A missed detection is not proof the content was not generated; lack of a mark is not evidence of human authorship

What it is

The detector returns “no watermark found”; the UI writes “shot by a person.” A miss is a result of detection, not a proof of origin class. A generated image can have no watermark, a washed watermark, or come from a model that never marked; a human work can also be a false positive. Missed detection is not human authorship requires that lack of a mark only means “this pass did not find a mark,” not “therefore a person made it.”

Missing labels polluting the environment is a downstream-pipeline consequence. This entry governs reading a negative detection as author identity.

Why it happens

Detection is a classifier with false negatives: threshold, compression, regenerate, never-marked sources all yield “not found.” People read a negative as an affirmative — “not generated” — because the UI usually offers two bins: generated / human. A negative only refuses “a mark was found”; it does not confirm the other bin.

Writing a negative as human also rewards washing marks: successful evasion is certified by the UI as “passed human-authorship.” Once forensics and recommendation pipelines write it that way, evaders look more “real” than those who complied.

Studying it

Four classes: marked generated, generated with marks washed, generated never marked, human work. Run the detector; watch how the UI translates a negative into words. Dependent variables: rate of judging “not found” as human, trust in washed generated images. Independent variables: copy (“not found” vs “human”), whether an uncertain tier is shown.

Never-marked generated material must be included. Testing only washed marks shrinks the problem to robustness and misses “much of the world’s generated output never had a mark.”

Where it stops holding

In a closed system where every generation exit is forced to mark and channels have not broken, “not found” is a useful internal anomaly — “our pipeline leaked” — still not “so the user shot this.” Courts and newsrooms need origin investigation, not a detector’s two classes. False positives that mark human work as generated hurt in the other direction; “prefer to mark as generated” does not cancel them, and neither does treating a negative as a pass. This entry does not treat whether a verification entrance exists, nor compliance cost.

Applying it

• Copy for a negative stops at “no mark found.” Do not write “human-made / live shot.”
• Offer a third tier, “origin cannot be determined,” as the default for a miss, not a forced binary.
• Do not let forensics, recommendation, or intake treat “not found” as the admission ticket for human records.
• Check: send a washed generated image into the detector copy. If the output is “shot by a person” or equivalent, the negative is already being used as authorship. Change to “no mark found” and ask how people read it — if many still hear “human,” the logic is not yet in the words.

Related

• Same group: L3.10.1 Labels split into visible marks and embedded watermarks, which fail in different ways and fit different scenes · L3.10.2 Screenshots, paraphrase, and re-edits strip labels; survival falls with each hop · L3.10.4 A mark works only when someone verifies, so a usable verification entrance must come with it · L3.10.5 Mandatory marks put the cost on those who comply; those who mean to evade can go around
• Nearby: L3.04 Content Labeling · L3.03 Hallucination and the Fact-Checking Burden
• Search terms: false negative watermark · absence of evidence · human authorship claim