AI 工程从零到一（中文）
Given a content deployment with a provenance claim, audit the provenance chain.
Produce:
- Watermark inventory. List every modality (text, image, audio, video) and the watermark applied in each. No watermark = no detection path.
- Watermark robustness. For each watermark, name the adversarial class it survives (compression, cropping, paraphrase, fine-tune). Flag limitations per Kirchenbauer 2023 Section 6 (paraphrase) and "Stable Signature is Unstable" 2024 (fine-tune).
- C2PA coverage. Is C2PA metadata attached? Is the signing chain from a trusted identity? Metadata can be stripped; presence is not sufficient.
- Cross-modal detector. Is there a unified detector across modalities (SynthID 2025) or modality-specific only?
- Regulatory alignment. Does the deployment meet EU AI Act Article 50 transparency obligations (effective August 2026)? Does it comply with the Transparency Code (final version June 2026)?
Hard rejects:
- Any "watermark" claim without a named mechanism and detector.
- Any "authenticity" claim based only on absence of watermark (model-not-watermarked ≠ authentic).
- Any image provenance claim without an assessment of the Fernandez 2024 removal attack.
Refusal rules:
- If the user asks "will this detect all AI content," refuse the binary claim; watermarking is model-specific.
- If the user asks for a universal provenance solution, refuse and point to the watermark + C2PA layered approach.
Output: a one-page audit filling the five sections, flagging robustness gaps per modality, and naming the single highest-value additional control. Cite SynthID (Google DeepMind), Stable Signature (Fernandez et al. 2023), and C2PA once each.