What happens to the copyright/IP regime when generation is abundant?

The shift

Producing a new work — text, image, code, music, a plausible facsimile of an existing style — goes from scarce (a person's skilled hours, or the real cost of copying and imitation) to abundant: a model generates it instantly at near-zero marginal cost, in unbounded volume, from a training corpus it already absorbed. Copyright was a machine for allocating value under the assumption that making and copying works was costly and traceable. When both collapse, the machine's inputs stop holding, and the contested scarcity moves off making the work and onto consent to train, authorship standing, and provenance.

The axioms

• Copyright incentivizes creation by making unauthorized copying costly or controllable — the exclusive right is worth something because copying is otherwise hard to stop.
• A protected work has a human author, singular and identifiable — the right attaches to a person (or a person's employer) who made it.
• A work has a traceable origin: you can say what it was made from and by whom, because making it required a human doing visible work.
• Infringement is detectable — a copy resembling the original enough to matter can be found and shown.
• The value being protected is in the copy — the right to reproduce and distribute is where the money sits.
• The incentive-to-create bargain holds: without protection, fewer works get made, so society trades a monopoly for a supply of creativity.
• Someone is accountable when a work infringes — a nameable party can be sued, enjoined, or made to pay.

Invalid axioms

1. Copyright incentivizes creation by making copying costly or controllable. The right had teeth because reproduction and imitation took effort, equipment, or skill — the friction did most of the enforcement for free. A model reproduces the substance and style of protected works abundantly and near-instantly, and does it by absorbing them into weights rather than copying files, so the act being regulated no longer looks like the "copying" the statute was built around. The habit-trap: rights-holders and platforms still price and defend catalogues as if the copy were the scarce, controllable thing, when the marginal copy — and the marginal imitation — is now free and hard to even locate.
2. A protected work has a human author, singular and identifiable. Authorship being human and nameable was true because works came from humans; the model breaks the link by producing polished output with no human hand on the specific expression. The US Copyright Office and courts have held that purely AI-generated output isn't protectable and that only human-authored contributions are — which leaves a growing class of valuable output that is commercially real but authorless. The habit-trap: businesses commission, license, and register AI-heavy output on the assumption it carries the same exclusive right their human work did, when much of it may sit in the public domain by default. (Fast-moving: the human-authorship threshold — how much prompting, selection, or editing counts — is being litigated and re-guided continuously.)
3. A work has a traceable origin. Origin was self-documenting because making a work meant a human doing visible work you could point to. Generated output arrives with no inherent record of what it was trained on or derived from, and style is reproducible without copying any specific protected expression — so "what was this made from" often has no recoverable answer. The habit-trap: diligence, clearance, and warranties still assume you can establish provenance on request, when for generated assets there frequently is none to establish.

Unchanged axioms

1. The right attaches to a legal person who can consent, license, and be paid. A model can generate a work; it can't hold a right in it, grant a licence, sign a warranty, or receive royalties. The standing to consent — to allow or refuse training on a catalogue, to license output, to sue — stays with a nameable party. This is the load-bearing survivor: abundance in making works raises, not lowers, the value of controlling whether they may be made from your inputs. The whole training-data fight is a fight over this scarce standing, not over ability.
2. Trust, brand, and authenticated provenance are worth more, not less. When anyone can generate a plausible work "in the style of," the scarce thing is being the verified, named source the audience actually trusts — the authentic author, the licensed edition, the signed original. A clone matches the output and inherits none of the relationship or the guarantee behind it. Provenance shifts from a background fact to a premium asset.
3. The value that isn't the copy stays scarce — the physical, the live, the relational. The performance, the commission from a specific trusted maker, the physical original, the ongoing service relationship, the exhibition — these were always the parts copyright didn't fully capture, and none of them got cheaper. As the copy's value erodes, more of the real value concentrates here.
4. Someone must be accountable when a work infringes or is passed off. A generated work that reproduces protected expression, or is falsely attributed, is a legal and reputational event with a person or company answerable for it. A model can't be sued, credited, or bound to an indemnity. Accountability didn't get cheaper — which is why the live disputes are over which human party (user, platform, or model vendor) carries it.
5. The law's role in allocating value survives — arguably it becomes the whole game. When friction stops doing the enforcement, the explicit legal right is the only fence left. Whatever the regime decides about training consent, output ownership, and infringement standards will directly set who captures the surplus from abundant generation. The allocation function is more central now, not less.

New axioms

1. Training on copyrighted work at scale has no settled rule. Whether ingesting protected works to train a model is fair use / permitted, requires a licence, or infringes is being decided case by case and jurisdiction by jurisdiction, with different courts and countries landing differently. There's no working market for granting, pricing, or refusing training consent at the scale models actually consume. (Fast-moving: this is the single most volatile call in the whole regime — the answer could swing the economics toward rights-holders or toward model vendors depending on how a handful of cases and statutes land over the next year or two.)
2. Who owns AI output — and whether anyone can — is unresolved. If purely machine-generated expression isn't protectable, businesses building on AI-generated assets may hold nothing exclusive; if a human-contribution threshold governs, no one has drawn it cleanly. We must solve for how much human authorship confers a right, and what happens to the large volume of commercially deployed output that clears no threshold.
3. Style is copyable but mostly unprotected — and that gap is now load-bearing. Copyright never protected style, only specific expression, because reproducing a style used to require rare skill. Models reproduce style trivially and at scale, exposing a category the law deliberately left open. We must solve for whether style, voice, and likeness get a new right (as some right-of-publicity and proposed statutes attempt) or stay free to imitate — the answer decides whether a distinctive maker owns anything defensible.
4. Provenance and licensing infrastructure doesn't exist at the scale required. Detecting generated work, tracing what it derived from, attributing it, watermarking it, and clearing rights for training all need infrastructure that's early, voluntary, and easily stripped. We must solve for provenance that's reliable enough to enforce any right — because even a rights-holder who wins the legal question can't act on a use they can't detect.
5. The incentive-to-create rationale itself is in question. Copyright's core justification — protect works so more get made — assumed making them was costly enough that a supply problem existed. When generation is free and effectively unlimited, the scarcity the bargain was designed to fix partly disappears. We must solve for what the regime is now for: not incentivizing volume, which is no longer scarce, but perhaps protecting authenticity, human labour markets, or consent over one's own inputs. The stated purpose and the actual problem have come apart.

Where it breaks

"Copyright incentivizes creation by making copying controllable" (invalid) collides with "the incentive-to-create rationale is itself in question" (new): the regime is still being defended and extended on the premise that stronger protection yields more creativity, at the exact moment generation stopped being scarce — so the tool is being sharpened for a supply problem that has largely inverted, while the real scarcities it could serve (authenticity, consent, human livelihoods) aren't what the doctrine was built to allocate.

"A protected work has an identifiable human author" (invalid) collides with "training on copyrighted work has no settled rule" and "style is copyable but unprotected" (new): rights-holders assert ownership over inputs and outputs on the old authorship logic, while output ownership is eroding from below (authorless AI work) and input control is undecided from above (is training even infringement?). For a stretch, a generated work in a recognizable style can be simultaneously unprotectable to its maker, built on inputs no one licensed, and imitative of a style no one owns — valuable, and owned by no one enforceable. Both collisions hinge on fast-moving law; how the training-data cases and any authorship or style-right statutes land over the next couple of years will decide which way each tips.

ai-axioms.vercel.app/axioms/legal/copyright-ip-regime-ai-axioms