Deepfake

A video shows a politician announcing a surrender. The face is right, the voice is right, the lips match the words. It was generated in an afternoon. The word for such a recording, deepfake, joins “deep learning” to “fake” and dates from 2017.

The obvious harm is that people will believe the fake. The philosophical literature has concentrated on a second harm, which does not depend on anyone being fooled.

Two accounts of the damage

Don Fallis argues that deepfakes reduce the amount of information that videos carry. A signal carries information about an event to the degree that it would not occur unless the event did. When convincing video can be produced whether or not the event happened, video in general tells us less, including the genuine video. People can respond by believing fakes, or by ceasing to believe real footage, and both are losses of knowledge.

Regina Rini’s account concerns testimony. Recordings, she argues, have served as an epistemic backstop: because a speaker might have been recorded, there was a standing check on what witnesses could get away with saying. If recordings lose their authority, the check goes, and ordinary testimony becomes less well regulated even where no recording is involved.

A related point is often called the liar’s dividend: once fakes are known to exist, a person caught on a genuine recording can say it is fake.

Where I stand

I think the second harm is the larger one, and that it shows something about recordings we had not needed to notice. A video was trusted as an instrument, like a thermometer: its content was fixed by the scene and not by anyone’s say-so. A deepfake has the form of an instrument reading and the origin of a claim, with no one standing behind it as its author. That is why the remedies that work look like the ones we use for testimony: provenance, a named source, a chain of custody. We are learning to ask of a video what we have always asked of a report, namely who says so.