The short answer is that AI will not suddenly stop improving. But it will have to change how it learns.

The argument that AI is running out of data is partly true and partly misunderstood. The problem is not that the internet is about to become empty. The problem is that the most useful part of it—the new, high-quality, diverse human knowledge that frontier models need—is finite.

That changes the next phase of AI development. The first generation learned mainly by reading what humans had already written. The next generation will increasingly learn by generating data, taking actions, running experiments, and receiving feedback.

The internet’s best text is finite

Most publicly available books, websites, research papers, forums, and code repositories have already been used heavily in training large models. There is still plenty of text online, but not all text is equally valuable.

The scarce resource is not text in general. It is new, high-quality, diverse text.

The internet contains an enormous amount of repetition, low-effort content, automated writing, duplicated documentation, and material that adds little new information. Adding more of it does not necessarily make a model smarter. In some cases, it may make the training signal noisier.

This is why the data question matters. As models become more capable, they need better examples, not simply larger piles of existing ones.

Models do not simply store entire books

It is common to say that models “steal” content. That is a vivid phrase, but it is not a precise description of how training works.

Models are trained on datasets containing licensed, public, or otherwise acquired content, depending on the developer and the dataset. The process has led to serious legal and ethical disputes about copyright, consent, and compensation. Those disputes are real and important.

But training a model is different from copying entire works into a searchable database. A trained model does not normally retrieve a complete book from memory on demand. It learns statistical patterns and relationships from its training material.

The deeper question is this:

What happens when there is not enough fresh human knowledge left to learn from?

Four changes are already pointing toward an answer.

1. Synthetic data becomes a new fuel source

The obvious next step is for models to generate training data for newer models.

This already works well in areas where answers can be checked reliably, including:

• Mathematics
• Coding
• Reasoning
• Translation
• Other structured tasks

An AI system can generate a problem, produce a solution, and verify the result. For code, it can write a program, run tests, inspect failures, and produce another attempt. For mathematics, a proof assistant or symbolic checker can distinguish a valid solution from an attractive mistake.

This creates a new supply of examples without waiting for humans to write every one of them.

There is a risk, though. Synthetic data is only as good as the process used to create and filter it. If a model generates flawed examples and those examples are used to train another model, the errors can compound. Repeating a mistake does not turn it into knowledge.

Poorly managed synthetic training can cause models to lose diversity and quality, a failure mode often described as model collapse. The solution is not to avoid synthetic data. It is to make its generation, verification, and mixing with human data much more deliberate.

2. AI starts generating experiences instead of just reading them

Reading about an activity is not the same as doing it.

An AI does not need to learn chess only by reading chess commentary. It can play millions of games, test strategies, lose, adjust, and keep the strongest lessons.

It does not need to learn software debugging only by reading forum discussions. It can write code, run the tests, observe failures, inspect logs, change the implementation, and try again.

This is a shift from passive learning to interactive learning. The model is no longer limited to examples that someone happened to publish. It can create situations, take actions, and learn from the consequences.

The quality of that learning depends on the environment. Chess has clear rules and a reliable score. A software repository can provide tests and build results. Many real-world problems are less cooperative. The feedback may be delayed, ambiguous, expensive, or wrong.

That is why environments matter so much. An agent with a reliable simulator, test suite, or evaluation harness has a much better opportunity to improve than one operating in a world where success cannot be measured.

3. Human feedback becomes more valuable

As raw internet data becomes less useful at the margin, expert judgment becomes more valuable.

A lawyer, doctor, engineer, or scientist can review outputs and explain which answer is safer, more accurate, more useful, or more aligned with the real task. A relatively small amount of expert feedback can teach a model distinctions that millions of random web pages never made clearly.

This makes human expertise a scarce resource in the training pipeline.

The expert does not need to write an entire textbook. They might rank several answers, correct a subtle mistake, label a failure mode, or demonstrate how a difficult task should be handled. Those decisions create high-quality preference and evaluation data.

The bottleneck moves from collecting examples to finding people who can reliably judge them.

That has practical consequences. Companies will need better tools for capturing expert feedback, protecting sensitive information, and turning individual judgments into reusable training signals. They will also need to compensate experts for the value of that judgment.

4. Real-world data becomes the moat

The most defensible advantage may increasingly come from data that cannot simply be downloaded from the public internet.

Examples include:

• Customer support conversations
• Medical records, with appropriate privacy protections
• Enterprise workflows
• Robotics data
• Scientific experiments
• Manufacturing processes

This data has something generic web text often lacks: it is connected to real outcomes. It shows what customers actually ask, which processes break, which treatments work, how machines behave, and whether a proposed solution survives contact with reality.

That makes proprietary data strategically important. Future models may be differentiated less by their basic architecture and more by the exclusive environments, feedback loops, and operational data available to their developers.

The company with the best model may not be the company with the biggest pile of public text. It may be the company that has spent years collecting high-quality interaction data and building reliable ways to learn from it.

The biggest shift: from knowledge to interaction

The first generation of AI learned primarily from human knowledge.

The next generation will increasingly learn from interaction with the world.

The rough progression looks like this:

• 2018–2025: Read the internet.
• 2025–2028: Generate and filter synthetic data.
• 2027 onward: Learn by acting, experimenting, and receiving feedback.

These stages overlap. Synthetic data is already in use, and models have been learning from environments for years. The point is not the exact calendar. It is the direction of travel.

This is also closer to how humans learn. We read what others know, but we also practice, make mistakes, run experiments, and update our beliefs based on what happens next.

AI may eventually create more valuable data than humans did

There is an ironic possibility here. AI may not merely consume the world’s useful data. It may eventually create more of it than humans ever produced.

Millions of AI agents could continuously:

• Run scientific experiments
• Test software
• Simulate engineering designs
• Discover new materials
• Explore mathematical proofs
• Optimize manufacturing processes

Every successful experiment can become a new training example. Every failed experiment can narrow the search space, if the failure is recorded and understood. Over time, the system produces knowledge that did not exist in its original training set.

In that future, the limiting factor will not be the amount of text left on the internet. It will be access to environments where AI can experiment safely, receive reliable feedback, and generate genuinely new knowledge.

AI is not reaching the end of learning. It is reaching the end of learning mainly by reading what humans have already written.

The next advantage will belong to systems that can do more than consume information. They will act, test, observe, and learn from the world itself.