Skip to content

Claude text and image checks

Claude AI Detector for Text and Images

The Text tab reads the writing itself, so it works on English Claude prose with or without a watermark: 99.5% of AI text caught at a 5% false-positive rate on RAID, a public benchmark we do not run, with 11 adversarial attacks in the test set. Slop or Not does not read Claude's SynthID-Text watermark, because Anthropic has not released the detector for it. The Photo tab reads C2PA Content Credentials, reports the source it recognizes, and scores the pixels when no readable provenance resolves the file. Online checks run on Numen's private Mac server, away from outside AI companies, and are deleted after processing; iPhone and Mac app checks run on-device.

Claude Watermark: How It Works and What It Doesn't Prove

Catches 99.5% of AI text at a 5% false-positive rate, scored by RAID, the public benchmark, not by us.

Paste a paragraph or a whole thesis. The iPhone and Mac apps have no word cap and no monthly word budget.

  • Paste English textSchool essays, product reviews, captions, articles, and social posts up to 20,000 characters.
  • ResultA probability verdict, AI score, and readability analysis.
  • PrivacyDeleted the moment the check finishes. We never keep your text.
English text only. Deleted after processing.0/20000

Try an example

AI detection is probabilistic. Results can vary with new AI models and evasion methods. Human writing that sounds formal can be flagged as AI.

Scored by a fine-tuned transformer model, not a word-pattern checker.

What it checks

Read the Writing Itself, No Watermark Key Required

A text check that reads the writing itself, so a missing or stripped watermark changes nothing: 99.5% of AI text caught at a 5% false-positive rate on RAID, adversarial attacks included

C2PA Content Credential reads on Claude images, reporting the source when Slop or Not recognizes one

Pixel-level visual detection when no readable provenance or watermark resolves the file

Good Uses

  • Checking an essay or assignment that may have been written with Claude
  • Checking whether a Claude image still carries C2PA Content Credentials
  • Checking a Claude image by its pixels when no readable provenance resolves the file

Limits to Know

  • Slop or Not does not read Claude's SynthID-Text watermark, so its text verdict is a classifier probability rather than a watermark read
  • Text detection is English only, and short, factual, code-heavy, proofread, or rewritten passages leave less classifier evidence
  • C2PA credentials can be stripped. Slop or Not reads them and reports the source it recognizes; the reader must decide whether that history points to Claude. Even then, it supports involvement; it does not prove authorship, ownership, content truth, every-pixel AI origin, or an unchanged history

FAQ

Claude AI Detector Questions

Does Claude currently watermark all text?

Not necessarily. Anthropic says future Claude models will generate watermarked text and that older models will be updated over the coming months. It has not published a model-by-model rollout table, so a response date alone does not prove the mark is present.

Can Slop or Not read Claude's text watermark?

No, and it does not need to. Anthropic says Claude's mark is a version of SynthID-Text: a secret statistical pattern created through ordinary token choices, readable only with Anthropic's own detector. A watermark read also stops working the moment the mark is absent, stripped, or paraphrased away. The Text tab reads the writing itself instead, which is why it caught 99.5% of AI text at a 5% false-positive rate on RAID with 11 adversarial attacks in the test set. The verdict is a probability, not proof.

Why are short, factual, or code-heavy answers harder to mark?

The mark has less room to work when Claude has few reasonable word choices. Anthropic says short samples, factual passages, code, and light proofreading can carry too little signal for a confident read. The same caution applies to classifier results on short or heavily edited passages.

Can the text watermark identify who used Claude?

Anthropic says the text watermark carries no identifying information and cannot be traced to a person, organization, or chat. It tests for likely Claude involvement, not who used Claude.

How does Slop or Not check Claude images?

Claude uses C2PA Content Credentials for supported image files, not image SynthID. A credential is signed provenance metadata attached to the file, not a watermark embedded in the pixels. Slop or Not reads the credential, reports the source it recognizes, and checks pixels when the file carries no readable provenance. The reader decides whether that history points to Claude; a missing credential proves nothing.

What does a Claude detector result prove?

A credential can support Claude involvement when the history it records points to Claude. Slop or Not reads the credential and reports the source it recognizes; the reader interprets it, and a dedicated C2PA viewer can show the full signed chain. C2PA does not prove who authored or owns the file, whether its contents are true, whether every pixel is AI-generated, or that the history is unchanged. A text-classifier result is a probability. A missing watermark or credential proves nothing.

Do Claude watermark cleaners work?

Be skeptical of any tool that reports a clean Claude watermark result today. WatermarkDetector.com scans hidden Unicode and spacing, but Claude's text mark contains no hidden characters. watermarks-remover performs real Unicode and metadata cleanup and can rewrite text, but without Anthropic's forthcoming detector it cannot prove the token-choice pattern is gone. Stripping C2PA from an image also leaves its pixels available for analysis.

Run the Next Check on Your iPhone or Mac

The Slop or Not app checks images and text on-device, so nothing you check ever leaves your device.