How Can We Detect AI-Generated and AI-Modified Images, and Where Does Detection Fall Short?
How Can We Detect AI-Generated and AI-Modified Images, and Where Does Detection Fall Short?
AI image detection tools can estimate whether a picture was generated or altered by AI, but they cannot prove it, and their reliability drops once an image is cropped, resized, retouched or edited several times. On Tuesday 20 October 2026 at 17:00 CET, CEPIC brings together Florian Barbaro, PhD (CEO of UncovAI) and Mete Zihnioglu (Directeur Général of Sipa Press) to put detection to the test on real images.
In short: a detector gives a probability, not a verdict. For newsrooms, photo agencies and fact-checkers, the safest approach combines detection (an estimate made after the fact) with provenance standards such as C2PA and CAWG (a record made at the time of creation and editing).
Why AI image detection matters for visual media
AI-generated and AI-modified imagery is becoming more sophisticated, and the question is no longer only "is this image real?" but also "what has been altered, and how?" A press photograph may be wholly authentic, lightly retouched, partly generated, or the product of several successive edits. Each case calls for a different editorial decision, and a single yes/no score cannot capture all of them.
What can AI image detection tools identify?
Detection tools are trained on large datasets of real and synthetic images and look for statistical patterns left by generative models. Used carefully, they can:
- Flag images that look fully AI-generated, giving an editor a reason to investigate further.
- Signal possible partial manipulation, when only some regions of an image appear synthetic.
- Help triage large volumes of submissions, so human verification time goes where it is most needed.
Where does AI image detection fall short?
Detection is an estimate, and its accuracy depends on what has happened to the image. The webinar will examine these cases with real examples:
| Image scenario | Why it is hard for detectors |
|---|---|
| Original, unmodified photo | A false positive is still possible: a genuine image can be wrongly flagged as AI-generated. |
| Photoshop or conventional retouching | Legitimate editing can be confused with, or hide, AI-based changes. |
| Fully AI-generated image | Generators evolve quickly, so detectors must keep up with new models. |
| Partially AI-modified image | A small synthetic region inside a real photograph is harder to spot than a fully generated image. |
| Cropped or resized image | Cropping and resizing can alter or remove the traces a detector relies on. |
| Image with several successive modifications | Layered edits blur the signals, and the result is difficult to interpret. |
The takeaway: a detection score is a lead to investigate, never a conclusion. When a photographer's credibility or an agency's reputation is at stake, an estimate is not enough.
Detection estimates, provenance documents
Detection examines a finished image and guesses how it was made. Provenance works the other way around: it records how content is made, as it is made, in a form that can be verified later. The two approaches are complementary.
- C2PA attaches a tamper-evident manifest to a file, recording the tools used, the edits made and any use of AI.
- CAWG builds on C2PA by adding verifiable identity, showing who claims to have created the content.
- EU AI Act, Article 50 has applied since 2 August 2026 and sets transparency obligations for AI-generated content, including machine-readable marking by providers of generative AI systems and labelling of deepfakes by deployers.
For a deeper look at these standards, read our analysis of why a detector score is a lead, not proof.
Webinar details: AI image detection put to the test
- Date and time: Tuesday 20 October 2026, 17:00 CET
- Organiser: CEPIC, the Coordination of European Picture Agencies Stock, Heritage and News
- Speaker, detection: Florian Barbaro, PhD, CEO of UncovAI, on what today's AI image detection can and cannot identify
- Speaker, press photography: Mete Zihnioglu, Directeur Général of Sipa Press, on the experience of an independent press agency
- Registration: details to follow
The session will explore what detection tools can identify, where their limits lie, and what happens when images are generated, retouched, cropped, edited or modified using AI.
CEPIC members: contribute images for the live test
CEPIC is looking for real-world images to use as practical examples. Contributions could include:
- Original, unmodified images
- Photoshop or otherwise retouched images
- Fully or partially AI-generated or AI-modified images
- Cropped or resized images
- Images that have undergone several different modifications
Your examples will help demonstrate both the possibilities and the limitations of AI detection. To contribute, get in touch with the CEPIC team.