Can AI Check Your Workout Form Accurately?
By Mac Burke, qualified sports nutritionist
The short answer
Partly. Phone-based AI tracks 33 body points in real time and is good at counting reps, judging range and spotting changes from rep to rep. It is weaker at exact joint angles, front-on views and subtle back position, with studies showing errors of up to about 19 degrees. Treat it as a useful second set of eyes.
AI form checking is useful, with limits. A phone camera and a pose model can follow your body well enough to count reps, measure depth and notice when your range shrinks. It is much less reliable for precise angles and for anything that happens towards or away from the camera. Knowing where the line sits lets you trust it for the right jobs.
How does AI form checking actually work?
Most phone apps use a pose estimation model. The model finds points on your body, such as shoulders, hips, knees and ankles, in each video frame. The app then turns those points into angles and movement patterns.
Google's BlazePose, described by Bazarevsky and colleagues in 2020, is a typical example. It finds 33 body keypoints for a single person and runs at over 30 frames per second on a Pixel 2 phone. The authors named fitness tracking as one of the main uses. Google's ML Kit pose detection, built on the same idea, also returns 33 points.
There is a catch in the fine print. ML Kit's documentation says the depth value for each point is experimental and not a true 3D value. In plain terms, a single phone camera sees you as a flat picture. It estimates depth, but it does not measure it.
How accurate is phone AI compared with a lab?
It depends on the task and the angle. A few data points:
- Nakano and colleagues (2020) tested a multi-camera OpenPose setup against marker-based motion capture. About 47% of position errors were under 20 mm and about 80% were under 30 mm, but around 10% were over 40 mm.
- Asaeda and colleagues (2024) compared MediaPipe Pose with a lab system during single-leg landings. MediaPipe read knee angles about 19 degrees higher than the lab system. When measured as change from the moment of landing, though, reliability and agreement were good.
- Schurr and colleagues (2017) found plain side-on 2D video was within 0.7-3.1 degrees of 3D capture at the trunk, hip, knee and ankle, while front-on video was off by up to about 9 degrees.
The pattern is consistent. AI is better at relative change, such as whether this rep was shallower than the last, than at absolute angles, such as exactly how many degrees your knee bent.
What can AI form checking do well?
- Count reps. Clear up-and-down patterns are easy for a pose model to follow.
- Judge range. Whether your hips dropped below your knees or your arms reached full lockout.
- Spot change across a set. Shorter reps, slower reps and extra lean as fatigue builds.
- Flag obvious faults from the right angle, such as knees caving in a front-on squat.
- Stay consistent. It applies the same rule to every rep, which tired eyes often do not.
Where does AI form checking fall short?
- Exact angles. Errors close to 20 degrees are possible, as the Asaeda study showed.
- Movement towards or away from the camera. A single phone guesses depth.
- Covered body parts. Barbells, plates, benches and baggy clothes hide joints. Nakano found the largest errors came from the software mistaking objects for body parts or swapping limbs.
- Subtle spinal position. Small changes in lower back curve are hard to see in 33 points.
- Context. AI cannot tell you whether a position is fine for your body or why a joint hurts.
ML Kit also lists practical limits: it works best with your whole body in frame, needs your face visible, and tracks one person only. Busy gym footage breaks those rules often.
Is a human coach more accurate than AI?
Not always, which surprises people. Whiteside and colleagues (2016) compared a certified screener's live grades on six movement tests with grades from sensor-based motion capture. Agreement was poor on all six tests. The authors blamed vague grading criteria.
The lesson is not that coaches are useless. It is that clear rules and recorded video beat a quick glance, whoever is looking. The best setup combines both: software for consistent counting and range checks, a person for judgement and context.
Is AI form checking better than reviewing video yourself?
They do different jobs. Reviewing your own side-on video is slow but detailed. You can pause on the bottom of a rep, compare the first and last reps, and notice things software ignores, like where the bar sits on your back or how your grip looks.
AI checking is fast and consistent but shallow. It gives instant feedback on every rep, during the set, without you stopping to scrub through footage. That makes it well suited to high-rep bodyweight and dumbbell work, home sessions and anyone who will not sit down and review video.
A sensible split: let the app count and police range on every session, and spend five minutes once a week reviewing a side-on video of your main lift yourself. If something keeps showing up in both, that is the fault worth fixing first.
How do you get the most accurate results from an AI form app?
- Film side-on for squats, deadlifts, push-ups and presses. Side views are the most accurate.
- Keep your whole body and face in frame, with nobody else in shot.
- Wear fitted clothing so knees and hips are visible.
- Use good front or side lighting, not a bright window behind you.
- Keep the phone still and at hip height.
This is the job Move IQ in the MBM app is built for: the phone camera counts reps and flags reps that miss range. For judgement calls, the app also has coach chat. It is launching soon, with early access at /app.
Common questions
Does AI form checking work for deadlifts?
It can track hip and knee movement from the side, but the bar and plates often hide the shins and hands. Expect it to be better at counting and lockout checks than at judging back position.
Is my video uploaded when an app checks my form?
It depends on the app. Some process video on the phone and never store it, while others upload footage. Read the privacy policy before you film.
Will AI form checking get more accurate?
Probably, as models and phone cameras improve. Current research already shows good tracking of change across reps, with absolute angles the main weakness.
Can AI tell me if an exercise will cause injury?
No. It can flag positions that differ from a target, but it cannot assess your joints or history. For pain or injury, see a GP or physio.
Sources
- Bazarevsky et al. (2020), arXiv - BlazePose: on-device real-time body pose tracking
- Google ML Kit - Pose detection documentation
- Nakano et al. (2020), Frontiers in Sports and Active Living - 3D markerless motion capture accuracy using OpenPose
- Asaeda et al. (2024), Heliyon - Knee valgus angle calculation by posture estimation using machine learning
- Schurr et al. (2017), International Journal of Sports Physical Therapy - Two-dimensional video analysis vs 3D motion capture
- Whiteside et al. (2016), Journal of Strength and Conditioning Research - Grading the Functional Movement Screen
Ready to apply this to your own program?
MBM builds 1:1 coaching programs around the exact principles above - weekly check-ins, weekly adjustments, and a plan that fits your life.