Skip to content

Draft. The steps on this page are not yet checked on the live app.

How GRAPE MARS uses AI, in plain words ​

GRAPE MARS makes 6 choices about how it uses AI. This page explains each one and where it stands today.

Before you start ​

  • These come from a study of the AI tools that suit this kind of research.
  • The study mapped 18 kinds of AI and weighed real models against each one.
  • 6 choices were left for the team. Each has a working answer for now, so work can go on.

1. Which licences we accept ​

Some strong models carry a licence that would force us to publish our own code. The working answer is to keep those out of the site and use a freer model instead. Research weights that ban paid use are fine for UJI work, but we do not pass them on.

2. How old a model may be ​

We include a model if someone worked on it in the last 24 months. We call it well kept if that was in the last 12. Some older research code is still the best there is, so we keep it by hand when it still runs.

3. What we build first ​

Language work is both the best served and the cheapest, because much of it runs inside your browser. So it goes first. The study set no formal ranking, and the lab's own plans should shape this.

4. The order of the next steps ​

The plan is: body, hand and face points; then telling speakers apart; then a plain description of facial movement; then where a person looks; then shot type. The first step serves 2 studies at once. See AI features announced.

5. Testing on our own footage ​

Every accuracy figure we have comes from clean, mostly English material. Nobody has yet tested these models on our own recordings, on Valencian or Catalan, on switched languages, on children, or on real gesture.

Until someone does, treat any AI result as help with a first pass. It is not a measurement. Check it as you would check the work of a new coder.

6. What we do with people's material ​

  • We do not offer to read emotion or engagement. The science behind it is contested.
  • Where a person looks, how they move, and the sound of their voice count as personal data.
  • GRAPE MARS does not match a person across 2 recordings.
  • Work runs in the most private place that can do it, and keeps your video on your own machine where it can.

The team still has questions for the university about consent. Nothing here is settled until they answer.

Check it worked ​

Nothing to do. Read this before you tell a colleague, or a paper, what the AI in GRAPE MARS does.

If something goes wrong ​

GRAPE MARS · Universitat Jaume I · Contact support