Accused of using AI?
AI detectors are stylistic classifiers with documented false positive rates, and they misfire particularly often on non-native English writers and on deliberately plain prose. A detector result is not evidence of authorship.
Watermarking is a separate matter: it can indicate that text came from a specific model, but only the provider holding the key can check, and no institution currently has that access for text.
The evidence that genuinely supports your case is process, not style: version history, drafts with timestamps, research notes and the ability to discuss your own argument in detail.
Frequently asked questions
Can a school prove I used Claude?
Not through watermarking, which requires Anthropic's private key that institutions do not have. AI detectors produce probability estimates with known false positive rates, not proof.
I only used AI for grammar. Am I at risk?
Light editing of your own text leaves it substantially yours, and a watermark would only apply to text the model generated. The practical risk comes from stylistic detectors, not watermarks. Keep your drafts and version history.