How AI Models Steal Creative Work — and What to Do About It | Ed Newton-Rex | TED

Summary

The script discusses the ethical and legal issues surrounding the use of unlicensed creative works in training generative AI models. It highlights how AI companies use vast amounts of copyrighted material without permission, which negatively impacts creators and their livelihoods. The argument is made that licensing training data is a fair solution, one that can benefit both AI developers and creators. Examples are provided of AI models competing with creative work, including music and art, and the call for more responsible AI development that respects creators' rights is emphasized. The script advocates for a future where AI and human creativity can coexist harmoniously.

Takeaways

• 😀 Generative AI technology is impressive, but using unlicensed creative work to train AI models is unfair and unsustainable.
• 😀 AI companies require three resources to build their models: people (engineers), compute (GPUs), and data (training datasets).
• 😀 While AI companies invest heavily in engineers and computing power, they often expect to use creative work for free without compensating creators.
• 😀 Many AI companies use web scraping to collect unlicensed content for training, including copyrighted works like newspaper articles and music.
• 😀 Generative AI competes with the very content it was trained on, creating a direct threat to creators' livelihoods by generating competing works.
• 😀 Creators, such as musicians and artists, are already seeing their income decrease as AI-generated content starts to replace their own work in the market.
• 😀 There is an ongoing debate over whether AI training falls under the fair use exception of copyright law, but many creators argue it is being exploited unfairly.
• 😀 Licensing training data, like how commercial entities license content, could provide a solution that benefits both AI companies and creators.
• 😀 AI companies claim that licensing data is impractical, but smaller companies are already proving that it can be done using various models, such as revenue sharing.
• 😀 The rise of unlicensed AI training is causing websites to restrict access to their content, which ultimately harms innovation and access to data for AI models.
• 😀 Public opinion is strongly against unlicensed AI training, with a large majority supporting compensation for data providers and permission to use their work.

Q & A

What are the three key resources AI companies need to build their models?

The three key resources are people (engineers), compute (GPUs), and data (training data).

Why is it considered unfair for AI companies to use creative work for free in building their models?

It is considered unfair because AI companies expect to use training data without compensation or permission, while they pay for the other resources like engineers and GPUs. This practice negatively impacts creators who aren't compensated for their work.

How do AI companies typically acquire training data for their models?

AI companies often use web scrapers to gather as much data as they can find, frequently without asking for permission or paying creators. They also often do not disclose the specifics of what they train on.

What does 'generative AI competes with its training data' mean?

It means that AI models, which are trained on creative works, can generate content that competes with the original creators' work. For example, an AI trained on short stories can create competing short stories.

What are some examples of industries where generative AI has already started competing with human-created work?

Examples include AI music competing with human-produced music (e.g., AI music being used in film production or hitting music charts), and AI-generated art outcompeting human artists, as seen with Kelly McKernan's work.

What is the legal debate surrounding AI's use of copyrighted work for training?

The legal debate centers on whether AI training falls under the fair use exception of copyright law. Creators argue that this exception does not justify the large-scale, unlicensed use of their works to build commercial AI models.

What solution do creators propose to address the issue of unlicensed AI training?

Creators propose licensing the training data, just as businesses do when they use copyrighted works for other commercial purposes. This would ensure creators are compensated for the use of their work in AI training.

Why do some AI companies argue that licensing training data is impractical?

AI companies argue that licensing training data is impractical because they use massive amounts of data, and individual payments to creators would be too small to make licensing feasible. However, this claim is increasingly being challenged as more companies find ways to license data.

What is the role of Fairly Trained in the AI industry?

Fairly Trained is a nonprofit organization that certifies generative AI companies that do not use unlicensed copyrighted work for training. It highlights companies that have found ways to license their data, promoting ethical practices in the AI industry.

What are the public's views on AI companies using publicly available data for training?

Public opinion largely rejects the idea of AI companies using publicly available data without compensation. A poll found that 60% of people believed AI companies should not be allowed to use such data, and 74% agreed that AI companies should compensate data providers.