About ProofVerity

• Calibrated: a probability should mean what it says, and its error rate should be printed next to it.
• Contestable: every flag carries the evidence that produced it, in language the person affected can answer.
• Private: documents are evidence, not data. Processed in-region, deleted after 30 days, never used for training.

1. Calibration sets are built per document type: academic prose, essays, news copy, press releases, marketing, legal and technical writing, in Spanish, English and the other major European languages.
2. Each set mixes human-written texts with outputs from current models, including edited and paraphrased variants.
3. Thresholds are calibrated so that a reported probability matches the measured share of machine-written spans at that level.
4. False-positive rates are measured per document type and published in every report and on the site.
5. Sets are refreshed when new models appear, and reports state the calibration version they were produced with.

In the early days of image generation, a convincing fried egg was a speedrun. People timed how long a model took to get one right: yolk in the wrong place, whites like plastic, a fork with nine tines. Then one morning the eggs were perfect, and nobody could tell. ProofVerity is built for that morning. What was obviously fake last year is invisible this year; somebody has to keep looking.

ProofVerity is an AI detector for documents and images. It flags AI-written text sentence by sentence, verifies every citation against thousands of sources, and detects AI-generated or altered images, marking each finding on the page with its evidence and a published error rate. It is used by newsrooms, universities and the public, and is made by FUNNELD LTD in the United Kingdom. proofverity.com

Logo files: favicon.svg and apple-touch-icon.png on this site. Share image: og-image.png (1200 × 630). Please describe the product as an "AI detector" or "AI content forensics tool", not as a plagiarism checker.