Almost every copyright question lands in one of three places: what you put in and what your tool is built on, and what comes out as outputs. AASHTO's online store, for example, now prohibits entering any AASHTO content into any AI tool. A subscription that lets you read the Green Book does not let you build a tool on it. Pasting or uploading may involve copying other uses beyond a reading license. Check the applicable publisher’s terms before entering a standard into an AI. Standards adopted into law by reference are still copyrighted. Courts have allowed them to be posted so the public can read the law, not so an organization or user can build an AI workflow on them.
The risk depends on how the material is used as explained below.
| Use | Example | Risk |
|---|---|---|
| One-off prompt or upload | Paste a section, ask for a summary | Lowest. Can still exceed your license. |
| Custom GPT, agent, or similar (a tool that answers or acts from documents you loaded into it) | A manual the tool draws on every time it runs | A separate, ongoing use. Needs rights that cover AI. |
| Model training or fine-tuning | The material is absorbed into the model | Highest. Needs explicit permission. |
You must also consider your own organizations data and its copyright protection and intellectual property (IP). Consumer accounts may train on your inputs and uploads. Enterprise tiers may offer different training commitments, retention terms, and controls. Confirm the terms and approved configuration before uploading material. Ask yourself the following questions:
| Question | Verdict |
|---|---|
| Is the material yours, your agency's, public domain (e.g. MUTCD), or openly licensed? | Proceed. Check for third-party content inside the material. |
| Is it third-party copyrighted, including a licensed standard? | Confirm the license covers this AI use. If it doesn't, stop. |
| Are you building a tool on, or training a model on, third-party content? | Needs review. If no permission covering AI use, then stop. |
| Are you adopting a vendor's tool? | Ask where its data comes from, whether it holds licenses for the standards it relies on, whether it can show its sources, and whether it will warrant non-infringement and indemnify you. A tool built on unlicensed standards hands its risk to you. Without answers, do not adopt. |
The rest of the section works through the details. It covers key terms, what happens to your own data, what you can put into a tool, building on or buying tools trained on copyrighted content, and an approval workflow, closing with an FAQ.
Purpose: the three IP touchpoints
Almost every IP question lands at one of these three touchpoints (WIPO 2024, 4–5):
Three Touchpoints
- In
Your Inputs
What you put in
- On
The Model or Custom GPT
What the tool is built or customized on
- Out
Your Outputs
What comes out
Parts of this area are settled law; others are still working through the courts.
IP glossary
Training
Building a base model by feeding it large amounts of material so it learns patterns. This is done by the frontier AI labs (OpenAI, Anthropic, Google, and the like), not by your organization.
Fine-tuning
Extra training of an existing model on a narrower set of material to specialize it. Unlike base-model training, this is something your organization might do when it builds a custom tool, and where the copyright question lands on you.
RAG (Retrieval-Augmented Generation)
Connecting a tool to a set of documents it looks things up in when it answers. Treated separately from training, with its own copying of the source.
Custom GPT
A tool you set up on your own uploaded documents or instructions. Equivalents include Microsoft Copilot agents and Copilot Studio, Claude Projects, and Gemini Gems. Custom GPTs typically rely on RAG. "Custom GPT" is used throughout as shorthand for this whole group, including agents built the same way.
Agent
A tool that does things rather than just answering: sends an email, files a form, queries a database. Agents can also pull from the web while they answer, so content nobody vetted can land in the output. Same IP questions as any tool, except the output can be an action. A Custom GPT can also be built as an agent.
Model vs. Tool
A model is the underlying AI system that generates a response. A tool is the application you use to interact with a model, such as ChatGPT, Microsoft Copilot, or a Custom GPT. A tool may also connect the model to uploaded documents, other data sources, or outside services. Check the terms and settings for the tool you use; knowing which model powers it does not tell you how your inputs are handled.
Input / Output
Input is what you put into the tool. Output is what it produces.
What happens to your data
A lot of the worry around AI and copyright comes from lumping together very different actions that all get called “using AI.” The three actions below differ, and each carries a different amount of risk (WIPO 2024, 7–8).
- A one-off prompt or upload in a chat session. You paste in a paragraph or attach a file and ask the tool to summarize or rewrite it.
- Building a Custom GPT, a knowledge upload, or a RAG tool. You load documents into a store the tool draws on again and again. That’s an ongoing use, and it counts as its own act, separate from training a model (U.S. Copyright Office 2025b, 30).
- Training or fine-tuning a model. The copyrighted material gets absorbed into the model itself.
Each use raises different rights and data-handling questions. Loading a document into a tool's knowledge base creates a stored copy that the tool draws on every time it runs, and copyright law treats that as its own act of copying, separate from training a model. Saying a tool was never trained on the material is therefore not an answer, because the stored copy needs permission of its own (U.S. Copyright Office 2025b, 28–29). A model or retrieval store can also reproduce protected text in its answers, so risk can surface at the output stage, not just at the input stage (NIST 2024b, 11).
How your own data gets used
The flip side is whether the tool keeps or learns from your own copyrighted work. Know which plan you're on and ask: how is our data used, and can we turn off training? Consumer tiers, free or personal-paid, may use your inputs to improve the model. Enterprise and business tiers usually don't and usually turn data training off by default, which can be confirmed in the settings and configuration for training, retention, access, and deletion. Procurement terms and the enterprise or government configuration often lock those settings in anyway, so confirm them in the contract rather than the app (WIPO 2024, 6).
Inputs: putting copyrighted material into AI tools
The most common day-to-day risk is feeding copyrighted third-party material into an AI tool, especially licensed engineering standards. A subscription that lets you read and use a standard (AASHTO, ASCE, ASTM, ITE, etc.) almost never includes the right to feed it into AI training or grounding. Those are separate uses, and the license usually doesn’t cover them (U.S. Copyright Office 2025b, 26–27). Always check the license and obtain any necessary permission before proceeding.
Some standards bodies say this outright: AASHTO’s online store states that “In order to protect the intellectual property rights of AASHTO to its content, AASHTO prohibits the use of any AASHTO content in conjunction with an artificial intelligence tool or program, including the training of models on AASHTO content or the entry of AASHTO content into any AI tool.” A subscription that lets you read the Green Book does not allow for any AI uses (AASHTO 2026).
Standards written into law are still copyrighted. Some standards get adopted into law by reference, like an AASHTO or ASTM specification. Courts have allowed those standards to be posted for free public access because people have a right to read the law (ASTM v. Public.Resource.Org 2023), but that ruling doesn’t establish a general right to upload them into an AI tool or any commercial workflow.
Lower-risk inputs: public-domain material like federal government works (the MUTCD is an FHWA publication), openly licensed material, and your own or your agency's work product may present fewer copyright concerns. Even then, check for third-party rights buried inside.
Where firm and agency guidance differs
- Firms: be careful with client-confidential and third party-licensed inputs. When in doubt, ask for permission prior to use, and confirm any AI use stays inside the input’s license before you upload it.
- Agencies: agency inputs can be subject to public records requests, and procurement or contract terms may limit what you can feed in. What you put into a tool may be subject to applicable public-records and records-retention requirements (WIPO 2024, 7–8).
Models and Custom GPTs: building on copyrighted content, and vetting vendors who do
Know what the sources are, and confirm you have rights that cover AI use.
Building on copyrighted standards is a separate use. Standing up a Custom GPT, agent, RAG system, or fine-tune on copyrighted specs is a separate use that an ordinary reading license almost never covers. Doing it without permission is an infringement risk (U.S. Copyright Office 2025b, 27).
A more controlled path is trusted, licensed data instead of scraped data. OpenEvidence, a medical AI startup tool, limits its data to "trusted medical sources" and reports formal partnerships with the New England Journal of Medicine and the American Medical Association (OpenEvidence 2026). It hit roughly a $12B valuation in January 2026 and is used daily by more than 40% of U.S. physicians across over 10,000 hospitals (Sriram 2026). Licensed content gives better accuracy and traceable citations, and it avoids infringement exposure. No transportation equivalent has been found, so for most organizations this is a reason to get AI-use rights for your own sources, or to favor vendors who license theirs.
Vendor due diligence: questions to ask any AI vendor
Most AI vendors won’t tell you where their training or grounding data comes from, and most model developers don’t disclose their data sources at all (NIST 2024b, 7). An organization that adopts a tool built on unlicensed copyrighted standards, say a vendor that quietly grounded its tool on the AASHTO Green Book or ITE resources, can inherit infringement and reliability problems. Most buyers can’t evaluate the technology themselves, so the protection is knowing what to ask. These questions apply to every vendor, including licensed-partnership models (WIPO 2024, 7, 14):
Questions to ask an AI vendor
NIST’s Generative AI Profile recommends supplier risk-assessment frameworks, procurement due diligence that covers intellectual property, and approved-tool lists before adoption (NIST 2024b, 20–21).
Approval workflow: the “Can I use this?” decision tree
Run this quick check using the decision tree below before any AI use that touches copyrighted material. It’s based on NIST’s map-and-manage approach which supports sending uncertain uses to review rather than either blocking them outright or waving them through (NIST 2024b, 20–21).
Text version of the decision tree
Start: what does your AI use involve?
Branch 1 · Inputs (what you're putting in)
- Is the material yours, your agency's, public-domain, or openly licensed? Yes → Proceed. Allowed. Still confirm no third-party rights are tucked inside.
- No → Is it third-party copyrighted material, including a licensed standard? Does your license expressly cover this AI use?
- Yes → Proceed within the license terms.
- No / unsure → Stop. Do not proceed without AI-use rights.
Branch 2 · Model / Custom GPT (what you're building on)
- Are you training, fine-tuning, or grounding a tool on third-party copyrighted content? Yes → Review. Needs review. Do not proceed without a license or permission that covers AI use.
- No → Are you adopting a third-party AI tool?
- Yes → Review. Run vendor vetting. If their training or grounding data isn't licensed, or they can't show it, do not proceed without answers.
- No → Proceed under your organization's standard AI policy.
FAQ
Can I upload a copyrighted standard to an LLM just to summarize it?
Typically no. Uploading is itself a use, and it can go past your reading license even for a quick summary. A license to read a standard rarely covers feeding it to an AI tool. Check the terms.
Can we build a “Custom GPT” on our AASHTO or ITE manuals and libraries for example?
Not without rights that cover AI use. Owning or subscribing to those standards isn’t permission to load them into a Custom GPT or RAG tool. That’s a separate use the license almost never includes, and AASHTO’s store, for example, now explicitly prohibits AI use of its content outright.
How do I vet an AI vendor, and are we on the hook if they used copyrighted standards without a license?
Most developers don’t disclose their data sources, so you have to ask: where their data came from, licensing, citations, infringement warranties, indemnity, and data retention. It matters because you can inherit infringement and reliability problems you can’t see, which is why warranties and indemnity are especially important before you adopt. A longer question list is provided above.
Can we copyright a report our AI helped write?
Often yes, to the extent the report contains sufficient original human-authored expression. The Copyright Office’s position is that copyright protects original human expression even in a work that also contains AI-generated material, but not purely AI-generated material, or material where there is insufficient human control over the expressive elements. Whether your contributions rise to authorship is a case-by-case question. Document which specific parts were the human contribution (U.S. Copyright Office 2025a, iii).
What does a public agency need to do differently?
Assume inputs and outputs may be public records. Route agency Custom GPTs and agents through data-governance review and add the vendor questions to your solicitation language.
Does our uploaded data get used to train the vendor’s model?
Depends on your tier and settings. Consumer tiers, free or personal-paid, may use inputs to improve the model. Enterprise tiers usually don't, and usually let you switch training off. This is why falling back on a personal account matters: a licensed standard or client-confidential material uploaded there can end up in a tool that keeps it.
Sources cited on this page are listed in full on the References page.