Contents
Maybe, and a detector can move you toward an answer, but it cannot settle the question on its own. The practical version of “is this image AI?” is a short evidence checklist in which the detector score is one input among several, and in which a confident number, in either direction, can still be wrong.
What a detector score actually tells you
A detector does not recognize fakery the way a person spots a wrong reflection. It measures whether an image carries a statistical trace left by the generators it was trained on, then reports how strongly. So a high score means “this image carries a trace I was trained to recognize,” and a low score means “I did not find my trace here.” On the models a tool has seen, that signal can be strong: Frank, Eisenhofer, Schönherr et al. (ICML 2020) reached 100% test accuracy separating real faces from StyleGAN faces, and AEROBLADE (Ricker, Lukovnikov, Fischer, CVPR 2024) reaches an average precision of 0.992 on Stable Diffusion, Kandinsky and Midjourney. But as the reference paper in the field put it, today’s generated images are “surprisingly easy to spot… for now” (Wang, Wang, Zhang et al., CVPR 2020), and the “for now” is the whole point: detection holds on the models and conditions a tool has seen, and slips outside them.
The checklist
| Check | What it tells you | Where it fails |
|---|---|---|
| Test the original file | Preserves the traces a detector reads | Screenshots and re-saves strip the signal |
| Run two or more detectors | Agreement across methods raises confidence | They often disagree on the same image |
| Look for C2PA or SynthID | A present credential is strong evidence | Most images carry no mark at all |
| Trace the source | Context and origin can settle it outright | A convincing image can be shared with no context |
Each row is a signal, not a verdict, and the point is to combine them rather than lean on any one.
Why a confident “real” can be wrong
The most dangerous result is a falsely reassuring one. Compression is the common cause: when an AI image is saved for the web, the same re-encoding that helps it blend in also erases the traces a detector needs, dropping a frequency detector from 70.5% to 52.7% AUC (Corvi, Cozzolino, Zingarini et al., ICASSP 2023). Bellingcat’s test of a popular detector found it labeled 7 of 10 compressed Midjourney images “real.” A clean “not AI” from a single tool on a heavily compressed file is close to no information at all.
Why a confident “AI” can be wrong
The opposite error is worse for the person on the receiving end, because it accuses. Detectors flag genuine images too. NewsGuard’s test of five tools found false-positive rates as high as 40 percent, with one widely used tool at 20 percent. Grommelt, Weiss, Pfreundt et al. (2024) explain one mechanism: because detection datasets carry “biases related to JPEG compression and image size,” a detector can mistake ordinary compression for a generation artifact and flag a real photo. This is the image version of the wrongly-accused problem that hits real writers in text, described in why AI text detectors falsely accuse real writers.
What if two detectors disagree?
Expect it, and read the split as information rather than a malfunction. Detectors disagree because they read different signals, a split examined in why do AI detectors give different results, and even the tools built to combine those signals do not resolve every case: AIDE (Yan, Li, Cai et al., ICLR 2025) reports gains of +3.5% and +4.6% over prior methods while concluding the problem “remains far from being solved.” So when one tool says AI and another says real, do not average them into a verdict. Ask which signal each read, whether the tested file was the original, and whether anything off the file agrees.
When the checklist actually resolves it
Sometimes you can get to a real answer. A readable Content Credential or SynthID watermark, several independent detectors agreeing, or a traceable original source can each move a “maybe” toward a confident conclusion. The cases that stay unresolved are the ones where the file is degraded, the generator is new (Ojha, Li, Lee, CVPR 2023), and no provenance exists, and there the right output is “undetermined,” not a guess dressed up as a percentage.
The bottom line
So, is the image AI-generated? A detector gives you a probability shaped by what it was trained on and how the file was handled, not a verdict, and the responsible answer combines that score with provenance, corroboration, and source. Treat a single number as a reason to look closer, not as the finding. For the step-by-step version of this check, see how to check an image for AI.
Sources
- Frank, Eisenhofer, Schönherr et al. (2020). Leveraging Frequency Analysis for Deep Fake Image Recognition. ICML 2020.
- Ricker, Lukovnikov, Fischer (2024). AEROBLADE: Training-Free Detection of Latent Diffusion Images Using Autoencoder Reconstruction Error. CVPR 2024.
- Wang, Wang, Zhang et al. (2020). CNN-generated images are surprisingly easy to spot… for now. CVPR 2020.
- Corvi, Cozzolino, Zingarini et al. (2023). On the Detection of Synthetic Images Generated by Diffusion Models. ICASSP 2023.
- Grommelt, Weiss, Pfreundt et al. (2024). Fake or JPEG? Revealing Common Biases in Generated Image Detection Datasets.
- Yan, Li, Cai et al. (2025). A Sanity Check for AI-Generated Image Detection. ICLR 2025.
- Ojha, Li, Lee (2023). Towards Universal Fake Image Detectors that Generalize Across Generative Models. CVPR 2023.
- 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?