Does Your Use of AI comply with the Copyright Fair Use Rules?
When does generative AI qualify for “fair use” under the copyright infringement exemption rule?
The following article will explain some new keywords in the world of Artificial Intelligence (AI), such as:
- Generative AI (Generative AI, GenAI, or GAI) is a type of artificial intelligence that creates new content, such as text, images, and music, by learning patterns from vast datasets. It differs from traditional AI by generating original material rather than just analyzing or classifying existing data. Popular examples include conversational AI like ChatGPT and image generators like DALL-E, which produce novel outputs in response to user prompts.
- Model Context Protocol (MCP) server. The MCP server is what allows brokers, agents, and their technology partners to securely ingest MLS records into their own AI workflows. Without it, you end up with a dangerous patchwork of duplicated databases, scraped data, and uncontrolled integrations.
- Inputs and Outputs –– Input is data sent to a computer system, while output is the data that a system sends back after processing it. Input and output (I/O) are essential for communication between a system and the outside world, whether it’s a user or another computer.
- Retrieval Augmented Generation (RAG) — policies that focus on security, data privacy, and factual accuracy, requiring access controls, data masking, and grounding responses in authoritative sources to prevent “hallucinations.”
- Entropy is the lack of order or predictability; gradual decline into disorder.
- Prompts — Generative AI prompt examples include asking it to write a story, summarize a text, or create a marketing email. For images, prompts can describe a scene like a “futuristic city skyline at sunset.” Other examples include role-based prompts, such as acting as a personal trainer or a lawyer, and providing a few-shot example to show the AI the desired output format or style.
- API endpoint refers to a specific network address (URL) and a defined set of rules (API) that allows other applications or systems to interact with an AI model or service.
The Fair Use critical question
Let’s start with taking a deep dive into some of the points made by Suchir Balaji in his memorandum as he suggests answers to the Fair Use critical question today in a memorandum published October 23, 2024. Suchir Balaji, was a former OpenAI researcher, who had publicly raised concerns about OpenAI’s practices and was considered a whistleblower before his death.
https://suchir.net/fair_use.html?utm_source=chatgpt.com
While generative models rarely produce outputs that are substantially similar to any of their training inputs, the process of training a generative model involves making copies of copyrighted data. If these copies are unauthorized, this could potentially be considered copyright infringement, depending on whether or not the specific use of the model qualifies as “fair use.” Because fair use is determined on a case-by-case basis, no broad statement can be made about when generative AI qualifies for fair use. Instead, he provides a specific analysis for ChatGPT’s use of its training data, but the same basic template will also apply for many other generative AI products.
Fair use is defined in Section 107 of the Copyright Act of 1976
In determining whether the use made of a work in any particular case is a fair use, the factors to be considered shall include—
- the purpose and character of the use, including whether such use is of a commercial nature or is for nonprofit educational purposes;
- the nature of the copyrighted work;
- the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and
- the effect of the use upon the potential market for or value of the copyrighted work.
Suchir Balaji Comments on #1 – and #4 — Model developers like OpenAI and Google have also signed many data licensing agreements to train their models on copyrighted data — for example with Stack Overflow, Reddit, The Associated Press, News Corp, etc. It’s unclear why these agreements would be signed if training on this data was “fair use,” but that’s besides the point. Given the existence of a data licensing market, training on copyrighted data without a similar licensing agreement is also a type of market harm, because it deprives the copyright holder of a source of revenue.
The distinction between Substitution and Transformation — between substituting and non-substituting uses — is actually the origin of “fair use” from the 1841 case Folsom v. Marsh, in which the defendant copied parts of a biography of George Washington to make a version of their own.
In many recent cases, factor (1) has been considered in terms of “transformativeness” — for example the Second Circuit’s findings in Authors Guild. v. Google on Google Books that: Google’s unauthorized digitizing of copyright-protected works, creation of a search functionality, and display of snippets from those works are non-infringing fair uses. The purpose of the copying is highly transformative, the public display of text is limited, and the revelations do not provide a significant market substitute for the protected aspects of the originals.
In sum, the first fair use factor considers whether the use of a copyrighted work has a further purpose or different character, which is a matter of degree, and the degree of difference must be balanced against the commercial nature of the use. If an original work and a secondary use share the same or highly similar purposes, and the secondary use is of a commercial nature, the first factor is likely to weigh against fair use, absent some other justification for copying. In practice, it’s hard to reason about the “purpose” of a product as broad as ChatGPT, or the “purpose” of the entire internet.
A better framing is: do the market harms from ChatGPT come from it producing substitutes that compete with the originals? Or is it an indirect effect, like what a book critic has on a book?
Suchir Balaji thinks it’s pretty obvious that the market harms from ChatGPT mostly come from it producing substitutes.
Factor (2): the nature of the copyrighted work
;The nature of copyrighted work — whether it’s a creative work that’s highly protected by copyright, or a factual work that’s mildly protected by copyright — will vary a lot on the internet. But most data on the internet is protected by copyright to some degree, so it’s unlikely that factor (2) will strongly support “fair use.” In practice, this factor tends to be the least important anyway.
There’s two interpretations of factor (3):
- The inputs of the model are full copies of copyrighted data, so the “amount used” is the entirety of the copyrighted work. This would weigh against “fair use.”
- The outputs of the model are almost never copies of copyrighted data, so the “amount used” is almost zero. This could potentially weigh in favor of “fair use.”
But interpretation (2) isn’t quite correct, because the purpose of copyright isn’t to protect the exact works produced by an author (otherwise, it’d be trivial to bypass by making small tweaks to a copyrighted work). What copyright really protects are the creative choices made by an author. Collage art is a simple example of this distinction — a collage artist won’t gain copyright protection for the underlying works they use, but they will gain copyright protection for the creative choices they made in the arrangement of those works. This is similar to the MLS assertion that they own the compilation of the listing data.
After the model trains on each datapoint once, its completions to the prompt “First Citizen” are high entropy and novel, although incoherent. But after training on each datapoint ten times, it ends up memorizing the beginning of a Shakespeare play and regurgitating it when prompted, thus repeating (information) without analyzing or comprehending it. At five repetitions, the model does something in between regurgitation and creative generation — some parts of its output are novel, some are memorized, and the two are mixed together in its output.
The main reason ChatGPT produces low-entropy outputs is because it is “post-trained” using reinforcement learning — in particular reinforcement learning from human feedback (RLHF).
None of the four factors seem to weigh in favor of ChatGPT being a fair use of its training data. That being said, none of the arguments here are fundamentally specific to ChatGPT either, and similar arguments could be made for many generative AI products in a wide variety of domains.
Another look at Generative AI
Generative AI is a type of artificial intelligence that creates new content, such as text, images, and music, by learning patterns from vast datasets. It differs from traditional AI by generating original material rather than just analyzing or classifying existing data. Popular examples include conversational AI like ChatGPT and image generators like DALL-E, which produce novel outputs in response to user prompts.
How it works
Training: Generative AI models are trained on massive amounts of existing data, like text, images, and audio.
Pattern recognition: During training, they learn complex patterns, structures, and relationships within the data.
Content creation: The model uses these learned patterns to generate new, original content that resembles the training data when given a prompt or input.
Applications and examples
Text generation: Creating stories, essays, articles, and code.
Image generation: Producing original images from text descriptions.
Audio and music: Composing music or generating realistic speech.
Video: Creating video clips from simple prompts.
Other uses: Simulating disease progression for medical research or generating synthetic medical data.
Key characteristics
Generates new content: It produces novel data that is often indistinguishable from human-created content.
Mimics human creativity: It can generate content in the style of a particular author, artist, or composer.
Relies on prompts: It generates outputs based on instructions or prompts provided by a user.
Learns from data: Unlike rule-based systems, generative AI learns and improves from the data it is trained on.
As defined in https://en.wikipedia.org/wiki/Generative_artificial_intelligence
Generative artificial intelligence is a subfield of artificial intelligence that uses generative models to produce text, images, videos, audio, software code or other forms of data. These models learn the underlying patterns and structures of their training data and use them to produce new data based on the input, which often comes in the form of natural language prompts.
Generative AI has raised many ethical questions and governance challenges as it can be used for cybercrime, or to deceive or manipulate people through fake news or deepfakes. Even if used ethically, it may lead to mass replacement of human jobs. The tools themselves have been criticized as violating intellectual property laws, since they are trained on copyrighted works. The material and energy intensity of the AI systems has raised concerns about the environmental impact of AI, especially in light of the challenges created by the energy transition.
MCP Servers (Model Context Protocol)
Now let’s focus on MCP Servers (Model Context Protocol) – in this section, we’ll draw from numerous articles written by Victor Lund, founder of the WAV Group and a recognized thought leader on AI applications in the real estate industry and MLSs. Disclosure: while the author shares education materials with Victor as a fellow real estate educator, we have no business relationship.
Some of Victor Lund’s related writings are:
The conversation around artificial intelligence in real estate has moved beyond speculation. MLSs are now facing practical questions: Which AI features should we deploy? How do we protect our data? And what infrastructure decisions will define our role in an AI-driven market? To explore these questions, we sat down in a video interview with Lucie Fortier and Jeff Fullbright from Intercontinental Exchange (ICE), the company behind Paragon® Connect MLS, a platform serving MLSs across the country. Our video discussion covered everything from agent-facing AI tools to emerging data access models like MCP servers, and what is “AI sovereignty.”
https://www.wavgroup.com/2025/08/28/the-compass-ai-demo-and-fluentes-path-forward-for-brokers/ – Compass has been showcasing a conversational AI demo that turns everyday agent tasks into a seamless, voice-driven workflow. Every broker should aim to stay ahead or keep up with Compass. This means owning your AI platform, and owning your data. We call this AI Soverenty and it’s the keynote of how WAV Group’s AI Division – Fluente – is coaching and building AI for brokers. When you work with us, you own your AI, not rent it.
Victor Lund points out that “the future of brokerage AI depends on seamless connections between CRMs, MLS systems, productivity suites, and third-party data services. That orchestration happens through MCP servers, a middleware orchestrator that translates a conversational request into discrete jobs routed to the right endpoints and specialty AI agents.”
Also: https://www.wavgroup.com/2025/09/16/utahrealestate-com-to-launch-an-mcp-server/ — To our knowledge, Utah Real Estate.com is the first MLS in the nation to launch a Model Context Protocol (MCP) server. The MCP server is what allows brokers, agents, and their technology partners to securely ingest MLS records into their own AI workflows using Retrieval Augmented Generation (RAG). Without it, we end up with a dangerous patchwork of duplicated databases, scraped data, and uncontrolled integrations. That’s not only inefficient, it puts data accuracy, compliance, and consumer trust at risk.
The real estate industry is consolidating. Brokerages are national. MLSs are regional. Now is the time to act like we’re on the same team. One of the most consistent points of tension among brokers is the IDX program. It may be time to reset the terms.
Picture this: What if brokers only advertise their own listings. When a consumer becomes a client through a signed buyer agency agreement, they are invited into a private portal. The public side of listings is handled only by Fair Display Guideline-compliant websites like HAR.com, Cribio.com, and UtahRealEstate.com. These platforms are governed by rules that prioritize broker control and consumer protection.
A call to MLS leaders: don’t wait. If your MLS is not actively developing an MCP server, now is the time to start. If your vendor does not offer one, press them for a roadmap. Because if the MLS community doesn’t step up, others will fill the void. And the role of the MLS will shrink along with it.
Every wave of technology gives us a chance to rethink how we work. MCP is that wave. MLSs need to lead. Brokerages need to act. And the industry must protect the very thing that makes it work. Broker cooperation is rare. It is valuable. It is worth defending. Let’s not lose it to an API endpoint.
https://www.wavgroup.com/product/why-mlss-need-mcp-servers/ —
The Biggest Shift in MLS Infrastructure Since IDX is just beginning—Model Context Protocol (MCP) servers are emerging as the new standard for powering AI-driven real estate tools. While some brokerages and technology companies are already exploring independent solutions, MLSs have a unique opportunity to lead this evolution by providing AI-ready access without giving up control.
This WAV Group report explains how MLSs can implement MCP servers to stay at the center of innovation, while maintaining the governance, compliance, and data protections that define their value.
RELATED COPYRIGHT ARTICLES
Title: Copyright infringement suit by CoStar Group, Inc. against Zillow.
https://thedataadvocate.com/real-estate-lawsuits-and-trends/
- In July 2025, CoStar sued Zillow in federal court (Southern District of New York) alleging that Zillow used tens of thousands of CoStar’s copyrighted (and watermarked) property photos without license, distributed them (including via syndication to other portals), and seeks damages potentially exceeding $1 billion. Zillow syndicated those property photos, although they have since removed them. Status: Litigation ongoing.
https://www.nar.realtor/copyright — with links to numerous helpful articles on topic
Improper use of listing content can create legal problems for agents, brokerages and MLSs. It’s crucial that real estate professionals know their rights regarding listing photos and listing content, as well as risk management strategies that can be used to avoid copyright infringement. Complying with the Digital Millennium Copyright Act (DMCA) will also protect website owners from a third party’s misuse of photographs within an IDX feed.
AUTHOR BACKGROUND:
John Reilly’s interest in the study of Copyright law began when he, as a young Honoulu real estate attorney and educator, self-published (1975) the 462 page The Language of Real Estate in Hawaii that started out as a 40-page glossary handout in a real estate licensing course. Back in the day (no internet, no computer skill), he used a ton of legal notepads and #2 lead pencils that was then typed into manuscript format by his girlfriend (now wife Patty). A few years later, he hooked up with the publisher Dearborn for the national version..
Dearborn recently informed John that the Ninth Edition of The Language of Real Estate will be celebrating its 50th anniversary next year. Sure did help funding the education of children Kristine, Sean Patrick, and Jennifer. Oh, by the way, Patty and John will be celebrating their 50th wedding anniversary in March 2027.
Now that you’ve read about the current status of the FAIR USE doctrine, you’ve probably noticed that I’ve used ChatGPT to write this article. Did I cross the line from Transformative into Substitution? That evaluation can be relevant to how YOU must be sensitive in your use of AI in today’s landscape. Finally, I have to confess that, although I heard about it, I never used the power of ChatGPT until last month when I yielded to the insistence of my fellow content creators as we did course work on an alternative to buyer broker compensation under the performance-based model of the NEGOTIATOR®system.
My reaction to the mind-blowing experience of ChatGPT was WOW! Where was ChatGPT through all my educator years!
Aloha
John
—
6 Responses
jOHN, Thanks for sharing thIS article on the intersection of AI use and copyright fair-use rules. I found it thoughtfully constructed and very timely given how rapidly the technology is changing in our industry.
I wanted to offer a supportive response—while also leaning into the key question you flagged: Did I cross the line from “transformative” into “substitution”?
What I found strong in the article
The piece clearly lays out the four fair use factors under Section 107 (purpose/character, nature of the work, amount/substantiality, effect on the market) and applies them to generative AI in a way that makes you pause and think.
The distinction the author draws between “transformation” (adding new purpose, meaning, or context) versus “substitution” (doing essentially what the original work did) is especially helpful.
The real‐world implications for our field (listing data, MLS content, marketing imagery) are well highlighted, and I appreciated the connection to infrastructure questions around data governance.
My personal reflection: Did I cross the line?
Since I’m working with AI in my own writing and marketing workflows, this article made me ask myself: Am I truly transforming the input, or am I leaning on it so heavily that I’m effectively substituting for the original? Here’s how I’m thinking about it:
Purpose/Character: My use of AI isn’t just to duplicate what already exists—it’s to generate new insights, frameworks, outlines, readability tools, and risk identifiers tailored to my book project. That suggests a transformative intent.
Nature of the Work: If the content I’m using or referencing comes from highly protected creative works (e.g., novels, full articles) then the bar is higher. If I’m working primarily with factual, public‐domain style material (e.g., writing about real-estate markets, processes) then there’s more leeway.
Amount/Substantiality: I’m careful not to rely on large fragments of any single work. I use AI to distill, reframe, ideate—not to republish or paraphrase large chunks of someone else’s text.
Effect on the Market: The big question: Could my output compete with the original work in a way that harms the copyright holder’s market? Since my work is original (my own book) and targeted at a niche real estate professional/educator audience, I believe the risk is lower—but it’s still something I monitor.
Based on this self‐audit, I believe I’ve remained on the “transformative” side—but only just. The line isn’t bright, and it’s helpful to have a framework like the article’s to check myself regularly.
Where I’ll tighten up going forward
Because I want to be confident I’m staying on the right side of the line, here are a few steps I’ve resolved to take:
Document each instance where I use AI on input material: what was the source, how much of it was used, and how I transformed it.
For any input that comes from a copyrighted work, I’ll ensure I’m using excerpts small enough, sufficiently revised, and clearly used for new purpose rather than replication.
Include a disclaimer in my workflow: “AI assisted; content generated/modified for new purpose.” That transparency supports the “purpose and character” factor.
Periodically ask: *Could someone look at this output and say: ‘That looks just like the original’?’ If the answer is “yes” then I’ll revise further.
Thanks for your input, Fran. Your thoughtful comments are solid enhancements to the article.
I fired up ChatGPT as a follow-up to our discussion here and asked it to create a Fair-Use Self-Audit Checklist … Here’s what it delivered (verbatim):
AI Fair-Use Self-Audit Checklist
A quick test to confirm you’re transforming—not substituting.
1. Purpose & Character (The “Why” Test)
* Ask yourself:
* Am I using this source to add new meaning, analysis, commentary, or investigative framing?
* Does the AI output reflect my own voice, judgment, narrative structure, or investigative approach?
Would a reader say, “This is Frances interpreting facts,” rather than “This looks like the original”?
If the answer is yes → you’re on the transformative side.
2. Nature of the Source (The “What” Test)
Check:
* Is the source largely factual (legal filings, public records, historical dates, MLS rules)?
* Or is it highly creative (unique phrasing, expressive commentary, someone’s polished prose)?
Facts = safer.
Creative expression = requires far more transformation.
3. Amount & Substantiality (The “How Much” Test)
Ask:
* Did I use only small, necessary portions of the original?
* Did I avoid copying the “heart” (i.e., the most distinctive or expressive sections)?
* Did the AI output move sufficiently far away from the source’s language and structure?
If you wouldn’t recognize the original in the AI output → safe.
If you can recognize it → revise further.
4. Effect on the Market (The “Could This Replace the Original?” Test)
Ask:
* Could my AI-assisted content serve as a substitute for the original work?
* Is my work aimed at a different purpose, audience, or market?
* Does my use harm, diminish, or compete with the original creator’s ability to sell their work?
If your book doesn’t replace or replicate the original—and is clearly your own investigative product—this factor works strongly in your favor.
BONUS: Quick Yes/No “Transformative vs. Substitution” Test
TRANSFORMATIVE if:
* You use the source to generate new insights, analysis, narrative, or commentary.
* You add substantial original thinking, structure, or interpretation.
* AI helps you reframe or organize—not replicate.
SUBSTITUTION if:
* The AI output mirrors the source’s structure or phrasing too closely.
* A reader could confuse the new text with the original.
* AI is used to reproduce rather than reinterpret.
What to Document (Minimal Paper Trail)
Just keep quick notes:
* Source used
* Purpose of use
* Portion used
* How the final version differs
* That you revised for original voice and context
This protects both your integrity and your insurer.
Fantastic article John and great feedback Frances. I believe this is a very important topic – especially in leu of the California law that requires that listing firms take care to watermark images that are AI altered and load a copy of the original into the MLS. It’s hard to know when the authorship of new creativity ends and when AI is the creator of something. Today, I think that AI is more like a paintbrush than a painter, more like a pen than an author, more like software that helps designers than a designer. Without the human prompt and revisions, and the context that humans provide to train and deprive AI outputs, those outputs would not exit today. But tomorrow could be different. For example, I can easily develop AI that composes community articles and publishes them to a broker’s website. The human gives the AI the job and sets the boundaries for the content – even tests the machine and connective tissue between the AI and the publishing platform, but once automated – the machine does a job as instructed. We can even design feedback loops that train the AI to get better and better with each iteration.
Appreciate your comments, Victor. I agree with your vision that “We can even design feedback loops that train the AI to get better and better with each iteration.”
From ChatGPT:
Authors Guild v. Google (Google Books) — Summary
Background
In 2004, Google launched the Google Books project, partnering with major libraries to digitize millions of books—including books still under copyright.
Google displayed “snippets” (a few lines of text) in search results so users could discover relevant books.
The Authors Guild sued Google in 2005, arguing that scanning and displaying snippets of copyrighted works without permission was copyright infringement.
This decision is considered one of the most important modern fair-use rulings. It solidified the principle that:
Mass digitization + search indexing = fair use, if not replacing the work.
It laid groundwork for later cases involving:
AI training on copyrighted data
Text and data mining
Internet search engines
Online content indexing (e.g., HathiTrust case)
Courts often cite it in contemporary AI litigation as an example where intermediate copying with a transformative informational purpose qualifies for fair use.