The tragic death of former OpenAI researcher Suchir Balaji has thrust a critical debate about AI training data and internet sustainability back into public consciousness. Balaji, who was found dead in late November in what San Francisco authorities ruled a suicide with no evidence of foul play, had become increasingly vocal about the potentially destructive impact of large language models (LLMs) on the internet ecosystem.
The core issue centers on how AI models are trained using vast amounts of internet data, then provide direct answers to users without driving traffic back to original content creators. This phenomenon, which Elon Musk has dubbed “Death by LLM,” threatens to undermine the economic foundations that support quality content creation online. When users get answers directly from ChatGPT or similar tools, they bypass the websites that originally created and verified that information, cutting off advertising and subscription revenue streams.
Stack Overflow serves as a cautionary tale. The popular coding Q&A platform experienced approximately 12% traffic decline following ChatGPT’s release, according to research Balaji cited. The study also found fewer new questions being posted and rising average account ages among question-askers, suggesting both reduced sign-ups and increased user departures. Developers who once relied on Stack Overflow’s community-driven knowledge base now simply ask ChatGPT for coding solutions.
About a month before his death, Balaji published an essay arguing that AI models like ChatGPT don’t qualify for fair use copyright protection. He contended that these tools fail one of the four critical tests for copyright infringement: whether a new work impacts the potential market or value of the original copyrighted work. “This is not a sustainable model for the internet ecosystem,” Balaji told The New York Times in October.
Tech reviewer Marques Brownlee (MKBHD) recently experienced this issue firsthand when testing OpenAI’s Sora video model. He discovered the AI generated content featuring a plant remarkably similar to one appearing in his own YouTube videos, raising questions about whether his work was used without permission in training data. Brownlee expressed frustration about the inability to opt out and uncertainty about whether it’s “too late” to prevent his content from being used.
OpenAI has defended its practices, stating that its AI model development is protected by fair use principles and supported by legal precedents, calling this “critical for US competitiveness.” However, Balaji’s analysis suggested none of the four fair use factors weigh in ChatGPT’s favor, and his arguments could apply to “many generative AI products in a wide variety of domains.”
Key Quotes
This is not a sustainable model for the internet ecosystem.
Suchir Balaji told The New York Times in October, expressing his disillusionment with how AI chatbots like ChatGPT strip away the commercial value of people’s work. This statement came from someone who had worked directly on collecting internet data for AI training at OpenAI, giving it particular weight and credibility.
Are my videos in that source material? Is this exact plant part of the source material? Is it just a coincidence?
Tech reviewer Marques Brownlee (MKBHD) raised these questions after discovering OpenAI’s Sora video model generated content remarkably similar to elements from his own YouTube videos. His concerns highlight the broader issue of content creators being unable to determine if their work has been used to train AI models without their knowledge or consent.
None of the four factors seem to weigh in favor of ChatGPT being a fair use of its training data.
Balaji wrote this in his essay published about a month before his death, directly challenging OpenAI’s legal position. He argued that AI models fail the copyright fair use test because they damage the market value of original works, and noted these arguments could apply to many generative AI products across various domains.
We view this principle as fair to creators, necessary for innovators, and critical for US competitiveness.
OpenAI’s statement to The New York Times defending its training data practices, asserting that fair use copyright principles protect how it builds AI models. This represents the company’s official position in the ongoing debate about whether AI training constitutes copyright infringement.
Our Take
Balaji’s death has become a tragic catalyst for a conversation the AI industry has actively avoided. The reluctance of tech companies to disclose training data sources—a transparency standard abandoned in recent years—suggests they recognize the legal and ethical vulnerabilities in their practices. The Stack Overflow data is particularly damning because it quantifies what many suspected: AI models are parasitic on the very ecosystems that created their training data. What makes this especially concerning is the potential feedback loop: as original sources lose traffic and revenue, content quality and diversity decline, which ultimately degrades the data available for future AI training. This could lead to what some researchers call “model collapse,” where AI systems trained on increasingly AI-generated content produce progressively worse outputs. The industry faces a reckoning between innovation and sustainability, and Balaji’s warnings suggest the current trajectory is untenable. His insider perspective—having worked directly on data collection at OpenAI before becoming disillusioned—lends credibility that external critics lack.
Why This Matters
This story highlights a fundamental tension at the heart of the AI revolution that could reshape the internet as we know it. The debate over AI training data isn’t merely academic—it threatens the economic viability of content creation across the web. If AI models continue to provide direct answers while siphoning traffic from original sources, the incentive structure that built today’s information-rich internet collapses.
The Stack Overflow case study demonstrates real-world consequences already underway. A 12% traffic decline may seem modest, but it represents a significant erosion of the community engagement that makes such platforms valuable. As advertising and subscription revenues decline, websites have fewer resources to invest in creating and verifying high-quality content, potentially leading to a less accurate, less diverse internet.
For businesses and content creators, this raises urgent questions about intellectual property rights, fair compensation, and the ability to control how their work is used. The lack of transparency from AI companies about training data sources—a practice that was once standard but has been abandoned—makes it nearly impossible for creators to protect their work or seek compensation. This could trigger a wave of litigation and regulatory action that fundamentally alters how AI models are developed and deployed, with major implications for AI competitiveness and innovation.
Recommended Reading
For those interested in learning more about artificial intelligence, machine learning, and effective AI communication, here are some excellent resources:
Recommended Reading
Related Stories
- OpenAI CEO Sam Altman Hints at Potential Restructuring in 2024
- Sam Altman’s Bold AI Predictions: AGI, Jobs, and the Future by 2025
- Photobucket is licensing your photos and images to train AI without your consent, and there’s no easy way to opt out
- The Impact of AI on Software Engineering Jobs and Market Outlook
- Elon Musk Drops Lawsuit Against ChatGPT Maker OpenAI, No Explanation