ERC-7517 Indicating consent to use for AI training data (Content Consent for AI/ML Data Mining)
Draft · Names & metadata · created 2023-09-12
A standard that adds a dataMiningPreference field to the metadata of digital content, letting creators declare in advance whether their content may be used for data mining or AI training.
Why it is needed
Digital content such as drawings and photos is widely used as AI training material, yet there is no common way for the creators themselves to state whether their work may be used for AI training. Creators find it hard to prevent their work from being used as training data without permission, and those building AI likewise find it hard to check which content is safe to collect. Just as websites use a robots.txt file to tell crawling bots the rules, content needs a similar means of expressing intent. This standard adds a field called dataMiningPreference to existing metadata such as NFT metadata (ERC-721, ERC-1155) and media provenance records (ERC-7053), leaving the creator's wishes in a form readable by both humans and machines, and allowing AI developers to check it systematically.
In depth
The dataMiningPreference object and its six categories
dataMiningPreference is an object holding one or more conditions. There are six categories. dataMining governs whether data mining to find patterns, trends, and correlations is permitted. aiInference governs whether the content can be used as input to a trained AI/ML model to produce inferences. aiGenerativeTraining governs whether the content can be used as training data for generative AI/ML models that can produce derivative assets, and aiGenerativeTrainingWithAuthorship is the same but requires attribution. aiTraining governs use as training data for both generative and non-generative models, and aiTrainingWithAuthorship adds an attribution requirement on top of that.
Three values: allowed, notAllowed, constrained
Each category takes one of three values. allowed means the use is free and unrestricted. notAllowed means the use is strictly prohibited. constrained means the use is permitted but subject to conditions. In the JSON schema, each property is defined as a string with an enum of these three values, and additionalProperties is true, so other properties can be added. For example, if aiInference is allowed, the content can be used as inference input without restriction, and if notAllowed, it cannot be used for AI inference.
constrained and license references
When marked constrained, the specific details of the condition are not written into the schema itself. This is to prevent conflicts with the content license; instead, a reference to the license must be included in the content metadata. Parties wishing to use the media file must follow the rules specified in the license. Per the Rationale, constrained is an intermediate state between allowed and notAllowed, signaling that additional permission or verification may be required, and the default is to treat it as notAllowed when such information is absent. The original example shows putting a name like CC-BY-4.0 and a document link into a license field, but the license field itself is not defined by this proposal.
Usage in ERC-721 and ERC-7053
For ERC-721 NFTs, the dataMiningPreference field is added to NFT metadata alongside name, description, and image. For on-chain media provenance registration under ERC-7053, there are two approaches. dataMiningPreference can be placed directly in Commit data retrieved via a decentralized content identifier (CID), or it can be placed inside custom metadata linked from the commit data through a field such as assetTreeCid. In both cases, following the CID lets you read the preference information and license reference.
C2PA alignment and design rationale
The Coalition for Content Provenance and Authenticity (C2PA) already has guidelines indicating whether an asset can be used for data mining or AI/ML training. This standard aligns on-chain metadata with those guidelines so that metadata embedded in a file and on-chain records map to each other. A JSON schema was chosen to make it easy to handle both by hand and programmatically. Training types were split into multiple categories to let creators exercise fine-grained control while accounting for both ethical and computational resource considerations, and the attribution option exists so creators can receive credit, addressing ethical and legal concerns.
Draft status
This proposal was created on September 12, 2023, and is currently in Draft status. Keep in mind that it is still a draft proposal and its contents may change.
Things to watch out for
- Data integrity: If dataMiningPreference is tampered with, it could be used for unauthorized data mining. Blockchain immutability plays a major role here, but additional security measures such as cryptographic signatures can further guarantee integrity.
- Verifying the authenticity of the party setting preferences: It must be verified that the person setting the preference is the legitimate owner of the asset or an authorized delegate. Unauthorized changes can lead to data misuse, so a service or smart contract is needed to verify the provenance and ownership of the asset before trusting its dataMiningPreference.
- Privacy protection: Care must be taken to ensure that recording preferences does not expose sensitive information about the creator or owner. Since the Ethereum blockchain is public, caution is needed to avoid unintended information leakage.