Understanding the Boundaries of Fair Use in AI

AI companies’ reliance on “fair use” to train their models is under fire, with lawsuits pushing for stricter copyright protection. If fair use fails, some models may even face complete shutdown, marking a major turning point for the industry.

Understanding the Boundaries of Fair Use in AI

by Mick Kiely, CEO of IAIAI Technologies

For years, AI companies have been leaning on the doctrine of “fair use” to justify their extensive
scraping of copyrighted material for training their models. They’ve argued that their use of
music, books, articles, and images is transformative and doesn’t harm the market for the original
works. But in recent months, this argument has begun to show more than a few cracks.

Over the past year, there has been a surge in AI-related lawsuits, with content creators, labels
and publishers taking a stand against what they see as blatant copyright infringement. Cases
like the New York Times vs. OpenAI lawsuit have rang out across multiple creative industries,
challenging the very foundation of how AI models are trained.

As courts consider the intricacies of AI training and its impact on intellectual property rights, the
“fair use” argument may well crumble. The sheer volume of copyrighted material used in training
AI models goes far beyond what has in the past been considered “fair use.” Also, many AI
models are used for commercial purposes, weakening the “fair use” claim, as profit is usually
not allowed under fair use. More important, however, is the fact that as AI-generated content
becomes more prevalent, it could directly compete with and devalue the very works on which it
was trained.

If the “fair use” argument collapses, many AI companies may find themselves in even deeper
waters. Their prized models, trained on vast amounts of what could arguably be considered
stolen copyrighted material, might become legal liabilities. Companies once hailed as pioneers
of AI innovation may suddenly look like old leaky pirate ships, drifting toward the edge of a world
they once insisted was flat, their AI models contaminated by copyright infringement.

The fall of “fair use” would likely force a massive restructuring of AI development. Companies
would have to develop new, ethically sourced datasets, invest heavily in legal defenses and
settlements, and even explore alternative training methods that don’t rely on copyrighted
material. This shift would be costly, time-consuming, and could push even leading AI companies
into an ice age of their own making.

But that’s only the tip of the proverbial iceberg. These models, like sponges, absorb every piece
of information they encounter during training. The neural networks that form the foundation of
these AI systems integrate this knowledge so deeply that it becomes an inseparable part of their
functionality. At present there’s no “undo” button—no way to selectively remove specific
influences without resorting to the drastic measure of destroying the entire model.

Imagine a student who has memorized a book word for word. Now, try to make that student
“unlearn” specific passages—it’s impossible. This is precisely the predicament we face with AI
models trained on unlicensed copyrighted material. While research is being done to develop
methods for selectively removing specific influences, there is no silver bullet yet, and time is
running out as damage continues to be done to artists.

All of this raises a huge problem for both AI companies and rights holders: the inability to “opt
out”. Once a work has been ingested by an AI model, its influence cannot be removed,
regardless of the copyright holder’s wishes.

We would see the emergence of hybrid generative AI models trained on both legally licensed
and unauthorized copyrighted material, thus presenting a whole new problem. These hybrid
models would straddle the line between legality and infringement and pose greater challenges
for developers, users, and copyright holders. The technical challenges of data separation and
output attribution, between legitimately sourced and illicitly obtained training data would be even
more challenging.

Considering the impossibility of selective unlearning and the continued violation of copyright
holders’ rights, we may be left with only one option: completely destroying and rebuilding
contaminated AI models from the ground up, using only permitted or licensed materials, as they
should have been in the first place.

This may seem extreme, but it could be the only way to set things right and uphold the integrity
of copyright law and ethical standards. An alternative might involve novel innovations, such as
developing systems that can detect, attribute, and compensate training sources at the time an
AI work is generated, or creating methods that truly reflect the concept of ‘opting out.’.

One thing is clear: the future of AI and intellectual property still hangs in the balance. In the
interim, we must act to ensure that the rights of creators are upheld while AI searches for firmer
ground, ensuring a future that truly benefits humanity.