Intellectual Property and Copyright in AI

Introduction
Intellectual property (IP) and copyright have become one of the most contested and fast-moving areas of AI risk. Every generative AI system sits on top of a stack of IP questions: Who owned the data the model was trained on? Who owns the model weights, the prompts, and the fine-tuning data? Who owns the output, and could that output infringe someone else's rights? For an AI risk professional, these are not abstract legal puzzles. They translate directly into litigation exposure, regulatory penalties, loss of competitive advantage, reputational damage, and contractual liability. Within the ISACA AAIR domain of AI Risk Governance and Framework Integration, IP and copyright are treated as a category of legal and compliance risk that must be identified, assessed, treated, and monitored like any other enterprise risk, and integrated into the organization's broader governance frameworks rather than left solely to the legal department.

Why Intellectual Property and Copyright in AI Is Important

1. Unsettled law creates uncertainty. Courts and legislators worldwide are still deciding whether training a model on copyrighted material is lawful, whether AI-generated works can be protected, and who is liable when an output infringes. High-profile cases involving news publishers, image libraries, authors, and software developers suing model providers illustrate that the exposure is real and that outcomes vary by jurisdiction. Uncertainty itself is a risk that governance must address.

2. Financial and operational exposure. Statutory damages for copyright infringement can be significant, and injunctions can force an organization to withdraw a model or product from the market. If a business process depends on an AI tool that must be taken offline, continuity is affected.

3. Loss of proprietary assets. Model weights, training datasets, system prompts, fine-tuning recipes, and evaluation benchmarks are often the most valuable intangible assets an AI-driven organization holds. If employees paste confidential source code or product plans into a public chatbot, trade secret protection can be lost permanently, because trade secrets only remain protected while reasonable confidentiality measures are maintained.

4. Regulatory obligations are growing. The EU AI Act requires providers of general-purpose AI models to adopt a policy to comply with EU copyright law, to respect rights-reservation (opt-out) mechanisms, and to publish a sufficiently detailed summary of training content. Copyright offices in several jurisdictions have issued guidance on registrability of AI-assisted works. Compliance now requires documented processes, not just good intentions.

5. Third-party and supply-chain dependency. Most organizations consume AI from vendors. The IP posture of those vendors, including what they trained on, what licenses govern their models, and what indemnities they offer, becomes the organization's inherited risk.

6. Reputational and ethical dimensions. Creators, artists, and the public increasingly scrutinize whether organizations respect the rights of those whose work underpins AI systems. Responsible AI principles frequently include respect for IP as a component of fairness and accountability.

What It Is: Core Concepts

Types of intellectual property relevant to AI
• Copyright protects original works of authorship fixed in a tangible medium, such as text, images, music, video, and software code. It arises automatically upon creation. Key AI touchpoints: training data, generated outputs, and the code and documentation of the model itself.
• Patents protect novel, non-obvious, useful inventions. Key AI touchpoints: novel architectures, training methods, and AI-enabled processes. Most jurisdictions require a human inventor, so an AI system cannot be named as inventor.
• Trademarks protect brand identifiers. Key AI touchpoints: outputs that reproduce logos or brand names, and AI product naming.
• Trade secrets protect confidential information that has commercial value because it is secret and is subject to reasonable protective measures. Key AI touchpoints: model weights, proprietary datasets, prompts, and internal data leaked through AI tools.
• Database and sui generis rights (notably in the EU) protect substantial investment in compiling databases, which is relevant to scraped and compiled training sets.
• Contractual rights and licenses are not IP per se but govern how IP may be used. Terms of service, open-source licenses, dataset licenses, and API agreements frequently decide what is permitted in practice.

The three IP exposure points in the AI lifecycle
1. Inputs (training and fine-tuning data): Was the data lawfully obtained and licensed for AI training? Did scraping violate terms of service or robots exclusion signals? Were opt-outs honored?
2. The model and associated artifacts: Who owns the weights, the architecture, the prompts? What license governs a foundation model that was fine-tuned? Does an open-weight license restrict commercial use, redistribution, or use above a certain user threshold?
3. Outputs: Can the output be owned, and by whom? Could the output reproduce protected material (memorization or regurgitation) or be substantially similar to an existing work? Does the output infringe a trademark or a person's publicity rights?

How It Works: Key Legal Mechanics and Risk Dynamics

Training data and the fair use / text-and-data-mining question
In the United States, model providers argue that ingesting copyrighted works to learn statistical patterns is transformative fair use. Rights holders argue it is unlicensed copying that substitutes for the original market. Courts weigh purpose and character, nature of the work, amount used, and market effect. In the EU, the Digital Single Market Directive provides a text-and-data-mining exception for research and a broader exception for commercial use that rights holders can opt out of through machine-readable reservations; the EU AI Act requires general-purpose model providers to honor those reservations. Japan has a relatively permissive statutory exception for information analysis. The United Kingdom has a narrow research-only exception. The practical risk lesson: the lawfulness of training is jurisdiction-dependent and still being litigated, so organizations should document provenance and licensing rather than rely on a single legal theory.

Ownership of AI-generated output
Most copyright systems require human authorship. The US Copyright Office has repeatedly stated that purely machine-generated material is not copyrightable, that prompts alone generally do not confer authorship, and that protection may extend only to the human-authored elements, such as creative selection, arrangement, or substantial modification of AI output. US courts have confirmed that a work created autonomously by an AI without human involvement cannot be registered. The UK is unusual in recognizing computer-generated works with the author deemed to be the person making the arrangements necessary for creation. For risk purposes, this means an organization may not be able to prevent competitors from copying purely AI-generated marketing copy, images, or code, which affects the value of AI-produced assets and should be reflected in business decisions about where human creative contribution is required.

Output infringement and memorization
Large models can memorize and reproduce portions of training data, especially content that appears many times in the corpus. A generated image may closely resemble a specific copyrighted photograph; generated code may reproduce licensed open-source snippets without the required attribution; generated text may reproduce paragraphs of a news article. Liability can attach to the user who publishes the output as well as to the provider. Controls include output similarity filters, duplicate-detection against reference sets, code-provenance scanning, prompt restrictions, and human review before publication.

Trade secret leakage
When employees enter confidential information into external AI services whose terms permit the provider to retain or train on inputs, the information may lose trade secret status because it is no longer subject to reasonable secrecy measures. Controls include acceptable-use policies, approved enterprise tiers with no-training and data-retention commitments, data loss prevention tooling, classification-based prompt restrictions, and training for staff.

Licensing of models and datasets
Foundation models arrive under diverse licenses: permissive open-source licenses, responsible AI licenses with use restrictions, community licenses with commercial thresholds, and proprietary API terms. Datasets similarly carry licenses that may prohibit commercial use or require attribution and share-alike obligations. A fine-tuned derivative may inherit restrictions. Governance requires an inventory of all models and datasets in use together with their license terms, and a review process before adoption.

Contractual risk transfer and indemnification
Several major providers now offer copyright indemnities for outputs of their enterprise products, typically conditioned on the customer using built-in safety filters and not deliberately generating infringing content. Indemnities are a form of risk transfer, but they have exclusions and caps, so they complement rather than replace preventive controls. Risk professionals should read the conditions carefully and ensure the organization actually satisfies them.

Integrating IP Risk into Governance Frameworks

Risk management cycle. Apply the standard ISACA approach: identify IP-related risk scenarios across inputs, model, and outputs; assess likelihood and impact considering jurisdiction, use case, and visibility; select a response (avoid by not using unlicensed data, mitigate through controls, transfer through indemnities and insurance, or accept with documented rationale); and monitor through metrics such as number of unlicensed datasets discovered, output-filter hit rates, and legal developments.

Framework alignment. The NIST AI Risk Management Framework addresses IP within its Govern function (legal and regulatory requirements are understood and managed) and Map function (understanding data provenance and third-party components). ISO/IEC 42001 requires organizations to determine legal requirements applicable to the AI management system and includes controls on data acquisition, provenance, and supplier relationships. ISO/IEC 23894 guides AI-specific risk management. COBIT provides the enterprise governance structure linking AI initiatives to business objectives and compliance. The EU AI Act imposes direct copyright-policy and transparency duties on general-purpose model providers and documentation expectations on deployers.

Roles and accountability. Legal counsel owns interpretation of law and contracts; data governance owns provenance and licensing records; procurement and third-party risk management own vendor due diligence; the AI risk function coordinates assessment and reporting; business owners decide where human authorship is needed and accept residual risk; the board provides oversight.

Core control set.
• Data provenance and lineage documentation for every training and fine-tuning dataset.
• Model and dataset license inventory with pre-adoption review.
• Acceptable-use policy covering what may and may not be entered into AI tools, and which tools are approved.
• Output review procedures, similarity and code-scanning filters, and attribution handling.
• Human-in-the-loop requirements for assets where copyright ownership matters.
• Vendor due diligence covering training-data practices, indemnities, no-training commitments, and content-provenance support such as watermarking or C2PA metadata.
• Records of opt-out compliance and training-content summaries where regulation requires them.
• Training and awareness for employees, especially in marketing, engineering, product, and legal.
• Monitoring of litigation and regulatory change with a defined process to update controls.

How to Answer Exam Questions on Intellectual Property and Copyright in AI

ISACA exams test judgment from the perspective of an AI risk or governance professional, not a litigator. Questions typically present a scenario and ask for the BEST, MOST important, FIRST, or PRIMARY action or concern. Use the following approach.

Step 1: Identify the exposure point. Ask whether the scenario concerns inputs (training data), the model and its artifacts (licenses, weights, prompts), or outputs (ownership or infringement). The right answer usually targets the correct layer. A question about an employee pasting source code into a chatbot is a trade secret and acceptable-use issue, not a training-data fair-use issue.

Step 2: Determine the type of IP at stake. Copyright, patent, trademark, trade secret, or contractual license. Remember the distinguishing features: copyright is automatic and requires human authorship; patents require a human inventor; trade secrets require ongoing reasonable confidentiality measures; licenses are contracts whose terms override assumptions.

Step 3: Think like a governance professional. Prefer answers that establish policy, accountability, documentation, legal consultation, and repeatable process over one-time technical fixes. Prefer preventive and detective controls over reliance on indemnity alone. When asked what to do FIRST, favor understanding and assessing the risk (inventory, provenance review, legal consultation) before implementing or accepting.

Step 4: Respect jurisdictional nuance. If an answer claims something is universally legal or illegal, be skeptical. Correct answers often acknowledge that the position depends on jurisdiction and that documentation and licensing reduce uncertainty.

Step 5: Eliminate distractors. Common wrong answers include: assuming the organization automatically owns copyright in purely AI-generated output; assuming publicly available data is free to use for any purpose; assuming vendor indemnification eliminates the need for controls; assuming open-source means unrestricted; treating IP as solely a legal department problem outside the risk framework.

Exam Tips: Answering Questions on Intellectual Property and Copyright in AI

Tip 1: Human authorship is the anchor. If a question asks who owns copyright in output generated entirely by AI from a simple prompt, the best answer is that it is generally not protectable by copyright in most jurisdictions and that human creative contribution is required for protection.

Tip 2: Publicly accessible does not mean public domain. Scraped web content is usually still copyrighted and may be subject to terms of service and opt-out reservations. Answers that treat scraped data as freely usable are typically wrong.

Tip 3: Trade secret questions hinge on reasonable measures. If confidential data is entered into a public AI tool that retains or trains on inputs, protection may be lost. The best control is an acceptable-use policy combined with approved enterprise tools and technical enforcement, not simply telling staff to be careful.

Tip 4: Licenses are controls. When a scenario involves adopting an open-weight model or a public dataset, the best first step is to review the license terms and record them in an inventory before deployment.

Tip 5: Indemnification is risk transfer, not elimination. Expect answers that recognize indemnities have conditions, exclusions, and caps, and that the organization must still implement preventive controls and satisfy the vendor's conditions.

Tip 6: Provenance and documentation are nearly always correct. Data lineage records, training-content summaries, and license registries support compliance, defensibility, and regulatory transparency obligations. If such an answer is offered, it is a strong candidate.

Tip 7: Output risk requires output controls. For memorization or regurgitation scenarios, look for similarity filtering, code-provenance scanning, attribution handling, and human review before publication.

Tip 8: When asked about regulatory requirements for general-purpose model providers, remember the EU AI Act themes: a copyright compliance policy, respect for machine-readable opt-outs, and a published summary of training content.

Tip 9: Map IP risk to frameworks. Recognize that NIST AI RMF Govern and Map functions, ISO/IEC 42001 legal-requirement and supplier controls, and COBIT governance objectives all accommodate IP risk. Answers that integrate IP into the enterprise risk register and existing frameworks beat answers that create isolated processes.

Tip 10: Watch the qualifier. FIRST usually points to assessment, inventory, or legal consultation. BEST usually points to the most comprehensive governance control. PRIMARY concern usually points to the highest-impact exposure, often legal liability or loss of proprietary assets, rather than a technical inconvenience.

Tip 11: Separate legal certainty from risk management. The exam does not expect you to predict court outcomes. It expects you to recognize uncertainty, reduce it through licensing and documentation, transfer part of it through contracts, and monitor the legal landscape for change.

Tip 12: Remember the stakeholder lens. Correct answers often involve collaboration between legal, data governance, procurement, business owners, and the AI risk function, with clear accountability and board-level oversight.

Quick Reference Summary
• IP types: copyright, patent, trademark, trade secret, database rights, contractual licenses.
• Exposure points: inputs, model and artifacts, outputs.
• Key principles: human authorship for copyright and inventorship; reasonable measures for trade secrets; jurisdiction-dependent training lawfulness; licenses govern practical use.
• Key controls: provenance records, license inventory, acceptable-use policy, approved tools, output filters, human review, vendor due diligence, indemnity review, training, legal monitoring.
• Framework hooks: NIST AI RMF (Govern, Map), ISO/IEC 42001, ISO/IEC 23894, COBIT, EU AI Act general-purpose model obligations.
• Exam mindset: governance over gadgets, assess before act, documentation is defensibility, transfer complements but does not replace prevention.

Mastering this topic means being able to look at any AI scenario, locate the IP exposure, name the right type of protection, and select the governance response that a prudent risk professional would recommend. That is precisely what the AAIR exam rewards.