Is this for creators?

[CC-signals]

Individual creators are not the primary audience for this tool: content repository and data stewards are. However, there is nothing preventing an individual from using the tool.

While CC signals are a framework available for anyone to use, we anticipate that the framework best applies to stewards of large collections of content. Because AI platforms ingest content at such a large scale, we anticipate an intervention at this level would ultimately be more effective in changing AI platform practices.

For more details on considerations for individual creators, see our report, (From Human Content to Machine Data: Introducing CC Signals, page 30):

“There are different ways in which stewards of content collections can use CC signals to give force to the expectations of individual creators, and we’re eager to engage further about this as part of our public consultation process. We’re observing efforts to understand and define community preferences with keen interest, such as Holly Herndon and Mat Dryhurst’s recent work with Serpentine Arts Technologies to bring groups of choirists together to determine how they’d like their works to be used by generative AI models.

Conceivably, stewards of content could enable individual contributors to select the signals they want associated with their contributions, similar to the approach being taken by the social media platform, Bluesky, to enable users to express their own preferences regarding the reuse of their public posts. Using this approach, the full collection would then be divisible into different datasets with different combinations of signals.”

[CC-signals]

@TimidRobot could you share more about the technical means for applying a CC signal to an individual work?

[TimidRobot]

Overview

ℹ️ Note: both the IETF AI Preferences and CC signals are still in draft (not yet ready for implementation).

I expect that once the IETF AI Preferences are released, additional attachment methods like, image metadata, will be developed. The initial attachment methods are robots.txt and the HTTP Content-Usage header.

If a creator has control over the web server, they will be able to mark individual works with CC signals by setting the desired HTTP Content-Usage header.

Example 1

From the README:

The following HTTP Content-Usage header example grants permission to the URL with a usage preference of no for generative AI training unless credit and an ecosystem contribution is provided per the CC signal.

200 OK
Date: Mon, 09 Jun 2025 12:42:03 UTC
Content-Type: text/plain
Content-Usage: train-genai=n;exceptions=cc-cr-ec

Example 2

1. Creates an opt-out regime by specifying yes for the Automated Processing (all) category
2. Allows AI training with credit per the CC signal
3. Disallows Generative AI training

200 OK
Date: Thu, 10 Jul 2025 13:15:10 UTC
Content-Type: text/plain
Content-Usage: all=y, train-ai=n;exceptions=cc-cr, train-genai=n

Example 3

1. Creates an opt-in regime by specifying no for the Automated Processing (all) category
   • This is far more conservative and will have potentially unexpected negative consequences
2. Allows AI training with credit per the CC signal
   • Since Generative AI training is not specified, it follows the same requirements as AI training
3. Allows AI usage with credit per the CC signal
4. Allow search

200 OK
Date: Thu, 10 Jul 2025 13:15:10 UTC
Content-Type: text/plain
Content-Usage: all=n, train-ai=n;exceptions=cc-cr, ai-use=n;exceptions=cc-cr, search=y

Best practices TBD

I expect that both AI Preferences and CC signals will eventually recommend that the top level category always be specified (all=n or all=y). Doing so will result in more clear and consistent behavior across legal jurisdictions.

Updates

• 2025-07-28: updated to reflect changes in IETF AIPREF drafts

[stkdiretto]

Could this allow, e.g. governmental agencies or other actors who provide Open Data to attach CC Signals to them and expect remuneration or credit at the level of individual facts that are used in classic TDM applications (which do not lead to trained generative models)?

[CC-signals]

This would allow declaring parties (e.g. governmental agencies) to attach CC signals, which would in turn require credit at the dataset level, rather than level of individual facts. For example, the Direct Contribution signal would require monetary or in-kind support to the declaring party, but this would not be based on individual facts, rather the collection of data as a whole. It’s worth noting that while CC signals offer the first step, declaring parties would still determine their best courses of action for stewardship, including how remuneration may be allocated.