OpenAI is once again in the spotlight—this time for allegedly training its AI models on copyrighted material without permission. While the company has long argued that its methods fall under fair use, a recent study casts serious doubt on that defense.
The paper, co-authored by researchers from Stanford, the University of Copenhagen, and the University of Washington, introduces a novel way of testing whether AI models “memorize” parts of their training data. Their findings? Some of OpenAI’s most advanced models, including GPT-4 and GPT-3.5, appear to have internalized portions of copyrighted books and articles.
OpenAI is already battling multiple lawsuits. Authors, software developers, and content creators have accused the AI lab of using their intellectual property—books, software code, journalistic content—to train models without their consent. Although OpenAI continues to lean on a fair use argument, plaintiffs maintain that U.S. copyright law doesn’t include a carve-out for training large-scale AI.
So, what's the truth? That’s where this new research gets interesting.
To get to the bottom of it, the study’s authors devised a clever technique involving “high-surprisal” words—those unexpected, contextually unique terms that don’t show up often in most writing. Think of a sentence like:
“Jack and I sat perfectly still with the radar humming.”
Here, the word “radar” is statistically less predictable than, say, “engine” or “radio,” making it a high-surprisal word.
The researchers masked such words in short passages from fiction and articles, then asked models like GPT-4 and GPT-3.5 to fill in the blanks. If the models guessed correctly, it suggested they had likely seen the exact snippet—or something strikingly similar—during training.
The results were revealing, to say the least. GPT-4 demonstrated an unusually high ability to recall high-surprisal words from sections of copyrighted novels and well-known New York Times articles.
Specifically, the model showed evidence of memorizing content from BookMIA, a dataset known to include copyrighted e-books. While memorization rates were lower for news content, the fact that it happened at all raises red flags.
According to study co-author Abhilasha Ravichander, a Ph.D. student at the University of Washington, the work provides an early step toward holding AI models accountable.
“If we want trustworthy large language models,” she told TechCrunch, “we need to be able to audit and scientifically evaluate how they learn and what they retain.”
Meanwhile, OpenAI continues to push back against tighter regulation. The company has signed some content licensing deals and rolled out opt-out options for publishers and rights-holders—but it’s also lobbying governments to support broader fair use laws for AI model training.
Their stance? Innovation in AI shouldn’t be stifled by restrictive copyright frameworks.
Critics argue this mindset risks setting a dangerous precedent—one where creative work can be repurposed by machines without meaningful consent or compensation.
As generative AI tools become increasingly mainstream—from writing assistants to image generators—the conversation around ethical data sourcing is heating up. Transparency, especially around training data, is becoming a central demand from policymakers, researchers, and creators alike.
If companies like OpenAI want to build public trust, they may have to move beyond legal defenses and start embracing open auditing practices. After all, no one wants a black-box system that quietly memorizes your novel—or your news article—and spits it out to someone else.