An AI tennis report needs evidence for each coaching claim
Video input and pose estimates can support analysis. A useful coaching report still needs observations tied to footage and a test against qualified review.

An AI report tells a tennis player that their timing is late. The statement sounds plausible, but the player needs to know which stroke the system means and what observation supports it. Without a reference to the footage, the advice is difficult to check or discuss with a coach.
That is a stronger test than asking whether the report uses convincing tennis vocabulary. A system can describe a common technical fault without having established that the fault occurred in the uploaded clip. The evaluation should connect each consequential claim to something visible at a specific time.
The technical starting point has also changed. It is inaccurate to say that chatbots generally cannot accept video. Google's current Gemini video documentation supports video input and timestamp-based questions. Whether a particular service exposes that capability depends on its implementation. Accepting the file is an input capability, with coaching accuracy left to establish.
Identify what the system actually examined
A video pipeline may sample frames instead of processing every frame at the source rate. Sampling is a normal engineering choice. Its suitability depends on the event being measured. A wide interval can preserve the course of a rally while missing the moment of racket contact.
Ask the system's operator to document the sampling configuration for the model and input path used. Keep that information with the analysis. A universal calculation based on sending every source frame to a model can badly misstate both cost and temporal coverage.
The distinction affects the language of the result. A report may be able to locate a rally or summarize visible movement. Claiming a precise contact time requires enough temporal evidence to support that precision. An answer with a decimal point can still exceed what the sampled footage allows.
For an evaluation, use clips with events that a reviewer can locate independently. Preserve the original file and the relevant timestamps. If a pipeline changes its sampling later, run the same clips again so the comparison has a known reference.
A pose estimate is one input to an interpretation
MediaPipe Pose Landmarker detects body landmarks in images and video, with outputs in image coordinates and estimated three-dimensional world coordinates. These outputs can help build an analysis tool. They do not supply tennis coaching rules or prove that a particular recommendation will help a player.
Camera placement affects what can be observed. A joint hidden behind the body is not directly visible in that frame. A racket can pass outside the image. The analysis should identify when the view is insufficient for the requested measurement, instead of filling the gap with a confident explanation.
A body-landmark pipeline also does not automatically track the ball. A claim about ball speed requires its own measurement method and reference. Adding several numbers to a report can create the appearance of precision while leaving each number's basis unexplained.
A useful output separates observation from interpretation. The observation might identify a frame where the player's position is visible. The interpretation explains how that position relates to the stroke being assessed. A coach can then examine the connection and disagree with it without having to reconstruct what the model looked at.
Test against review on unfamiliar footage
The NIST AI risk framework's discussion of validity ties evaluation to the intended context of use. Applied here, an attractive report from a prepared demonstration clip is limited evidence. A tool intended for ordinary phone recordings needs to be tested on those recordings, including footage with imperfect views.
Have qualified reviewers assess a set of clips before comparing the system's claims with their findings. Include cases where the footage is insufficient to judge the requested event. Correctly declining to estimate can be a useful result. It keeps uncertainty visible to the person who will act on the report.
Disagreement should be examined at the claim level. If the system identifies the right rally but gives an unsupported explanation of the player's movement, those are separate outcomes. A single rating for whether the whole report seems useful can hide the unsupported part.
Keep a separate set of clips for the final check after adjusting prompts or processing. Repeatedly tuning against the same demonstration can improve that example while telling little about performance on new footage. Document the version that produced each report so a later change can be evaluated against the same standard.
For a first trial, ask the tool to support one narrow observation with a timestamp, then review that observation beside the original recording. Broader coaching recommendations can follow only where the evidence and qualified assessment support them. The player should be able to see where the observation ends and the advice begins.