Yesterday I talked about computer graphic systems that are able to synthesize realistic faces of people who don’t exist. They do this by looking at large numbers of real peoples’ faces, and computing all sorts of statistical data from those faces.
When faces are analyzed this way, each real person’s face ends up being represented in the computer as a single point in a high dimensional mathematical space. If two data points are near each other in that space, then they will tend to produce faces that look similar, but not quite the same.
Once this framework is set up, then to create a new realistic face you can just choose a new point in that high dimensional space. The face that you create will end up looking more like some real people and less like others, but it won’t look exactly like the face of anybody who actually exists.
To me this raises an interesting ethical question: If you are training this model on the faces of all people, then are you violating anybody’s privacy or ownership over their own appearance? Or are you just doing what we all do every day inside our heads — building a model of what a face looks like based on the people we see around us?
And if that is the case, do you then have free license to use that data to create any new faces you want? And if not, then why not?