Suchir Balaji spent almost four years as an artificial intelligence researcher at OpenAI, where he played a role in collecting and organizing vast amounts of internet data for the development of its chatbot, ChatGPT.

At that time, he didn’t seriously question whether the company had the legal authority to use this data. He assumed that the San Francisco-based startup could utilize any internet data, regardless of copyright status.

However, after ChatGPT’s launch in late 2022, he reconsidered the company’s practices. He concluded that OpenAI’s use of copyrighted data was unlawful and that technologies like ChatGPT were harming the internet.

In August, he left OpenAI because he no longer wished to contribute to technologies he believed would cause more societal harm than good.

“If you share my concerns, the only choice is to leave the company,” he stated in a recent interview with The New York Times.

At 25 years old, Mr. Balaji has not taken another position and is focusing on what he calls “personal projects.” He is among the first employees from a significant A.I. company to publicly criticize how these firms have utilized copyrighted material to develop their technologies. A former vice president at Stability AI, a London startup specializing in image and audio generation, has expressed similar views.

In the past two years, several individuals and businesses have sued various A.I. companies, including OpenAI, alleging the illegal use of copyrighted materials to train their technologies. Plaintiffs include software developers, artists, record labels, authors, and news organizations.

In December, The New York Times sued OpenAI and its main partner, Microsoft, claiming they used millions of articles from The Times to develop chatbots that now compete with the publication as a reliable information source. Both companies have denied these allegations.

Many researchers who have worked at OpenAI and other tech firms have warned that A.I. technologies could pose serious risks. However, most of these warnings concern potential future dangers, such as A.I. systems that might help create new bioweapons or even threaten humanity.

Mr. Balaji believes the threats are more immediate. He argues that ChatGPT and similar chatbots undermine the commercial viability of individuals, businesses, and online services that generated the digital data used to train these A.I. systems.

“This is not a sustainable model for the entire internet ecosystem,” he told The Times.

OpenAI disagrees, stating, “We develop our A.I. models using publicly available data in a manner consistent with fair use and established legal precedents. We believe this approach is fair to creators, essential for innovators, and critical for U.S. competitiveness.”

In 2013, a London-based A.I. startup called DeepMind introduced technology that learned to play classic Atari games independently, including Space Invaders, Pong, and Breakout.

When Mr. Balaji was a teenager in Cupertino, California, he encountered a news story about this technology that fascinated him, as did DeepMind’s later achievement of mastering the ancient game of Go.

“I saw A.I. as a solution for problems that seemed unsolvable, like curing diseases and halting aging,” he recalled. “I envisioned it as a tool that could act like a scientist to help solve these issues.”

During a gap year after high school and as a computer science student at the University of California, Berkeley, Mr. Balaji began exploring the core concept behind DeepMind’s technologies: a mathematical framework known as a neural network capable of learning from digital data.

In 2020, he joined a group of Berkeley graduates who began working at OpenAI. In early 2022, he started collecting digital data for a new initiative called GPT-4, a neural network that spent months analyzing nearly all English-language text available on the internet.

He and his colleagues approached this as a research endeavor. Even though OpenAI had recently shifted to a profit-driven model and started selling access to similar technology known as GPT-3, they did not view their work as competing with existing online services. GPT-3 was not a chatbot; it was a technology that enabled businesses and developers to create other software applications.

“With a research project, you can generally train on any data,” Mr. Balaji explained. “That was the mindset at the time.”

Then OpenAI launched ChatGPT. Initially powered by a precursor to GPT-4 and subsequently by GPT-4 itself, the chatbot captured the attention of millions and quickly became a source of revenue.

OpenAI, Microsoft, and other companies contend that utilizing internet data to train their A.I. systems aligns with the principles of “fair use.” The fair use doctrine includes four criteria, and the companies argue that these factors—such as the substantial transformation of copyrighted works and not competing directly with those works—support their case.

Mr. Balaji disagrees, asserting that these criteria are not satisfied. When a system like GPT-4 learns from data, he contends, it effectively creates a complete copy of that data. Subsequently, a company like OpenAI can teach the system to either reproduce that data verbatim or generate entirely original text. In practice, he argues, companies train the systems to produce outputs that fall somewhere in between.

“The outputs aren’t exact copies of the inputs, but they’re not entirely original either,” he noted. Recently, he shared an essay on his personal website, presenting a mathematical analysis that he claims supports this viewpoint.

Stanford University law professor Mark Lemley argues otherwise. He asserts that the majority of what chatbots produce is sufficiently distinct from the training data.

“There are instances where an output resembles an input,” he acknowledged. “However, most content generated by a ChatGPT or image generation system does not heavily rely on any specific piece of content.”

Mr. Balaji argues that the technology breaches the law because it often directly competes with the copyrighted works from which it learns. Generative models are designed to replicate online data, making them capable of substituting for “essentially anything” found on the internet, from news articles to online discussions.

He points out a larger issue: as A.I. technologies replace existing internet services, they are generating false information and even completely fabricated content—what researchers term “hallucinations.” He believes the internet is deteriorating as a result.

Bradley J. Hulbert, an intellectual property attorney specializing in copyright law, noted that existing copyright laws were established long before the advent of A.I., and no court has yet determined whether A.I. technologies like ChatGPT violate these laws.

He advocates for Congress to create new legislation to address this technology. “Given the rapid evolution of A.I.,” he stated, “it is time for Congress to take action.”

Mr. Balaji concurs. “The only viable solution to this situation is regulation,” he concluded.