Data in Dispute: The Conflict of Training AI with Intellectual Property

View on LinkedIn

During a recent Thanksgiving dinner, I found myself wrestling with an old recipe of my grandmother's. The ingredients listed belonged to brands that had disappeared decades ago, the instructions mentioned a hand mill that my modern kitchen has never known, and the directions called for an oven requiring 20 minutes of preheating. "Do people still do that?" I asked myself, while trying to translate this culinary legacy into the age of smart ovens and electric food processors.

This seemingly trivial experience reflects a far deeper dilemma facing our society: how do we adapt the past to our present? And more specifically, how do we manage the collision between traditional copyright law and artificial intelligence's voracious appetite for data?

The numbers are simply overwhelming.

Meta has just announced that LLaMA 3, its latest AI model, was trained on more than 15 trillion tokens. To put this astronomical figure in perspective, imagine 193 complete Vasconcelos Libraries — every book, every page, every word — processed and analyzed.

GPT-4, according to recent leaks, is not far behind with its 13 trillion tokens. We are talking about the largest compilation and processing of human knowledge in history.

But here is the irony: just when we need more data than ever, we face a double crisis.

According to projections from the Stanford AI Index, we could exhaust the stock of high-quality linguistic data by 2026. It is as if we were burning digital fossil fuel at an unsustainable rate.

Meanwhile, legal battles multiply. The New York Times has sued OpenAI, arguing that the unauthorized use of its articles to train ChatGPT constitutes a massive copyright infringement. Getty Images is fighting Stability AI over the use of its photographs. It is a war being waged on multiple fronts, with implications that go far beyond simple corporate disputes.

Some have found a middle ground. The Associated Press signed a multimillion-dollar agreement with OpenAI, setting a precedent for content monetization in the AI era. Other media giants such as News Corp, the Financial Times, and Axel Springer have followed suit. But this raises another concern: are we creating a world where only big tech companies can afford to train AI models?

History offers some clues as to how this conflict might be resolved. In the 1970s, the arrival of the photocopier sparked similar debates about access to knowledge. A landmark case was Williams & Wilkins Co. v. United States (1973), in which a publisher sued the government for allowing the photocopying of medical publications. The final decision favored access to knowledge over particular commercial interests.

Decades later, the music industry faced its own crisis with the arrival of Napster and peer-to-peer networks. What appeared to be the end of the music industry eventually led to innovative business models such as Spotify and Apple Music. However, the current situation with AI presents unique challenges in scale and complexity.

In the end... what do I think will happen? History has repeatedly taught us that technological progress and the common good ultimately prevail over particular interests. As happened with the photocopier and digital music, AI represents an unprecedented inflection point that will inevitably find its way toward collective benefit.

The future will be shaped along three fundamental dimensions:

1. Changes to the legal framework

New AI-specific legislation that responds to the unique challenges of large-scale data training Fair compensation mechanisms that balance the interests of creators and developers Legal concepts adapted to the digital age, recognizing new forms of content use and transformation

• New AI-specific legislation that responds to the unique challenges of large-scale data training
• Fair compensation mechanisms that balance the interests of creators and developers
• Legal concepts adapted to the digital age, recognizing new forms of content use and transformation

2. New business models

Collective licensing that democratizes access to quality training data Automated compensation through transparent and verifiable systems New specialized intermediaries that facilitate rights management in the AI era

• Collective licensing that democratizes access to quality training data
• Automated compensation through transparent and verifiable systems
• New specialized intermediaries that facilitate rights management in the AI era

3. Technological adjustments and advances

Rights-respecting training that integrates attribution by design Better attribution and tracking through blockchain systems and smart contracts Transparency standards that allow for auditing and verifying the ethical use of data

• Rights-respecting training that integrates attribution by design
• Better attribution and tracking through blockchain systems and smart contracts
• Transparency standards that allow for auditing and verifying the ethical use of data

Like that old recipe of my grandmother's, we will find a way to adapt what is essential to the new times. And although the path will be complex, history suggests that the final result will not only preserve the value of human knowledge, but amplify it in ways we can barely imagine today.