Checking the Medium Itself Instead of Just Trusting the Platform Label
Many apps today automatically show a note such as AI-generated. This is well intentioned, but it does not work as the sole source of truth. Such labels depend on a manufacturer setting them correctly and on them surviving uploading, compression, and re-sharing. In practice they often disappear, or they never appear in the first place because the image reached the web through indirect paths.
That is why aiorauthentic.com starts one level deeper. We analyze the file itself and look for traces that sit directly in the image or video. This way you do not have to rely on a platform labeling content honestly and completely. Our actual check rests on three mutually independent pillars, which we describe one by one below.
Pillar 1: Provenance Records Following the C2PA Standard
The first and strongest pillar consists of cryptographically signed provenance records following the open C2PA standard. Put simply, this is a tamper-evident history that can be embedded directly into the file: what the image was captured or created with, when, and which editing steps it went through afterwards. Because these details are cryptographically signed, they cannot be altered unnoticed.
The advantage: wherever a valid C2PA record exists, we can establish authenticity or editing with high confidence. The catch: such records are only present if the camera or the software used embedded them in the first place. Many devices and programs do not yet do this, and when content is shared across social networks the data is often lost. So a missing record proves nothing. It only means that we cannot use this one strong source of information and have to rely on the other pillars.
Pillar 2: Digital Watermarks Such as SynthID
The second pillar consists of digital watermarks that some AI generators embed when creating an image or video. A well-known example is SynthID from Google DeepMind. Such watermarks are invisible to the human eye, but they can be read out technically and can be a clear indication that a medium originates from a particular AI system.
Here too an important limitation applies, which we will not conceal: not every AI generator adds a watermark, and the various methods are not compatible with one another. The absence of a watermark is therefore not proof of authenticity. An image can be AI-generated without carrying any watermark at all. If, on the other hand, a valid watermark is found, that is a strong signal pointing toward AI generation.
Pillar 3: An Ensemble of Forensic AI Models
When neither a provenance record nor a watermark is present, the third pillar comes into play: an ensemble of several forensic AI models. These models are trained to detect statistical traces that typically arise during AI generation, such as anomalies in fine textures, in light gradients, at edges, or in the noise structure that a genuine photo does not exhibit in this way.
We deliberately do not rely on a single model, but on several model families for image and for video. The reason is simple: different models have different strengths and weaknesses. When several independent models agree in their assessment, the result is more robust than when only one model reaches a verdict. From these signals and the results of the first two pillars we ultimately derive an overall probability.
The Result: Probability Classes and the Abstention-First Principle
From all the signals we form a result that we do not output as a bare yes or no, but as a probability with clear classes. This way you see not only the verdict, but also how certain it is.
The abstention-first principle is especially important to us here. If the available evidence is not sufficient for a reliable verdict, we deliberately output not decidable instead of guessing. This happens, for example, with heavily compressed reposts in which many forensic traces have already been destroyed. We consider it more responsible to admit uncertainty honestly than to falsely accuse someone or to let a fake wrongly pass as authentic.
- verified authentic: established by a valid provenance record
- probably authentic: forensic signals predominantly point to a genuine medium
- AI-edited: the medium was subsequently altered with AI
- AI-generated: the medium was created by an AI with high probability
- not decidable: the available evidence is not sufficient for a robust verdict
Limits We Name Openly
No technology in the world detects AI content with absolute certainty, and our methodology should not claim that either. Several factors make a forensic check more difficult or impossible. We name them here deliberately so that you can put our results in the right context.
Compression and subsequent editing destroy the fine traces that the forensic models rely on. Video is generally harder to assess than a single image, because video compression loses even more detail. On top of that, new generators keep getting better and leave ever fewer telltale traces, which turns detection into a constant race. And once again the central point: the absence of a watermark or a provenance record never proves authenticity. It only means that we lack strong evidence and that the verdict is accordingly more cautious.
Privacy: What Happens to Your Medium
So that you can use our check without hesitation, we handle your data sparingly. The image or video you upload for the check is processed exclusively for the analysis and then deleted again. We do not build a permanent media collection from it.
Only the result of the check and a small preview image are stored permanently, so that you can recognize your request again later. The actual medium in full resolution does not stay with us. This keeps it traceable what you checked, without us retaining more of your content than necessary.