By Dr. Tamara Patzer
Key Takeaways
- Epoch AI, a nonprofit research institute, estimated in a June 2024 paper that the usable supply of public, human-written text is about 300 trillion tokens, and projected that AI developers could use it up between 2026 and 2032.
- Stanford’s 2026 AI Index, released April 13, 2026, covers the same question in a section titled “Will Models Run Out of Data?” and reports that synthetic data is not yet replacing real data in pre-training.
- A July 2024 study from the MIT-led Data Provenance Initiative found that 25 percent of the data from the highest-quality sources in three major training sets had been restricted by the websites that publish it.
- A July 2024 study in the journal Nature found that models trained over and over on AI-made text degraded into nonsense.
- For business leaders, the lesson is plain: original human knowledge is getting harder for machines to replace.
The Last Shelf in the Library
Picture the largest library ever assembled, built from books, articles, forum posts, manuals, encyclopedias, and public records. For about a decade, AI developers have been walking its aisles and reading everything on the shelves. Now they are nearing the last one.
That is the picture painted by Epoch AI, a nonprofit research institute that tracks the inputs behind artificial intelligence. In a paper published June 6, 2024, its researchers estimated that the effective stock of quality-adjusted, public human-written text is around 300 trillion tokens, with a 90 percent confidence range of 100 trillion to 1,000 trillion. A token is a small chunk of text, roughly a word or part of one.
The researchers then set that supply against how fast developers are consuming it. Their projection: if current trends continue, the largest models will be trained on datasets roughly equal to the entire stock sometime between 2026 and 2032. Training on the same material repeatedly or building models that are deliberately overtrained would bring that date closer.
The forecast has moved before. Epoch’s 2022 paper expected high-quality text to run out around 2024. The 2024 update pushed the central estimate back toward 2028. Epoch is also careful to say that running out of public text would not end progress. Its researchers point to synthetic data, learning from other kinds of data such as images and video, and more efficient training as ways forward.
What Stanford Found
The question has now reached the field’s annual report card. Stanford’s Institute for Human-Centered Artificial Intelligence released its 2026 AI Index on April 13, 2026. Chapter 1, on research and development, includes a highlight section asking whether models will run out of data. The report notes concern among leading researchers about “peak data,” the point at which high-quality human-generated text and web data is effectively used up, and cites projections that place it between 2026 and 2032.
The report’s more practical finding is about the workaround. Synthetic data, meaning text and images produced by AI itself, is not yet replacing real data in pre-training, the stage in which a model first learns language. Mixing real and synthetic data can speed up training, the report says, but it does not beat models built on high-quality real data. Where gains are coming, according to the Index, is from cleaning data more carefully and from improvements made after pre-training.
The report also points to estimates that more than half of newly published online content is now AI-generated. That means the open web, the library’s biggest wing, is increasingly filled with machine writing.
The Doors Are Closing
Scarcity is not only about what exists. It is also about what is open to read.
In July 2024, the Data Provenance Initiative, a volunteer research group led by researchers at the Massachusetts Institute of Technology, published a study called “Consent in Crisis.” Lead author Shayne Longpre and his colleagues audited 14,000 web domains behind three widely used training sets: C4, RefinedWeb, and Dolma. They tracked two kinds of restrictions: instructions in a site’s robots.txt file, which tell automated crawlers to stay out, and limits written into a site’s terms of service.
The results, as reported by The New York Times on July 19, 2024: restrictions covered 5 percent of all data in those sets and 25 percent of the data from the highest-quality sources. In one of the sets, C4, terms of service restricted as much as 45 percent of the data. The researchers warned that the shift would affect not only AI companies but also academics and noncommercial researchers who depend on open web data.
Website owners had their reasons. Many publishers wanted a say in how their work is used. Whatever the motive, the effect was the same: fewer open doors to the library.
The Copy Of A Copy Problem
If human text is running short, why not let AI write its own study material? A team led by Ilia Shumailov of the University of Oxford tested that idea and published the results in Nature on July 24, 2024.
The researchers started with a language model trained on Wikipedia articles. They then trained a new model on the first model’s output, and repeated the process, each generation learning from the one before. By the ninth generation, the output had slid into gibberish. In one example, a prompt about medieval church architecture produced a final answer that was little more than a list of animals.
Shumailov compared the effect to repeatedly photocopying a photocopy: each pass adds errors, and rare details disappear first. The study called the effect “model collapse.”
The news was not all bad. Axios reported on July 27, 2024, that when the researchers kept 10 percent of the original human data in the mix, performance did not suffer. Shumailov told Live Science that synthetic data will need to be better designed, and that a model builder holding a human-produced copy of the internet is better placed to build capable systems.
Outside experts urged caution about the study’s reach. Reviewers quoted by the Science Media Centre in Spain noted that the experiment assumed models would train only on AI output, whereas real training will keep including human data.
What It Means For Business Leaders
None of this says AI is about to stall. Epoch, Stanford, and the Oxford team all describe workarounds, and none predicts a sudden stop.
What the research does show is a shift in what is scarce. For years, the limiting resources in AI were computing power and talent. The 2026 AI Index and the Epoch forecast point to a third constraint: high-quality material written, recorded, and checked by real people.
For a CEO, that has a few practical consequences:
- Original work stands out. As machine-made text spreads, writing that comes from a named person with real experience becomes easier to tell apart.
- Records matter. The Data Provenance Initiative’s findings show that who made something, and what permission came with it, now shapes what can be used.
- Transparency is shrinking. The Foundation Model Transparency Index reported that the average score fell from 58 in 2024 to 40 in 2025. Training data was among the least transparent areas.
Leaders do not need to settle the forecast debate to act on it. They can document how their own organization works, keep authorship clear on what it publishes, and expect more questions from customers about where information comes from.
Frequently Asked Questions
Is AI really running out of data?
Of one kind, yes, eventually. Epoch AI projects that public human-written text could be fully used between 2026 and 2032. Other kinds of data, including images, video, and AI-generated text, are not subject to the same limit.
Can synthetic data fill the gap?
Not fully, at least not yet. Stanford’s 2026 AI Index says synthetic data is not replacing real data in pre-training, and the 2024 Nature study found that training repeatedly on AI output degrades models.
Why are websites blocking AI crawlers?
The MIT-led study found rapid growth in restrictions, with 25 percent of data from the highest-quality sources in three major datasets restricted. The researchers noted that owners want control over how their work is used.
What is a token?
A token is a small unit of text that AI systems read, usually a word or a piece of a word.
Does this mean progress will stop?
No forecast cited here says so. Epoch lists synthetic data, other data types, and efficiency gains as ways forward.
About The Author
Dr. Tamara “Tami” Patzer is a publisher and former newspaper editor who was on the staff of The Charlotte Sun when it was named a 2005 Pulitzer Prize finalist in Breaking News Reporting. She writes about AI and media literacy. Connect with her on LinkedIn. Contact: info@dailysuccessinstitute.com.
