What does an AI image-detector score mean?

An AI image-detector score is a number that says how strongly a picture matches the generated examples a model was trained on, read at a threshold you usually cannot see. It is not the probability that the image is AI. These tools report their behavior as accuracy or area under the curve across a benchmark, and that figure describes the whole test set, not your file, which is why the same score can mean very different things from one image to the next.

What is the detector actually outputting?

A single number for how closely your image resembles the generated images the model learned from, mapped through a threshold into a label. It is not a measurement of reality. Wang, Wang, Zhang et al. (CVPR 2020) built one of the standard pixel-based detectors and showed it reads a generator’s trace well inside the family it trained on. Ojha, Li, Lee (CVPR 2023) report the ceiling of that approach as mean average precision, 100 on the GAN family a detector trained on and 61.32 on an autoregressive model it had not seen. Those figures describe how a model behaves across a benchmark, not how sure you should be about one photo.

Does “94 percent AI” mean a 94 percent chance the image is fake?

No. The output is a similarity-to-training score passed through a function, not a calibrated probability that the picture is generated. Reading “94” as “94 percent likely AI” assumes the tool is calibrated on images like yours, and it usually is not. The gap appears the moment the image leaves the benchmark: Corvi, Cozzolino, Zingarini et al. (ICASSP 2023) found a spectral detector fell to between 44 and 49 percent AUC, essentially chance, on diffusion families such as ADM and Stable Diffusion once ordinary compression was applied. The same displayed confidence can sit anywhere along that drop.

What is the threshold, and why does it matter?

Every detector ends in one decision, is the score high enough to call “AI,” and that line is a choice. A tool set to catch more fakes will also flag more real photos, and a tool tuned to avoid false accusations will pass more fakes. The costs are asymmetric: a false “AI” label on a genuine news photograph harms a real person, while a missed fake spreads a lie. NewsGuard tested five detectors on the same 45 images and recorded false-positive rates from 0 percent (Hive, Sightengine) to 40 percent (ScamAI), with ZeroGPT at 20 percent and AI or Not at 6.67 percent, the false-alarm side examined in AI image detector false positives. Same images, same task, very different lines, and the vendor rarely shows you which one produced your score.

Why does the same score mean different things on different images?

Because the accuracy behind the score depends on where the image came from and what happened to it. Corvi and colleagues showed that social-media compression and resizing drop a frequency detector from 70.5% to 52.7% AUC, close to a coin toss, on the very same detector. A “90 percent” on a clean file from a known generator and a “90 percent” on a re-saved screenshot from an unknown one are not the same claim, even though the interface prints one number.

What can make a high score misleading?

A high score can be reading the wrong thing. Grommelt, Weiss, Pfreundt et al. (2024) showed standard detection datasets are biased by JPEG compression and image size, and that detectors “indeed learn from these undesired factors,” with cross-generator accuracy shifting by more than 11 percentage points once those biases were removed. So a confident score can be a statement about how a file was compressed or resized, dressed up as a statement about whether it was generated. It can also be a genuine upsampling-fingerprint match, the frequency trace Frank, Eisenhofer, Schönherr et al. (ICML 2020) tied to how generators build an image, which is useful only when the detector’s training covered your generator.

What can make a low score misleading?

A low score can mean the image is real, or that the detector never learned this generator, or that compression stripped the trace before the detector saw it. Bellingcat, testing a popular detector, found it called 7 of 10 compressed Midjourney images “real,” so a clean result on a re-encoded fake carries little information. The reverse error is just as common: the same test flagged 6 of 20 genuine contest photographs as AI, so neither a low nor a high score is self-justifying on a file outside the tool’s training.

How should you read a score, then?

As a similarity reading at an operating point you are usually not shown, and as an invitation to ask four questions: what file did you test, which generator might have made it, what threshold was applied, and what is the false-positive rate on images like this. Even the newest multi-signal detectors keep this humility: AIDE (Yan, Li, Cai et al., ICLR 2025) combines a semantic expert and a patch-statistics expert for gains of 3.5% and 4.6% over prior methods, while stating that detecting AI-generated images “remains far from being solved.” Preserve the original file, run a second tool that reads a different signal, look for provenance, and weigh the source. Putting that reading to work on one image is the practical question in is this image AI-generated.

Sources

• 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.
• 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.
• Frank, Eisenhofer, Schönherr et al. (2020). Leveraging Frequency Analysis for Deep Fake Image Recognition. ICML 2020.
• Yan, Li, Cai et al. (2025). A Sanity Check for AI-Generated Image Detection. ICLR 2025.
• 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?