Five ways organizations can make consent, credit, compensation, and recourse part of how they build and buy AI.
Imagine spending years developing a recognizable artistic style, only to find a service generating images that look remarkably like your work. Was your portfolio part of its training data? Did anyone license it? Would you even know whom to ask?
Artists deserve more than an acknowledgment that AI is changing the world. For companies building and buying AI, the practical response is to make permission, provenance, compensation, and accountability part of the architecture, not an afterthought.
Consent, Credit, Compensation, and Recourse
Artists and their advocates often frame the ask as consent, credit, and compensation. We add a fourth: recourse. Knowing your work was used without authorization means little without a way to challenge that use, and even a well-designed system can make mistakes.
| Principle | Practical standard | What good looks like in a system |
|---|---|---|
| Consent | Demonstrate the permission, license, or other legal basis for using the work. | Documented rights for training, fine-tuning, and retrieval. Applicable rights reservations honored in data collection. |
| Credit | Preserve appropriate creator attribution and provenance information. | Provenance records on data entering the system, with Content Credentials on supported media outputs. |
| Compensation | Establish fair licensing or payment terms where applicable. | Licensed stock, commissioned creative work, and vendor arrangements that compensate creators. |
| Recourse | Give creators a way to challenge unauthorized use, receive a response, and seek appropriate remedies. | A published complaint process with a responsible owner, review timelines, documented outcomes, and removal or suppression where technically feasible. |
Credit is not consent. Provenance is not permission. Compensation does not automatically establish authorization.
Five Controls Organizations Can Put in Place Now
Responsible AI cannot rest on good intentions or vendor promises. It needs controls that can be tested, documented, and reviewed.
1. A Data Provenance Register With Permitted Uses
For every dataset used in training, fine-tuning, or retrieval, record its source, legal basis, collection date, owner, and permitted uses (training, retrieval, publication, retention). If a team cannot show where data came from or a defensible basis for its use, the data stays out until reviewed.
Consider an illustrator’s portfolio found in a web crawl. Your pipeline can record where the images came from, but public access does not establish permission for every use. With no documented authorization, the images are quarantined until their permitted uses are established.
The system should not assume that publicly visible means freely reusable. That distinction is what turns an ethical principle into an architectural control.
Permission to index a document does not necessarily authorize fine-tuning on it. The US Copyright Office examines these stages separately, and your register should too.
2. Vendor Due Diligence on Training Data
Even organizations that never train a model still choose vendors. Ask them:
- What information can you provide about your training-data sources?
- What copyright compliance policies apply?
- How are rights reservations identified and respected?
- What processes exist for complaints or disputed uses?
- What contractual protections and remedies are available?
Document the answers, review them with counsel, and write the commitments into vendor agreements. An assurance that an AI is responsible is not evidence.
3. Respect Rights Reservations in Your Own Pipelines
If you collect web content for retrieval or fine-tuning, honor applicable rights reservations, including robots.txt and other appropriate machine-readable mechanisms, as the European Commission’s General-Purpose AI Code of Practice asks of its signatories. A good pipeline does more than check whether a URL can be fetched:
Discover → Check rights → Record permitted uses → Approve or quarantine → Monitor and review.
Make each decision auditable, so a challenged use can be traced to what was known and why. Respecting creators should be routine, not an emergency exception.
4. Label What You Make, and Know the Limits of Labels
The C2PA Content Credentials standard is an open, royalty-free way to attach signed, tamper-evident provenance to supported media, including whether AI was involved. That is useful, but it is not a complete rights-management system.
Metadata can be removed, and the specification‘s optional watermark and fingerprint soft bindings do not guarantee recovery. More importantly, C2PA establishes mechanisms for recording and validating provenance claims. It does not provide value judgments about whether provenance data is true. Nor does a credential, on its own, show that the underlying activity was authorized.
A system can accurately label an image as AI-generated without showing that the works used to train that AI were licensed. Provenance should inform rights decisions, not replace them.
5. A Clear Use Policy for Style, Likeness, and Complaints
Decide what your AI systems will not be used to create, and review it with counsel. One policy we recommend:
Do not intentionally use generative AI to imitate the recognizable style of a specific living artist, whether or not that artist is explicitly named in a prompt. Do not reproduce a real person’s voice or likeness without appropriate authorization.
This is a recommended safeguard, not a claim that every such act is unlawful. It sets a clearer standard than waiting for courts, and it needs an approval process and a named owner.
Creators also need a clear complaint process: who reviews it, when they hear back, and what remedies exist. Be honest about limits: removing material from a retrieval index is not the same as removing its influence from a trained model. Accountability should not depend on whether a creator can attract enough public attention to force a response.
Five Questions Artists Should Be Able to Ask Any AI Company
Any organization using creative AI should be ready to answer these, and artists should not need to understand machine learning to ask them.
- Was my work used? Was it included in training or retrieval data, and what can you disclose about its source?
- What gave you the right to use it? What permission, license, or other legal basis supports that use?
- Can I refuse future use? How can I reserve my rights or object to particular uses, and what limitations apply to material already processed?
- What happens to my attribution and compensation? How are creator credit, licensing, and payment handled where applicable?
- What happens if something goes wrong? Who reviews my complaint, how will I receive a response, and what remedies are available?
A company may have legitimate limits on what it can disclose or remedy, but it should explain them rather than treat the questions as unreasonable.
Why Companies Can No Longer Ignore These Questions
The US Copyright Office Has Clarified the Questions
In May 2025, the US Copyright Office issued a pre-publication report on generative AI training. It finds neither that all AI training is fair use nor that all of it infringes, but it raises concerns about commercial training that produces competing expressive content, particularly from unlawfully accessed sources. Part 3 is still a pre-publication report, not a binding decision.
The EU Has Established Copyright Obligations for AI Model Providers
Under Article 53 of the EU AI Act, providers of general-purpose AI models must maintain a copyright policy that addresses rights reservations under EU text and data mining rules and publish a sufficiently detailed summary of their training content. This has applied to new models since August 2, 2025; under Article 111, models placed on the market before that date have until August 2, 2027.
Most organizations that use AI are deployers rather than model providers, so for them these rules mainly shape which vendors and contracts they choose.
These rules have specific scopes and are not a universal right to inspect training data. But they make copyright compliance a matter of operational governance.
A Major Settlement Underscored the Risks of Pirated Training Data
On July 20, 2026, a US federal court granted final approval of a $1.5 billion class-action settlement in Bartz v. Anthropic. It resolved claims over books downloaded from pirate libraries. The court estimated about $3,000 per work before costs and fees. That figure was part of a negotiated settlement, not a damages amount calculated by the court after a finding of liability.
In an earlier June 23, 2025, ruling, Judge William Alsup found that the training use of the books at issue was fair use, as was digitizing lawfully purchased print books. He reached a different conclusion on the unauthorized acquisition and retention of pirated copies. That ruling came from a single trial court and does not bind other courts.
The settlement did not establish that all AI training infringes copyright. It did reinforce a principle: how you obtain information matters, even when the eventual technology has legitimate uses.
Why This Is Good Architecture, Not Only Good Manners
- Lower legal exposure. Documented sources and decisions are easier to defend than undocumented collections.
- Fewer expensive corrections. Content that must later be excluded from a fine-tuned model may force retraining. Catching it before deployment can avoid that cost.
- Stronger governance. Tracked rights support vendor review, audits, and future regulation.
- Trust you can show. Creators and customers can examine decisions, not just statements of principle.
None of this guarantees compliance, and none of it should be the only reason to respect artists.
The Responsibility Belongs With Those Building the Systems
Artists should not have to become forensic investigators to learn how their work is used. The organizations building and buying AI have the purchasing power to demand better answers. We can document where data comes from, choose vendors on more than token costs, treat unclear permissions as a reason for review rather than an invitation to proceed, and listen to creators first.
If you create art, write, compose, or photograph professionally, what would meaningful protection look like to you? Consent before training? Better licensing? Fair compensation? Reliable attribution? The ability to challenge unauthorized use? Technology companies need to hear those answers before designing the next generation of AI systems.
Redrock Systems architects help organizations map where their AI data comes from and design rights controls into their systems. To talk through your own, book a scope call.
This article is general information, not legal advice. For advice on your situation, talk to your own counsel.
Sources
- US Copyright Office. Copyright and Artificial Intelligence, Part 3: Generative AI Training. Pre-publication report, May 2025. https://copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-3-Generative-AI-Training-Report-Pre-Publication-Version.pdf
- US Copyright Office. Copyright and Artificial Intelligence. Report status and supporting materials. https://copyright.gov/ai/
- European Union. AI Act, Article 53: Obligations for Providers of General-Purpose AI Models. https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-53
- European Union. AI Act, Article 111: Transitional Provisions. https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-111
- European Commission. General-Purpose AI Code of Practice. Published July 10, 2025. https://digital-strategy.ec.europa.eu/en/policies/contents-code-gpai
- US District Court, Northern District of California. Bartz v. Anthropic PBC, No. 3:24-cv-05417, Order Granting Final Approval of Class Action Settlement, Dkt. 680, July 20, 2026. https://storage.courtlistener.com/recap/gov.uscourts.cand.434709/gov.uscourts.cand.434709.680.0_1.pdf
- US District Court, Northern District of California. Bartz v. Anthropic PBC, No. 3:24-cv-05417, Order on Fair Use, Dkt. 231, June 23, 2025. https://cases.justia.com/federal/district-courts/california/candce/3:2024cv05417/434709/231/0.pdf
- US District Court, Northern District of California. Bartz v. Anthropic PBC, No. 3:24-cv-05417, Memorandum Opinion on Preliminary Approval, Dkt. 437, October 17, 2025. https://storage.courtlistener.com/recap/gov.uscourts.cand.434709/gov.uscourts.cand.434709.437.0.pdf
- The Authors Guild. Court Grants Final Approval of $1.5 Billion Anthropic Copyright Settlement. July 21, 2026. https://authorsguild.org/news/court-grants-final-approval-anthropic-copyright-settlement/
- Coalition for Content Provenance and Authenticity (C2PA). Content Credentials Explainer, Version 2.2. https://spec.c2pa.org/specifications/specifications/2.2/explainer/Explainer.html
- C2PA. Technical Specification, Version 2.2, Section 9.3: Soft Bindings. https://spec.c2pa.org/specifications/specifications/2.2/specs/C2PA_Specification.html