Methodology, in plain language.

What we check in images

First, we look for Content Credentials (C2PA), a signed record that can describe a file’s origin and editing history. A valid, trusted credential is the strongest signal we show. We also inspect EXIF, XMP, and PNG metadata for recognized generator markers. Metadata is easy to remove or change, so it is supporting evidence only.

We then run a Community Forensics image model trained across thousands of image generators. This is a statistical signal, not a proof. Compression, screenshots, editing, and generators outside its evaluation data can all affect its result.

What we check in text

Text is scored in overlapping sections with a classifier trained to identify language patterns associated with AI-generated writing. Short, translated, heavily edited, mixed-authorship, and specialist text can be especially difficult to classify. We report disagreement and uncertainty instead of forcing a label.

We do not relabel this classifier as Google SynthID Text. SynthID requires a compatible watermark configuration and tokenizer; Google’s open reference implementation is intended for research and its hashing does not automatically transfer to Gemini App output. See our AI watermark guide for the current provider-by-provider snapshot.

What we check on a website

We fetch one public HTML page and inspect its visible English copy, title, main heading, indexing directive, link density, common hidden-text techniques, and unusually dominant repeated terms. If enough visible text is available, the same AI-text classifier checks it in overlapping sections.

These page-level checks are aligned with observable patterns described in Google Search spam policies, but they do not reproduce Google’s private systems or reveal a ranking penalty. One page cannot establish scaled content abuse, doorway networks, cloaking, site reputation abuse, link schemes, or a manual action. Site owners should use Google Search Console for confirmed manual actions and security issues.

What “no signal” means

It means the supported checks did not find evidence. It does not mean that a file is human-made or a website complies with every search policy. Credentials can be stripped, providers may use unsupported watermarks, and a model can miss unfamiliar content.

Models and versions

V1 uses the MIT-licensed OwensLab/commfor-model-384 image checkpoint and the Apache-2.0 fakespot-ai/roberta-base-ai-text-detection-v1 text checkpoint. Exact revisions and calibration notes are exposed through the capabilities endpoint and pinned in deployment. We do not claim generic SynthID Text detection.