You've undoubtedly seen "watermarked" photos on the web, such as the one above. These marks are typically designed from discouraging you from using the photos without paying. However, the word "watermark" is used more generally as some kind of indicator about something on the web, e.g., an indicator that a piece of music was AI generated rather than composed by a human.
Educators in particular have been concerned about computer-generated text that might be handed in by students. Researchers have been concerned about research papers that were computer-generated. There have been programs available that are supposed to detect computer-generated text. But there are concerns about the possibility of false positives resulting in, say, discipline of a student for handing in a term paper that was machine-written when the student denies it.
On November 3rd (Election Day), Claude rolled out a watermarking system for Claude-generated text in the US. See:
https://x.com/petesena/status/2116243133956796806
Apparently, this rollout was at least partially in response to EU legislation. The EU legislation was effective in August. So it seems likely that other AI firms will follow the Claude example.
From Claude: https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content
...How Claude marks content
Claude uses two complementary techniques to mark content generated and processed by Claude: (1) watermarks embedded in text, and (2) Content Credentials (C2PA) attached to files.
1. Embedded watermarks in text
When a supported Claude model generates text, it weaves an imperceptible watermark directly into the text itself. You won’t see it, and it doesn’t change the meaning, quality, or readability of Claude’s response.
Because the watermark is part of the text, it will travel with the text when it’s copied and pasted elsewhere, and may persist through some editing. Watermarking will be applied at the model level, which means it will be present no matter which Claude product or surface the text comes from.
2. Content Credentials (C2PA)
When Claude generates a supported file type such as a PNG or JPEG, it will attach signed provenance metadata. This metadata is called a Content Credential and follows the Coalition for Content Provenance and Authenticity (C2PA) open standard, which is used across the industry to record information about content provenance. If a signed metadata label is present, it signals that a file was processed by Claude.
Detect Claude’s marks
Detection checks whether a piece of text or a file carries a supported Claude mark. If a supported mark is found, it indicates that the content may have been generated or processed by Claude.
To check whether a file contains a Claude-issued Content Credential, use the free Claude Content Checker. To learn more about how Claude marks files and how to verify Claude-issued Content Credentials, see Content Credentials on generated files.
Watermark detection is currently in private preview, available to eligible organizations as required under EU law (such as regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations, and EU civil society groups). It is also available for enterprises who are similarly obligated to verify watermarking for their own compliance with the Act...
Limitations
Machine-readable marks provide important signals about content, but it’s worth understanding their limitations across all content types.
- A detected mark provides a signal that content was processed by Claude, but is not fully conclusive. Detecting a Claude mark tells you that the content may have been processed by Claude. It does not, on its own, confirm the full provenance of the content. For example:
- Claude may not be the original author. People often use Claude to proofread, translate, summarize, or convert files. The output can carry a Claude mark even if the underlying ideas, text, or data originated from another source;
- The content may have changed after Claude processed it. Marked content may be modified, excerpted, or combined with other material after Claude processed it.
- Lack of a detected mark doesn’t mean the content wasn’t AI-generated or processed. Content generated by Claude may not carry a detectable mark if, for example:
- It was generated by a model before marking was supported for that model;
- The text has been heavily edited, paraphrased, translated, or mixed into other writing;
- The passage is very short, leaving too little text for a reliable signal;
- A file’s metadata was stripped through format conversion, re-saving, screenshots, or other means;
- It was produced through a platform, feature, or file type where a particular marking type wasn’t supported...
===
Unlike the kind of watermarking on the photo above, the text watermarking of most concern to educators is invisible to the author and reader (without a detector) and doesn't affect the quality of the generated text. More detail on how this is done can be found at:
https://www.anthropic.com/news/claude-text-watermark.
===
Whether all this will solve the education and research problems is unclear. There are currently programs available that remove the watermarks from photographs such as the one above. It would be harder to come up with programs that remove the text watermarks. But the material above suggests that such techniques are possible. Moreover, universities - not wanting to be left behind - have been making AI available to students and staff. Even students who entirely do their own research and write a basic draft of term papers might well use AI for proofing and grammar, thus inserting a watermark. Is AI usage just for proofing and grammar illegitimate? If it is, why did the university make AI readily available? In short, there will need to be tight standards for what is and what isn't legitimate.
