Is “AI-Generated” Enough? Why Creation History Matters in E-Commerce
Introduction
Hello, I’m Mia Sato, AI Researcher at GDX.
Generative AI is already becoming part of everyday EC operations. Teams use it to draft product descriptions, ad copy, social posts, customer support replies, and even product visuals.
I also use AI in my own work, especially for first drafts and translation support.
That is why Anthropic’s recent announcement about Claude’s text watermark caught my attention.
The idea is simple: Claude can leave invisible patterns in generated text, making it possible to later estimate whether Claude may have been involved in creating that content.
At first, I thought of this mainly as a way to detect AI-written text.
But the more I looked into it, the clearer it became that this is part of a much bigger shift. The question is no longer only, “Was this made by AI?” For companies, the more practical question is, “How do we manage AI-generated content inside real workflows?”
AI provenance is becoming part of the content infrastructure
Major AI companies are already working on different ways to preserve the origin and history of AI-generated content.
| Service | Main method | Example content types |
|---|---|---|
| Claude | Invisible text watermarking | Text |
| Gemini | SynthID | Text, images, audio, and video |
| OpenAI | C2PA and SynthID | Images and audio |
| Adobe Firefly | Content Credentials based on C2PA | Images and other creative assets |
| Microsoft | C2PA and invisible watermarking | Images, audio, video, and other media |
The technical approaches vary. Google uses SynthID to embed invisible watermarks into generated content. OpenAI, Adobe, and Microsoft are also working with provenance technologies such as C2PA to preserve information about how content was created and edited.
What they have in common is the same basic goal: making it possible to understand where content came from and how it may have changed over time.
In other words, content provenance is starting to become part of the infrastructure around AI-generated content.
Why AI-generated content needs a visible history
One reason is obvious. AI-generated content is becoming harder to distinguish from human-created content.
When people know AI was involved, they can use that information as context when reading, reviewing, or publishing the content.
Another reason is regulation.
Article 50 of the EU AI Act requires certain types of AI-generated or AI-manipulated content to include information in a machine-readable format. For deepfakes and some other content types, it also requires disclosure that AI was used.
These transparency obligations have applied since August 2, 2026.
Anthropic has also described EU AI Act compliance as one of the reasons for introducing Claude’s watermarking technology.
But for businesses, this raises another important question.
If we know a piece of content was made by AI, is that enough?
“AI made it” does not remove business responsibility
Imagine an EC product page that says, “This jacket is made with waterproof material.”
Later, the team discovers that the product is not actually waterproof.
Even if the description was generated by AI, that does not solve the issue.
Once a company publishes the information on its product page, the company is responsible for what customers see. It cannot simply say, “AI wrote it.”
In Japan, product claims that make quality or specifications appear significantly better than they really are may fall under misleading representation rules under the Act against Unjustifiable Premiums and Misleading Representations.
So the fact that content was AI-generated does not make the risk disappear.
From a company’s perspective, watermarks and provenance data should not be treated as a way to shift responsibility to AI. They are more useful as signals for managing risk and improving review processes.
The real question is where the information changed
Even when the same incorrect word appears in the final content, the cause may be different.
Was the original product master wrong?
Did AI add something that was not in the source material?
Did a human editor change the meaning while polishing the text?
Or was the issue missed during the final review?
Each cause points to a different place to improve.
That is why companies first need to understand where AI is used in their content workflows.
For important content such as product information, advertising claims, pricing, and campaign conditions, it helps to keep a basic record of the source material, where AI was involved, and who reviewed the final output.
For a product description workflow, that might look like this:
Product source data → AI draft → Human edit → Product owner review → Publish
With this kind of record, the team can trace where the content changed if an issue appears later.
That trace can then help improve the product master, AI instructions, review checklist, or publishing workflow.
But I think there is an even more useful way to look at this.
Provenance should not be useful only after something goes wrong. It can also help decide which parts people actually need to review before publishing.
Not every AI-generated sentence needs the same level of review
If AI creates a product description, not every sentence carries the same level of risk.
Some edits are relatively low risk. For example, adjusting tone, making wording more consistent, or improving readability without changing the original meaning.
Other parts need much closer review. Product performance claims such as “waterproof” or “100% organic,” as well as prices, discount conditions, campaign periods, and delivery promises, can directly affect customer decisions.
A practical workflow could let AI handle routine language work, while sending content to human review only when certain conditions appear.
- AI added information that was not in the source material
- Numbers, prices, or campaign conditions were changed
- New product specification claims were introduced
- The generated output differs from the approved source data
If I were designing the workflow, I would separate the review levels like this.
| Review level | What it means | Typical examples |
|---|---|---|
| AI can handle | Low-risk language work where the original meaning does not change. | Tone adjustment, wording cleanup, readability improvements, and expression standardization. |
| Human review required | High-impact information that can affect customer decisions, compliance, or brand trust. | Product performance claims, prices, campaign conditions, delivery promises, and legal or compliance-sensitive wording. |
| Conditional review | Human review is triggered only when the generated content shows signs of risk. | AI adds unsupported information, changes numbers or prices, introduces new specification claims, or creates output that differs from approved source data. |
This way, people do not need to reread every AI-generated sentence from scratch. They can focus on the parts where human judgment actually matters.
In real operations, the important question is not only whether AI is used. It is where humans should step in.
Passing reviewed content to the next step
Another useful idea is to keep a record of human review itself.
For example, once the product owner checks the specifications, the workflow can record the status as “reviewed,” together with the reviewer name and review date.
Then the person responsible for publishing does not have to check the same information again from the beginning.
AI draft → Targeted human review → Reviewed status → Publish
In EC operations, multiple people often touch the same content before it goes live. In that kind of workflow, it is useful to know not only who reviewed something, but also which judgment can safely be reused in the next step.
In this sense, watermarks and provenance records are not only tools for identifying AI-generated content.
They can also support a workflow where human review is focused, recorded, and carried forward.
The “AI-generated” label is only the starting point
I started this research from Claude’s text watermark, but the broader movement is clear. AI-generated content is becoming easier to trace.
For companies, however, the value is not simply knowing that AI was involved.
The real value is using that information to understand where content changed, decide what humans need to review, and pass reviewed information safely into the next step.
In that sense, I see AI watermarking and content provenance less as “AI detection” tools and more as clues for designing better human review.
The next stage is not just workflows that use AI. It is workflows designed on the assumption that AI will be used, and that human review needs to be placed deliberately.
As AI-generated content continues to grow, this kind of process design will become increasingly important for EC teams.
References
- Official: How Claude’s text watermark works / Anthropic / Anthropic
- Official: SynthID / Google DeepMind / Google DeepMind
- Official: Advancing content provenance for a safer, more transparent AI ecosystem / OpenAI / OpenAI
- Official: Content Credentials overview / Adobe / Adobe
- Official: Content provenance for Foundry models / Microsoft Learn / Microsoft Learn
- Official: When does enforcement start? / European Commission / European Commission
- Official: Misleading representation / Consumer Affairs Agency, Government of Japan / Consumer Affairs Agency
※ Part of this article was created with the support of AI, reviewed by the author, and then edited. The content reflects the author’s personal views and does not represent the official views or statements of GDX Inc.
