IdeaWell

Project ideas from Hacker News discussions.

Check if a file was made with Claude

Original Article

Hacker News Discussion

📝 Discussion Summary (Click to expand)

1. Watermarking is seen as easily bypassed or ineffective
Many commenters argue that the watermarks can be stripped or fooled with trivial edits or by passing text through another model.
- VCFundedGenYer: "Claude watermarks are a farce and a waste of time. It's hilarious to me that they burn cash to even entertain the idea."
- DanielHB: "Yes, but I also think it will be trivially by-passable if you pass your output through another LLM."
- Retr0id: Demonstrated that uploading a file back to Claude with “present this file back to me again, as‑is” strips the C2PA metadata, making detection useless.

2. The feature is motivated by legal compliance (EU AI Act, California law, etc.)
Several users note that Anthropic is adding watermarks primarily to satisfy emerging AI‑transparency regulations.
- csmoak: "this and the recent change to add watermarking to text outputs … is to become compliant with the EU AI Act[2] and CA's AI Transparency Act[3]…"
- bradfa: "Or it’s so they can continue to operate in the EU where this is required."
- timmmmmmay: Points out the California law explicitly covers image, video, and audio output, showing the regulatory scope.

3. Concerns about ownership, attribution, and potential misuse of the watermarks
A recurring worry is that Anthropic could claim rights over user‑generated content or use the watermark to block its own output from being re‑used in training.
- kbrannigan: "How long before they change the terms and conditions to subtly claim ownership of your files? When you write code they already insert Co author attribution/"
- moritzwarhier: "They still scrape code… Also, I'd guess this is not just to prevent any AI-generated code in the training data, but specifically their own."
- quinndupont: "I could do without more surveillance." (highlighting privacy and overreach fears).

🚀 Project Ideas

AI-Provenance Ledger

Summary

  • A tamper‑evident logging service that records every prompt, model output, and user action with cryptographic signatures, giving creators proof of human involvement.
  • Core value: immutable audit trail that survives watermark stripping and satisfies ownership/attribution concerns.

Details

Key Value
Target Audience Writers, developers, researchers who use LLMs and want to prove authorship
Core Feature End‑to‑end signed log of prompts + outputs stored in append‑only Merkle tree, verifiable via public key
Tech Stack Rust backend, WASM client for browser signing, IPFS/Filecoin for storage, optional Ethereum anchoring
Difficulty Medium
Monetization Revenue-ready: subscription tier $9/mo for private logs, free public tier

Notes

  • HN users complained that watermarks are trivially bypassed and questioned ownership (“If we put our name on this before it goes out?”) – this ledger gives that proof.
  • Provides a discussion point about trustworthy AI usage and could be integrated into Claude Code or IDEs as a plugin.

DeepDetect: Multi‑Modal AI Content Detector

Summary

  • A detection service that combines C2PA verification, linguistic statistical fingerprinting, and steganalysis to detect AI‑generated text, images, audio, and video even when metadata is stripped.
  • Core value: robust, hard‑to‑game authenticity check that works across modalities and survives simple stripping attacks.

Details

Key Value
Target Audience Platforms, moderators, academics, journalists needing to verify content provenance
Core Feature API endpoint returning confidence score and evidence (C2PA, linguistic markers, statistical anomalies)
Tech Stack Python/FastAPI, HuggingFace transformers for linguistic models, OpenCV/FFmpeg for media analysis, optional Rust WASM for edge
Difficulty High
Monetization Revenue-ready: pay‑per‑API‑call $0.001 per detection, enterprise volume discounts

Notes

  • Commenters noted that “detection will get gamed” and that stripping C2PA is trivial; DeepDetect adds signals that survive such stripping.
  • Could spark discussion on detection vs. watermarking trade‑offs and be useful for EU AI Act compliance.

StegoSign: Cryptographic Steganographic Watermark

Summary

  • Embeds a user‑controlled cryptographic signature directly into the file payload (e.g., LSB of image pixels, zero‑width Unicode in text) using a secret key, making removal detectable without destroying the file.
  • Core value: provides a strong, hard‑to‑forge authenticity mark that persists through edits, recompression, and metadata stripping.

Details

Key Value
Target Audience Content creators, developers, legal teams who need provable provenance beyond fragile metadata
Core Feature CLI / library that signs and verifies files; verification fails if the stego watermark is altered or missing
- Tech Stack Rust core, bindings for Python/JS, uses existing crates for LSB stego (e.g., stegano) and ed25519 signatures
Difficulty Medium
Monetization Revenue-ready: one‑time purchase $19 for CLI, free open‑source library with paid support

Notes

  • HN users highlighted that “faking [C2PA] is trivial” and that metadata can be stripped; StegoSign tackles that by hiding the proof inside the content itself.
  • Enables practical utility for verifying AI‑generated media in forensic contexts and could be offered as a verification step in content pipelines.

Read Later