Copyright — Module 20.1

How This Lesson Fits the Module & Volume

After AI safety and privacy, engineers still ship models trained on text, code, and media they did not author. Copyright in this module is a product-risk literacy lecture: training data, generated output, and licensing trade-offs. It is educational, not legal advice. It does not tell you to copy protected works, scrape in violation of terms, or “get around” licenses.

Document licenses on the data card. Escalate real disputes to counsel via governance and compliance. Next lectures shift to adversarial integrity: prompt injection and beyond.

Learning Objectives

By the end of this lesson, students should be able to:

Separate three copyright-relevant layers: training/RAG corpus rights, generated output, and user-upload / vendor ToS constraints.
Compare licensing postures (owned, openly licensed, vendor ToS, unknown web-scale) as engineering trade-offs.
State why near-verbatim reproduction is a product risk without giving copy instructions.
Put license and attribution fields on a data card / manifest.
Know what this lecture is not: legal advice, fair-use determinations, or how-to infringement.
Link copyright constraints to RAG corpora, fine-tunes, and user-uploaded files.

Definition

Copyright (as used here) is the exclusive-rights regime that typically covers original works—text, code, images, audio, video—and restricts copying, distribution, and some derivative uses unless a license, exception, or owner permission applies. For AI engineers, three questions dominate: (1) May we use this corpus to train or retrieve? (2) Who owns or may use model outputs? (3) What must we disclose and retain as provenance? Laws and cases differ by country and keep moving. This lecture does not decide fair use, ownership, or infringement for your product. When unsure, do not copy; ask counsel.

Three Layers Engineers Actually Touch

Layer	Typical question	Engineer lever
Training / fine-tune data	Is this corpus licensed, owned, or unknown?	Prefer owned or clearly licensed sets; record provenance
RAG / eval corpora	May we index and quote this to users?	Index rights + snippet policy + ACLs (privacy)
Generated output	Can output be too close to a source? Who may commercially use it?	Similarity review, style constraints, vendor ToS, user terms
User-uploaded files	Did the user have rights to upload?	ToS, filters, do not train unless permitted
Third-party APIs	What does the provider ToS allow for outputs and logs?	Read ToS; no-train flags; pass constraints to customers

Open-source software licenses (MIT, Apache, GPL family) are not automatically the same as licenses on books, news, or images inside a scrape. A permissive code license can still sit next to copyrighted comments, assets, or data files. Treat “it was on the internet” as a risk flag, not a permission slip.

Licensing Postures (Trade-offs, Not a Verdict)

Owned / commissioned

Highest clarity for training and RAG.
Costly; coverage may be narrow.
Still respect employee/contractor agreements.

Clearly licensed open data

Follow the license (attribution, share-alike, non-commercial).
Some licenses forbid training or require downstream terms.
Put obligations on the data card.

Vendor model + your data

ToS governs output use and logging.
Your uploads can create dual risk (your rights + provider).
Do not assume you own everything the model emits.

Disciplined licensing buys

Fewer surprise takedowns and contract failures
Honest procurement answers
Cleaner fine-tune and RAG stories

“Ship first, ask later” costs

Unclear commercial rights on outputs
Customers inherit your corpus risk
Pressure to reproduce famous works (do not)

Manifest + Product Guardrails (Educational)

The code below tracks licenses you already have and refuses to treat “unknown” as approved. It is not a scraper, not a bypass, and not a how-to for reproducing protected works. Near-duplicate checks, if you add them later, are for reducing accidental regurgitation in your own product—not for extracting copyrighted text.

# Educational license manifest — not legal advice, not a scrape tool.
# Do not use this pattern to copy or harvest protected works.

APPROVED = {"owner-internal", "cc-by-4.0", "apache-2.0", "mit"}
NEEDS_REVIEW = {"unknown", "tos-unclear", "mixed"}

corpus = [
    {"id": "kb-internal", "license": "owner-internal", "may_train": True, "may_rag": True},
    {"id": "docs-cc-by", "license": "cc-by-4.0", "may_train": True, "may_rag": True, "attribution": "Required"},
    {"id": "web-unknown", "license": "unknown", "may_train": False, "may_rag": False},
]

def gate(item, action: str) -> str:
    lic = item["license"]
    if lic in NEEDS_REVIEW or not item.get(f"may_{action}", False):
        return "block_pending_counsel"
    if lic == "cc-by-4.0":
        return "allow_with_attribution"
    return "allow"

print([(c["id"], gate(c, "train"), gate(c, "rag")) for c in corpus])

# Product policy stub (customer-facing, high level):
POLICY = """
We do not use this product to copy or redistribute third-party books, films, or paywalled
articles. User uploads remain subject to your rights and our terms. Model outputs may be
inaccurate and are not guaranteed to be free of third-party claims. This is not legal advice.
"""
print(POLICY.strip())

# Data-card fields to fill (transparency.html):
# provenance, license, attribution_required, training_allowed, rag_allowed, owner, review_date

Related Lectures

Lecture	Role
Transparency	License and provenance on data/model cards
Privacy	User uploads can be both personal and copyrighted
AI Safety	Misuse includes asking the model to reproduce protected catalogs
Governance / Compliance	When to involve counsel; record decisions
Responsible AI	Org norms beyond the letter of a license
Knowledge base (Vol. 14)	RAG corpora need index rights, not only embeddings

Common Misconception

“If it is online, we can train.” Public availability ≠ a license. Second: this lecture is a substitute for a lawyer. Third: open-source software licenses automatically cover all nearby media. Fourth: the user owns every model output in every jurisdiction. Fifth: asking the model to reproduce a copyrighted book “for learning” is in-scope for this curriculum—it is not; do not do that. Sixth: ignoring vendor ToS because “we only call the API.”

Knowledge Check

Short Answer: Name the three copyright-relevant layers for AI engineers in this lecture. Answer: Training/fine-tune data (and RAG corpora), generated output, and (related) user uploads / vendor ToS constraints.
True/False: This lecture is legal advice that determines fair use for your company. Answer: False—educational only.
Multiple Choice: “It was on the public web” means: (a) a risk flag, not automatic permission, (b) a full license, (c) GPL for all images. Answer: (a).
Short Answer: Where should license and provenance be recorded? Answer: On the data card / license manifest (transparency artifacts).
True/False: Open-source code licenses always cover every book and image in a scrape. Answer: False.
Multiple Choice: Unknown license in the manifest should: (a) block pending review, (b) silently train, (c) be published as a demographic study. Answer: (a).
Short Answer: Why do vendor ToS matter if you do not train your own model? Answer: They govern output use, logging, and what you may pass to customers.
True/False: Students should use this page as a guide to copy protected works into a dataset. Answer: False—do not copy protected works.
Multiple Choice: User-uploaded files add risk because: (a) the uploader may lack rights, (b) UTF-8 is illegal, (c) batch size is copyrighted. Answer: (a).
Short Answer: Which sibling lecture comes next in this module? Answer: Prompt injection.

Key Takeaways

Engineers juggle corpus rights, output use, user uploads, and vendor ToS—not a single slogan.
Prefer owned or clearly licensed data; treat unknown web-scale corpora as unapproved until reviewed.
This is not legal advice and not a license to copy protected works.
Put provenance on the data card; escalate disputes through governance.
Next: Prompt Injection (adversarial control of models and tools).

Trainer’s Guide

Lab: Students complete a license manifest for three fictional sources (internal wiki, CC-BY docs, “unknown web dump”). Only the first two may train/RAG. Write a four-line customer notice. No scraping exercises.

Whiteboard: Three boxes—Train / Retrieve / Generate—each with a license question. Arrow to counsel when status is unknown. Remind: curriculum ≠ permission to reproduce books or media.

Recap: Copyright literacy for AI products is about licensed corpora, cautious outputs, and honest cards—not legal conclusions and not copying protected works. Continue to Prompt Injection.