# Is this AI-generated? URL: https://nader.io/projects/cest-claude/ Find clues in text and images. Inspect hidden text, image metadata and camera-processing details. Free, private, and explained. [Check text]() [Check an image]() Free to use. No account. Text and image checks stay on your device. A finding, explainedExample ▧ image.png Metadata inspection Software ComfyUI AI-related metadata found A generator reference is a clue to investigate. It is not a probability score or proof of authenticity. ◎ Evidence you can inspect yourself Read the evidence. Metadata can be removed or changed. Writing style is only a clue. A clean result does not prove authenticity. Take a closer look ## What would you like to check? Local text & image analysis Text to analyse 0 characters Page to fetch Public pages only. Anything behind a login won't load. Image to inspect **Look for the traces an image carries.** Choose a file or drop it here. JPEG, PNG or WebP, up to 20 MB. Your image stays on this device. Checks metadata for generator references and generation parameters. Visual checks start automatically using two experimental models on this device. First use downloads about 290 MB from Hugging Face; your image is not uploaded. You can cancel the visual checks. Downloaded reports include inspected metadata, which may contain prompts or other private details. 0:00 1. Metadata 2. Credentials 3. Visual model 1 4. Visual model 2 ### Evidence found ### Content Credentials Credential validation details ### Visual AI checks Experimental Pixel analysis starts automatically when you select an image, even when metadata is missing. First use downloads about 290 MB of model weights from Hugging Face, plus the local runtime. Your image stays on this device. Checks may take a few minutes. Visual checks start automatically after an image is selected. These models do not verify who or what a picture depicts. Face swaps and other local edits can be missed. Model scores are not probabilities that an image is fake. Model outputs and limitations Outputs are model scores, not the probability this image is AI-generated. Agreement is not proof. Model agreement is not independent corroboration. New generators and edited or compressed images may fool both. Models: [SwinV2 image detector]() and [Deep-Fake Detector v2](). Heuristic thresholds: AI score ≥ 0.8 is a higher signal; ≤ 0.2 is lower. These are not independently calibrated accuracy claims. ### Watermarks and source tracing Invisible watermark scan: **not performed**. Open an official verifier and choose the file there if you want to share it with that provider. A missing watermark does not rule out AI. [Official SynthID tools]() · [Content Credentials verifier]() · [Google Lens reverse image search]() · [TinEye]() Inspect metadata (may contain private details) ### Before you trust or share it Ask for the original file and its source. Check earlier publications with a reverse image search, and corroborate the scene with independent sources. Camera metadata does not prove a scene is real. AI editing can affect just part of a photograph. 0 findings 0 agent tags 0 invisible characters 0 decoded payloads 0 style score ## Highlighted text Invisible characters have no shape of their own, so each one is drawn as a labelled pill. Everything else is your text, untouched. **Copy cleaned text** strips the markup and the invisible characters in the enabled categories. The visible words stay exactly as they are, including anything a stripped tag was wrapped around. ## Hidden payloads, decoded Runs of invisible characters that carry a readable message. Tag-block characters mirror plain ASCII one-for-one; variation selectors encode raw bytes. Both survive copy/paste, and neither shows up on screen. ## Style read Habits that turn up far more often in model prose than in human writing. This is circumstantial, not proof: a fluent human writer can trip several of these, and a model told to write plainly can trip none. Every signal is listed with its own hits so you can disagree with any of them. ## What turned up | What | Category | Count | First seen | | --- | --- | --- | --- | A practical starting point ## How to check if content is AI-generated You deserve more than a mysterious percentage. Each finding comes with the underlying clue and its limits. 1. 01 ### Start with the original Paste the text, choose an image, or fetch a public page. Original files carry more useful evidence than screenshots or recompressed social posts. 2. 02 ### Inspect what turned up Read highlighted text, decoded hidden messages and image metadata. Expand HDR maker notes only when you want the camera-processing details. 3. 03 ### Check the source, too Ask where the content came from. Use a reverse image search and independent sources. Treat the result as part of your investigation. What this tool actually checks ## Two kinds of content. Different kinds of clues. Aa ### AI text clues Inspect text for leaked agent tags such as ``, invisible Unicode characters, hidden payloads and model-like writing habits. Review each match and copy cleaned text. Style scores describe the signals found, not the probability that AI wrote it. [Check text for AI clues]() ▧ ### AI image metadata Look for recognised generator names, generation parameters and workflow data in JPEG, PNG and WebP files. Decode supported Google HDR maker notes. Embedded Content Credentials receive local signature and integrity checks. Visual models automatically inspect pixels for AI-like signals, with disagreement and uncertainty shown. Signer trust and invisible watermarks need separate verification. [Inspect an image’s metadata]() Before you draw a conclusion ## AI content and authenticity, explained. Can this tell if an image is AI-generated? It can find recognised AI generator references, generation parameters and workflow data in JPEG, PNG and WebP metadata. Two experimental visual models also run automatically and locally to look for pixel-based AI signals. Their outputs are not calibrated probabilities, and they may disagree or be wrong. If the metadata is missing or unrecognised, the result is inconclusive. An AI-generated image can have no metadata, and metadata can be forged. Can I check whether text was written by ChatGPT or Claude? Yes. Paste the text to inspect agent markup, invisible Unicode characters, hidden messages and writing-style clues associated with language models. The tool cannot reliably identify the author or model. Quoted tags, ordinary typography and human writing can trigger findings too. Does a clean result mean the content is authentic? No. A clean result means the enabled checks did not find recognised clues. It does not prove human authorship, camera capture or an unedited scene. Check the original source and corroborate important claims independently. Are my images or pasted text uploaded? Selected images and pasted text are analysed locally in your browser. They are not uploaded by the checker. Reports are downloaded to your device and can contain private metadata. Fetching a web page is different: its URL is sent to public proxy services so the page can be retrieved. What does an HDRPlus maker note reveal? Supported Google HDRP and HDRPlus maker notes can contain image-record labels, device and camera-app details, frame counts, exposure settings and processing notes. The checker decodes supported values automatically, with the details collapsed until you expand them. A partial or damaged payload is labelled clearly. HDR processing is not evidence of generative AI. Does the checker verify Content Credentials? It reads embedded Content Credentials and checks their signatures and content integrity locally. Signer trust, revocation and remote manifests are not checked. For full verification, open the original file in an official verifier. Invisible watermarks such as SynthID require a separate provider check and are not scanned here. Can I check an article or website URL? Choose Web page and enter a public URL. The checker retrieves the page through public proxy services and inspects its extracted text. Login-protected pages may not load. Extraction can change or remove hidden characters, so pasting the original text is preferable when those details matter. ## A small tool for a more careful internet. C’est Claude!? is built by [Nader Cserny](). Explore more independent tools and projects at [nader.io](). Technical references: [Google HDR maker-note decoding]() and [Content Credentials]().