Screenshots, paraphrase, and re-edits strip labels; survival falls with each hop

What it is

The original carries visible words and a watermark. Hop one: a screenshot into a chat, visible words cropped. Hop two: someone paraphrases the image into text, the watermark has no carrier. Hop three: another model redraws from the paraphrase; both channels are empty. Stripping across hops means screenshot, paraphrase, and re-edit are not accidental vandalism but the ordinary deaths of a mark; survival falls with each hop, it is not “sometimes lost once.”

That a label should persist through circulation is a product duty. This entry lists how it dies, and how fast.

Why it happens

The three acts hit different layers. A screenshot turns view into pixels: words on chrome and some metadata die first; a crop-robust watermark may still live. Paraphrase drops the carrier and keeps only semantics; watermark and visible layer die together. Re-edit (crop, filter, local redraw, regenerate through another model) sits between, each step lowering watermark SNR. Public chains include all three by default, so decay is the expectation, not a surprise.

Survival measured “still in the original app” does not generalise to hop three. Decay curves must be reported by hop and by channel.

Studying it

Fix a three-to-five-hop chain: in-app → screenshot → thread → save-as → paraphrase or regenerate. At each hop measure visible readability and watermark detection. Independent variables: order of acts, watermark strength, visible-mark placement (corner vs burned into the picture). Dependent variables: survival by hop, hop of first death, hop at which both channels are dead.

Swap order. Paraphrase-then-screenshot and screenshot-then-paraphrase die at different points. Screenshot-only studies will undercount paraphrase.

Where it stops holding

Closed sessions and enterprise containers that forbid screenshots hold hop count down by policy — until someone photographs the screen with a phone, and the chain reopens. Print-then-scan is another re-edit; if the watermark was not designed for it, do not claim cross-medium survival. Words burned into the picture resist screenshot-crop, not paraphrase. This entry does not treat how missing labels pollute corpora; it treats physical decay of channels across hops.

Applying it

• Measure survival along the hop counts target users actually walk; do not tick a box at the moment of export.
• For screenshot-heavy products, burn visible words into the picture, not a cropable margin. For paraphrase-heavy products, do not count on a watermark; force a relabel on the new text.
• On the share menu, preview which channel this hop will drop.
• Check: walk the three most common export chains; at the end test both channels. Both empty at hop three means decay was ignored as if it were free.

Related

• Same group: L3.10.1 Labels split into visible marks and embedded watermarks, which fail in different ways and fit different scenes · L3.10.3 A missed detection is not proof the content was not generated; lack of a mark is not evidence of human authorship · 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: screenshot stripping · watermark robustness · hop decay