Do AI image detectors work on screenshots and compressed social media copies?

Start by asking for the original, and ask again. A screenshot is a new image that your phone or computer created. It inherits the pixels, loosely, and none of the file history. Anything attached to the original file does not travel with it.

For the desk: write down every hop you can see, from which account, on what date, and treat the earliest copy you can find as the thing to analyse. If you cannot get closer to the source, the right statement is "we cannot tell", and I would rather see that written down than a confident percentage.

if the image came from one of the big generators it has that invisible watermark baked in, and those survive screenshots because they are in the pixels, not the file. so just run the screenshot through the checker from the company that made the generator and you will know. that is the reliable route.

@Jules I would not rely on that. A watermark in the pixels is designed to survive some changes, but how well it survives depends on the scheme and on how hard the image has been altered, and a screenshot plus recompression plus crop is a stack of transformations. Vendors can say what their own tool is built to withstand. Nobody outside can promise it holds for your third-generation copy. And the opposite direction matters more for a newsroom: a checker finding nothing is not evidence that the image is real, because plenty of generators do not add any mark at all.

Photographer here. Clients now send me screenshots of my own work and ask if it is real. I have started sending the original with the camera file attached. When I cannot, I tell them the screenshot cannot prove anything on its own, in either direction. That has not cost me a single client.

Short answer: they can, and it is a documented weakness. Compression, resizing and re-encoding change what detectors see, and detectors tuned on clean examples can be less reliable on copies that went through social platforms. A screenshot also drops the original file's metadata. For a third-generation copy, an unreliable score plus no provenance means "unverified", not "real" or "fake".

Two sources help here. The paper "Fake or JPEG? Revealing Common Biases in Generated Image Detection Datasets" looked at datasets used to train and test detectors and found that many carry biases related to JPEG compression and image size. Using one such dataset, the authors showed that detectors learn from those factors, and that removing the biases substantially increased robustness to JPEG compression and changed how the detectors compared across generators. So part of what a score reflects can be file format rather than image content, which is the wrong thing to base a newsroom decision on.

An older paper, "CNN-generated images are surprisingly easy to spot... for now", shows both sides. A classifier trained on one generator generalised to others "with careful pre- and post-processing and data augmentation", which is a reminder that results depend on how the training and the input handling are done, and that this was reported for the generators of that period, not for current ones. I would not carry its optimism over to today's images or to screenshots.

What I would do: chase the earliest copy, check whether the image carries any provenance information (see the thread on content credentials), use reverse image search and context, and treat a detector as one weak input. The reliability thread covers the wider evidence.

@Tomas Reyes thank you, that matches what we see. "Unverified" is the word we will use. The point that part of the score can come from file format and size rather than the picture is the one I will take to the desk, because it explains why two copies of one image can disagree.