Could the way your generative artificial intelligence (GenAI) tools are trained expose you or your clients to infringement claims? If this question unsettles you, you’re not alone. The popularity of AI tools continues to increase, to the point that even regulators and bar associations are providing guidance on navigating the associated legal and ethical challenges. And, while many of us are eagerly embracing the benefits of these tools, their adoption has also sparked a wave of intellectual property questions and disputes over pretraining data. These debates highlight questions about ownership, authorship and ethical use, and could impact both AI liability and intellectual property frameworks––making it critical for legal professionals to pay attention.
Before GenAI tools reach end users, they are pretrained on large datasets. But where does the training data behind your AI come from? And does its use infringe on the rights of others?
Improperly sourced training data could lead to copyright infringement claims, privacy violations or ethical breaches. Additionally, secondary liability is becoming an increasingly prominent issue. Courts and regulators are scrutinizing the platforms, developers and corporate officers involved in creating and deploying these tools. With the rapid inclusion of GenAI features in legal technology tools, this expanding scope highlights questions about contributory infringement, vicarious liability and corporate accountability.
As the legal landscape surrounding these issues evolves, identifying and proactively addressing these risks is critical. Attorneys must grapple with how to counsel clients using or building AI tools while ensuring their own AI use remains above board.
Before GenAI models reach end users, they must undergo pretraining––or, the “P” in GPT generative pretrained processor). Pretraining provides the data-based knowledge needed for the model to generate human-like text, create images or perform other sophisticated tasks.
So, how does pretraining work? It involves collecting and providing data to an AI model. The large nature of these datasets enables a model to identify patterns, define context and predict favorable outputs. Analyzing billions of textual examples from books, articles, websites and anything else placed in the training set enables large language models to identify how words, phrases and sentences relate to one another.
Training data provenance refers to the origin of the datasets used in pretraining. In other words, where did the data come from, and was its use authorized? For attorneys, this issue strikes at the heart of their obligations to practice law ethically and competently.
What makes pretraining so powerful is also what makes it precarious: the nature of the data itself. Collecting the large datasets needed for pretraining has largely taken place within a legal gray area, with huge quantities of data being scraped from publicly available sources without clear licensing or consent.
Models are often trained on publicly available and acquired datasets that can include:
The scale of this data allows models to generalize across topics, but it also opens the door to serious intellectual property and ethical concerns. Though AI developers do sometimes acquire datasets through licensing agreements with publishers or data aggregators, licensing can be prohibitively expensive, leading many to opt for scraping, instead.
Scraped datasets can include copyrighted works, proprietary business information, sensitive personal data or even data from minors. So, what happens when the training data includes copyrighted materials or sensitive personal information? Who is accountable if the outputs of the model closely mimic a protected work?
When it comes to training data, ignorance isn’t a defense—it’s a liability. Training datasets that are unlicensed, sensitive or ethically questionable pose risks for everyone involved in the AI ecosystem, affecting both legal compliance and ethical accountability.
When training datasets include copyrighted or unlicensed material, they risk producing outputs that infringe on intellectual property (IP) rights. This exposure affects all participants in the AI ecosystem:
Unregulated use of personal or sensitive data in training datasets raises ethical dilemmas.
These concerns have real-world implications that attorneys must address when using, recommending or providing counsel surrounding AI technologies.
Disputes over training data are raising pressing questions that could redefine intellectual property and liability frameworks. Does scraping copyrighted material for training purposes constitute infringement, or does it fall under the fair use exception? How can we ensure that training data is legally and ethically sourced? And what happens when AI outputs closely mimic protected works?
Several high-profile cases are beginning to define the boundaries of these issues:
The outcomes of these cases may have far-reaching implications, influencing the evolution of IP law and defining how compliance and liability are assigned across the entire AI ecosystem.
Fair use is the legal doctrine that permits limited use of copyrighted material without permission under certain circumstances, primarily for purposes like commentary, criticism, education or research. Courts assess fair use based on four factors.
Developers may argue that using copyrighted material for AI pretraining qualifies as transformative. But whether training transforms the data enough to meet the standard is up for debate.
Alarmingly, we are running out of authentic human-generated data to use during pretraining. The availability of high-quality training data is becoming a bottleneck for AI development. Early models thrived on the abundance of freely accessible internet data, but the pool of usable training data is shrinking. Not only can we not generate it quickly enough, but awareness around intellectual property and privacy issues is growing.
Companies like OpenAI are now paying for licensed datasets from publishers—a significant shift that highlights the increasing difficulty of acquiring training data legally and ethically. So, what happens when the well of high-quality, original data runs dry? The answer could fundamentally change the trajectory of AI development—and not for the better.
As the pool of authentic data diminishes, many AI developers are turning to synthetic data—data generated by other AI models—to fill the gap. While this approach can work for a time, it introduces a significant risk: model collapse.
Unlike training on original, diverse datasets, synthetic data can create a feedback loop of diminishing quality, as errors and biases compound over successive generations. Without access to fresh, reliable data, the robustness and reliability of AI systems could erode, undermining their utility.
To address the challenges posed by dwindling authentic training data, researchers are exploring techniques like Monte Carlo tree search (MCTS). MCTS, widely used in game theory and decision making, is a heuristic algorithm that searches for the best outcomes by simulating various decision paths and evaluating their potential results. This approach can introduce variability into AI training without relying on external datasets.
Unlike traditional Monte Carlo simulations that rely on randomness to model outcomes, MCTS builds a "tree" of possible decisions and outcomes. Each node in the tree represents a possible state, while branches represent decisions leading to other states. The algorithm iteratively simulates these paths, updates the tree with the most promising outcomes and refines its decision making over time. Key steps include:
MCTS could be used to simulate complex decision trees or scenarios, allowing models to learn from synthetic yet structured data. There are, however, some criticisms of this technique:
What legal professionals should take away from this is that the discussion of these concepts signals a shift in how AI models will be trained in the future, and how liability questions may play out. If pretraining doesn’t rely solely on human-generated data, then IP considerations will be due for yet another legal shift.
As we embrace the benefits of AI, we must also grapple with its challenges, particularly around training data provenance, intellectual property and ethical standards.
The way AI models are pretrained isn’t so different from how humans acquire knowledge: they process patterns, identify relationships and build from experience. People grow and benefit from their experience––we cannot keep them from relying on it when they make art or generate intellectual property. But there are also temporal and emotional aspects that come along with human learning—we breathe, feel, bleed and suffer as we acquire knowledge and experience. AI, on the other hand, operates at immense scale and speed, eating vast datasets without pause. This ability magnifies its potential, but also its risks. Creating legal and ethical frameworks that govern our use of this technology demands immediate and careful consideration.
As attorneys, we can influence sustainable AI adoption frameworks.
As regulators demand greater transparency and companies turn to licensed datasets, the legal profession will be essential in shaping how these tools are adopted and used responsibly. Attorneys have a unique opportunity to aim for innovation that balances progress with protection—advancing technology responsibly while safeguarding intellectual property, data privacy and ethical standards. AI may be rewriting the rules, but the legal profession must stay the author of accountability.
