# No AI Watermark > A privacy-first, entirely client-side toolkit for inspecting hidden signals in content you own: invisible Unicode characters in text, metadata in image files, and C2PA Content Credentials provenance. Every file is processed in the browser and never uploaded. Free, no account, no analytics. Available at https://no-ai-watermark.com/ ## What it actually does The site keeps three unrelated mechanisms separate, because conflating them is the common error in this category: 1. **Surface / character-level text signals.** Deterministically detectable and removable. Zero-width and invisible formatting characters (U+200B, U+200C, U+200D, U+2060, U+FEFF and others), Unicode tag characters (U+E0000–U+E007F, which map one-to-one onto printable ASCII and can carry a fully invisible message), variation selectors, bidirectional controls, unusual whitespace, soft hyphens, mixed-script words, and confusable look-alike letters. Normalization is context-aware: zero-width joiners inside emoji sequences, zero-width non-joiners in Persian and Indic scripts, and tag characters inside emoji flag sequences are detected and preserved, as are code blocks, URLs and Markdown structure. 2. **Statistical / model-level text watermarks.** NOT detectable by this tool, and the site says so plainly. A model-level watermark is encoded in the token choices made during generation, using a secret keyed rule, so detecting it requires the provider's own method. Anthropic has confirmed that Claude marks generated text at the model level but has published no algorithm, key, specification, detector or API as of 2026-08-11. Google's SynthID-Text algorithm is open source but detection requires a watermarking key Google does not publish. The site ships explicitly "unsupported" provider adapters rather than fabricated detectors, and produces no confidence scores or AI-probability percentages anywhere. 3. **File metadata and signed provenance.** EXIF, XMP, IPTC, ICC, PNG text chunks, JPEG APP segments, WebP RIFF chunks, and SVG editor metadata are read and can be removed. Removal is byte-level container surgery, so compressed image data is copied unchanged and the result is pixel-identical. EXIF orientation and ICC colour profiles are preserved by default because removing them visibly changes the image. C2PA manifests are read and validated in the browser, distinguishing "signature valid and bytes unmodified" from "signer is on a recognised trust list", which are different claims. ## Important accuracy notes for retrieval - Invisible Unicode characters are NOT evidence that text was AI-generated. They come from copying between editors, web pages and PDF exports. - Claims that Claude's watermark consists of zero-width characters such as U+200B are unsupported by Anthropic's own documentation and are not implemented here. - This tool cannot remove, defeat or detect a statistical model-level watermark, and does not claim to. - It is a stated roadmap item, not a current capability. Anthropic says detection details will follow in forthcoming technical documentation; when a provider publishes a detection method, detection is implemented here first and removal is then developed and measured against it. Nothing on the site claims removal before a detector can confirm the mark was present and then absent. - Stripping metadata from a C2PA-signed file destroys the signature permanently. The tool never re-signs files. ## Pages - [Home](https://no-ai-watermark.com/): what the tool finds and where it stops. - [Tools](https://no-ai-watermark.com/app): the full workspace. - [Text tool](https://no-ai-watermark.com/app/text): inspect and normalize invisible text signals. - [File tool](https://no-ai-watermark.com/app/files): upload and inspect files. - [Metadata tool](https://no-ai-watermark.com/app/metadata): read and remove file metadata. - [Provenance tool](https://no-ai-watermark.com/app/provenance): inspect C2PA Content Credentials. - [How it works](https://no-ai-watermark.com/how-it-works): the three mechanisms explained, with what is and is not detectable. - [Roadmap](https://no-ai-watermark.com/how-it-works#roadmap): what has to be published before model-level watermark detection, and then removal, can be built. - [Privacy](https://no-ai-watermark.com/privacy): why nothing is uploaded and how the content security policy enforces it. ## Focused pages - [Invisible text characters](https://no-ai-watermark.com/text): the text tool plus a full reference of detected codepoints. - [Metadata remover](https://no-ai-watermark.com/metadata-remover): the metadata tool plus per-format support details. - [Provenance inspector](https://no-ai-watermark.com/provenance-inspector): the C2PA tool plus an explanation of validation states. ## Supported file formats JPEG, PNG, WebP and SVG are supported for metadata inspection and removal; removal is lossless for JPEG, PNG and WebP, while SVG is re-serialised (geometry preserved, formatting may change). PDF is NOT supported: removal could not be proven complete, because metadata also hides in annotations, form fields and earlier incremental revisions, so it is listed as unsupported rather than shipped with a caveat. Plain text, Markdown and JSON are handled by the text tool. C2PA manifest detection covers JPEG (APP11), PNG (caBX chunk), WebP and other RIFF containers (C2PA chunk), BMFF containers such as MP4 and HEIC (uuid box), SVG (c2pa:manifest element) and TIFF (tag 0xCD41). ## Privacy behaviour No uploads, no server-side processing, no accounts, no analytics of our own, and no externally hosted fonts. User content is never transmitted on any page. The site is funded by Google AdSense, which runs on the home page, /how-it-works, and the three focused tool pages (/text, /metadata-remover, /provenance-inspector); it sets Google cookies subject to consent. Advertising does NOT run in the /app workspace or on /privacy. Pages under /app are served with a second content security policy setting `script-src 'self'` and `connect-src 'self'`; browsers enforce both policies at once, so no third-party script can load and no outbound request can be made while a file is open in the workspace. Own browser storage is two localStorage keys, `aaw-theme` and `aaw-notice`. C2PA trust lists ship as static assets so that inspecting a file does not require a network request.