Understanding Fair Use in the Age of Artificial Intelligence

The rapid advancement of artificial intelligence has introduced complex questions about copyright law and fair use. As AI systems train on vast datasets of existing content and generate outputs that may reference, incorporate, or be inspired by copyrighted works, creators and companies are navigating uncertain legal territory. Our AI Fair Use Analyzer helps evaluate whether specific uses of copyrighted material might qualify as fair use, but it’s essential to understand the legal framework behind this analysis.

What is Fair Use?

Fair use is a legal doctrine that permits limited use of copyrighted material without permission from the rights holder. It serves as a crucial exception to the exclusive rights granted to copyright owners, allowing for uses that benefit the public interest through commentary, criticism, news reporting, teaching, scholarship, or research.

In the United States, fair use is codified in Section 107 of the Copyright Act, which establishes four factors courts must consider when determining if a use qualifies as fair:

1. Purpose and character of the use: Whether the use is commercial or nonprofit/educational, and whether it is “transformative” (adding new meaning, message, or purpose)
2. Nature of the copyrighted work: Whether the original work is factual or creative
3. Amount and substantiality of the portion used: How much of the original work is used, both quantitatively and qualitatively
4. Effect on the potential market: Whether the use harms the existing or potential market for the original work

These factors are not a simple checklist but must be weighed together, with courts often giving different emphasis to different factors depending on the circumstances.

The Intersection of AI and Fair Use

Artificial intelligence introduces new complexities to fair use analysis. AI systems interact with copyrighted material in several distinct ways:

AI Training Data

Machine learning models require vast datasets for training. Companies often scrape the internet or use collections of books, articles, images, and other media—much of which is copyrighted—to train their systems. This raises questions about whether this use constitutes fair use, particularly when:

• The AI company is a commercial enterprise
• The AI learns patterns from works without explicitly storing or reproducing them
• The training process is arguably transformative as it converts creative works into mathematical representations

Recent lawsuits against companies like OpenAI, Meta, and Stability AI challenge the assumption that all AI training falls under fair use. The outcomes of these cases will significantly shape the legal landscape.

AI-Generated Outputs

When AI systems produce content inspired by, similar to, or incorporating elements of copyrighted works, additional fair use questions arise:

• Is AI-generated content that stylistically resembles a particular artist’s work infringing?
• When does AI output cross the line from being “inspired by” to becoming a derivative work?
• If an AI reproduces portions of copyrighted text or images, does this qualify as fair use?

The Four Fair Use Factors in AI Contexts

Let’s examine how each fair use factor specifically applies to AI use cases:

Factor 1: Purpose and Character of Use

For AI, this factor often centers on whether the use is transformative. Courts consider whether the new use adds something new, with a different purpose or character, rather than merely superseding the original.

AI Training Considerations:

• Training models may be considered transformative because the purpose is to teach the AI patterns rather than to consume or display the work
• Commercial AI companies face a tougher standard than academic researchers
• Using works to extract facts and patterns rather than to engage with expressive content strengthens the fair use argument

AI Output Considerations:

• AI-generated content that parodies, criticizes, or comments on original works stands on stronger fair use ground
• Outputs that serve the same purpose as the original (entertainment, information, etc.) without adding new context or meaning may weaken fair use claims
• Purely commercial generation of content similar to copyrighted works weighs against fair use

The landmark Google Books case (Authors Guild v. Google) suggested that mass digitization of books for search and analysis could be transformative. However, the 2023 Supreme Court decision in Andy Warhol Foundation v. Goldsmith narrowed the transformative use doctrine, emphasizing that works serving the same purpose as the original are less likely to qualify as fair use.

Factor 2: Nature of the Copyrighted Work

This factor examines the characteristics of the original work being used.

AI Considerations:

• Using factual works (news, scientific information) generally receives more fair use protection than using highly creative works (novels, poetry, music)
• Published works receive less protection than unpublished works
• AI systems trained on diverse data may have stronger fair use arguments for factual works than for creative content

Factor 3: Amount and Substantiality Used

This factor looks at both how much of the original work is used and how important that portion is to the original.

AI Training Considerations:

• AI training often uses entire works, which typically weighs against fair use
• However, courts have found that using entire works can be fair use when necessary for a transformative purpose (as in Google Books)
• The transformation of works into mathematical representations rather than readable copies may strengthen the fair use argument

AI Output Considerations:

• AI systems that reproduce substantial portions of original works face greater copyright risk
• Outputs containing small, less-central portions of works have stronger fair use claims
• Even small portions can be problematic if they constitute the “heart” of the original work

Factor 4: Effect on the Potential Market

This factor examines whether the use harms the existing or potential market for the original work. Courts consider not just actual market harm but also potential markets the copyright holder might reasonably develop.

AI Training Considerations:

• If AI training doesn’t directly compete with or reduce demand for original works, this strengthens fair use claims
• However, if AI systems can generate content that substitutes for or competes with original works, this could indicate market harm
• Emerging licensing markets for AI training data complicate this analysis—if creators could reasonably license their works for AI training, using works without permission might harm this potential market

AI Output Considerations:

• AI-generated content that directly competes with or substitutes for the original works weakens fair use arguments
• Outputs that enhance or drive interest in the original works may strengthen fair use claims
• Courts consider whether the use would harm licensing opportunities if the practice became widespread

Recent Legal Developments in AI Fair Use

The legal landscape around AI and fair use is rapidly evolving. Recent developments include:

The GitHub Copilot Lawsuit (Doe v. GitHub, Inc.) – This case challenges Microsoft/GitHub’s AI code generator, alleging it reproduces substantial portions of open-source code without proper attribution or compliance with license terms. The outcome may clarify how fair use applies to code-generating AI.

Visual Artists’ Litigation – Several lawsuits have been filed by visual artists against companies like Stability AI, Midjourney, and DeviantArt, claiming their image-generating AI systems were trained on artists’ works without permission. These cases may establish whether AI art generation based on existing works constitutes fair use.

Authors’ Litigation Against OpenAI – Authors including Paul Tremblay, Mona Awad, and others have sued OpenAI for allegedly training on their books without permission. The resolution of these cases will provide important guidance on whether using books to train generative AI is fair use.

The New York Times v. OpenAI and Microsoft – This lawsuit alleges copyright infringement through both training and output reproduction. The Times argues that ChatGPT can produce content that closely mimics its articles, potentially affecting its business model.

Practical AI Fair Use Considerations

When evaluating AI fair use, consider these practical aspects:

Data Transparency and Provenance

Courts may look more favorably on AI systems with transparent data collection practices. Organizations should:

• Document the sources of training data
• Implement data filtering processes when appropriate
• Consider obtaining licenses for high-risk content categories
• Respect opt-out mechanisms from content creators

Output Controls

AI systems that include safeguards against problematic outputs may have stronger fair use arguments:

• Implement filters to prevent reproduction of substantial copyrighted content
• Design systems that create transformative outputs rather than near-copies
• Include attribution capabilities when appropriate
• Develop procedures to handle copyright complaints about outputs

Risk Tiers for Different Content Types

Not all content carries the same fair use risk:

Lower Risk:

• Factual, informational content (news articles, scientific papers)
• Works with expired copyrights or open licenses
• Content consisting primarily of ideas, facts, or methods

Medium Risk:

• Mixed factual and creative content
• Older creative works with established markets
• Works where small portions are used in a transformative context

Higher Risk:

• Highly creative works (fiction, poetry, art, music)
• Recent commercial works with active markets
• Works used in ways that could substitute for the original

Best Practices for AI Developers and Users

To minimize legal risk while using copyrighted materials in AI contexts:

1. Start with proper licensing when possible. Obtain permission or use content with appropriate licenses (like Creative Commons) when available.
2. Build transformative systems. Design AI that transforms inputs into genuinely new insights, applications, or forms of expression rather than reproducing original content.
3. Implement technical safeguards. Use content filters, opt-out mechanisms, and attribution systems to respect copyright holders’ interests.
4. Document your fair use analysis. Keep records of your reasoning for why specific uses constitute fair use, which can demonstrate good faith.
5. Stay informed about legal developments. This area of law is evolving quickly, and what constitutes fair use today might change tomorrow.
6. Be responsive to concerns. Implement takedown procedures and be willing to address legitimate copyright concerns.
7. Consider the broader impact. If your practice became widespread, would it substantially harm creators’ incentives or markets?

The Fair Use Analyzer Tool: A Guide to Usage

Our AI Fair Use Analyzer walks you through an evaluation of your specific use case against the four fair use factors. Here’s how to use the tool effectively:

Step 1: Define Your AI Use Case

Begin by selecting the option that best describes how you’re using copyrighted material in relation to AI:

• Training AI Models: Using copyrighted works as part of a dataset to train AI systems
• AI-Generated Content Inspired By Works: Using AI to create content similar to or in the style of existing works
• AI-Generated Derivative Works: Using AI to adapt, transform, or build upon specific copyrighted works
• AI Reproduction of Works: Using AI to reproduce substantial portions of specific copyrighted works

Your selection helps tailor subsequent questions and the final analysis to your specific scenario.

Step 2: Evaluate Purpose and Character of Use

Next, you’ll assess the first fair use factor by considering:

• Whether your use is commercial, educational/research, nonprofit, or personal
• How transformative your use is, from highly transformative (adding significant new expression or meaning) to not transformative (using with little change for similar purposes)

Commercial uses generally face a higher standard for fair use, while highly transformative uses are more likely to be considered fair. For example, using copyrighted materials to train an AI for public research purposes receives more favorable treatment than using the same materials to train a commercial AI product that generates content similar to the originals.

Step 3: Assess the Nature of the Copyrighted Work

This step examines the type of work being used and its protection level:

• Whether the content is primarily factual/informational, mixed, or highly creative
• If the work is published or unpublished

Factual works like news articles receive less copyright protection than highly creative works like novels or music. Similarly, published works are more available for fair use than unpublished works where the creator hasn’t chosen to release them.

Step 4: Analyze Amount and Substantiality Used

Here you’ll consider how much of the original work is being used:

• The quantity (small amount to entire work)
• The quality or importance (peripheral parts vs. the “heart” of the work)

Using small portions of works generally supports fair use more than using entire works. However, even small portions can be problematic if they constitute the most valuable or memorable parts of the original.

Step 5: Evaluate Effect on Potential Market

The final step addresses how your use impacts existing or potential markets for the original work:

• The level of market effect (from positive/none to substantial negative)
• Whether licensing is available for your type of use

Uses that don’t harm—or potentially enhance—the market for the original work support fair use, while uses that substitute for or directly compete with the original generally don’t. The availability of reasonable licensing options is also relevant; if you could easily license the content but choose not to, this may weigh against fair use.

Understanding Your Results

After completing the assessment, you’ll receive:

1. Overall Fair Use Score: A percentage indicating the relative strength of your fair use argument, from weak to strong
2. Factor Analysis: Detailed ratings of how each factor applies to your case
3. Summary Analysis: An explanation of your results in the context of your specific AI use
4. Legal Considerations: Important legal points relevant to your scenario
5. Risk Mitigation Strategies: Practical suggestions to reduce your legal risk

Remember that this analysis is informational and cannot guarantee how a court would rule in your specific case. Fair use determinations are notoriously complex and unpredictable, and AI applications add further uncertainty to this area of law.

FAQ: AI Fair Use Questions

How is fair use for AI different from traditional fair use?

Traditional fair use developed around human uses of copyrighted material for purposes like commentary, criticism, and education. AI introduces novel questions because machines, not humans, are “reading” works, and the purpose is often to extract patterns rather than engage with the expressive content. Additionally, AI can process thousands of works simultaneously, raising questions about scale that weren’t contemplated in traditional fair use cases.

The transformative nature of AI use is also distinctive—works are converted into mathematical representations rather than being reproduced in recognizable form. This technical transformation may support fair use, but courts are still determining how to apply existing precedents to these new technological contexts.

Does training an AI model on copyrighted works constitute copyright infringement?

This remains unsettled law. While copying works for AI training technically involves reproduction (a copyright owner’s exclusive right), proponents argue this is fair use because:

The purpose is typically transformative—to extract patterns and information rather than to consume or display the work The AI doesn’t generally store or output complete copies of original works The training process often doesn’t harm traditional markets for the works

However, recent lawsuits challenge these assumptions, especially as AI systems improve at reproducing content similar to their training data. The outcomes of cases against companies like OpenAI and Stability AI will provide crucial guidance on this question.

Can AI-generated content infringe copyright if it “sounds like” or is “in the style of” a particular creator?

Generally, copyright protects specific expression but not styles, ideas, or methods. Creating content “in the style of” a particular artist, author, or musician may not constitute copyright infringement if it doesn’t copy specific elements of their works.

However, the line between style and expression can be blurry. If an AI output closely mimics specific works rather than just their general style, this could potentially infringe copyright. Additionally, marketing AI content as being “in the style of” specific creators could raise other legal issues like right of publicity or trademark concerns, even if copyright fair use applies.

What’s the difference between an AI output being “inspired by” versus creating a “derivative work”?

This is a crucial distinction in copyright law:

Being “inspired by” existing works generally means taking ideas, concepts, or non-copyrightable elements and creating something new with them. Copyright doesn’t protect ideas, only their specific expression, so being inspired by works is generally permitted.

A “derivative work” is a new, creative work based on or incorporating elements of a preexisting copyrighted work. Copyright owners have the exclusive right to create derivative works based on their originals. Examples include translations, adaptations, or new versions that include recognizable elements of the original.

The line between inspiration and derivative works can be subjective. Key factors include how much the new work takes from the original and whether the average person would recognize the connection to the original work.

Does having an AI system attribute or cite sources strengthen a fair use claim?

Attribution alone doesn’t make an unauthorized use legal—you can’t copy an entire novel and avoid copyright issues just by crediting the author. However, attribution can be relevant to fair use analysis in several ways:

It may demonstrate good faith, which courts sometimes consider favorably It can support transformative use arguments, particularly for commentary or scholarship It might reduce market harm if it drives traffic or interest to the original

For AI systems, implementing attribution mechanisms may be seen as responsible practice and could potentially strengthen fair use arguments in some contexts, though it’s not a shield against infringement claims.

Are there different fair use standards for different types of AI-generated content?

Yes, the fair use analysis varies significantly based on the type of content and how it’s used:

Text Generation: AI-generated text that reproduces substantial portions of specific copyrighted works (like books or articles) faces higher copyright risk than text summarizing factual information from multiple sources.

Image Generation: AI art based on training from many artists’ styles generally poses less risk than systems that can create images closely mimicking specific artists’ recognizable works.

Code Generation: Functional code generally receives thinner copyright protection than highly creative works, but verbatim reproduction of significant code segments could still raise infringement concerns.

Music Generation: Music involves multiple copyright elements (composition, lyrics, performance), and AI systems creating music “in the style of” specific artists navigate complex legal territory.

If I pay for access to an AI system, does that make my use “commercial”?

Not necessarily. While you’re engaging in a commercial transaction with the AI provider, whether your own use is “commercial” depends on how you use the outputs.

If you’re using the AI-generated content for personal projects, education, or non-commercial research, your use may be considered non-commercial despite paying for the AI service. However, if you’re using the outputs in products or services you sell, this would likely be considered commercial use.

The AI provider’s commercial status and your commercial status as a user are separate considerations in fair use analysis, though both may be relevant.

What happens if my AI use doesn’t qualify as fair use?

If your use doesn’t qualify as fair use, it could potentially constitute copyright infringement unless another exception applies. This could lead to:

Legal Liability: Potential lawsuits from copyright holders seeking damages, which can include statutory damages up to $150,000 per work for willful infringement Takedown Notices: Requests to remove infringing content from platforms or services Cease and Desist Letters: Demands to stop the infringing activity Licensing Requirements: Need to negotiate licenses for continued use

To avoid these outcomes, consider:

• Obtaining proper licenses when fair use doesn’t apply
• Using alternative content sources (public domain or openly licensed works)
• Redesigning your AI system to use less copyrighted material or use it more transformatively
• Consulting with an attorney to evaluate your specific situation

How does the recent Supreme Court decision in Andy Warhol Foundation v. Goldsmith affect AI fair use?

The 2023 Supreme Court decision in Andy Warhol Foundation v. Goldsmith narrowed the scope of transformative use in fair use analysis. The Court ruled that Warhol’s adaptation of Lynn Goldsmith’s photograph of Prince was not fair use, despite its artistic transformation, because it served the same commercial purpose as the original (licensing for publication).

This decision has significant implications for AI fair use:

It suggests courts should focus not just on whether a use transforms the original work’s appearance, but also whether it serves a fundamentally different purpose or function AI systems that transform works visually or stylistically but serve similar market functions to the originals may face greater fair use challenges The decision emphasizes the fourth fair use factor (market effect), suggesting AI uses that impact potential licensing markets will face heavier scrutiny

AI companies and users should carefully consider whether their transformations go beyond surface-level changes to serve genuinely new purposes that don’t compete with the original works’ markets.

How can I determine if my AI training dataset might raise copyright concerns?

Evaluating the copyright risk of your training dataset involves considering:

Content Sources: Data scraped from the internet generally includes copyrighted material, while public domain works, openly licensed content, or licensed datasets pose lower risk

Documentation: Well-documented datasets with clear sources and rights information reduce uncertainty and demonstrate due diligence

Opt-Out Mechanisms: Implementing ways for creators to exclude their works shows respect for copyright holders’ interests

Content Types: Datasets heavy on highly creative works (fiction, art, music) generally pose higher risk than those focused on factual content

Transformative Processing: How the data is processed, stored, and used affects fair use analysis—more transformation generally strengthens fair use arguments

Commercial Context: Commercial use of training data faces higher scrutiny than academic or research uses

For existing datasets, review documentation about their creation and composition. For custom datasets, consider implementing copyright filters and documentation processes. When possible, obtain licenses for high-risk content categories.

Can I be liable for copyright infringement if I use an AI tool that was trained on copyrighted materials?

This remains a developing area of law, but several principles are relevant:

End User vs. Developer Liability: Generally, liability for how an AI system was trained would more likely fall on the developers who conducted the training rather than end users

Output-Based Liability: However, if you prompt an AI system to create outputs that substantially reproduce copyrighted works, you might share liability for that specific output

Knowledge and Intent: Your awareness of potential infringement and your intentions in using the system could affect liability

Commercial Context: Commercial uses face higher scrutiny than personal or educational uses

Most typical uses of commercial AI tools likely pose minimal direct liability risk for end users regarding training data. However, using AI to deliberately circumvent copyright (such as requesting verbatim reproduction of books or articles) could potentially create liability.

If you’re concerned about potential liability, consider:

• Using AI tools from reputable companies with clear terms of service
• Being cautious about using AI to generate content very similar to specific copyrighted works
• Reviewing AI outputs before using them commercially
• Obtaining appropriate licenses when necessary

While legal questions around AI and copyright continue to evolve, understanding fair use principles and implementing thoughtful practices can help navigate this complex landscape. Use our AI Fair Use Analyzer as a starting point, but remember that professional legal advice tailored to your specific situation is invaluable for high-stakes decisions involving copyrighted materials and AI systems.