Contents
You can get a useful first-pass read on an image in a minute or two, but a detector gives you a probability, not a verdict. The steps below make that read as reliable as it can be, and, just as importantly, tell you when to distrust the number it returns.
Start with the original file
Test the file as it came to you, not a screenshot or a re-save. Detectors rely on faint traces in the pixels, and re-encoding an image strips those traces. Frank, Eisenhofer, Schönherr et al. (ICML 2020) showed that the frequency-domain artifacts many detectors read come from the generation process itself, so compressing or resizing the file blurs exactly what the detector was going to measure. Corvi, Cozzolino, Zingarini et al. (ICASSP 2023) put a number on the cost: social-media-style compression and resizing drop a frequency detector from 70.5% to 52.7% AUC, close to a coin toss. A screenshot of a screenshot is nearly worthless to a detector, so find the highest-quality version of the file you can.
Read the confidence score, not the label
Drop the original into a detector and look at the number, not the headline word. A result of “AI, 92 percent” and a result of “AI, 51 percent” are very different findings even though both say “AI.” The score measures how strongly this one detector found the trace it was trained on, at a threshold the vendor chose, and different families read different signals: a frequency detector, a pixel-CNN like Wang, Wang, Zhang et al. (CVPR 2020), and a feature-space detector like Ojha, Li, Lee (CVPR 2023) are not interchangeable. The score is not a probability that a machine made the image. Treat a borderline number as unresolved, not as a weak yes.
Look for provenance signals
Some images carry a provenance mark you can read directly. Content Credentials (the C2PA standard) and Google’s SynthID watermark can both be checked with the right viewer, and when one is present it is far stronger evidence than any detector score. The catch runs the other way: most images carry no such mark, so an absent credential proves nothing. “No watermark found” means only that, not that the image is real.
Get a second opinion, and corroborate off the file
No single detector is authoritative, so run at least one more that works differently. A reconstruction-based tool such as AEROBLADE (Ricker, Lukovnikov, Fischer, CVPR 2024) reads a different signal than a frequency detector, so agreement between them means more than two readings of the same trace. Detectors disagree often: NewsGuard, testing five tools on the same 45 images, found at least one tool dissenting on 35 of them. Then look beyond the pixels entirely. A reverse image search can place the picture earlier than the event it claims to show, or surface the original, and context like that can settle a case a detector cannot. If two independent signals agree, your confidence should rise; if they split, the image is sitting at the edge of what these tools resolve, and you should say so rather than pick the answer you prefer.
Know when to distrust the result
A confident answer can still be wrong in both directions. On the false-negative side, compression can hide a fake: Bellingcat found a popular detector called 7 of 10 compressed Midjourney images “real.” On the false-positive side, a detector can flag genuine work, especially images from a generator it never trained on (Ojha, Li, Lee, CVPR 2023) or files whose compression it misreads as a generation artifact, the dataset bias documented by Grommelt, Weiss, Pfreundt et al. (2024). Unusual content, screenshots, artwork, and low-resolution files are where false alarms cluster.
The mindset that makes it work
So the procedure is simple, but the mindset matters more than the tool. Use the original file, read the score, check for a credential, get a second opinion, corroborate off the file, and stay honest about the cases where none of it is reliable. A detector is a prompt to look closer, not a verdict. For why two detectors so often disagree, see why do AI detectors give different results; for what a score does and does not prove, see is this image AI-generated; and for how far these tools can be trusted overall, see are AI detectors accurate.
Sources
- Frank, Eisenhofer, Schönherr et al. (2020). Leveraging Frequency Analysis for Deep Fake Image Recognition. ICML 2020.
- Corvi, Cozzolino, Zingarini et al. (2023). On the Detection of Synthetic Images Generated by Diffusion Models. ICASSP 2023.
- Wang, Wang, Zhang et al. (2020). CNN-generated images are surprisingly easy to spot… for now. CVPR 2020.
- Ojha, Li, Lee (2023). Towards Universal Fake Image Detectors that Generalize Across Generative Models. CVPR 2023.
- Ricker, Lukovnikov, Fischer (2024). AEROBLADE: Training-Free Detection of Latent Diffusion Images Using Autoencoder Reconstruction Error. CVPR 2024.
- Grommelt, Weiss, Pfreundt et al. (2024). Fake or JPEG? Revealing Common Biases in Generated Image Detection Datasets.
- NewsGuard (2024). Leading AI Image Detection Tools Mislead Online Users, Often Declaring Authentic Content Fake.
- Bellingcat (2023). Testing ‘AI or Not’: How Well Does an AI Image Detector Do Its Job?