A revolution is happening in how AI models get trained. Two years ago, training a model meant collecting real data from real users, then paying human labellers to annotate it.

Thousands sitting at screens, classifying images, rating responses, marking text as positive or negative. It was slow, expensive, and created a bottleneck that decided how fast AI products could improve.

Today, that bottleneck is disappearing. The training data for many of the best models in production is not written by humans at all. Instead, other AI models are writing it.

A large, capable "teacher" model generates examples, such as instructions, responses, preference pairs, and domain-specific scenarios.

Then, a smaller, cheaper "student" model learns from them. It is synthetic data. The training data is artificially generated rather than collected from the real world.

The shift has been dramatic. For instruction tuning, the stage where a model learns to follow instructions and behave helpfully, synthetic data generation has won.

Distillation from stronger models now produces higher-quality training examples than most human writers can provide at scale.

Gartner estimates that by 2030, synthetic data will surpass real data in AI model training. But the transition is already well underway.

The core idea is straightforward. Use a strong model to produce the data that trains a weaker model. The simplest version is distillation.

You take a frontier model, say, Claude Opus or GPT-4, and have it generate thousands of high-quality responses to a diverse set of prompts.

Those prompt-response pairs become the training data for a smaller, cheaper model.

The smaller model learns to imitate the larger one's behaviour, inheriting its capabilities at a fraction of the inference cost. Think of it like a master chef writing a cookbook.

The chef's experience is distilled into recipes that a home cook can follow.

The home cook will never match the master in every situation, but for the most common dishes, the recipes produce close results. A more refined version is self-instructed.

Instead of a human writing the initial prompts, the model generates its own prompts, answers them, and then evaluates its answers.

The best prompt-response pairs are filtered and used for training. This approach bootstraps an entire training dataset from a small set of seed examples.

Stanford's Alpaca project confirmed this by generating 52,000 synthetic instruction examples to fine-tune Meta's LLaMA model.

It also reduced the need for human-created data. A third approach ties into what we covered in the post-training piece: RLAIF (Reinforcement Learning from AI Feedback).

Instead of paying human labellers to judge which response is better, an AI model does the judging. You give the model a set of principles, "be helpful, be honest, avoid harmful content," and it evaluates responses against those principles.

The judgments become the preference data that trains the next model.

Anthropic's Constitutional AI is built on this approach.

Also read: Data Labelling: The Authoritative Guide

The economics are overwhelming.

Human labelling costs $1-10 per annotation, depending on complexity.

A preference labelling task, where a human compares two responses and chooses the better one, might cost $5-15 per judgement.

You need thousands of these to train a reward model or run DPO.

A dataset of 10,000 preference pairs could cost $50,000 to $150,000 in human labelling alone. Synthetic generation costs a fraction of that.

Running a frontier model to generate 10,000 high-quality preference pairs costs a few hundred dollars in API fees. The quality is often comparable.

In some cases, it is better because the frontier model is more consistent than humans, who bring varying standards, attention levels, and biases.

Speed is equally important. Human labelling takes weeks or months.

Synthetic generation takes hours.

For teams that improve models quickly, the ability to generate a new training dataset in an afternoon rather than wait six weeks for annotations is transformative.

And there is a consistency advantage. Human labellers disagree with each other roughly 20-30% of the time on subjective judgements.

Ask five people if a response is "helpful," and you will get various answers depending on their background, mood, and interpretation of "helpful."

A model applying a consistent set of principles produces more uniform labels. That consistency can actually improve training stability, because the model is not trying to learn from contradictory signals.

This mix of cost, speed, and consistency is why synthetic data has moved from research curiosity to production default in under two years.

Synthetic data is not a free lunch. Three failure modes matter for PMs.

When models train on other models' data, mainly their own, there is a risk of progressive quality degradation. Each generation of synthetic data loses some of the nuance and diversity of the original training data.

Train on synthetic data for too many rounds, and the model's outputs become generic, repetitive, or detached from reality. Researchers call this model collapse.

It is the single biggest risk of over-reliance on synthetic data.

Synthetic data inherits the biases of the model that generated it.

If the teacher model produces overly formal responses or subtly favours certain viewpoints, those tendencies get baked into the training data and amplified in the student model.

Also, human data has biases, but they are diverse and often cancel each other out. Synthetic data can concentrate a single model's biases.

Distilling one company's model to train your own is a grey area.

OpenAI's terms of service, for example, have historically restricted relying on their API outputs to train competing models.

The Alpaca project, which distilled GPT-3.5 outputs to train an open-source model, raised debate about whether this violated OpenAI's terms.

PMs need to understand constraints around the data their models train on, especially when competitor's API generate that data.

The most sophisticated teams do not choose between human and synthetic data.

They combine them. The pattern that works best in 2026 is what researchers call "human-in-the-loop synthetic generation."

A frontier model generates first-pass outputs. Human reviewers do not write from scratch. They accept, reject, or edit the synthetic outputs.

Each human decision is a supervision signal, and the volume of human work drops so much because they are reviewing rather than creating.

This scales human judgment rather than replacing it.

A labelling team that previously produced 500 annotations per day can now review and validate 5,000 synthetic annotations in the same time.

The quality stays anchored in human judgement. The volume scales with compute. Anthropic's Constitutional AI follows a similar principle.

The constitution, the set of principles the model evaluates against, is human-written. The model does the actual evaluation and revision.

Humans define the standards. The model applies them at scale.

When you evaluate a model vendor, understand their data pipeline.

A model trained primarily on human-curated data may have different characteristics than one trained on distilled synthetic data.

Neither is inherently better, but they fail in different ways. When you fine-tune a model for your product, synthetic data is probably the fastest path to a working training set.

Generate examples with a strong model, review a sample for quality, and fine-tune your target model on the results. The cost is low, the iteration speed is fast, and for most use cases, the quality is sufficient.

When you build feedback systems into your product, such as ratings and corrections, you are generating the human data that synthetic generation cannot fully replace.

That data is valuable because it comes from your actual users in your actual domain.

It is the ground truth that keeps synthetic data honest. The future of training data is not human or synthetic. It is human and synthetic, each doing what they do best.