Developers using Claude’s API face questions about whether Anthropic embeds steganographic request markers in outputs, what data those markers encode, and implications for privacy, security, and compliance. Steganographic marking in LLM outputs is not theoretical; technical mechanisms are understood, provider incentives are clear, and community investigations have begun. Steganography means embedding hidden information within text that appears normal to human readers, unlike visible watermarking or statistical watermarking, which biases token distributions. Statistical watermarking, explored by Scott Aaronson during his time at OpenAI, partitions token vocabulary into “green” and “red” lists at each generation step, nudging the model to prefer green tokens, creating a statistical signature detectable by those knowing the scheme. Steganographic encoding would require a separate system selecting among equiprobable tokens based on a payload, a different architectural component not described in Aaronson’s published work. This exploits language models’ frequent choices between tokens of nearly identical probability, where synonyms, phrasings, or punctuation patterns are equally valid, allowing specific selections to encode bits of a hidden payload. Such signals are harder to detect and strip than statistical watermarks because variation resembles natural language randomness. Output watermarking answers whether text was AI-generated, while request-level marking answers which API key, account, or session generated it.