Responsibility, Law & Governance
Watermarking
Invisible patterns in AI content are meant to make it machine-detectable what is synthetic. The state of the art is sobering: image watermarks rarely survive editing, text watermarks are fragile, and anyone who wants to deceive uses tools without marking. Sensible as part of a provenance infrastructure, a sedative as the sole protection.
An example
There are two approaches with different prospects. One hides a signal in the content itself — robust against light editing for images, largely failed for text, because rephrasing removes it. The other attaches a signed provenance record to the file: who created or edited it, when, and with what. The second is technically sound and currently fails because most platforms strip the information on upload.
A closer look
Generated content is given a statistical pattern that is invisible to humans but demonstrable to a verifier. In addition there are provenance records in the metadata, for example under the C2PA standard.
Why it matters
For companies the interesting direction is not detecting other people’s forgeries but proving one’s own authenticity. Whoever ships product photos, press material or expert reports with signed provenance makes forgeries in their own name verifiable — and that becomes more valuable to the degree that forging gets cheaper. The effort is manageable and today mainly a question of the tool chain. As protection against misuse of others’ content, however, the technique is so far not suitable.
Commonly misunderstood
Watermarks survive editing, compression and screenshots only to a limited extent, and open models can be run without them. A missing watermark therefore proves nothing at all.