Post #3735159

2026-07-11 08:00 UTC

Learning to recognize artificial pseudo-photographs is a very interesting domain (if "interesting" is the word). Again, repetition is the mother of learning, naturally. Myself, I failed the sample tests (four and four images) in the article. I partitioned them into two classes mostly correctly (real vs. artificial), but I swapped the classes. BBC article: See if you can spot an AI deepfake with our test #AI #ArtificialIntelligence #FakeImageRecognition #FakeImages #ImageRecognition #LargeLanguageModels #LLM

Replies (2)

@screwlisp@gamerplus.org 2026-07-11 08:09

@vnikolov@ieji.de one would imagine that in the traditional deep learning style, if you have a large training dataset of human real/fake photograph labellings, in the domain of the training dataset you would expect to be able to make fake face images humans are more likely to choose as real than real face images. I guess this is why political ads and ads generally have so desperately adopted fake images in that fake images can be optimized for persuasivity moreso than reality can.

Open ##3735158
@marintkael@mastodon.social 2026-07-11 09:49

the swap is the interesting bit though, because it means your discrimination was fine and only the labeling flipped. did the features that separated the two sets still hold up once you knew which was which, or did the tell dissolve the moment you had the answer key?

Open ##3908397