Generative AI

Ethics Law And Trust

Copyright and Training Data

Discussion of "AI and copyright" usually collapses two independent questions. Keeping them apart is the single most clarifying move available.

JrCodex·8 min read

Jr Codex Generative AI Notes

Level: Intermediate Prerequisites: Module 1, Chapter 1 Time to complete: ~20 minutes

Table of Contents

  1. Two Separate Questions
  2. Question One: Was Training Lawful?
  3. Question Two: Does the Output Infringe?
  4. Memorisation and Regurgitation
  5. Who Owns the Output?
  6. A Practical Posture
  7. Summary & Next Steps

1. Two Separate Questions

Discussion of "AI and copyright" usually collapses two independent questions. Keeping them apart is the single most clarifying move available.

The Two Questions
─────────────────────────────────────────
  1. TRAINING
     Was it lawful to COPY copyrighted works in order
     to train the model?
     ── a question about the model's creator

  2. OUTPUT
     Does a specific generated piece infringe a
     specific existing work?
     ── a question about YOU, the user
─────────────────────────────────────────

Why the Split Matters to You
─────────────────────────────────────────
  Question 1 is being decided in courts and
  legislatures. You cannot influence it and mostly
  cannot resolve it — you can only choose providers
  and keep their terms.

  Question 2 is about the specific image or text YOU
  publish, and it is entirely within your control.

  Most practical risk for a builder sits in question 2.
─────────────────────────────────────────

A necessary note: this chapter is a technical practitioner's orientation, not legal advice. It describes the shape of the questions and standard industry practice. Anything commercial and consequential needs a lawyer in your jurisdiction.

2. Question One: Was Training Lawful?

The Core Dispute
─────────────────────────────────────────
  Training requires making COPIES of works —
  downloading, storing, processing.

  RIGHTS HOLDERS argue: that is unlicensed
  reproduction at industrial scale, and the resulting
  model competes commercially with the works it
  consumed.

  MODEL DEVELOPERS argue: training is
  TRANSFORMATIVE — the model learns statistical
  relationships and does not store or redistribute
  the works, analogous to a person learning from
  reading.
─────────────────────────────────────────

Where It Stands
─────────────────────────────────────────
  UNSETTLED, and diverging by jurisdiction. Litigation
  is active across text, image, music and code. Some
  jurisdictions have introduced text-and-data-mining
  exceptions, often with an opt-out for rights
  holders; others have not.

  Do NOT rely on a summary of the legal position from
  any source, including this one, being current. Check
  the state of play for your jurisdiction and your
  modality when it matters.
─────────────────────────────────────────

What Has Changed in Practice
─────────────────────────────────────────
  Regardless of outcome, the market has moved:

  - Licensed training corpora are now a real product
    category, and several major providers train on
    them deliberately
  - Enterprise offerings commonly include COPYRIGHT
    INDEMNIFICATION for output — a contractual answer
    to an unresolved legal question
  - Opt-out registries and robots.txt-style signals
    for training have become common
─────────────────────────────────────────

3. Question Two: Does the Output Infringe?

This question is not new law. Generated output is assessed by the same standard as anything else.

The Standard
─────────────────────────────────────────
  Infringement generally requires SUBSTANTIAL
  SIMILARITY to a specific protected work.

  The fact that a machine produced it is not a
  defence, and not an aggravator. The question is
  what the output IS, not how it was made.
─────────────────────────────────────────

Where the Risk Actually Concentrates
─────────────────────────────────────────
  HIGH RISK
    Prompting with a living artist's name for
    commercial output
    Recreating a recognisable character (these are
    protected independently of any single image)
    Reproducing a logo, trade dress, or brand mark
    Regenerating a specific well-known photograph

  LOWER RISK
    Generic subjects, generic styles
    A style described by its ATTRIBUTES rather than by
    an artist's name ("flat vector, limited palette,
    thick outlines" rather than "in the style of X")
─────────────────────────────────────────

The Style Distinction
─────────────────────────────────────────
  In most jurisdictions, an artistic STYLE is not
  itself copyrightable — only specific works are.

  But this is a narrow shield, not a wide one:
  - "in the style of [living artist]" invites output
    close enough to specific works to be a problem
  - some jurisdictions protect personality or
    publicity rights separately from copyright
  - and it may be lawful yet still cost you a
    reputational fight you did not want

  Practical rule: DESCRIBE the style, do not NAME the
  artist. You lose almost nothing and avoid the entire
  category.
─────────────────────────────────────────

4. Memorisation and Regurgitation

Module 1, Chapter 1 noted that models store parameters, not files — but that verbatim reproduction still happens. The mechanism matters here.

Why Memorisation Happens
─────────────────────────────────────────
  A work that appears THOUSANDS of times in a training
  corpus — a famous photograph, a stock image, a
  widely-quoted passage, a popular code snippet — is
  reinforced so heavily that the model can reproduce
  it near-exactly.

  This is not retrieval. It is the extreme end of
  OVERFITTING (ML Notes, Module 1) on
  over-represented data.
─────────────────────────────────────────

The Practical Consequences
─────────────────────────────────────────
  - The MOST FAMOUS works are the MOST likely to be
    regurgitated — precisely the ones whose owners
    will notice
  - Duplicate-heavy corpora make it worse, which is
    why deduplication is now standard practice
  - Highly specific prompts ("the photograph of the
    girl with the green eyes on the magazine cover")
    raise the probability substantially

  Consequence for you: a check is worthwhile before
  publishing anything commercially. A reverse image
  search on generated output is cheap.
─────────────────────────────────────────

5. Who Owns the Output?

The Human Authorship Requirement
─────────────────────────────────────────
  Several jurisdictions — notably the United States —
  require HUMAN AUTHORSHIP for copyright protection.

  Consequence: purely AI-generated content may not be
  protectable AT ALL. Not that someone else owns it —
  that nobody does.

  The practical test applied is how much human
  creative control shaped the specific result.
  Selecting, arranging, editing and substantially
  modifying output strengthens the claim; typing one
  prompt and publishing the first result weakens it.
─────────────────────────────────────────

What Follows for a Product
─────────────────────────────────────────
  If your business depends on OWNING the output —
  a character, a brand asset, a distributed work —
  build meaningful human authorship into the process
  and DOCUMENT it: iterations, selections, edits.

  If your output is transient or functional — a draft,
  an internal summary, a background texture —
  ownership rarely matters and this is a non-issue.

  Decide which case you are in BEFORE you build.
─────────────────────────────────────────

6. A Practical Posture

Given genuine unresolved questions, this is the defensible position — and it is mostly about process, not prediction.

Provider and Contract
─────────────────────────────────────────
  - Prefer providers who disclose their training data
    posture and offer output indemnification
  - Read the licence: several consumer tiers grant NO
    commercial rights by default
  - Keep the terms you agreed to, versioned, on file

Prompting Discipline
─────────────────────────────────────────
  - Do not name living artists, brands, or characters
    in prompts for commercial output
  - Describe styles by attribute
  - Keep prompt logs — they evidence intent, which
    matters if anything is ever questioned

Output Review
─────────────────────────────────────────
  - Reverse image search anything published widely
  - Check generated code against known licences,
    especially for distinctive blocks
  - Escalate anything recognisably close to an
    existing work

Disclosure and Records
─────────────────────────────────────────
  - Attach provenance metadata (Chapter 3)
  - Disclose AI involvement where your users, your
    platform, or your regulator expects it
  - Record the human contribution where ownership
    matters (Section 5)
─────────────────────────────────────────

The Underlying Principle
─────────────────────────────────────────
  You cannot resolve the open legal questions. You CAN
  demonstrate that you took reasonable care.

  Every item above is cheap, and together they are the
  difference between an explainable process and an
  indefensible one.
─────────────────────────────────────────

7. Summary & Next Steps

Key Takeaways

  • Separate the two questions: whether training was lawful (unsettled, not yours to resolve) and whether a specific output infringes (ordinary law, entirely within your control).
  • Generated output is judged by the normal substantial-similarity standard; risk concentrates on named artists, recognisable characters, brands, and famous specific works.
  • Memorisation is extreme overfitting on over-represented data, so the most famous works are the most likely to be regurgitated — check before publishing commercially.
  • Human authorship requirements mean purely AI-generated work may be unprotectable; if ownership matters, build and document real human creative control.

Concept Check

  1. Why does separating the training question from the output question change what a builder should actually worry about?
  2. Explain the mechanism by which a model reproduces a famous photograph near-exactly without storing it.
  3. Your company wants to own a generated mascot as a trademarkable brand asset. What does that requirement change about how you produce it?

Next Chapter

→ Chapter 2: Deepfakes, Consent & Misuse

Jr Codex — 1-on-1 Personalized Coaching | Back to Module Index | Back to Generative AI Index