Claude output is now statistically watermarked via Google's SynthID-Text, embedded in token-level word choices and detectable through a forthcoming API—surviving light editing but not full rewrites, and largely absent from code except comments. If you're shipping products that pass Claude text off as human-authored or feed it into pipelines where provenance matters, assume it's now flagged detectable; the practical escape hatches are heavy paraphrasing or the fact that code generation carries near-zero watermark signal. This is EU AI Act compliance, so expect the same from every major provider, not just Anthropic.