking — knee spacing versus ankle spacing at its narrowest on the way up, the measurement that catches knees caving in — and stance width relative to your shoulders.

From either: descent time, and how evenly your reps repeat across the set.

These aren't estimates dressed as data — they're computed from tracked motion, in body-relative units so tall and short lifters compare fairly. And the split matters: the side camera physically cannot see knee cave; the front camera flattens depth. A tool that reports knee tracking from a side view is inventing it.

What it honestly can't see — and shouldn't pretend to

  • Your spine's shape. Tracking gives one trunk segment, hip to shoulder — angle, yes; rounding or butt wink within it, no. Freeze-frames let your eyes check; no honest number exists.
  • The bar, or any load. There's no barbell tracking, and the tracker can't tell bodyweight from a loaded bar — you declare the load, because a front squat's textbook torso angle and a low-bar squat's differ wildly, and grading one by the other would be noise.
  • How heavy it was. No percentages of max, no load prescriptions, ever. A slow rep might be a grind or deliberate tempo work; the video doesn't know and won't pretend to.
  • Foot pressure and bracing. No force plates, no view inside your trunk.

There's calibration honesty too. The depth read compares tracked markers, and the hip marker sits above your actual hip crease — so the app never declares "above parallel" from a number alone; verdicts come from the number plus the bottom keyframe, hedged like a good coach hedges. The knee-cave read errs lenient by geometry, so a flagged cave is trusted and a clean read is "probably clean," never a certificate.

Why the limits are the feature

A measured read on ten real checkpoints, plus a coach's-eye read of the keyframes for everything else, turns out to be exactly what a form check needs: numbers where numbers exist, honest hedging where they don't, and a trend line — same angle, week over week — that no mirror or memory can give you. What it replaces isn't a great human coach. It's guessing.

So: can AI check your squat? The measurable half, better than your training partner's eyeball — and it will tell you plainly which half is which. Film a set, side-on, phone at hip height. See what the honest version looks like.