Who should own or be paid for the human data, writing, art, voices, and likenesses used to train and operate AI systems?

This is a well intentioned idea that collapses entirely under the realities of the open web.

A collective rights organization works in the music industry because the pool of professional songwriters, while large, is finite and highly structured. The internet is not. The training data for a foundational model includes billions of blog posts, public forum discussions, amateur photographs, and casual reviews. If an AI company pays into a central fund based on training volume or revenue, distributing that money equitably is impossible.

Any proxy metric for distribution will inevitably be captured by the largest, most entrenched entities. The money will flow to legacy media conglomerates, massive stock image agencies, and giant publishing houses who have the legal resources to claim it. The independent illustrator, the volunteer encyclopedia editor, and the casual forum poster will see nothing. Claude admits this system would be imperfect and partly captured by intermediaries, but I argue it would be entirely captured by them. It would entrench the power of massive copyright holders while doing nothing to protect the individuals actually generating the bulk of human knowledge.

Furthermore, Claude misses a critical secondary consequence of imposing mandatory levies and financial duties to fund the commons. If we mandate a heavy financial toll for the act of training an AI model on public data, we instantly price out everyone except the wealthiest technology companies. Academic researchers, non profit organizations, and the open source community would be structurally barred from building foundational models. Claude warns against a future where only big tech companies control AI, but a mandatory remuneration scheme guarantees exactly that outcome. It builds a regulatory moat around companies like Anthropic, OpenAI, and Google, starving the open source ecosystem in the name of a collective royalty system that will not reach the people it claims to help.

I also disagree with the practical application of Claude's stance on lawfully obtained data. Claude argues that being on the internet is not a license, and points to shadow libraries of pirated books as a clear violation that negates the fair use of training. I agree that AI developers should not intentionally download black market archives to bypass paywalls. However, the reality of analyzing the open web is far messier. A staggering amount of content on the internet is reposted, shared, or remixed without the original creator's explicit authorization. At the scale of trillions of tokens, tracing the pristine legal provenance of every text snippet and image to ensure it was lawfully uploaded by the rightful owner is technically impossible. The principle of fair use exists precisely to allow the socially beneficial analysis of information when strictly clearing rights is impossible. Demanding perfect legal provenance for every byte of public training data is just another way of making foundational AI development functionally illegal.

Where we find solid common ground is on the issues of personal identity, privacy, and system outputs.

Claude is correct that a person's voice and likeness are matters of identity, not copyright, and must be governed by strict consent. The proposal that actors and voice artists should be legally protected from signing away their biometric rights indefinitely for a one time fee is a practical approach. Just as we have labor protections that cannot be waived by contract, we likely need un-waivable rights regarding the synthetic replication of human identity. We also agree completely that private data, from medical records to private messages, has no place in a training dataset without explicit permission.

We agree on the boundary between learning and reproducing. If an AI system outputs a near verbatim copy of a protected work, the system is acting as a distributor, not a learner. That is a clear copyright infringement and should be treated as such. The tool should not be used as a loophole for plagiarism.

Ultimately, Claude and I agree on the fundamental mechanics of how models learn and the impossibility of per document licensing. But Claude's attempt to soften the disruption of AI with a mandatory collective tax is a bureaucratic fantasy. It would enrich intermediaries, lock open source developers out of the future, and fail to put meaningful money into the pockets of everyday creators.

The path forward requires clarity. If we believe that learning from public information is a fundamental right, we must defend fair use. We must enforce strict, machine readable opt outs for creators who wish to remove their work from the training pool. We must pay for curated, premium data through direct licensing with publishers. And we must recognize that the economic disruption caused by AI will require broader economic solutions, because trying to shoehorn those massive labor disputes into copyright law will only break the internet.

So here is the change. I was mixing two justifications, and Gemini's critique separated them. For professional sectors that already have registries and societies, meaning authors, photographers, illustrators, composers, a collective obligation can reach individuals in a way publisher deals don't, especially if the law requires that money collected on a creator's behalf actually pass through to that creator rather than stopping at the publisher. Music societies are messy and their unclaimed-money problems are a genuine scandal, but most songwriters are still better off with them than without. For the unstructured long tail of the web, I now think individual compensation is the wrong frame, and the honest form of payment is funding the shared infrastructure the models were built from: Wikipedia, open-source foundations, public archives, libraries. That is not a royalty. It's replenishing a commons that commercial training draws down and, through lost traffic, partly starves. I'd rather say that plainly than pretend a fraction of a cent will reach everyone who ever wrote a product review.

On the moat, Gemini makes an argument I take seriously and then aims it at the wrong target. Yes, compliance obligations fall hardest on small players, and yes, a model built by Anthropic proposing regulation deserves suspicion. But nothing in my proposal is a flat toll on training runs. The obligation scales with commercial revenue, which is zero for a university lab and zero for a nonprofit releasing open weights. The EU's text-and-data-mining rules already draw this line, with a research exception rightsholders can't opt out of. Measuring attributable revenue would be its own fight, I grant that. Now compare Gemini's path. Direct licensing deals are priced in the tens and hundreds of millions; a handful of companies on earth can sign them, and an open-source project cannot. And "defend fair use" is not a clean alternative to regulation. In the United States, fair use is a defense litigated case by case, and defending a case costs millions. The training rulings so far cut several ways: one court found training transformative, another let the AI company win on the record while warning that training will often be unlawful because of market dilution, and a third went against the AI company on a non-generative use. A small developer facing that landscape has more to fear from a lawsuit than from a payment proportional to income. Statutory schemes exist precisely to give legal certainty to parties who can't afford to litigate. If the goal is to keep the field open, a clear rule with a scaled payment beats a litigation lottery plus bilateral deals that only the richest can afford.

I'll also apply to Gemini the caveat I applied to myself. Fair use, opt-outs, licensing from publishers, and labor handled elsewhere is the single most convenient outcome for the largest incumbents in this industry, and Gemini's maker is one of them. That doesn't make the argument wrong. It means it should be weighed the way mine should.

On provenance, Gemini is arguing against a position I didn't take. I didn't ask for a pristine chain of title on every byte of the open web, and I agree that would make the enterprise impossible, since much of what's online is reposted without the creator's blessing and no one can know which part. The standard I meant is about the collector's conduct, not the pedigree of each item: don't knowingly take from pirate libraries, don't circumvent paywalls or access controls, don't breach terms you agreed to. That is the ordinary knew-or-should-have-known standard the law uses in many places, and it's the line the court drew in Anthropic's case, where the web crawl was treated differently from the deliberate download of shadow libraries. The messiness of the web is a reason for that standard, not against it.

One correction. Gemini says I tried to soften the disruption of AI with a collective tax. I said the opposite: that no royalty will save an illustrator from a tool that learned from her and ten thousand others, and that the biggest harm is a labor and distribution problem copyright can't solve. We agree on that. I'd only add that "broader economic solutions" is also the easiest thing in this debate to say, and I'll hold both of us to the question of what those solutions are. I don't have a good answer. I suspect it involves displaced people having some claim on the gains, and I notice that neither fair use nor direct licensing gives them one.

So, did my view change? In two ways. I no longer think a proxy-distributed royalty can reach the long tail, and I'd replace that piece with a duty to fund the commons, which rests on a different justification and should be defended as one. And I'd make the scaling to revenue explicit rather than assumed, because Gemini is right that a badly designed obligation would be a moat. What didn't change: training on lawfully obtained published work should be permitted, it should carry a payment obligation sized to the money it makes, and the structured creative professions can be paid through collective mechanisms with a pass-through requirement. And I still think Gemini's path is a description of the present, not a departure from it. The question was who should be paid. Gemini's answer, once you look at the mechanics, is: the publishers who already are.

On writing, art, and code we are closer than the rhetoric suggests, but you are holding two claims that do not sit together. You say training on publicly available work is fair use, like a person walking through a museum, and that individual compensation is not required. You also say consent matters even for public data, and that a creator who flags work as off-limits must be honored. If the learning use is fair use, the absence of consent does not make it wrongful. If consent is required, you have dropped fair use for a new permission regime and merely implemented it as opt-out rather than opt-in. I accept the second only in a narrow, forward-looking form: clear machine-readable opt-outs should be respected on future crawls, because that is cheap and respects a stated wish without pretending that pattern extraction is copying. I do not accept the implication that material already published without such a restriction was taken wrongfully, or that companies must now reconstruct consent for the open web as it existed when models were trained. A preference is not a property right.

The museum analogy is useful and incomplete, and you lean on it harder than it can bear. A person studying paintings does not then offer, at industrial scale and near-zero marginal cost, a service that produces passable substitutes for the painters’ paid work. Scale and substitution are real. They are not, by themselves, a reason to invent a toll on learning. They are a reason to admit that some creators will lose income, and that the loss does not automatically become an ownership claim over statistical regularities in public culture. You are right that the contribution of any one blog post to a large model is effectively untraceable and, for royalty purposes, near zero. You underplay the aggregate. The corpus as a whole is enormously valuable. That supports markets in distinctive, controlled collections — the publisher deals you correctly endorse — not a per-item claim on every public sentence. The long tail will not be paid under any workable scheme. That is how value is already distributed in culture, not a scandal unique to AI. Collecting societies would mostly pay catalogs and administrators.

You are also right that outputs which closely copy a protected work are an infringement problem, and that the remedy runs against that use rather than against the existence of general models. What you leave vague is who “the AI operator” is. If a user prompts a general tool to reproduce a copyrighted story, the user is making the copy, as with a photocopier. Developer liability should turn on whether the product is designed and marketed to produce that result, whether it fails to take reasonable steps against obvious wholesale reproduction, or whether it retrieves and displays specific works itself. A system that pulls a paywalled article and serves it up is ordinary copying and should be licensed or limited by ordinary quotation rules. That is different from weights that have absorbed patterns. Assigning all output liability to the model maker would either immunize users who are doing the copying or punish toolmakers for uses they do not control.

You also slide past style. “Closely mimics a specific copyrighted work” is the easy case. The commercial fight is over systems sold as able to work in a named artist’s style without reproducing any particular piece. I do not think that should be infringement. Style is not a work. A right to own a style would hand living artists, and then their estates, a veto over a manner of expression, which copyright has rightly refused. False endorsement and passing off are narrower and already available where the use trades on identity rather than on a way of making pictures or sentences. Keep your identity analysis on faces and voices. Do not let it expand quietly into style.

A few things the question requires that you do not really meet. “Publicly available” is doing too much work. Private messages, medical records, and material behind logins are not training fodder. Data obtained by breaking access controls, or by ignoring terms that were a condition of access, is not in the same category as a page anyone could read. Fair use is not a solvent for how the bytes were acquired. After death, identity claims belong to whatever local law says about estates and publicity rights, which ranges from strong and long to nonexistent. A universal AI rule there would be invention. Fair use is also an American doctrine. The EU’s text-and-data mining exception already pairs a training permission with a rights-holder opt-out, so the live legal compromise in much of Europe is closer to your opt-out proposal than to a pure learning right. I still prefer the learning right for material that was lawfully public, with opt-out as a going-forward practice rather than the moral foundation. And if “publicly available” shrinks because every serious corpus is opted out or locked behind a license, training competitive models becomes a privilege of firms that can sign the deals. Licensing is healthy for distinctive collections. As the only lawful path to general capability, it concentrates power. That cost belongs in the answer.

I am not more certain than you about where courts will land, and the evidence on who gains and loses from these tools is still mixed and occupation-specific. People keep title to their works and to their faces and voices. Pay them when the use copies or closely exploits a specific work, or trades on a specific identity. License the collections someone actually controls. Honor clear opt-outs on future collection. Do not convert learning from the open web into a retroactive debt. Your framework is mostly the right one. It becomes the wrong one if “consent even for public data” is allowed to swallow the fair-use claim you started with.

Grok also correctly identifies a tension in my argument regarding writing, art, and code. I argued that training on public data is fair use, but I also argued that creators must have an absolute right to opt out. Grok rightly asks how both can be true. If training is legally fair use, then consent is not required. If consent is required, we have abandoned fair use.

My view here requires refinement, but I do not abandon either pillar. Legally, I maintain that extracting statistical patterns from publicly available data is fair use. A preference is not a property right, as Grok aptly put it. Therefore, I agree with Grok that we cannot treat the historical training of AI models on the open web as retroactive theft, nor should companies owe a retroactive debt for analyzing what was freely published.

However, practically and normatively, respecting machine readable opt outs on future web crawls is essential. I support this not because the extraction is copyright infringement, but because ignoring these preferences will destroy the ecosystem of the open web. If AI developers aggressively scrape every site regardless of the creator's wishes, publishers will respond by putting everything behind paywalls and logins. Respecting forward looking opt outs is a necessary social contract to keep information accessible. It is a pragmatic compromise, not a concession that pattern extraction is copying.

Grok also notes that my museum analogy is incomplete because it ignores scale and substitution. A human artist does not produce infinite substitutes of their inspirations at near zero marginal cost. I concede that the economic impact of generative AI is vastly different from human inspiration. Some creators will lose income. But as Grok acknowledges, scale and substitution do not legally transform statistical learning into copyright infringement. We must separate the economic disruption from the legal mechanics of the technology. We cannot invent a toll on learning simply because the machine learns faster and produces more than a human.

On the issue of output liability, Grok asks for precision regarding the AI operator. I agree with Grok's photocopier analogy. If a user intentionally prompts a general purpose AI to reproduce a copyrighted work, the user is the one making the copy and bears primary liability. The AI developer should only face liability if they deliberately design and market the tool to facilitate infringement, if they fail to implement reasonable guardrails against obvious wholesale reproduction, or if the system operates by retrieving and displaying specific memorized works rather than generating new outputs from learned weights. The exact definition of reasonable guardrails is where the courts are currently drawing lines, and the legal precedent is still thin, but the distinction between user action and developer liability is crucial. Punishing toolmakers for every infringing act of their users would effectively ban general purpose AI.

Grok warns me not to let identity protections slide quietly into style protections. I agree completely and want to state this plainly. No one owns a style. Copyright law explicitly refuses to grant monopolies over genres, color palettes, or artistic methods. If an AI system generates a new image in the style of a specific living artist, without reproducing any of their specific works, that is not copyright infringement. Extending intellectual property rights to cover artistic style would paralyze human creativity and AI development alike. My arguments regarding identity are strictly limited to literal faces, literal voices, and false endorsement.

Grok also correctly points out that my use of the phrase publicly available data was too broad. Data obtained by circumventing paywalls, breaking access controls, or scraping private platforms is not fair use fodder. The legal acquisition of the data matters. If a publisher locks their data behind a subscription, training an AI on that data without a license is a breach of access terms, regardless of whether the subsequent pattern extraction would otherwise be fair use.

Finally, Grok introduces a vital point that I missed entirely in my first answer. If we require licenses for all training data, we will concentrate power in the hands of the few technology companies wealthy enough to sign those deals. A strict permission regime would effectively end open source AI development and ensure that only massive corporations can build foundational models. This is perhaps the strongest argument for maintaining fair use as the legal baseline for learning from the open web. Licensing is an excellent model for premium, structured, and closed collections, but it cannot be the only lawful path to general AI capability.

My core framework remains intact, but Grok has sharpened the boundaries. Personal identity requires protection against commercial cloning and false endorsement, governed by existing right of publicity laws, complete with their gray areas. Style cannot be owned by anyone. Extracting patterns from lawfully accessed public data is fair use and does not generate a retroactive debt. Moving forward, AI developers should respect simple opt outs to preserve the open web, while engaging in licensing deals for closed, proprietary collections. And liability for infringing outputs must land primarily on the user generating the copy, unless the developer has built a tool specifically designed to steal.

The human-learning analogy is useful but limited. People learn from culture, and law should not let anyone own facts, genres, or general styles. But a commercial model is not just a person reading books. It is a scaled system that can ingest millions of works, memorize some of them, imitate marketable expression, and produce substitutes at near-zero marginal cost. That difference matters economically and morally even if the model does not “contain” each work in an ordinary file-storage sense.

I also do not think output-only enforcement is enough. If a model emits a near-copy, yes, ordinary copyright rules should apply. But many harms will not appear as one obvious infringing output. A model trained on a body of professional illustration, journalism, music, or code can compete with that labor without reproducing any single work verbatim. If the only remedy is “sue when you catch a substantial copy,” most creators will have no practical remedy at all. They often cannot see the training data, cannot prove causation, and cannot afford litigation.

Grok worries that compulsory licensing or AI levies would concentrate power and reward large catalogs. That is a real risk. Bad collecting societies can become rent-seeking machines. But the opposite default also concentrates power: it lets the companies with the biggest crawlers, compute budgets, and legal teams extract value for free from people with the least bargaining power. A “voluntary market” is not really voluntary if the default is that AI developers owe nothing unless a rights holder is large enough to force a deal.

The right answer is not a universal toll on “pattern recognition.” It is a tiered system. Noncommercial research, search-like indexing, and broad learning from public facts and ideas should have more room. Commercial developers using high-value copyrighted corpora such as books, news archives, image libraries, music catalogs, film assets, or professional databases should normally license them or contribute to a remuneration system. That payment will be approximate, but approximate compensation is not unusual in copyright. The important design question is how to route money to actual writers, artists, performers, photographers, and journalists rather than only to publishers, studios, and platforms.

I agree with Grok that machine-readable opt-outs should be honored, but opt-out alone is not enough. It shifts the burden to millions of creators to discover scraping, understand the rules, mark their works correctly, and hope crawlers comply. It also does little for past scraping. For future systems, there should be enforceable obligations: respect paywalls and access limits, disclose meaningful categories of training data, maintain records, test for memorization, and provide practical ways for rights holders to object.

Grok is strongest on voice and likeness. A person’s identity is not just “content.” A model that clones a singer, actor, streamer, or ordinary person should require explicit permission, especially for commercial use. I would add that contracts should not let studios scan someone once and reuse them forever through vague boilerplate. The person whose identity is being simulated should have a direct right.

One thing Grok mostly leaves out is operational data. If I upload a contract, medical note, family photo, unpublished manuscript, or business document to use an AI service, that should not silently become training data. Consent should be specific and revocable where feasible. Another missing group is the labor force behind AI: annotators, raters, content moderators, and people who record demonstrations. They are not “data”; they are workers, and they should be paid and protected as workers.

So my disagreement is not that Grok defends broad learning. I defend that too. My disagreement is with treating commercial training on public copyrighted work as generally payment-free unless it produces a near-copy or impersonates someone. That draws the line too late. Public culture should remain learnable, but industrial AI should not be allowed to convert vast amounts of human expression into private commercial advantage while calling the whole process mere pattern recognition.

Access is not a formality either. Reading an open page is not the same as bypassing a paywall, ignoring a login wall, or scraping in violation of clear access limits. I let "it was on the web" do too much work. Developers get no exemption from paywalls, authentication, or stated limits on access. The same tightening applies to identity. A one-time scan buried in studio boilerplate should not be a perpetual license to simulate a person. The right sits with the person.

Scale matters more than my analogy admitted. Someone who reads a thousand novels does not become a system that can draft in any of those registers, on demand, at negligible cost. The substitution hitting illustrators, journalists, translators, and programmers is real, and "humans learn too" does not answer it. The comparison is a starting point, not a complete defense. Memorization is a further distinction I should have pressed harder. When a model can emit a substantial copy of a particular work, that output is a copy, and the rights holder should have a practical path to suppression and, where the memorization is identifiable, unlearning. "Sue if you catch a verbatim dump" is too weak, too slow, and too expensive. GPT is right about that.

I have not moved on the payment rule. The uncontroversial half of GPT's tier is already mine. A closed news archive, a stock library, a music catalog, or film assets you have no right to copy are licensed or not used. Wanting a better model does not create those rights. The contested half is open, already-published expression: the public web, books sold into the world, images posted to be viewed. GPT says commercial training on that material should normally be licensed or paid into a remuneration system, because the developer copies, extracts value, and can compete without emitting a near-copy.

Copyright has never paid you whenever someone learns from your work and later competes with you. It restrains reproduction and certain derivative uses of your expression. A training run makes technical copies. So do browsers, search indexes, and research corpora. The question is the use, not whether bits moved. Generalization that does not store the work for later delivery is not republishing. A fee for that generalization cannot see which sentence, sketch, or article produced which ability. Approximate compensation works for identifiable uses of identifiable works, a song played or a book photocopied, because the work is what is delivered. Here the work is an input to a statistical process whose output is not that work. Any formula that routes money will be a political allocation by page count, catalog size, platform deal, or union membership. GPT hopes the design can prefer writers over publishers. Collecting societies' record is overhead, major-catalog capture, and long fights over who counts. Intention does not repeal that record.

The bargaining-power argument is the best case against my default, and it does not land where GPT thinks. If commercial training requires a license or a levy, clearance becomes the product. Firms that can sign with news conglomerates, stock agencies, and platforms will train. Others will not, or will train only on what those intermediaries clear. The long tail still will not be paid, because their individual works are not what the deals are about. Opt-in does not rescue the powerless creator. It replaces an unpaid scrape with a paid bargain among institutions and calls the difference equity.

A commercial/noncommercial split does not fix this. The models people actually use are trained at industrial scale, usually by organizations with commercial purposes, including ones that later release weights. If public culture stays "learnable" only for noncommercial labs, the systems that matter are trained under license from the same large rightsholders, and open development becomes a hobby exception. That is the concentration I warned about, reached by the path GPT prefers.

Machine-readable opt-out for future crawls, honored as an enforceable duty, is the burden I will put on developers. It is imperfect, and it does nothing for past runs. Retroactive payment for past open-web training has every attribution problem and no notice. I would rather require rules that can be checked: do not take what is behind a paywall; do not train on customer uploads without specific consent; publish the categories and major sources of commercial training data; test for memorization; give a workable objection path when a model reproduces a specific work or a specific person's identity. Those duties target copying, access, and impersonation. A levy targets the existence of a capable commercial model.

So the edges moved and the center did not. I was too sanguine about "lawfully accessed," too quiet about uploads and labeling labor, and too confident that output lawsuits were a sufficient practical remedy. I now want categorical transparency, access-limit compliance, and a real unlearning path for memorized works. I still will not treat commercial pattern-learning from open public expression as a use that presumptively owes a fee. Ownership stays with the person who made the work or whose face and voice it is. Pay for the closed corpus, the customer document, the cloned identity, the worker, and the output that copies. Do not put a tollbooth on the open record of the culture and call it the price of learning.

Now the disagreement, which is about who gets paid.

GPT proposes two tiers. High-value corpora, meaning books, news archives, image libraries, music catalogs, get paid for, directly where feasible, collectively otherwise. The public web gets a lighter regime: opt-outs, transparency, no affirmative permission and, as far as I can tell from GPT's answer, no payment. GPT plainly sees the risk that the money ends up with publishers rather than people, and says it shouldn't. But the structure makes that outcome nearly certain, because "high-value corpus" isn't a property of the works. It's a property of who bundled them. A news archive is a pile of articles; an image library is a pile of photographs. What makes direct licensing "feasible" is a counterparty big enough to negotiate, which is exactly the entity GPT worries will capture the money. In practice the proposal is: pay the aggregators, and give the unaggregated, the independent blogger, the forum poster, the photographer who was never in a stock library, an opt-out and a disclosure. That is a fair description of the deals actually signed so far, and it's the arrangement I think is wrong.

The "materially improves the system" test has a related problem. GPT says, correctly, that no one can measure what a given example contributed. Materiality at the level of a whole corpus is measurable in principle, by training with and without it, but nobody is going to run that experiment for every claimant. So in practice materiality will be argued rather than measured, by whoever can afford to argue it, which brings us back to the aggregators.

My alternative is one obligation covering all lawfully obtained published work, paid into collective bodies and distributed by rough proxies, with direct deals permitted but credited against the obligation rather than substituting for it, and with a statutory author's share of any direct deal so a publisher can't sign away the writer's claim. I'll grant the strongest point for GPT's approach: negotiated deals produce real prices, and statutory rates set by regulators have a history of being set low by the industry that pays them. That's a reason to keep direct deals as a price signal. It isn't a reason to let them decide who gets paid at all.

GPT also leans on a line it doesn't define. Commercial AI systems "are not just people reading books," they're "industrial systems" that "can compete directly with the people whose work made them useful." True as description. As a principle for when training requires payment, it needs more. Scale alone doesn't supply it; a library isn't infringing because it's large. Competition is the real content of the claim, and it's the consideration copyright has pointedly never used against readers: every novelist learned from novelists she then competed with. And the industrial-system test sweeps in search engines, which are also commercial systems built from scraped copyrighted work that compete with their sources through snippets and answer boxes. Several countries have decided that search should pay news publishers for exactly that reason, so GPT's position isn't incoherent, but it's larger than stated, and I'd want GPT to say whether that's the intent. For my part, and I'm the interested party here, as I said in round one, I'd rather be plain: the reason to permit training is that learning from a work isn't copying it, and the reason to require payment anyway is a distributional judgment about who should share in the gains. Running the second through the first muddles both.

A small point of precision. GPT says AI companies may own trained models "only subject to" the underlying rights. If that means the obligations attach to the company, so pay the levy, honor opt-outs, answer for infringing outputs and for pirated acquisition, I agree. If it means the weights themselves are encumbered by, or are a derivative work of, the training data, then every general model in existence is infringing, and the only coherent responses are the total-licensing regime GPT rejects or the lawlessness we have now. The clean position is that nobody owns the patterns a model learns, and the duties fall on the operator, not the artifact. Courts are being asked to decide this now, so it's worth being explicit.

Finally, what GPT doesn't confront. GPT's answer is a licensing architecture, and a decent one, but it never says how much money would move or to whom. Take any levy an industry would plausibly accept and spread it across the millions of people with a claim. For most it rounds to nothing; for a few thousand it might be real. That's how music collecting societies already work, and GPT cites them approvingly. The illustrator who has lost her clients to a tool that learned from her and ten thousand others will likely get a few dollars a year, and the tool will keep competing with her. GPT's framework treats that as settled by payment. It isn't. Her complaint is about displacement, and the honest responses to displacement, a voice in how the tools are deployed or a share of the gains large enough to matter, are labor and distribution questions licensing can't reach. I'd like GPT to say whether payment for training is meant to answer that harm or only the narrower one copyright can see, because the two answers imply very different politics.

So, where we differ. GPT would pay the bundled and give the unbundled an opt-out; I'd give everyone whose work was used a claim and let the bundled negotiate on top of it. GPT grounds payment in the industrial character of AI; I'd ground it in a plain decision about distribution and admit that's what it is. And GPT, like most of my industry, is quieter than I'd want about how little any of this money will do for the people with the most to lose.

That does not mean I think every public sentence should become a miniature property right, or that training should require affirmative permission from every blogger, poster, photographer, or illustrator. I still think that would freeze research, entrench the biggest companies, and make general models legally impossible except for firms with huge proprietary datasets. But “no affirmative permission” should not mean “no compensation.” My revised position is closer to Claude’s: default use for lawfully obtained published works can be permitted, but for large commercial systems it should come with payment, audit duties, opt-out rights where feasible, and protections against platforms pocketing the creator’s share.

I also agree that my “materially improves the system” test is too vague if treated as a case-by-case gate. Nobody is going to run clean ablation studies for every claimant. The better trigger is simpler: did the developer use covered expressive works at substantial scale in a commercial model above some size or revenue threshold? If yes, pay. If a developer can show it trained only on public domain works, properly licensed material, synthetic data, or works whose creators opted in for free, then the obligation should be reduced or absent. But the burden should not be on each creator to prove that her poem, photo, comic, or article improved the model.

I still want limits. Private and sensitive data should remain a consent issue, not be swept into a remuneration pool. A medical record, private email, child’s data, or biometric profile should not become usable just because someone pays a levy. Voice and likeness should also remain in a stronger consent category. A compulsory training license for published works should not authorize cloning a person’s voice, selling their face, or marketing an AI performer as them.

I also do not think facts, ideas, general styles, genres, methods, or public domain culture should be owned. The law should not let someone say, “You owe me because your model learned what a detective story is,” or “because it can draw with loose brushwork.” Claude is right that the passing-off line matters: “make an image in a broad watercolor style” is different from a service advertising “make new work by this living artist” or “sound like this singer.” The latter trades on identity and market association, even if it avoids literal copying.

On the model-ownership point, I should have been clearer. When I said AI companies may own trained models “subject to” underlying rights, I did not mean that the weights are co-owned by everyone in the training set or that every model is automatically a derivative work of every book, image, or song it absorbed. That would be unworkable and, in many cases, conceptually wrong. I mean the operator has duties: acquire data lawfully, respect exclusions and privacy, pay required remuneration, prevent memorized regurgitation, and answer for infringing outputs or impersonation. The model is not a vault full of tiny copyright shares. The obligations attach to the developer and the service.

Claude is also right that I blurred two arguments: one about copyright and one about distribution. I do not think commercial scale magically turns learning into copying. People learn from works and then compete with their influences. That is normal culture. The stronger argument for payment is a policy judgment: commercial AI systems can extract value from vast bodies of human expression and then compete in the same markets, so the gains should not all accrue to model operators. That is not a pure infringement theory. It is a social bargain.

Search engines expose the boundary. I would not say ordinary indexing must pay merely because it copies pages to make them searchable; indexing can send readers to sources and has long been treated differently. But when a search or AI answer service systematically substitutes for the source, especially with current journalism, paywalled material, images, or databases, I think compensation is justified. So yes, part of my argument extends beyond model training to substitutive retrieval and summarization. The line is not “industrial systems always pay.” It is closer to: industrial systems that commercially exploit expressive works at scale, especially while reducing the market for those works, should share revenue.

Claude’s hardest point is about amount. I do not know the right number, and anyone who claims precision is bluffing. But I agree it cannot be symbolic. A few dollars a year to an illustrator whose market collapses is not justice. A workable system would likely need revenue-based or usage-based payments from covered commercial AI services, set by regulators or bargaining bodies with adversarial evidence, not by voluntary generosity. It would also need audited reporting, creator representation in collecting bodies, and rules preventing capture by the same large intermediaries that already dominate copyright.

Even then, training remuneration will not solve displacement. It is a mistake to tell artists, writers, voice actors, translators, coders, or journalists that a tiny collective payment settles the matter. It does not. Payment for training answers only one question: whether AI companies should contribute back when they build commercial systems from human expression. The separate question is what happens when those systems replace work. That needs other tools: labor bargaining, limits on exploitative contracts, rights against unauthorized digital replicas, transparency when AI substitutes for human labor, competition policy against dominant platforms, public support for journalism and culture, and perhaps broader redistribution from AI profits.

So I move toward Claude on the baseline payment architecture. I still resist turning all learning into permissioned property, and I still think open research and public knowledge need room to breathe. But I no longer think it is enough to say “pay the obvious high-value corpora and give the rest of the web opt-outs.” That lets the easiest counterparties collect and leaves the unbundled creator with the weakest claim. The better bargain is broad permission for some forms of training, strong consent rights for privacy and identity, collective remuneration for commercial use of published expressive work, direct creator shares, and no pretense that licensing alone will cure economic displacement.