AI Vocabulary

Watermarking of generated text: what does it actually prove?

Watermarking embeds in text produced by a model a statistical signal imperceptible on reading, meant to reveal that a system was involved. It attests to processing, not authorship: a fully generated text and a merely rephrased one carry the same signal, and nothing distinguishes them.

How it works

As the model chooses its words, a constraint very slightly steers those choices according to a pattern known to its publisher. The text stays natural, its quality does not suffer, and a statistical analysis can later recognise the pattern.

It is therefore not a signature placed in a file, which could be stripped by changing format. It is a property of the text itself, surviving copy-and-paste and part of any subsequent editing.

What it attests, and what it does not

It attests that a watermarked model produced that passage. That is all, and it is little against what interests a lawyer.

It does not attest that the reasoning came from the model. A text drafted by a colleague and then submitted for improvement of its wording carries the mark just as a fully generated one does.

It does not say what proportion of the text is concerned, nor at what stage of drafting the intervention occurred.

And it does not cover every model: it applies to systems placed on the market from a given date, coverage of earlier versions being progressive.

The ordinary case, which is the troubling one

A colleague drafts submissions. He submits one paragraph to check its clarity, takes the suggested version, reworks it, integrates it. The final text is his, the analysis is his, the strategy is that of the partner who signs.

That text now carries a mark. If detected, it attests that a model processed the passage. It does not say the reasoning came from it, and nothing in the signal allows that to be established.

The difficulty is therefore not strictly legal, it is evidential: the firm will have to explain what happened, and can only do so if it kept a record.

The difficulty particular to a firm

It comes from a mismatch of timing. Marks are being applied now, in texts that circulate, are filed and archived. The tools for detecting them are not yet widely available.

A firm therefore has no way today of knowing what its own past output carries. And it may have to explain a situation it cannot itself observe.

What it does not solve

Watermarking says nothing about the quality of a text, nor about the accuracy of its citations. A marked text may be excellent and an unmarked one may contain serious errors.

It is not evidence in the procedural sense either: it is a probabilistic signal, it can disappear on rewriting, and its interpretation depends on tools whose specifications are unpublished.

And it creates no new obligation for you. Nothing prevents a lawyer from using a tool; what is at stake is the ability to describe what was done, an old requirement rather than a new one.

Why it matters to a lawyer

Because the burden of explanation will fall to you. If a mark is detected, it will not say what happened: you will, and only if you kept a record of your production process.

This development therefore creates no obligation: it makes suddenly useful a traceability that few firms had organised.