AI-detection tools have a well-known problem: they generate too many false positives. A human-written story can easily be flagged as machine-generated, and a lightly edited AI draft can slip through. For sites that care about provenance—whether a piece was written by a person or produced by a model—the current methods are simply not reliable enough.
During a thread on the Stories Online forums I suggested a different approach. Instead of trying to detect AI text after the fact by statistical quirks, why not have the generating systems deliberately embed a lightweight, invisible watermark in the prose itself? If the major AI writing tools agreed to apply such a mark, sites like SOL could check for it and know with reasonable confidence that a story had come from one of those tools.
The reaction was skeptical. Fair enough. So I went away and built a small prototype in Python to see whether the idea was even feasible.
The system works by making a series of tiny, meaning-preserving choices while the text is written or lightly rewritten. It currently uses three independent channels:
Synonym selection — choosing among close alternatives for certain content words.
Function-word preferences — small closed-class choices such as that versus which.
Adverb and modifier placement — preferring one natural position over another for a limited set of adverbs.
None of these changes is meant to be noticeable to a reader. Taken together, however, they create a recoverable statistical pattern.
The story I just posted, “Steganography Sample,” was processed with this prototype. It contains 29 encoded points. When the detector is run on the finished text it recovers the watermark with high confidence.
Only after building the prototype did I discover that the general technique already has a name and a substantial research literature: linguistic steganography (and the closely related field of natural-language watermarking). Researchers have been exploring synonym substitution, syntactic variation, and similar carriers for years. My implementation is a modest, practical demonstration rather than a novel algorithm. Still, it was useful to work through the problem by hand and see that a lightweight, multi-channel watermark can be embedded in literary prose without obvious damage to voice or meaning.
Whether AI companies would ever adopt something like this is a separate question. The experiment does suggest that, if they wanted to, a cooperative watermarking scheme for attribution is at least technically plausible.