EvidPic methodology
How EvidPic checks images
EvidPic checks the image itself and the data inside its file. Neural networks look for signs of generation, while metadata can point to the camera or software that was used. Below, we explain how to read these results and where errors are possible.
Updated
01 / What the check includes
What we check
We analyze an image in several ways. Some look for signs in the pixels, while others read data about how the file was created and processed.
How neural networks analyze an image
For analysis, we use several neural-network detector models that look for signs of generation in an image. Their scores are shown in the report. If one model gives a high score and another a low score, they do not support a definite conclusion.
Several models can make mistakes on the same image. That is why it is important to consider the other check results too.
Metadata and generator markers
Metadata is information stored inside a file, such as a camera model or an editor's name. Some of these records are stored in EXIF format. We also look for markers left by image-generation software.
These records can be changed or removed. For example, an editor name in EXIF is a clue that processing may have happened, but it does not confirm what was done to the image.
C2PA and watermarks
C2PA is a standard for recording a file's creation and editing history. These records are called Content Credentials. They may contain a digital signature. When checking them, it is important to determine separately whether a record exists, whether its signature verified, and whether the person or organization that signed the data is trustworthy.
A digital signature helps verify that the signed data has not changed and is linked to this file. The scene in a photo can still be staged: a signature does not check that. Learn more about C2PA.
Some generators add invisible markers, or watermarks, to an image. SynthID is one such technology. Our model may detect signs of a watermark. This result should be distinguished from a response from an official verification service. If a separate SynthID/C2PA check through OpenAI is available, the image is sent there only after you press that check's button.
02 / Reading the report
How to read the report
Open an example report — see where to find the conclusion, the evidence behind it, and detailed check results.
- See what the conclusion is based on. Did a model find signs of generation? Is there an editor's name in the file? Was signed data found? These are different results.
- See which checks were performed. “Not found” means the check ran but did not find the relevant signs. The statuses “Check unavailable” and “not enough data” mean that this particular check cannot support a conclusion.
- Read the explanations. See whether the models agree with each other and whether there are warnings about the file format, compression, or other processing.
A score is not a percentage probability
The score shows how strong the signs of AI use found by the service are. For example, 80 out of 100 does not mean there is an 80% probability that an image was created by a neural network. To understand the score, look at which results influenced it.
03 / In practice
What the results mean
Let's look at three hypothetical examples. They show how to read a report, not how accurate a detector is.
A model gave a high score, but there is no metadata
The model found signs of generation in the image itself. But there is no information on how the file was created. It may have been stripped during sharing, so it is worth requesting and checking the original.
An editor is listed in the metadata
A program name does not explain what was done in it. The image could have been cropped, color-corrected, or processed with AI. One record alone cannot distinguish between these possibilities.
The check found no signs of AI
The completed checks found no signs of AI. That still does not mean you are looking at an in-camera photo or that no one edited the image.
04 / Limitations
Where errors are possible
- Screenshots and sharing. A screenshot does not preserve the original metadata. Sharing services can also remove it or compress the image.
- Compression, cropping, and resizing. After this kind of processing, models may score an image differently from the original.
- New generators and unusual images. Detectors may miss unfamiliar signs of generation or mistake features of an ordinary image for traces of AI.
- Changing one part of an image. The overall score may not reflect a small edit. If the report includes an experimental map of suspicious areas, it shows where to take a closer look. A highlight does not prove that area was changed.
- An unavailable check. Not every method works for every file. A separate model or external service may also be unavailable; this is marked in the report.
We do not give one percentage accuracy figure for every image. For such a number to be meaningful, it must say which files, model versions, and conditions produced it.
05 / Next step
If the result affects a decision
Keep the original file. Try to find the first publication: when and where the image appeared, who posted it, and what they wrote about it. If you only have a screenshot or a forwarded copy, request the original.
A report can help you understand a file. By itself, it cannot establish who created an image or whether the event it shows actually happened.
Go to image checkWhat happens to the file
For the check, the image is uploaded to the server. The original file is stored for up to 30 minutes so that you can take additional actions with it.
Data handling terms