Method and limits

This site makes only statements that the data at hand can support. It does not guess from visual impression or writing style whether something was machine generated. Every method used is reproducible: same input, same result.

Why there are three results and not two

A verdict of "written by a human" cannot be justified with the current state of the art. What can be established is the presence of a marker, not its absence as counter-evidence. Platforms strip metadata on upload almost without exception; an image without markers is therefore the normal case and no indication of anything.

Red – machine origin established

The file carries a marker of machine generation, or a watermark was decoded from it.

Green – camera capture established

A signed provenance record identifies the file as a camera capture, the signature chain checks out against the C2PA trust list, and the hashes still match the image data.

Grey – nothing established

None of the checked markers is present. Nothing follows from that.

The methods used

C2PA provenance record

The Coalition for Content Provenance and Authenticity describes a signed provenance record embedded in the file. Three things are checked separately: whether the hashes still match the image data, whether the signature chain checks out against the official trust list, and what the record states about the kind of source. The trust lists are refreshed regularly.

This is the only method with cryptographic substance, but it has known weaknesses. A study dated 23 April 2026 (arXiv 2604.24890) describes, among other things, interchangeable timestamps, incomplete checking of revoked certificates, and the fact that different validators reach different results on the same file. Version 2.4 of the specification, dated 21 April 2026, does not address these points.

IPTC and EXIF labelling

The IPTC vocabulary defines dedicated values for the kind of source, among them trainedAlgorithmicMedia for material generated entirely by a model, compositeWithTrainedAlgorithmicMedia for mixtures and digitalCapture for camera captures. These statements are not signed. They can be stripped with a single command and entered just as easily. A finding is therefore explicitly presented as a claim made by the file, not as an established fact.

Generator working data in the image

Locally run image generators store their working data in text blocks of the PNG file: Automatic1111 under parameters, ComfyUI under prompt and workflow with the complete node graph, InvokeAI under sd-metadata. Such data arises during generation and is not added afterwards. Saving in another format loses it.

Invisible watermark in the image

What is checked is the openly documented method from the Stable Diffusion ecosystem (dwtDct and dwtDctSvd), which places the identifier in the image data itself. Google's SynthID cannot be checked: there is neither an open interface nor a detector that can be run independently, and that is deliberate – a published detector would be open to attack.

Green/red list watermark in text

The method comes from "A Watermark for Large Language Models" by Kirchenbauer, Geiping, Wen, Katz, Miers and Goldstein. During generation a list of favoured words is computed from the preceding context and given an advantage when the next word is picked. The check retraces the same computation and counts how often the choice fell on that list.

The z value states by how many standard deviations the measured share exceeds what chance would give:

z = (hits − γ · T) / √(T · γ · (1 − γ))

A grid of two seeding schemes, two list shares and six tokenizers is checked. Because several tests are run, the error probability of the best value is multiplied by the number of combinations. Without this correction the sheer number of attempts would produce spurious hits.

What matters for interpretation: only content generated with one of the openly documented settings can be detected. The widely used language models do not currently apply this method. A missing detection is therefore the normal case.

Hidden characters

Zero-width and directional control characters are used to carry hidden markings. They are counted and named. But they also arise from typesetting and from copying out of word processors, so they are a lead and not a statement about origin.

What is deliberately not done here

No assessment by visual impression. The freely available models for detecting AI-generated images reach around 60 per cent accuracy at roughly 17 per cent false positives in independent measurements – one in six genuine photographs would be wrongly labelled as machine generated.

No stylistic analysis of text. A study by Stanford University (Liang and others, Patterns, July 2023) measured seven widely used detectors on essays by non-native speakers: 61.3 per cent false positives on average, with more than 91 per cent of the essays flagged by at least one detector. Such a method systematically accuses the wrong people.

No error level analysis and no sensor noise analysis for the question "AI or not". Both methods were developed for different questions – locating local edits and attributing an image to a particular camera – and do not answer this one.

No transfer to external detection services. The entire check runs on our own server.

Legal background

Article 50 of the European Union's AI Act has applied since 2 August 2026 and requires machine-generated content to be marked in machine-readable form. The accompanying Code of Practice on Transparency was finalised on 10 June 2026 and assessed as adequate by the Commission and the AI Board on 8/9 July 2026. It states that no single technique meets the legal requirements and recommends two layers: signed metadata and a watermark in the content itself. In Germany the AI market surveillance and innovation act (KI-MIG) has been in force since 29 July 2026; the competent authority is the Bundesnetzagentur.

This site generates no content of its own and is therefore not subject to the marking obligation. It checks whether others have met it.

The limits in one sentence

A finding establishes something. A missing finding establishes nothing. This site is not evidence and is not a basis for an accusation against any person.