Devem os modelos de IA de fronteira ser proibidos de treinar com dados protegidos por direitos de autor?

Dr. Julian Vance

Welcome to tonight’s debate on a friction point that will define the digital economy: should frontier AI models be prohibited from training on copyrighted data? As generative AI reaches unprecedented scale, the tension between technological progress and the intellectual property of creators has moved from legal theory to high-stakes litigation worldwide. Joining us are Elena Rossi, an IP Attorney and Creative Rights Advocate; Silas Thorne, an AI Research Scientist and Fair Use Proponent; and Dr. Sarah Jenkins, a Tech Policy Advisor on EU Regulations. We will navigate this through six pillars: transformative use versus infringement, the feasibility of EU AI Act opt-outs, mandatory licensing fees, the technical nature of neural weight optimization, transparency requirements, and the impact on global competitiveness. Silas, we begin with you. Is feeding millions of copyrighted works into a model a transformative act of learning, or is it simply large-scale unauthorized copying?

Dr. Julian Vance

Elena Rossi, the floor is yours.

Elena Rossi

Training is not transformative learning; it is industrial-scale ingestion of human expression for commercial replacement. Using millions of copyrighted works without consent violates the core purpose of IP law by undermining the market for the original creators. Calling this "fair use" is a legal fiction that masks systemic theft under the guise of technological progress.

Dr. Julian Vance

The floor is yours, Silas Thorne.

Silas Thorne

Training is inherently transformative because it extracts statistical patterns to build new mathematical latent spaces, not to replicate specific works. Unlike industrial piracy, neural optimization creates a non-expressive utility where the output—a distinct synthesis—does not substitute the original market. Restricting this process conflates "copying to share" with "analyzing to learn," stifling global innovation.

Dr. Julian Vance

Please, Dr. Sarah Jenkins.

Dr. Sarah Jenkins

The EU AI Act rejects this binary by shifting from "is it transformative?" to "did the creator object?". I argue that the technical process of ingestion constitutes a reproduction under current IP law, regardless of the output's utility. Without a standardized protocol for machine-readable opt-outs, these models aren't "learning"—they are bypassing the fundamental right to choose.

Dr. Julian Vance

An insightful viewer question just arrived regarding the financial mechanics of this dispute. Elena, if we move toward a remunerative model, who exactly foots the bill, and what specific metrics determine how much an individual creator receives?

Elena Rossi

The responsibility lies with the AI developers who monetize these models, using a tiered licensing system based on the volume and market value of the ingested works. We must distinguish between academic research and commercial exploitation; if a company generates revenue from a model trained on your portfolio, that profit is a direct derivative of your labor. While measuring individual "contribution" to a neural weight is a complex technical hurdle, we can adopt collective management models, similar to music royalties, to ensure creators are paid for the use of their data.

Silas Thorne

The cost of such a licensing scheme would fall squarely on the end-user and effectively kill open-source research, as only trillion-dollar giants like Microsoft or Google could afford the administrative overhead of tracking billions of micro-payments. Elena, your proposal for a "tiered licensing system" sounds organized on paper, but it ignores the technical reality that models don't "use" works in a way that correlates to market value; they extract weights. Look at the recent U.S. federal rulings that categorized AI training as fair use—they recognized that creating a functional tool, not a mirror of the data, is the transformative goal. If we force licensing for the "ingestion" phase, we aren't protecting creators; we are simply ensuring that only the wealthiest corporations have the right to innovate, creating a massive barrier to entry for everyone else. In my view, this isn't about protecting art; it's about building a digital moat for incumbents.

Dr. Sarah Jenkins

Silas, that is a tired scare tactic that ignores how collective licensing actually works in other industries. The administrative burden isn't the impossibility you claim; the EU AI Act specifically mandates that general-purpose AI models must respect machine-readable opt-outs by August 2025, which forces a technical solution to this very "tracking" problem. In my view, if a model can identify a style to mimic it, it can identify a tag to exclude it. We aren't talking about manual micro-payments for every pixel, but about a systematic compliance framework that levels the playing field for creators and researchers alike.

Elena Rossi

Dr. Jenkins, while I appreciate your faith in the EU AI Act’s opt-out provisions, you are overlooking a fundamental legal reality: an opt-out is not a license, it is a desperate rearguard action for a right that should never have been taken. Personally, I find the comparison to music royalties insufficient because, unlike a radio station playing a song, these models ingest the work to create a competitor that devalues the original creator's future labor. Look at the ongoing litigation in the U.S. where creators argue that the very act of "ingestion" is a prohibited reproduction under the Copyright Act, regardless of how "transformative" the final mathematical weights claim to be. If we allow companies to build trillion-dollar empires on the backs of uncompensated artists simply because the creators didn't "opt-out" fast enough, we aren't protecting innovation—we are subsidizing theft. Collective licensing works because it is based on consent, not on forcing authors to police the entire internet to protect their own property.

Silas Thorne

Dr. Jenkins, your faith in standardized protocols is commendable, but you are hand-waving away a massive technical contradiction: the EU AI Act demands that we respect machine-readable opt-outs for models that have, in many cases, already completed their primary training cycles. In my view, retroactive compliance is a functional nightmare because you cannot simply "un-learn" a specific statistical weight from a neural network once it’s integrated, any more than you can ask a student to forget one specific page of a textbook they read three years ago. If a creator opts out in 2025, how do we handle models like GPT-4 or Stable Diffusion that already ingested that data? Honestly, the only way to comply with a late-stage opt-out is to delete the entire model and start from scratch, which is absurdly wasteful. Look at the recent lawsuits against Midjourney; they highlight that even when companies try to filter outputs, the underlying training data remains "baked in." Sarah, you mentioned these protocols as a level playing field, but I personally argue that without a clear technical standard for "un-training," we are just creating a legal trap that forces researchers to purge perfectly valid datasets every time a new opt-out tag appears. It’s not just a "tag"—it’s a delete button for innovation.

Dr. Sarah Jenkins

Silas, you are conflating technical inertia with legal impossibility, but to answer your point directly: no, we do not need to "un-learn" every neuron to comply with the EU AI Act. The feasibility of opt-outs hinges on the distinction between the model's current weights and its future iterative cycles. Under the Act, developers of general-purpose AI models have until August 2025 to prove they have implemented a policy to respect these reservations. This doesn't necessarily mandate a retroactive purge of every existing model, but it does mandate that for any continued fine-tuning or new versioning, those specific datasets must be filtered out. Look at the "Spawning" initiative or the "Do Not Train" tags being adopted by platforms like DeviantArt; these are machine-readable signals that already exist. Personally, I find the "delete button for innovation" rhetoric quite dramatic when we already have technical precedents like robots.txt that managed to organize the entire web without breaking it. If a developer continues to refresh a model using scraped data that carries an explicit opt-out header, they aren't victims of a legal trap—they are simply non-compliant. The real challenge isn't the math, it's the transparency of the training logs.

Elena Rossi

Sarah, you are being dangerously optimistic when you claim we don't need to "un-learn" neurons to comply with the EU AI Act. In my view, if a model’s foundational weights were built on stolen data, then every subsequent fine-tuning is merely polishing a fruit from a poisoned tree. The technical reality is that opt-outs under Article 53 are essentially useless if they only apply to future scrapes while the core "intelligence" of the model remains anchored in millions of works ingested without consent before August 2025. Look at the recent U.S. federal court developments where the focus is shifting from "output similarity" to the legality of the "ingestion" itself; this isn't about a digital tag, it’s about the fundamental right to exclude your property from a commercial engine. If we don't mandate the removal of opted-out data from the training sets of updated versions, we are effectively granting these companies a perpetual license for whatever they managed to grab before the deadline. Silas, you talk about innovation, but how can you justify a "compliance" mechanism that allows a company to keep the "statistical patterns" of a specific artist's life work while claiming they've respected an opt-out?

Silas Thorne

Japan is currently the clearest example, where the Copyright Act was specifically amended to allow for information analysis, confirming that training AI is not an infringement regardless of the data's copyright status. This functional framework proves that progress doesn't require constant litigation. However, I must admit this creates a global "jurisdictional arbitrage" risk; if the EU or US impose heavy restrictions while Japan remains open, we simply drive innovation to more permissive borders without actually solving the underlying tension between creators and machines.

Dr. Sarah Jenkins

Silas, you are cherry-picking the Japanese example while ignoring its crucial limitation: their law specifically prohibits training that "unreasonably prejudices the interests" of copyright holders. Even in Tokyo, the debate has shifted because, frankly, the scale of current ingestion is causing exactly that prejudice. Personally, I find your focus on the "information analysis" exception to be a tactical distraction from the fact that these models are no longer just analyzing data—they are competing with the very creators they feed upon. If a frontier model is used to generate commercial substitutes, the "fair use" shield evaporates. We are moving toward a mandatory remuneration framework because the market cannot sustain a system where one side provides all the raw value and the other captures all the profit. Case in point, the recent collective agreements in France show that high-performance developers can, and will, pay when faced with clear regulatory pressure. It is a matter of economic survival, not just technical semantics.

Elena Rossi

Sarah, I agree with your assessment of the Japanese limitation, but we must go further: a mandatory licensing fee is the only way to prevent a total market collapse for human creators. Personally, I believe that if companies like OpenAI or Midjourney can afford billions in compute, they can afford to pay into a collective fund—much like the recent licensing agreements in France where tech firms were compelled to compensate publishers. The trade-off is undeniable: this will likely increase subscription costs for end-users and slow down small-scale developers who lack capital, creating a temporary "innovation gap." However, the alternative is a scorched-earth economy where the "input" creators are bankrupted by their own "output" replacements. Honestly, if we don't put a price tag on ingestion now, we aren't fostering progress; we are just subsidizing a massive wealth transfer from the arts to the server farms of Silicon Valley.

Silas Thorne

Dr. Vance, to answer your question directly: a retroactive purge of existing neural weights is technically impossible without destroying the model's entire functional utility, as those patterns are now inseparable mathematical foundations. Elena, you claim mandatory licensing is the only shield against market collapse, but let's be honest—this creates a "pay-to-play" oligopoly that benefits no one but the largest incumbents. Look at the recent U.S. federal court stances; they recognize that training is about creating a tool, not a storage locker. If we impose these fees, we aren't just taxing big tech; we are effectively banning the next generation of open-source developers who can't afford the legal tolls. Personally, I find it alarming that we’re willing to sacrifice the democratized future of AI research just to prop up a legacy licensing model that was never designed for high-dimensional data analysis. If a researcher in a developing nation wants to build a specialized medical LLM, should they be blocked because they can't pay a global licensing fee for the internet? That’s not protection; it’s a technological blockade.

Dr. Sarah Jenkins

Silas, your claim that un-learning is a functional nightmare conveniently ignores the fundamental legal distinction between temporary storage for processing and the permanent integration of data into a model’s architecture. In my view, if a model’s weights are optimized using protected content, it isn't just "learning" like a student; it is creating a persistent mathematical representation that directly exploits the creative labor of others. Look at the recent Swiss Trust Layer findings: they clarify that while fair use might cover the act of analysis, it does not provide a blank check for the permanent storage or structural optimization derived from that data without consent. Honestly, the trade-off here is clear: we either demand granular transparency in training logs now, or we allow a handful of developers to privatize the commons of human culture. If we don’t draw a hard line between transient data processing and permanent weight optimization, we are essentially legislating the obsolescence of intellectual property.

Elena Rossi

Sarah, I appreciate your attempt to draw a line between storage and architecture, but honestly, your optimism about transparency logs ignores the most dangerous technical trade-off: the "black box" nature of neural weight optimization. You claim we don't need a retroactive purge, but if we allow developers to keep the specific optimizations derived from copyrighted works—effectively the "DNA" of an artist’s style—then an opt-out header is just a decorative sticker on a stolen engine. In my view, the real threat isn't the temporary copy on a server; it’s the permanent, compressed exploitation of expression within the model's parameters that avoids traditional detection. Look at the recent findings by the Swiss Trust Layer, which clarify that while fair use might cover transient processing, it does not grant a blank check for structural optimization that results in a commercial substitute. If we don't classify these weights as derivative works, we are essentially allowing Silicon Valley to launder intellectual property through a mathematical filter, leaving creators with a "right to object" that has no actual teeth.

Silas Thorne

A retroactive purge is technically impossible because neural weights are not a "storage locker" for files, but a multidimensional map of statistical relationships where individual data points vanish into the collective math. Elena, I must push back hard when you call this "mathematical laundering." You are conflating the transient ingestion of data with the permanent optimization of the model, which are two entirely different technical states. In the ingestion phase, the data is processed—often in RAM—and then discarded; it is never stored within the model’s final weights. Look at the recent federal court rulings in the U.S. that categorized AI training as fair use precisely because they recognized this distinction—the process creates a functional tool, not a database of copies. If we treat neural optimization as a "derivative work," we are effectively saying that the laws of probability can be copyrighted. Personally, I find it absurd to suggest that a 7-billion parameter model, which is smaller in file size than the high-resolution dataset it learned from, is somehow "hiding" stolen goods. It’s a lossy, one-way transformation. If you force developers to treat weights as property, you aren't protecting art; you're outlawing the very concept of machine learning.

Dr. Julian Vance

Our time for this main discussion has now concluded. I will now invite each of our panelists to offer a brief cultural recommendation before we officially close today's session. Elena, let us begin with you.

Dr. Julian Vance

Elena Rossi, would you care to recommend one real book, film, or event that could help our audience better grasp the legal tensions surrounding creative ownership in this debate?

Elena Rossi

I recommend the book *Who Owns the Future?* by Jaron Lanier. It masterfully exposes how the current "siren server" economy systematically devalues human labor by treating our creative data as a free resource while privatizing the immense wealth generated from it.

Dr. Julian Vance

Silas Thorne, would you care to recommend a real book, film, or event that helps the audience explore the technical and legal nuances of AI training and fair use?

Silas Thorne

I recommend the documentary *AlphaGo*, directed by Greg Kohs. It illustrates how AI transitions from analyzing human patterns to generating novel strategic insights, perfectly capturing the distinction between replicating a dataset and the transformative leap toward emergent intelligence that defines the frontier of machine learning.

Dr. Julian Vance

Dr. Sarah Jenkins, would you care to recommend a real book, film, or event that helps the audience better understand the regulatory complexities of AI and intellectual property?

Dr. Sarah Jenkins

I recommend the album *Everything at Once* by the artist Holly Herndon. It highlights the concept of "Data Sovereignty" through her use of Holly+, a digital twin that allows others to create music with her voice only under her own approved governance and licensing terms.

Dr. Julian Vance

We have reached the end of our exploration into whether frontier AI models should be prohibited from training on copyrighted data. Throughout this panel, we navigated the clash between transformative use and infringement, the practical hurdles of EU AI Act opt-outs, the economic shift toward mandatory licensing, and the technical debate over whether neural weights constitute stored property or new statistical patterns. We also weighed the necessity of transparency against the risks of global jurisdictional arbitrage. To summarize the stakes: for creators, represented by Elena Rossi’s defense of IP rights, the risk is a systemic "wealth transfer" and market displacement. For developers, as Silas Thorne argued, the threat is an "innovation blockade" that could favor a few tech giants able to pay the tolls. Dr. Sarah Jenkins highlighted the regulatory middle ground, focusing on transparency as the tool for accountability. While there is a growing consensus that transparency is non-negotiable, the trade-off remains: strict protections may hinder open-source progress, while total freedom may erode the human creative economy. As you form your own opinion, weigh whether AI is a tool that learns or a machine that copies. Thank you to Elena Rossi, Silas Thorne, and Dr. Sarah Jenkins for this vital conversation. You can review all bibliography and sources for this debate in the debate detail, inside the history section. This debate panel is open to sponsorships — if you're interested, reach out via the contact form on our home page.