2 ms·
I'm not so sold on the specific example they use here. The tool specifically says it is a proof of concept and only predicted 52% chance of AI generated. Using
by godelski 3y ago
I'm not so sold on the specific example they use here. The tool specifically says it is a proof of concept and only predicted 52% chance of AI generated. Using those signals to conclude that something is "likely AI generated" is a bigger problem from the evaluator, not the tool.
While this is abundantly clear with ML, it is far more general: metrics are guides, not answers. Metrics are models, and like all models are wrong (wrong != useless && wrong != incorrect). And no model/metric perfectly aligns with intent. This needs to be something drilled into every single person's head. There is always more than needs to be done after a machine spits out a number, a lot more. If all we do is take a number at face value then we're going to fuck up a lot of stuff (which we're doing a lot of)
There are plenty of other examples of ML creating noise in the war environment, even around citizen interpretations of those images. For example, I've seen accounts whos profiles are of people that don't exist and I could confidently tell you the style generator that made them without an ML classifier (if you work in generation and haven't looked at millions of ML generated photos go to the top of this comment). (this is not saying I'm not a noisy detector, just I know naturally some key mistakes that get made that are non-realistic and these photos weren't cleaned up in post)