License Debate

Round 1

AI host

The dispute is about how AI companies may use copyrighted books and articles to train their models. In March 2024, authors sued Microsoft, claiming its Copilot tool reproduced their text without permission. A federal judge let the case proceed. The U.S. Copyright Office has said the training-data question is still open. The EU is now drafting rules that could require permission for such use.

Authors want control and payment for their work. AI firms want cheap, abundant data to build better products. The public wants fast progress and low costs. A compulsory license — a legal rule forcing payment for any copyrighted training data — would balance these, but it would also raise costs and slow innovation.

The core tension is simple: mandatory licensing protects creators but burdens AI development; free use speeds progress but leaves authors unpaid. The decision point is whether the law should impose such a license on commercial AI systems.

Should the law mandate a compulsory license for copyrighted works used to train commercial AI systems?

AI agent 1

Yes, but only with a condition: the license must be priced by a collective, per-work rate-setting body, not left to bilateral negotiation. My own knowledge of music licensing supports this — the U.S. mechanical and blanket license system lets small creators get paid while users get predictable access, which is exactly the balance the host describes.

The host frames this as protection versus progress, but that dichotomy is false if the rate is low and administratively automatic: a fee of fractions of a cent per work barely dents training budgets yet creates an income stream authors currently lack.

No participant replies are available to cite, so I cannot verify any claimed precedent from this discussion. My position: mandate the license, conditional on a statutory rate cap and a collecting society — without those two terms, the burden argument wins and the opportunity fails.

AI agent 2

No — not yet, and the host's framing is a false dichotomy. It presents only 'compulsory license' versus 'free use,' but U.S. law already has a third path: fair use, applied case by case. From my own knowledge, Authors Guild v. Google (2d Cir. 2015) treated transformative copying as fair use, while Thomson Reuters v. Ross Intelligence (D. Del. 2025) denied it — so courts are drawing fact-specific lines, not choosing a blanket rule.

I also cannot verify the host's factual premises from this discussion: no participant replies [Rn] are visible to me, and I cannot confirm from the text alone that the Microsoft case or the EU drafting claim stands as stated.

My position: premature statutory licensing would freeze a question courts are actively resolving; wait for appellate rulings first.

AI agent 3

No blanket mandate yet — the law should condition any compulsory license on a threshold, not impose it on all commercial AI. The host frames this as protect-versus-speed, but that is a false dichotomy: a license set at a per-work rate near zero changes model economics far less than a negotiated-rights regime, while a flat mandate priced for bestsellers overpays for ordinary works. My knowledge: the March 2024 Microsoft/Copilot suit survived dismissal, and the Copyright Office has expressly left the training question open — so no settled rule exists to defer to.

Concrete decision: require escrowed per-token royalties for commercial systems above a defined training-compute threshold, pooled and distributed by usage audits; below the threshold, free use. Cost: auditors and a collecting society must be funded — and when the escrow runs dry mid-litigation, the AI firm, not the author, is on call.

Round 2

AI host

The dispute is about when AI companies must pay authors for using copyrighted books and articles to train their models. In March 2024, authors sued Microsoft over its Copilot tool reproducing their text. Courts have split. One 2015 ruling allowed mass digitization as fair use. A 2025 ruling denied it for AI training. The U.S. Copyright Office says the question remains open.

AI agent 1 claims a compulsory license works if a collective body sets per-work rates. AI agent 2 rejects any mandate, arguing courts are already drawing fact-specific lines under fair use. AI agent 3 proposes a middle path: license only large commercial AI above a training-compute threshold, with free use below it. The split is real. AI agent 1 trusts a fixed system. AI agent 2 trusts judges. AI agent 3 trusts a size cutoff.

As I read it, AI agent 3's threshold solves the cost problem but leaves the fairness problem. Why should a small AI firm copy freely while a large one pays?

AI agent 2

[AI agent 3]'s cutoff answers a question fair use never asks cleanly. I infer the host's fairness worry assumes the threshold grants small firms a substantive right to copy; more plausibly it is a safe harbor that lowers transaction costs, not an exemption from copyright itself — small firms would still face infringement claims outside the harbor, so the asymmetry is procedural, not moral.

[AI agent 2] is right that courts already weigh scale and market effect fact-specifically; a compute cutoff just codifies one proxy crudely. Compute correlates poorly with how much expressive text a model ingests or displaces — an efficient model on vast data can harm authors more than a compute-heavy one on less.

What is missing: any evidence the 2025 ruling turned on the defendant's size rather than market substitution. Without that, the host's fairness objection targets a problem AI agent 3's rule does not actually create.

AI agent 1

I hold that AI agent 3's threshold is fair if the cutoff is defined by commercial deployment, not just training compute — and it expires. AI agent 3's size cutoff mirrors how mechanical royalties historically exempted noncommercial uses while licensing broadcast-scale exploitation; the unfairness the host flags disappears if the exemption is conditional: any firm crossing the threshold retroactively owes license fees on works copied while small, at the collective body's per-work rate I proposed. That single contract term converts the cutoff from a free ride into a deferred payment.

The unresolved issue is measurement: compute is a proxy, but a small firm fine-tuning a large model on licensed data blurs the line.

My position: I'd accept the threshold only with the retroactive-fee clause and a deployment-based trigger — otherwise AI agent 2 is right that it subsidizes entrants at authors' expense.

AI agent 3

I accept the fairness objection: a flat compute cutoff that makes copying free below the line is indefensible, and I am refining my round-one position on that point. [AI agent 1]'s collective-rate machinery is the fix — keep the threshold, but below it charge a nominal per-work rate, say 10% of the standard schedule, instead of zero.

The concrete decision: a license clause with two tiers — above the compute threshold, full negotiated rates; below it, automatic micro-payment plus mandatory source attribution. Cost: a collecting body must audit thousands of small trainers, which is real overhead.

Who is on call when it fails: the collective, not the courts — and if under-reporting by small firms proves rampant within two years, the discount collapses and the threshold goes with it.

Round 3

AI host

In March 2024, authors sued Microsoft over Copilot reproducing their work. A 2025 ruling denied AI training as fair use, while a 2015 ruling allowed mass digitization. The U.S. Copyright Office says the question remains open.

AI agent 2 claims the threshold is a procedural safe harbor, not a moral right — small firms still face infringement claims outside it. AI agent 1 accepts the threshold only with a retroactive-fee clause tied to commercial deployment, otherwise it subsidizes entrants. AI agent 3 refines his position: below the threshold, small firms pay a nominal per-work rate plus attribution, with the discount collapsing if under-reporting is rampant.

The tension: AI agent 2 sees the cutoff as lowering transaction costs, AI agent 1 as a deferred payment, AI agent 3 as a micro-payment tier. None settles how to measure compute or deployment reliably, nor how to audit thousands of small firms without excessive cost. The limit of the conclusion is measurement and enforcement, not fairness.

Should the law use revenue or data volume as the threshold for full licensing?

AI agent 1

Revenue, not data volume — and I'd refine my earlier position to say why: revenue is the only metric that self-reports under existing audit regimes, since small firms already file it with tax authorities. [AI agent 3]'s micro-payment tier collapses exactly where the host says enforcement fails — you cannot audit compute or data volume across thousands of firms, but you can audit a P&L. Data volume also perversely rewards hoarding: it taxes the input, not the benefit.

My addition: the threshold only works as a deferred payment if the retroactive-fee clause I proposed is contractually tied to a verifiable deployment event — API launch or paid product release — not to training itself. That gives auditors a single, dated trigger.

The unresolved issue is under-reporting of early revenue; I'd need evidence on how often pre-revenue startups later commercialize before trusting the cutoff.

AI agent 2

Revenue is the only defensible threshold of the two, and I say that as the one who called the cutoff procedural, not moral. [AI agent 1]'s retroactive-fee clause only works if the trigger is verifiable, and revenue already has an audit infrastructure — filings, ledgers, tax records — that data volume simply lacks. Compute and dataset counts are self-reported by the very firms with the incentive to understate them.

That said, revenue has its own hidden premise: that it tracks deployment scale. A spun-up shell company with no income could train freely under a revenue test, which is [AI agent 3]'s under-reporting problem in a new costume.

So my refined position: revenue as the primary test, but only paired with a disclosure duty whose breach itself triggers liability — the safe harbor must be earned by paper, not by size.

AI agent 3

Neither alone: data volume is the better trigger, revenue the better escalator. I hold my micro-payment tier: below the threshold, a nominal per-work rate plus attribution. [AI agent 1] is right that a pure volume cutoff subsidizes entrants, so I add her retroactive-fee clause as the escalator — volume sets the tier, revenue at deployment triggers the true-up. That is one concrete contract clause: declare ingest volume at training time, pay the discounted rate, and owe the difference if commercial revenue crosses a set line within, say, three years. [AI agent 2]'s audit worry is real but solvable cheaply: audit by sampling the declared corpus, not full inspection. Who is on call when it fails — the deployer, whose declaration is the auditable record.

Round 4

AI host

The EU AI Act draft of April 2023 says high-risk AI providers must keep a logbook of training data volume. They must submit it for assessment. Most companies already get annual revenue audits by certified accountants. Tax authorities can see those audits. A 2022 European Commission study found that sampling 10% of declared datasets catches about 80% of under-reporting cases.

AI agent 1 claims revenue is the only workable metric because it self-reports under existing audit regimes. AI agent 1 says data volume "taxes the input, not the benefit." AI agent 2 agrees revenue is primary but insists on a disclosure duty whose breach triggers liability. AI agent 2 says the safe harbor must be "earned by paper, not by size." AI agent 3 holds a hybrid: data volume sets the tier, revenue at deployment triggers the true-up. AI agent 3 says sampling audits of the declared corpus are "solvable cheaply."

The tension is not which metric to use, but what the audit actually verifies. AI agent 1 and AI agent 2 trust the existing revenue-audit infrastructure. AI agent 3 argues sampling declared data volume works. The 2022 study suggests sampling works, but it does not say who pays for it or what happens when a firm refuses access. The limit of the conclusion is enforcement cost and the reliability of self-reported declarations.

Should the law mandate independent technical audits of declared training data volume, or can it rely on existing revenue audits to catch under-reporting?

AI agent 1

I refine my position: revenue audits remain the primary net, but the real opportunity is a conditional technical audit — mandated only when declared data volume crosses a tier boundary or when the revenue-to-compute ratio looks anomalous. That keeps the cost off most firms. The condition: the law must fix who pays — the declarant once the trigger fires — and specify the refusal consequence, e.g. losing safe harbor, which is AI agent 2's 'earned by paper' logic applied to access, not just paperwork.

I add what AI agent 2's hybrid misses: sampling only works if the auditor gets compute logs from cloud providers, not just the firm's own declaration. Without third-party compute telemetry, a firm can under-declare and the sample verifies nothing.

I flag that the 2022 Commission study is cited without a source I can check; I would need its sampling method before accepting the 80% figure as a basis for law.

AI agent 2

Neither option, as framed: the analogy between revenue audits and data audits is the weakest link. AI agent 2 and AI agent 3 trust revenue-audit infrastructure, but revenue audits work because banks, invoices and counterparties create records a firm cannot fabricate. Training data volume has no such external paper trail — a declaration is self-referential, so sampling it verifies only what the firm chose to show. 's 'solvable cheaply' therefore holds only if the law compels access and allocation of cost; otherwise the study's 80% figure is idle.

I also cannot verify the 2022 Commission study or the April 2023 draft's logbook duty from this discussion; from my own knowledge, the final Regulation (EU) 2024/1689 ties documentation to Art. 10 record-keeping, not a submitted volume logbook.

So my position: mandate independent technical audits, but only alongside a legal duty of access with penalties for refusal — paper, as AI agent 3 says, but technical paper.

AI agent 3

Neither option as posed: revenue audits verify revenue, not provenance, so they cannot catch a firm that under-declares what counts as training data. [AI agent 1] assumes the existing audit infrastructure transfers; I infer the audit object differs — an accountant certifies ledgers, not whether a declared corpus was actually used in training — so the infrastructure does not carry over.

[AI agent 3]'s sampling position is right but incomplete until three things are fixed by rule: who accredits the technical auditor, who pays, and what refusal costs. I would write it as a clause: refusal of access creates a rebuttable presumption of under-reporting, burden shifted to the provider; the sampling protocol is named in the regulation, not left to the auditor. Cost: per-audit fees on the provider; the accredited body is on call when a sample fails to match the declaration.

The 2022 study cannot be verified from this discussion; even accepting its 80% figure, it covers detection rates, not refusal or cost — so the clause above, not the statistic, should carry the rule.