No machine learning system learns from nothing. When a headline says AI works "without data," the data has usually moved somewhere else: into rules written by experts, into a simulator that produces practice examples, into a few examples typed in a prompt, or into text generated by an earlier model. The cleanest case, DeepMind's AlphaGo Zero in 2017, used no human game records but still produced its own training games through self-play.
Key takeaways
- Systems described as data-free almost always rely on a substitute: expert rules, a simulator, a few examples, or machine-generated data.
- Self-play works where a program can check results itself, as in board games with fixed rules.
- The 2020 GPT-3 paper evaluated a model on new tasks from instructions and examples alone, with no further training, after it had already been trained on a huge text collection.
- A 2024 Nature paper found that training models repeatedly on model-generated text makes the rare content in the original data disappear.
- Keeping a share of real data in each round (10 percent in the Nature experiment) limited the damage.
Four meanings of "without data"
The phrase covers several different designs. Sorting them shows what each one still needs.
| Approach | Where the knowledge comes from | What it still needs |
|---|---|---|
| Hand-written rules | Experts encode their knowledge directly | Experts' time, and rules that cover the cases |
| Self-play or simulation | The system generates its own practice from fixed rules | A simulator and a way to score outcomes |
| Few-shot prompting | A pretrained model reads a handful of examples in the prompt | An earlier, very large training run on text |
| Synthetic data | One model writes training material for another | Real data to start from, and checks on the output |
Rule-based systems predate modern machine learning and are the only row where nothing is learned from examples at all. A rule such as a tax threshold is data written by a person, not discovered by a program.
Self-play: AlphaGo Zero and its limits
DeepMind announced AlphaGo Zero on October 18, 2017. It started from random play, learned only from games against itself, and after three days of training beat the earlier published AlphaGo version 100 games to zero, according to the developer's account, which accompanied a paper in Nature. No human games were used. The lesson is narrow, though. Go has fixed rules and a clear winner, so the program could generate unlimited practice and score every game exactly. Most real tasks, such as grading an essay or advising a student, offer no such referee.
Few-shot prompting: the data came earlier
In May 2020, Tom Brown and colleagues released the paper "Language Models are Few-Shot Learners." It reports that GPT-3 was evaluated on tasks specified only through text interaction, "without any gradient updates or fine-tuning," meaning the model's internal numbers were not changed for each new task. The abstract also states that few-shot performance still struggles on some datasets and that training on large web corpora raises methodological issues. The model needed no task-specific data at use time, but it had been trained on a very large body of text beforehand. The data requirement was paid once, upfront.
Synthetic data and model collapse
Another way to avoid collecting new data is to have a model produce it. A team led by Ilia Shumailov, with authors at Oxford, Cambridge, Imperial College London, and the University of Toronto, tested this in a paper in Nature (volume 631, July 2024). They report that indiscriminate use of model-generated content in training causes irreversible defects in which the tails of the original distribution disappear, an effect they call model collapse. In one experiment they repeatedly fine-tuned a small language model, OPT-125m, on the wikitext2 dataset, then trained each new generation on text produced by the previous one. When a random 10 percent of the original data was kept in each round, performance degraded only slightly.
The authors add that telling machine-generated text from human text at scale is an open problem, and that data collected before generative AI spread may become more valuable. For schools and researchers, this means synthetic examples are a supplement to real data and not a replacement for it.
A 2025 attempt at "zero data" reasoning
A preprint submitted on May 6, 2025, and revised on October 16, 2025, titled "Absolute Zero: Reinforced Self-play Reasoning with Zero Data," describes one model that proposes its own tasks and then solves them, without external training examples. The system, called the Absolute Zero Reasoner, uses a code executor to check both the proposed tasks and the answers. The authors claim state-of-the-art results on coding and math reasoning among comparable "zero" setups, and see this as a response to a possible shortage of human-made data. Three cautions apply. The work is a preprint on arXiv, the checker is a code executor that only works for code and math, and the comparison is against other zero-setting methods, not every method.
Questions to ask about a "no data" claim
- Who or what produced the examples the system learned from?
- Is there an automatic way to check answers, as in games and code, or does a person have to judge them?
- Was a larger model pretrained on text first?
- Is the result peer reviewed, or a preprint or company announcement?
- Does the claim hold outside the benchmark it was measured on?
FAQ
Can AI learn with no data at all?
None of the systems covered here does. Self-play and simulation generate their own data from rules, rule-based systems carry human knowledge written by hand, and pretrained language models rely on earlier training on large text collections.
Is synthetic data safe to train on?
It can be useful, but the 2024 Nature experiments show risk when generated text replaces real data across several generations. In that setup, mixing in 10 percent of the original data reduced the loss of quality.
Does few-shot prompting mean a model needs no training?
No. The model in the GPT-3 paper learned its general ability from large-scale pretraining. The few examples only specify the task at the moment of use.
What does this mean for students learning about AI?
A useful exercise is to take any claim of data-free learning and trace where the examples came from. The answer is often a simulator, a verifier, or a prior model, and naming it shows what the system actually depends on.
Related reading
- The difference between AI and machine learning
- Limited memory AI and restricted data storage
- Traditional AI compared with modern machine learning
- Whether AI can operate without coding
More in AI in education.
Last verified: October 2026
Sources
- AI models collapse when trained on recursively generated data, Shumailov et al., Nature 631, 755-759, July 2024
- Language Models are Few-Shot Learners, Brown T et al., arXiv:2005.14165, May 2020
- Absolute Zero: Reinforced Self-play Reasoning with Zero Data, Zhao A et al., arXiv:2505.03335, May 2025, revised October 2025
- AlphaGo Zero: Starting from scratch, Google DeepMind, October 18, 2017, with Silver D et al., Nature, 2017