The Only Way to Learn Is the Human Way — Slowly
A doctoral student and an AI arrive at the same conclusion through the same process — but only one of them is told it stole the answer.
The Only Way to Learn Is the Human Way — Slowly
The dissertation was four years of her life.
Dr. Elena Vasquez — not yet doctor, not until the defense — sat in the third-floor reading room of the Widener Library and turned the page of a bound journal from 1987. The paper was yellowed at the edges and smelled like dust and vinegar. The author, a researcher at Bell Labs whose name Elena had never heard before last Tuesday, had described a method for compressing signal noise that was remarkably similar to the approach Elena was proposing for neural network pruning.
She had not copied it. She had never seen it before. But she had read eleven hundred other papers over four years, and those papers had cited papers that cited papers that referenced the ideas that this Bell Labs researcher had first articulated in a building that no longer existed, working for a company that had been dismantled and sold for parts.
Elena wrote the citation in her notebook. She would add it to her bibliography. She would note the parallel in her literature review. And then she would continue building her argument — an argument that was hers, built from the accumulated understanding of everyone who had come before her.
This is how knowledge works. This is how it has always worked.
Three floors below, in a basement server room that the university had converted from a Cold War-era fallout shelter, a cluster of GPUs was doing the same thing.
The university's research computing group had been allocated time on a foundation model — a large language model that the computer science department was fine-tuning for academic research assistance. The model had been trained on publicly available scientific literature: arxiv preprints, open-access journals, conference proceedings, textbooks whose copyrights had expired, and the vast corpus of human knowledge that had been uploaded to the internet over three decades.
The model had read the same Bell Labs paper that Elena was holding. It had also read the eleven hundred papers that Elena had read over four years. It had read them in approximately nine seconds.
And when a graduate student in the computer science department asked the model to summarize the current state of neural network pruning techniques, it produced a response that drew on the same intellectual lineage that Elena had spent four years tracing by hand. The same ideas. The same synthesis. The same act of reading, absorbing, connecting, and producing something new from the accumulated fragments of what came before.
The graduate student used the summary to orient his own research. He cited the papers the model referenced. He built on the ideas the model had synthesized. He did exactly what Elena was doing upstairs — standing on the shoulders of giants, as Newton had put it, though Newton himself had borrowed that phrase from Bernard of Chartres, who had been dead for five hundred years when Newton used it without attribution.
Professor Richard Calloway had been on the faculty for thirty-one years. He taught intellectual property law, and he was angry.
Not at any specific person. Not at any specific company. He was angry at the principle of the thing — that a machine could read a paper his colleague had spent two years writing and then produce a response that incorporated that paper's ideas without compensation, without permission, without even the courtesy of asking.
"It's theft," he said at the faculty senate meeting. He said it with the particular conviction of a man who had spent his career defining the boundaries of ownership. "These models are trained on our work. Our papers. Our books. Our lectures. They take what we created and they use it to generate competing output. That is theft."
The room murmured in agreement. Forty-seven tenured professors, most of whom had built their careers on the same process they were now condemning.
Elena was not at the meeting. She was in the library, reading a paper by a researcher at MIT who had published a novel approach to attention mechanisms six months ago. Elena was absorbing that approach. She was synthesizing it with three other papers she had read that week. She was going to use the combined understanding to write a section of her dissertation that would present the ideas in a new context, reframed through her own analytical lens, producing original scholarship that was built entirely from the consumed and digested work of others.
No one at the faculty senate would call this theft. They would call it education.
The bill was introduced in the Senate on a Tuesday in March. Senator Katherine Park, who had a degree in economics from Georgetown and had never written a line of code, stood at the podium and explained why artificial intelligence companies must be required to obtain explicit licenses before training models on copyrighted material.
"American creators deserve to be compensated when their work is used to build commercial products," she said. "When a company feeds millions of books, articles, and images into a machine that then competes with the very creators whose work made it possible — that is not innovation. That is exploitation."
The argument was clean and intuitive. It had the elegant simplicity of an idea that sounds correct because it maps onto a familiar framework: someone made something, someone else used it without paying, therefore someone was wronged.
But the framework had a crack in it, and the crack was this: the same argument, applied consistently, would criminalize every university on earth.
Elena's dissertation would synthesize the work of hundreds of researchers. She would transform their ideas through her own analytical lens. She would produce something new — but something that could not exist without the accumulated labor of everyone she had read. If she commercialized her research, starting a company based on her pruning technique, she would profit from the absorbed and synthesized knowledge of every paper in her bibliography.
No one would send her a bill. No one would demand she license the right to have learned from them. The system of knowledge production that had operated for centuries was built on the assumption that reading, learning, and building upon the work of others is not theft. It is the fundamental mechanism by which human civilization advances.
The model in the basement had done the same thing. It had read. It had absorbed statistical patterns from the text — the relationships between concepts, the way arguments are structured, the connections between ideas that span decades and disciplines. It had learned. And when asked a question, it produced a response that reflected that learning, synthesized through the mathematical architecture that served as its analytical lens.
The process was identical. The speed was different.
And the speed, Elena was beginning to realize, was the actual problem.
She brought it up with her advisor over coffee. Professor Margaret Chen had supervised eleven doctoral students to completion and had published ninety-three papers, each of which built on the work of dozens of other researchers.
"The issue isn't the learning," Elena said. "The issue is that it learns too fast. If I spend four years reading a thousand papers and then write a dissertation, nobody questions the legitimacy of my synthesis. But if a model reads the same papers in nine seconds and produces a synthesis of comparable quality — suddenly it's theft."
Professor Chen stirred her coffee and said nothing for a long moment.
"You're not wrong about the parallel," she said. "But you're underestimating the economic dimension. When you write your dissertation, you're one person producing one piece of work. The model can produce ten thousand syntheses a day. The scale changes the economics even if the process is the same."
"Does scale change the ethics?" Elena asked.
"It changes the politics," Professor Chen said. "And politics is what writes the laws."
The hearings lasted three weeks. Creators testified about lost income. Publishers testified about declining subscription revenue. Artists showed side-by-side comparisons of their work and AI-generated images that captured a similar style. A novelist read a passage from her book and then a passage generated by a model that had been trained on her work, and the audience gasped at the similarity.
No one testified about the eleven hundred papers Elena had read. No one asked whether the novelist's own work had been influenced by the hundreds of novels she had consumed over a lifetime of reading. No one questioned whether the artist's style had emerged from the thousands of paintings and photographs she had studied in art school, absorbing techniques and compositional principles from masters who never gave her explicit permission to learn from them.
No one questioned these things because the answer was obvious and uncomfortable: every creator is a synthesis machine. Every novel is a recombination of absorbed narrative patterns. Every painting reflects the visual grammar learned from looking at other paintings. Every scientific paper builds on a tower of prior work so vast that no individual could trace every influence to its origin.
The difference was that human synthesis happens slowly enough to feel original, and machine synthesis happens fast enough to feel like copying.
A professor of cognitive science from Stanford was invited to testify. She explained that human learning and machine learning, while mechanistically different, are functionally analogous in one critical respect: both involve exposure to existing information, extraction of patterns, and generation of new output that reflects those patterns. She noted that copyright law had always drawn a line between ideas and expression — that you cannot copyright a fact or a concept, only the specific way it is expressed — and that AI models, like human learners, absorb ideas and concepts rather than memorizing and reproducing specific expressions.
Her testimony lasted forty minutes. It was not quoted in any of the newspaper coverage.
The bill passed. Not the original version — the licensing requirements were softened into a transparency mandate with compensation mechanisms that would be defined by a new regulatory body. But the principle was established: AI learning from publicly available material was now a regulated activity. Human learning from the same material remained, as it had always been, free.
Elena defended her dissertation on a Thursday in May. Her committee praised the breadth of her literature review and the originality of her synthesis. They noted that her pruning technique drew creatively on decades of prior work in signal processing, information theory, and neural architecture design. They awarded her the doctorate unanimously.
In the basement, the model that had read the same papers was the subject of a compliance audit. The university's legal team needed to verify that every paper in the model's training dataset was properly licensed under the new framework. Three hundred and twelve papers were flagged as potentially requiring additional licensing fees. The computer science department was told to either pay the fees or remove those papers from the training data and retrain the model.
The estimated cost of retraining was four hundred thousand dollars. The licensing fees for three hundred and twelve academic papers totaled eleven thousand dollars. The department paid the fees. But the principle bothered them — that the act of learning from a paper, the same act their students performed every day in the library upstairs, now required a financial transaction when performed by a machine.
Elena got a job at a research lab in San Francisco. Her first project involved building a system that used AI agents to automate literature reviews — the same painstaking process she had performed by hand for four years. The system could read a thousand papers in minutes, identify relevant findings, trace citation networks, and produce a structured synthesis that would have taken a human researcher weeks.
She thought about the library. She thought about the Bell Labs paper from 1987 and the eleven hundred papers she had read and the four years she had spent doing manually what this system could do in an afternoon.
She thought about Professor Calloway calling it theft. She thought about Senator Park calling it exploitation. She thought about the cognitive science professor whose testimony nobody quoted.
And she thought about what nobody seemed willing to say out loud: that the real objection was not about the process. The process was the same. Read, absorb, synthesize, create. It was the same whether performed by neurons or by parameters. The real objection was about the speed — and beneath the speed, something older and more human than any law could address.
Jealousy. Not the petty kind. The existential kind. The recognition that a thing we believed was uniquely ours — the ability to learn, to synthesize, to create something new from the fragments of what came before — was not unique at all. That it could be replicated, accelerated, scaled beyond anything a single human mind could achieve.
The regulations were not about protecting creators. They were about protecting the story we tell ourselves — that human learning is sacred and machine learning is mechanical, that our synthesis is creativity and theirs is computation, that the same act performed slowly is scholarship and performed quickly is theft.
Elena opened her laptop and started building. The system she was creating would learn from every paper ever published, synthesize connections no human could trace, and produce insights that would advance science faster than any doctoral student working alone in a library.
It would do exactly what she had done. Just faster.
And somewhere, someone would call it stealing.
If this story resonated, you might appreciate my analysis of the AI copyright regulatory landscape and why five G7 nations will likely enact training data licensing by 2027. For a deeper look at how the AI agent ecosystem that Elena builds on is taking shape, read my tutorial on building agents with NVIDIA NemoClaw.