AI progress over the past decade has mostly followed one formula. Feed a model more data, make it bigger, and it gets better. Nobody spent much time worrying about what happens if the data runs out, because for years there was always more of it sitting online, waiting to be scrapped. That assumption is no longer safe, and researchers have recently put a real timeline on when it stops holding.

How Bad Is the AI Training Data Shortage

In 2022, a team of researchers began modelling something nobody had rigorously measured before, how much high-quality human-written text actually exists, and how fast AI labs were consuming it relative to how fast new text gets created. Their paper, "Will We Run Out of Data?", became one of the most cited pieces of AI forecasting research in the years that followed. Epoch AI later sharpened the estimate:

Depending on how aggressively labs continue to train, the world's stock of high-quality text data could be exhausted somewhere between 2026 and 2032.

The situation was described by one of the study's authors as a kind of literal gold rush, one that depletes a finite resource the same way any real mining operation does. The comparison is worth sitting with. AI progress has largely come from scaling two things at once, computing power and the volume of text fed into models, and text is growing far more slowly than the models trying to consume it.

At the same time, not everyone treats the deadline with the same urgency. Anthropic CEO Dario Amodei has put the odds of an actual data-driven slowdown at around ten percent, a real risk worth planning for, but far from a certainty.

That range of opinion, from an approaching wall to a manageable risk, is itself part of the story. Even the researchers closest to the data don't fully agree on how hard the ceiling is, only that it exists.

The Internet Is Also Restricting Access to AI Training Data

The exhaustion timeline assumes the internet stays roughly as open as it's been. That assumption is already wrong, and it's wrong in a way that's making the problem arrive faster than the raw math suggests.

In 2024 MIT's Data Provenance Initiative audited 14,000 websites to see how content owners were responding to AI crawlers. In the twelve months studied, roughly a quarter of tokens from the domains most critical to popular AI training datasets became restricted through robots.txt, the small text file that tells automated crawlers what they can and can't access. Some of the most heavily used, actively maintained sources saw restrictions climb past a quarter of their content in that single year alone. Separate reporting on the same trend put blocked tokens in key datasets as high as a third.

What makes this compounding rather than simply additive is that it's a moving target layered on top of a shrinking one. Even if the total pool of human-written text were static, which it isn't, the portion AI companies are actually permitted to use keeps shrinking as more publishers, news organizations, and platforms wall off their content specifically because it's being used to train the models competing with them.

Why Synthetic Data Can't Solve the AI Training Data Shortage Alone

The most commonly proposed solution is synthetic data, text, images, or audio generated by AI models themselves and fed back in as training material. In theory it solves the scarcity problem outright, because if a model can generate its own examples, the supply becomes limitless.

In practice, the research is far less comforting. A study by Shumailov and colleagues documented what's now called model collapse, a pattern where models trained repeatedly on their own outputs or the outputs of other models gradually lose the rare, unusual, long-tail information that made the original human data valuable in the first place. Errors and blind spots compound across generations until outputs drift noticeably from anything a human would have actually written. One widely quoted description called it something close to the computer-science version of inbreeding.

The research community hasn't fully settled the question, though. Another paper argues the early alarm was overstated in important ways, and that collapse mainly shows up under specific, avoidable conditions, particularly when synthetic data is used in raw, unfiltered form and allowed to recursively retrain on itself without any fresh human data mixed back in. Researchers have proposed mitigation techniques that keep synthetic data useful without letting quality erode, largely by verifying synthetic outputs before they re-enter the training pool and capping how much of any dataset is allowed to be synthetic at all. The emerging consensus sits somewhere in the middle, synthetic data is a genuinely useful supplement rather than a genuine replacement, and the ratio matters enormously.

Video and Multimodal Data: The Next Source for AI Training

If text is running out and synthetic text is a limited patch rather than a real fix, the more interesting question is where researchers are actually turning instead.

Some argue that text was always a lossy compression of reality in the first place, a description of the world rather than the world itself. For example, computer scientist Yann LeCun, before his departure from Meta, trained a multimodal model from scratch and found that huge volumes of unlabelled video, largely untapped by comparison to text, could be folded into training without hurting the model's language performance at all. The 2026 International AI Safety Report reached a similar conclusion from a different angle, noting that with roughly ten trillion tokens of high-quality public text already close to exhausted by current models, multimodal data of other kinds is one of the most immediate levers available for extending the runway.

OpenAI and Google have both signed commercial deals to license text from platforms like Reddit and major news publishers rather than rely on open scraping, an implicit admission that freely available web text is no longer sufficient or, in many cases, no longer permitted. The same logic is extending into video and audio, where the rights and structure questions are even less settled than they are for text, and where a smaller number of large, high-quality libraries, film, television, and archival footage among them, hold a lot of exactly the range and volume that open web scraping can no longer reliably provide.

What the AI Training Data Shortage Means for AI Labs and Rights Holders

Pulled together, the research points to a conclusion that's more specific than "AI is running out of data." Three separate pressures are converging on the same timeline, a genuinely finite supply of fresh human text, a web that's actively restricting access to what remains, and a synthetic-data fix that helps at the margins but can't carry the full load on its own. None of the three alone would force much change. Together, they're already changing where labs look for training material and how they acquire it.

For AI labs, the data strategy is becoming as central to competitive advantage as model architecture itself, and it increasingly means turning to sources, video, audio, and other underused modalities, that were treated as secondary a few years ago. For publishers, rights holders, and archives, it means the content sitting in their libraries has taken on a value that didn't exist five years ago, provided it can be licensed and delivered on terms that actually hold up. And for the field more broadly, it means the next real breakthroughs may depend less on finding smarter algorithms and more on finding, and properly clearing, the raw material to train them on in the first place.

ContentX Labs is built around exactly that kind of supply, real, human-made, and cleared at the source. If that's the direction your own data strategy is heading, a free sample set is the fastest way to see what it actually looks like.