The biggest controversy in AI training data last year wasn't about the lack of materials, but about the inability to prove their origin—models would be criticized for "data infringement" as soon as they went live. The recently launched Seal module by Walrus ecosystem offers a new approach: when uploading files, they are first split into hundreds of fragments using erasure coding, then access permissions are directly written into objects on the Sui chain via threshold keys. The inference service can only access authorized fragments in an isolated environment, and the original files are completely inaccessible. This way, the source of model training data can stand firm legally, the community can verify it, and there's no way to shift blame.

Compared to other solutions—such as a major storage platform that only provides content hashes, or another platform that uses publicly permanent storage for a one-time deal—Seal combines privacy protection, controllable deletion, and tamper-proof features, which were originally conflicting requirements.

The downside is also obvious: key rotation costs are not low. If you need to replace 200GB of data keys at once, you must first unlock the old keys and then authorize new ones. The operation process is more complex than conventional solutions, especially when rushing to meet project deadlines, which can be a headache for development teams. However, community-provided script templates are already circulating, and future integration into CI/CD workflows should improve this significantly.