Why use an AI image detector?
AI-generated images are becoming increasingly realistic and can be difficult to identify by eye alone. ImageSignals helps you screen suspicious images for visual signals associated with synthetic media, giving you additional evidence before you trust, share or act on an image.
Misinformation & fake news
Check suspicious images before sharing them, publishing them or using them as evidence.
Fake profiles & impersonation
Screen profile pictures and identity images that may have been synthetically generated.
Fraud & fabricated evidence
Analyse images used in claims, listings, reports or other situations where authenticity matters.
Marketplaces & user submissions
Screen potentially synthetic listings, submissions and other user-generated media.
Research & verification
Use visual AI detection together with metadata and provenance checks as part of a broader image-verification workflow.
AI detection is a screening signal, not proof of provenance. Important decisions should not rely on a single automated result.
Detection performance
The visual detector used by ImageSignals has been evaluated across clean, web-compressed and degraded image conditions.
Balanced accuracy
ROC AUC
Clean-image balanced accuracy
These figures are development and calibration results, not an untouched competition-test estimate. ROC AUC measures ranking across thresholds; it is not the percentage of images classified correctly.
What do these numbers mean?
Balanced accuracy — 93.3%
Balanced accuracy measures how well the detector distinguishes real and AI-generated images while giving equal importance to both classes.
This helps avoid misleading results where a model performs very well on one class but poorly on the other.
The reported out-of-fold macro balanced accuracy is 93.28% across clean, web-compressed and harder image conditions.
ROC AUC — 97.9%
ROC AUC measures how well the detector separates real and AI-generated images across different decision thresholds.
A score of 0.5 is roughly equivalent to random ranking, while 1.0 represents perfect separation.
The reported raw ROC AUC is 0.9790, indicating strong separation between real and synthetic images in the evaluation data.
Clean-image balanced accuracy — 95.7%
On clean, minimally altered images, the detector achieved 95.68% balanced accuracy.
Performance across real-world conditions
AI images are often resized, recompressed or shared through websites and messaging platforms. The detector was therefore evaluated under several image-quality conditions rather than only on pristine files.
- Clean
- 95.7%
Minimal processing beyond standard decoding and resizing.
- Web-compressed
- 93.9%
Simulates common web redistribution with reduced resolution and JPEG compression.
- Hard / degraded
- 90.5%
Uses stronger downscaling and compression to simulate more difficult real-world conditions.
- Worst of clean, web and hard per image
- 88.3%
Uses the least favourable result across the three conditions for each image.
Performance can still fall further under more extreme transformations or unfamiliar image-processing pipelines.
Strong improvement over the baseline
The current detector was fine-tuned and evaluated against its original baseline using the same image loader, evaluation images and calibration procedure.
Balanced accuracy
ROC AUC
Hard-image balanced accuracy
The largest gains appeared under compressed and degraded image conditions, where synthetic-image detectors often struggle most.
These figures are development and calibration results, not an untouched competition-test estimate.
Built for modern AI generators
The detector was trained on a broad mixture of real images and synthetic images produced by many generations of AI systems.
The training data includes 68,247 synthetic images across 131 source buckets, combining older generator families with newer frontier models.
- GPT Image
- DALL·E 3
- FLUX 1
- FLUX 2
- Imagen 3
- Imagen 4
- Seedream
- Qwen Image
- Hunyuan Image
- Midjourney
- Recraft
- Ideogram
- Janus
- Sana
- Z-Image
- Stable Diffusion
- GLIDE
- Wukong
- and others
The detector is not tied to a single AI generator. Training across many generator families helps it learn broader visual patterns associated with synthetic imagery.
Detection quality can vary between generators, and newly released systems may behave differently from the models represented in the training data.
Important limitations
No AI image detector is perfect. ImageSignals is designed to provide evidence, not absolute proof.
Very small images are harder to analyse
Severely downscaled or very low-resolution images can reduce detection quality.
Small AI-edited regions may be missed
The visual detector analyses the image as a whole. A small synthetic element inside an otherwise real photograph may not produce a strong AI signal.
Heavy processing can reduce confidence
Aggressive compression, repeated re-encoding, resizing and other transformations can remove or weaken useful visual signals.
New generators can behave differently
AI image generators evolve quickly. New systems and new processing pipelines may differ from the data used during training and evaluation.
A detector result is not proof of origin
A strong or weak visual signal should be interpreted together with context, metadata and provenance information whenever authenticity matters.
For higher-confidence verification, ImageSignals also checks available metadata and C2PA provenance separately from the visual detector.
Frequently Asked Questions
How does AI image detection work?
The visual detector analyses patterns in the image pixels themselves.
AI-generated images can contain statistical, structural and rendering patterns that differ from photographs or conventionally created digital images.
ImageSignals separately analyses metadata and provenance information where available, but these signals are not mixed into the visual detector score.
Does ImageSignals rely on metadata?
No. The visual AI detector analyses the image content itself and can run even when EXIF or other metadata is missing.
Metadata and provenance are analysed separately because they provide a different type of evidence.
Removing metadata therefore does not prevent the visual analysis from running.
What does the result mean?
The result describes how strongly the image matches patterns associated with real or AI-generated images.
For example, “No notable AI signals” means the detector found relatively little visual evidence associated with AI generation.
It does not prove that the image is authentic.
Can an AI-generated image be classified as real?
Yes. False negatives are possible.
They may be more likely with:
- new or unfamiliar generators
- very small images
- heavily compressed images
- repeated screenshots or re-encoding
- strongly edited images
- small AI-generated regions inside an otherwise real image
For important cases, use the visual result together with metadata and provenance checks.
Can a real image be classified as AI-generated?
Yes. False positives are possible with any probabilistic detector.
This is why ImageSignals presents AI detection as evidence rather than proof.
When authenticity matters, review the visual result together with the file metadata, C2PA provenance and any relevant external verification tools.
Can ImageSignals tell which AI generator created the image?
Not reliably. The detector is designed primarily to determine whether an image contains visual signals associated with AI generation.
It is not intended to identify the exact generator that created the image.
Does ImageSignals detect AI-edited images?
The visual detector analyses the entire image and may detect strong synthetic content.
However, it is not a localization model and is not designed to reliably identify a very small AI-generated or AI-edited region inside an otherwise real image.
When available, C2PA or other provenance information may provide stronger evidence of AI-assisted editing.
Does it work on screenshots and compressed images?
The detector was evaluated under both web-compressed and more heavily degraded image conditions.
Reported balanced accuracy was approximately:
- 95.7% on clean images
- 93.9% on web-compressed images
- 90.5% under harder degradation
More extreme processing can still reduce performance.
What is balanced accuracy?
Balanced accuracy measures performance on real and AI-generated images while giving equal importance to both classes.
This is useful because a normal accuracy figure can sometimes look artificially high if one class is easier or more common than the other.
The reported out-of-fold macro balanced accuracy is 93.28%.
What is ROC AUC?
ROC AUC measures how well a detector separates two classes across different decision thresholds.
A value around 0.5 corresponds to random ranking. A value of 1.0 represents perfect separation.
The reported raw ROC AUC for the detector used by ImageSignals is 0.9790.
This should not be interpreted as “97.9% of images are classified correctly”.
Why does ImageSignals also check metadata and provenance?
Visual detection, metadata and provenance answer different questions.
Visual detection looks for patterns inside the image itself.
Metadata may contain information about cameras, editing software or generation workflows.
Provenance systems such as C2PA can sometimes provide cryptographically verifiable information about how an image was created or edited.
ImageSignals keeps these evidence sources separate rather than combining them into a single misleading score.
What does “No verified provenance detected” mean?
It means ImageSignals did not find a locally verifiable provenance signal proving how the image was created or edited.
It does not mean the image is real.
Many authentic images contain no C2PA credentials, and many AI-generated images may also have their provenance information removed.
How many free analyses do I get?
ImageSignals provides a limited number of free image analyses.
Your current allowance is shown directly in the tool before you run an analysis.
What happens when I use all my free analyses?
ImageSignals is designed to remain accessible without requiring a subscription for occasional use.
An ad-supported credit system is planned to allow additional analyses after the free allowance has been used.
Until that feature is enabled, the interface will show the analyses currently available to you.
How will the ad-supported credit system work?
This feature is not yet active.
Once enabled, watching one eligible video advertisement will unlock one additional image analysis.
The advertisement only provides the analysis credit. It has no influence on the image result.
Why use ads?
Image analysis has a real computing and infrastructure cost.
The planned ad-supported model is intended to help cover those costs while keeping occasional use available without a paid subscription.
Do advertisements affect the analysis result?
No. Advertising and image analysis are completely separate.
The result is produced by the ImageSignals analysis pipeline and is not influenced by advertisers.