Detection Methodology

Content Credentials (C2PA)

Reads the provenance record some cameras, editors and AI tools attach to a file. A credential declaring AI generation settles the question on its own.

C2PA is an open standard for attaching a signed record of how a file was made. Some cameras, editing apps and AI image tools write one, including tools from Adobe, Microsoft and OpenAI. FauxLens reads the manifest and reports who issued it and what it says about the file. The two directions are treated differently. A credential that declares AI generation ends the scan even if it is unsigned, since nobody forges an AI label onto their own file. A credential that claims a real camera counts only if its signature chains to the published C2PA trust list.

A credential declaring AI generation is the strongest evidence there is, so it can end the scan early: the report then says a match was found instantly, so the full scan wasn't needed. The same goes for the AI label a platform put on a post shared from the Android app.

The limit is that credentials live in metadata, and metadata rarely survives the journey. Screenshots, messaging apps and most social platforms strip it. So a missing credential proves nothing, and FauxLens never treats its absence as evidence in either direction.

Metadata (EXIF and XMP)

Looks for a positive generator signature in the file's metadata. A missing camera record is never treated as evidence.

Beyond C2PA, many AI tools leave plain metadata behind: a software field, an XMP marker, a credit line such as "Made with Google AI". When the metadata names a generation tool, the report says so, and says that this is what decided the verdict, rather than describing a visual inspection that did not happen.

The rule we hold to is that only a positive signature counts. A photo with no camera make or model, or with its metadata stripped, is not suspicious: that is what every screenshot, every messaging-app forward and most downloads look like. Treating absence as evidence is the classic way a detector ends up calling real photos fake, and when our own pipeline stopped doing it, the number of real photos wrongly flagged in our held-out testing fell sharply.

The other direction exists too, and you should know about it: a photo carrying a full camera record (make, model and sensor settings) can be cleared as likely authentic on that metadata alone, without the pixel check. Metadata can be edited, so treat that result as strong but not conclusive.

Invisible AI Watermarks

Detects the invisible watermark OpenAI and Google embed in the pixels of images their tools make. Unlike metadata, it is designed to survive screenshots and compression.

Metadata is fragile; a watermark carried in the pixels is not. Google marks images from its AI tools with SynthID, and since May 2026 every image from ChatGPT and OpenAI's API carries an invisible watermark alongside its content credentials. These watermarks are designed to survive cropping, filters, screenshots and recompression, which is exactly what a forwarded image goes through.

When a watermark is found with high confidence, the report names the vendor family, OpenAI or Google, and not a specific model. The watermark identifies whose tool made the image, not which version, and we print no more than the evidence supports.

One case is handled specially. If the detection model says an image looks real but a high-confidence watermark is present, the image was most likely generated and then edited. The verdict then reads "Likely AI-Generated" at a moderate confidence, and never "authentic", because a real watermark is the stronger evidence. Watermarks only cover the vendors who add them: an image from a tool that does not watermark has to be judged on its pixels.

AI-Image Detection Model

A specialist model trained on real and AI-generated images scores the pixels, and tells AI generation and AI editing apart from real images.

Most images carry no credential and no watermark, so most verdicts come from the pixels. FauxLens sends the image to a specialist third-party detection model built for photographs and output from current generators. It returns a reading with a score: real, AI-edited (part of a real image altered with an AI tool), or AI-generated. Its separate "digitally edited" reading is treated as real: filters, crops and colour correction are not an AI claim, and that reading also fires on untouched screenshots.

For "AI-edited", the evidence line says exactly that: an AI tool altered part of the image, and the rest may be a genuine capture.

The primary model receives the file as uploaded, not resized (the fallback below receives a reduced copy), because recompression and downscaling measurably push real photos toward an AI score and wash out the traces in AI ones. That is also why a heavily compressed or forwarded image is our weakest case.

Multimodal Visual Review

When the detection model cannot answer, a multimodal vision model reviews anatomy, lighting and composition, and writes out what it found.

When the detection model cannot answer, because of an outage, a timeout or a frame decoded from a video, the scan falls back to a multimodal vision model. It looks at the image the way a careful reviewer would: anatomy, hands and teeth, the direction of the light, reflections, text in the background, and whether the scene obeys physics. What it finds becomes the evidence line in the report.

This pass is good at explaining a verdict and less certain than a model trained only for detection, which is why it is the fallback rather than the first choice. Its sampling temperature is set low so repeat scans of the same image stay consistent, and its output is held to the same verdict rules as everything else.

Video and Audio

Video is judged on its picture, from the whole file or from frames sampled across it. The soundtrack is checked too, but cannot clear or condemn a video on its own.

A video is judged on its visual content, either as a whole file or from frames sampled across it, which is how long clips and slow connections are handled without making anyone wait for a full upload. The soundtrack is checked for synthetic voice as well, but today that check can only strengthen a verdict the picture already points to. It never turns an authentic verdict into a fake one on its own.

Video detection has been evaluated far less than image detection, and a single convincing frame can mislead, so video verdicts deserve extra caution.

Verdict and Evidence Chain

The checks run in a fixed order and each is reported separately, so you can see what decided the answer.

The order is deliberate: provenance first, because a signed credential or a metadata signature can settle the question without a guess; then the watermark; then the pixels. The report lists the checks that produced a finding and what each found, so a verdict decided by metadata looks visibly different from one decided by the model.

The confidence score is a weight of evidence, not a calibrated probability. A score of 90% does not mean nine in ten such images are AI-generated, and we do not present it that way.

The detector is built for photographs and photo-like images. On most scans the detection model will still give a verdict on a drawing or a graphic, so treat one with caution; when the fallback reviewer handles a scan, it declines to judge drawings at all. A scan that fails is reported as a failure, with no verdict. It is not charged when the failure is ours, such as an outage or an error (after a few provider time-outs in one day, the next ones on that account or device are charged), but it is charged when the model ran and could not give a usable answer, or found the file was a drawing.