Check if a file was made with Claude

Original Article ↗ Hacker News Discussion ↗

Anthropic’s new “Check if a file was made with Claude” tool, which verifies C2PA provenance metadata on media files, has prompted debate over how meaningful such watermarks really are when they can be easily stripped or spoofed yet are hard to fake cryptographically. Commenters see the move largely as regulatory compliance (e.g., EU AI Act, California rules) and a way to keep AI-generated content out of future training data, but raise concerns about surveillance, unclear copyright implications, limited file support, and the risk that watermarking degrades quality or is quickly bypassed by simple post‑processing or other models.

What the tool is and how it works

  • The checker targets media (images, audio, video) and inspects C2PA provenance metadata client‑side via WebAssembly.
  • It does not yet support text or common document formats like PDF; text watermark detection is a separate API in private preview.
  • Users observe that Claude-generated images downloaded directly include C2PA data, but that metadata disappears if files are recompressed, zipped, or otherwise reprocessed.

C2PA vs text watermarking

  • C2PA is described as cryptographic metadata for authenticity/provenance, easy to strip but hard to forge without signing keys.
  • Text watermarking is described as biasing token choices via a secret key instead of pure randomness; no invisible Unicode tricks.
  • Some note that watermarking for code is harder, possibly limited to comments.

Bypassability and technical limits

  • Many say C2PA can be removed with simple tools or re-saving; faking it is harder but can be “outsourced” by asking Claude to re-emit arbitrary files.
  • Text watermarks might be weakened by paraphrasing, translation, or passing through other LLMs, though others argue that fully removing them without distortion is non‑trivial.

Regulatory and legal angles

  • Several comments tie the feature to compliance with the EU AI Act and (for media) California’s AI transparency law.
  • Debate over copyright analogies: LLM as “monkey” vs “camera,” with broad agreement that vendor ownership of outputs would lack legal basis, though broad licenses are possible.

Trust, surveillance, and ownership concerns

  • Some see this as necessary transparency against AI-generated fraud; others as increased surveillance with limited deterrent value.
  • Worries include privacy (storing hashes or outputs), lack of offline validators, and potential for false positives making detection legally risky.

Impact on quality and use

  • Several assert that text watermarking must degrade writing quality; others suggest the impact can be minimal but acknowledge trade‑offs.
  • There is tension between automation and human responsibility: a watermark only shows “model involvement,” not whether a human stands behind the content.

Training data and self-contamination

  • Many believe a key motive is to avoid re-training on Claude’s own outputs and to satisfy rules about labeling AI content, especially in the EU.

Security and authenticity debates

  • Some argue C2PA can meaningfully prove origin from a given device/model; others point to past camera‑auth systems being broken and note that compromised keys or malicious signers undermine guarantees.

Overall sentiment

  • Reactions are mixed: some view watermarking/provenance as a pragmatic, imperfect “good enough” step; others consider it easy to bypass, legally fraught, or mostly a box‑ticking exercise.