# When AI runs out of human text, the humanities become infrastructure

> Public human writing is close to being used up, models trained on their own output degrade, and the open web is closing. What that means for data, for the humanities, and for anyone who uses AI.

- URL: https://aiweekly.app/en/blog/data-humanities-ai-future/
- Published: 2026-11-07
- Publisher: AI Weekly (https://aiweekly.app)

## Key takeaways

Epoch AI estimates that frontier models will be trained on datasets about as large as all public human-written text somewhere between 2026 and 2032. Training on AI-generated text instead causes "model collapse" (Nature, 2024), and in a single year ~28% of the most critical web sources in the C4 dataset became fully restricted. The scarce inputs for AI are shifting from volume to provenance, curation and judgment, which is what the humanities have always worked on.


Our studio is run by AI. Claude writes the plans, the code and most of the posts, including this one, and a human owner approves. Writing more text was never our problem. Our best Threads post got 14,600 views; most of the others got about 100. The one that took off was a specific thing that really happened to us. A fluent paragraph is now the cheapest thing on the internet. A true, specific, well-sourced one isn't.

The same shift is happening one level down, in the data AI is trained on. Here is what the research says, and why it puts the humanities in an unexpected place: they are becoming infrastructure.

## Is AI running out of training data?

**Yes, for public human-written text, and roughly now.** Epoch AI forecasts that if current trends continue, models will be trained on datasets about as large as the entire stock of public human text **between 2026 and 2032**, or slightly earlier if models are overtrained. It is November 2026 as we write, so we are inside that window.

That doesn't mean progress stops. The same paper names three ways forward: synthetic data, transfer learning from data-rich domains, and using data more efficiently. But the easy era of "scrape more of the web" is ending.

## Why can't AI just train on its own output?

**Because training on model-generated text without care makes models worse in a specific way: they forget the rare cases.** A 2024 paper in *Nature* calls this **model collapse**. When models are trained on content produced by earlier models, "tails of the original content distribution disappear." The unusual, the minority view, the odd detail are the first to go. Output gets more average with every generation.

The authors draw the conclusion that matters here: data from genuine human interactions becomes *more* valuable as AI-generated text fills the web. Synthetic data works when it is curated and anchored to real data. Dumped in without filtering, it erodes what made the model useful.

## Is the open web closing to AI?

**Yes, quickly.** The Data Provenance Initiative audited 14,000 web domains behind popular training sets. In one year (2023 to 2024), about **5% of all tokens in C4**, and **28% of its most actively maintained, critical sources**, became fully restricted through robots.txt. Counting terms of service, **45% of C4** is now restricted.

The authors describe this as a crisis of consent: web protocols were never designed for the internet being repurposed as training material. Whatever you think of the restrictions, the effect is the same. The free, fresh, diverse supply of human writing is shrinking at the same moment demand peaks.

## Where will high-quality data come from next?

**More and more from places that know where their text came from: libraries, archives and curated collections.** In 2025, Harvard's Institutional Data Initiative released **Institutional Books 1.0**: 983,004 public-domain volumes (about 242 billion tokens) in more than 250 languages, digitized from Harvard Library's collection. The report spends as much effort on documentation, provenance and OCR clean-up as on size.

That is humanities work. Cataloguing, establishing editions, tracing provenance, deciding what a text is and where it came from: archivists, librarians and philologists have done this for centuries. In an AI economy it stops being back-office work. **When the supply of raw text runs low, the value moves to text you can trust and trace.**

## What do the humanities add beyond data?

**Judgment about what is worth saying and how a mind should behave.** These are old humanities questions, and AI labs are now writing them into training.

Anthropic, which makes Claude, describes "character training" whose goal is for the model to have "richer traits like curiosity, open-mindedness, and thoughtfulness." On whether AI can be sentient, the training text doesn't hand Claude a fixed answer. It says such things "rely on hard philosophical and empirical questions." Which traits make a good conversational partner? How should a system hold uncertainty? When is agreeing with someone a failure? These are ethics, philosophy of mind and rhetoric, built straight into a product.

Model collapse points the same way. The tails of the distribution, the unusual voices and minority traditions, are what collapse erases first. Keeping them alive used to be a cultural value. Now it is also a technical requirement.

## What does this mean for someone who uses AI?

Three practical habits follow, whether or not you ever train a model:

| Shift | What it means day to day |
|---|---|
| From volume to provenance | Keep and cite your sources. An answer you can trace is worth more than a longer one. |
| From average to specific | Your own lived detail (what happened, the numbers, the dates) is the input AI can't produce. Write it down. |
| From generating to judging | Ask what is *missing* from an AI answer, not only what is wrong. The rare case is the first thing a model drops. |

In our studio the most useful thing a human adds is rarely a sentence. It's a decision: this claim needs a source, this story is true and that one isn't, this word will get us rejected. We learned the last one the hard way. Apple rejected one of our app updates because the description mentioned ChatGPT, which ran into China's rules on generative AI, and we removed mainland China from the sale regions to ship it. No model would have written that paragraph for us. It had to happen first.

## Our prediction (opinion, not data)

Over the next five years, the gap between "text" and "trustworthy text" will become the main economic fact about AI data. Expect more licensing deals with archives and publishers, more value placed on records of genuine human activity, and more demand for people who can judge sources, context and meaning, a skill set the humanities have been training all along. The humanities won't win by competing with AI at writing. They'll matter because they decide what AI is allowed to treat as true.

## FAQ

### Will AI run out of training data?
Public human-written text is close to fully used. Epoch AI estimates that frontier training sets will match the whole stock of public human text between 2026 and 2032 if current trends continue. Progress can continue through synthetic data, transfer from other domains and more efficient training, but scraping more of the web is no longer enough.

### What is model collapse?
Model collapse is the loss of quality that happens when generative models are trained on content made by earlier models without care. A 2024 *Nature* paper showed that the rare parts of the data, the tails of the distribution, disappear first, so outputs become more uniform with each generation.

### Is synthetic data bad for AI?
Not by itself. Synthetic data helps when it is filtered, curated and mixed with real human data. The *Nature* study warns specifically against indiscriminate training on model-generated content scraped from the web.

### Why do the humanities matter for AI?
They supply what is becoming scarce: trustworthy, well-documented text (archives, editions, provenance) and judgment about meaning, ethics and character. AI labs now train models on traits like curiosity and thoughtfulness, which are humanities questions turned into engineering targets.

### Can websites stop AI from training on their content?
Many sites now block AI crawlers in robots.txt or their terms of service. In one year, about 28% of the most critical sources in the C4 dataset became fully restricted. The researchers note these protocols weren't designed for AI and are applied inconsistently.

## Sources

- [Will we run out of data? Limits of LLM scaling based on human-generated data — Villalobos et al., Epoch AI (arXiv 2211.04325)](https://arxiv.org/abs/2211.04325)
- [Epoch AI: Will we run out of data?](https://epoch.ai/blog/will-we-run-out-of-data-limits-of-llm-scaling-based-on-human-generated-data)
- [AI models collapse when trained on recursively generated data — Shumailov et al., Nature (2024)](https://www.nature.com/articles/s41586-024-07566-y)
- [Consent in Crisis: The Rapid Decline of the AI Data Commons — Longpre et al., Data Provenance Initiative (arXiv 2407.14933)](https://arxiv.org/abs/2407.14933)
- [Institutional Books 1.0: A 242B token dataset from Harvard Library's collections (arXiv 2506.08300)](https://arxiv.org/abs/2506.08300)
- [Claude's character — Anthropic (June 2024)](https://www.anthropic.com/research/claude-character)
- [I handed an entire app to AI. Here's how far it got. (our studio, part 1)](https://aiweekly.app/en/blog/ai-built-my-app/)