Copyright control for generative AI starts with the production process, not the final output

Japan’s Agency for Cultural Affairs guidance distinguishes between the AI development/training stage and the generation/utilization stage, individually examining factors such as the purpose of use of copyrighted works, similarity and dependence on existing works, and human creative intent and contribution. Therefore, selecting a model with an appropriate license does not mean the generated content can be published as-is; these are not the same judgment. [1]

In practice, we link the rights basis of input materials, service terms, prompts, candidate generations, selection, addition, and final scope of use into a single record. Recording where humans selected expressions and where they made corrections is not to automatically determine copyrightability, but to allow for explaining and re-evaluating factual relationships later. [1]

The intensity of case verification is divided not only by the format of the output but also by the origin of input materials, whether existing works were specified, the scope of publication, and plans for modification or redistribution. When moving materials approved with low risk to different uses such as advertisements or products, previous verifications are not automatically inherited. We place the examination elements indicated by Japan’s Agency for Cultural Affairs guidance in the factual verification column, leaving legal conclusions as individual judgments. [1]

• Record the rights holder, license, source, and purpose of use for input materials
• For generation, save the model name/version, terms of service, prompts, and reference materials
• For editing, record the content selected, structured, and added by humans, along with the responsible person
• For publication, finalize similarity checks, scope of use, disclosures, and approvers

Article 53 of the EU AI Act requires providers of general-purpose AI models to maintain a policy compliant with EU copyright law and publish a sufficiently detailed summary of the content used for training. The article does not uniformly exempt obligations regarding copyright policies and training content summaries, even with certain exceptions for open source. [2]

However, the summary is not a ledger proving the legality of individual works. Procurement parties should not pass based solely on the presence of a summary; they must inquire about mechanisms for detecting rights reservations, responses to deletion requests, classification of data sources, additional training data during model updates, and contact channels for disputes, evaluating these against their own regions of use and purposes. [2]

C2PA 2.4 uses assertions linked to assets, signed claims, and manifests to handle provenance information such as creation, editing, and constituent materials in a tamper-evident manner. Attaching Content Credentials to published items provides clues for verifying who claimed what in which process. [3]

On the other hand, as C2PA itself explicitly states, verification confirms whether claims are linked to assets, in the correct format, and unaltered, not whether the content of the claims is legally or factually correct. It is important to separately verify the authority of signers and the rights of original data, conduct similarity reviews, and avoid treating provenance information as a form of liability exemption. [3]

The absence of Content Credentials is not evidence of fraud, nor does its presence prove a license. [3] In provenance verification, we separately record the signature verification result, the claiming entity, the chain of materials, and points where information is interrupted. If there are interruptions, we do not uniformly reject them; instead, we send them for additional verification of original materials and editing history, narrowing the scope of publication if the rights basis cannot be recovered.

Since Japan’s Agency for Cultural Affairs materials indicate different examination elements for each process, and C2PA distinguishes between the integrity of provenance claims and the truth of their content, there is room for inconsistencies between pieces of evidence even in the same case. [1] [3] Therefore, we maintain an adjudication table that lists claims, supporting materials, opposing materials, and unverified items, rather than collapsing verification results into a single pass stamp.

Each row of the adjudication table becomes a specific claim such as 'This material can be used for this purpose, within this scope.' We link the version of supporting materials, the verifier, opposing information, deadlines for additional verification, and permitted provisional use, ensuring that when materials increase, we can identify which claims need to be reopened. We do not score the weight of evidence but adjudicate by purpose.

As a hypothetical example, if there is a license record for input materials but similarity checks for generated candidates are pending, we reach different conclusions for internal review use and external publication. We do not extend 'some evidence exists' to approval for all uses; instead, we fix in the adjudication table, rather than in the responsible person's memory, which behaviors are permitted and which are prohibited until conflicts are resolved. If the conditions of the hypothetical example change, the conclusion is not automatically inherited.

When using the same generated content in different media, regions, or scopes of publication, do not re-explain the production process from scratch; instead, list the conditions that have changed since the previous approval in a delta sheet. Fix unchanged evidence as references, and explicitly state which rights claims are affected by differences in purpose, target, modification, redistribution, and disclosure methods. Explicitly state that columns with no differences are not subject to re-review.

In delta review, test counterfactuals one by one: removing editing, reverting to original materials, narrowing the scope of publication, and prohibiting secondary distribution. If we can identify which changes eliminate the reasons for holding, we can compare the options of rejecting the entire case or replacing only the doubtful parts as explainable choices. We pair the points remaining and those disappearing for each counterfactual.

When closing the review, we save not only the adopted use conditions but also the conditions considered and rejected, along with their reasons. This is to prevent later personnel from misusing the fact that 'it was previously published' for different uses. We mechanically check if the new use exceeds the previous scope of allowance, assigning re-verification responsible persons and deadlines only to columns that exceed it.

• CLAIM: Record materials, generated content, and scope of use as a single sentence rights claim
• CONFLICT: Place supporting materials, opposing materials, and unverified items in the same row
• DELTA: Extract only changes in media, region, modification, and redistribution
• RULING: Finalize permitted behaviors, prohibited behaviors, re-review conditions, and deadlines

To solidify actions, rather than abstractly telling personnel to 'be careful about copyright,' we place verification items immediately before the publish button. We automatically display the rights basis, model terms, human editing records, similarity checks, and Content Credentials status from the case ledger, creating a mechanism that prevents publication if any are missing.

For high-risk publications, Catalysphere proposes separating production and approval responsibilities and allowing time to replace doubtful materials. Link records under the same identifier from the start of generation so that the information needed for explanation, correction, deletion, and republication decisions can be traced. This is an operating proposal; the cited authorities do not demonstrate a reduction in decision time. [1] [2] [3]

We also include correction paths after publication in the audit trail design. When a rights holder raises a doubt, we halt the target version, identify the publication site and secondary distribution sites, execute replacement or deletion, and track the process under the same case number until approval for republication. We also record changes in the procurement source's training content summary or copyright policy as triggers for evaluating the impact on existing published content. [2]

• Do not include input materials with unknown rights basis in the generation process
• Return outputs with suspected similarity to source exploration and expert verification
• Re-sign or replace derivatives with missing provenance information before publication
• Do not automatically inherit past approvals when systems, terms, or model versions change
• Production personnel: Save facts regarding input, candidate selection, and addition
• Approval personnel: Determine missing scope of use and rights basis
• Publication administrator: Close stops, replacements, and corrections to secondary distribution sites