AI Watermark · Content Provenance
AI Watermark / C2PA
What is an AI watermark?
- Embedding markers and provenance history in AI-generated content to tell real from fake
- Watermark = AI label; C2PA = creation and edit history
- Linked to deepfake defense and AI-labeling rules
What this technology actually changes is the order in which authenticity gets disputed. Until now, you ran a detector on a suspicious image and weighed the probabilities. Watermarks and provenance history flip that approach to attaching the evidence at the moment of creation. The center of the debate shifts from 'is this fake?' to 'can you show where this file came from?' For practitioners, one setting at the production stage reduces the cost of explaining things after the fact.
There are three main approaches: visible labels shown on screen; invisible watermarks hidden in the pixel or frequency domain that withstand cropping and recompression to a degree; and C2PA, which attaches creation and editing history as signed metadata. The first two address 'is this pixel AI-made?' while C2PA addresses 'who touched it, with which tool, and when?' so they complement each other rather than replace each other. C2PA only takes effect when signed with an issued authentication certificate, and if the signature breaks it simply shows 'no history'; it does not tell you the content is fake.
Markers on text are a different story. Rewrite a sentence a few times and the marker fades, so text does not hold up like images or video. That is why many companies handle text through an operating rule, an 'AI use disclosure statement,' rather than technology. Regulation points the same way. Both the EU AI Act's transparency provisions and Korea's AI Basic Act require an outcome, that users can tell content is AI-generated, rather than specifying a particular watermark technology.
The point that breaks most often in practice is distribution. Most social networks and messengers strip metadata and recompress images on upload, so provenance history attached in-house is often already gone by the time customers see the content. It is safer to assume a single screenshot leaves nothing behind. Treating detection results as conclusive is just as risky. If you reject a partner's deliverable based only on a verdict like '90% likelihood AI-generated' and it turns out to be a real photo, both the schedule and the relationship suffer. The habit of keeping signed originals in-house is ultimately the last line of defense.
John KimInstructor John Kim, author of "B2B Marketing Design" and CEO of Nanum Consulting
Seungtae ParkAX instructor who teaches by building AI business development and marketing automation himselfIt began with the need to counter deepfakes and disinformation, and spread through the C2PA (Coalition for Content Provenance and Authenticity) standard, backed by Adobe, Microsoft, and others, and through AI-labeling regulations in various countries.
Consider a hypothetical case. Company A, a Korean color cosmetics brand (180 employees, annual revenue of about KRW 90 billion (≈ US$67,335,000)), creates 200 ad assets a month for Instagram and Naver, and about 60% of them are produced with generative AI. Within six months it received more than 40 customer inquiries asking 'is this model a real person?', and platform ad review repeatedly rejected the same assets.
- Marketing Team Lead Park traced back 210 assets from the last three months, but could tell AI-made images from real shots for fewer than half of them. In a meeting, someone said, 'We have nothing to go on but the file names.'
- The design team standardized on a version of its generation tool that attaches C2PA Content Credentials, obtained a signing certificate, and installed it on only 3 work PCs. It also wrote the same conditions into contracts for external freelancers’ deliverables.
- When Assistant Manager Jung ran an upload test, all metadata was stripped the moment the image was posted to social media. The team then created a dual-labeling rule: every post also carries an 'AI-generated image' line at the top.
- Kim, a senior manager in Legal, split labeling standards into three tiers (fully AI-generated, AI-retouched, and live-shot) and banned outright any material that composites a real model's face.
- The signed original files are kept in internal storage for 180 days, separate from the delivered versions, and every quarter we randomly pick 20 images to confirm the provenance history is still intact.
- Six months later, ad review rejections had fallen from 12 to 2 per quarter, and answering 'is this a real model?' inquiries dropped from an average of 3 days to the same day. Team Lead Park says, 'One provenance lookup link settles it now.'