Tel Aviv University: AI that trains without human data
Researchers at Tel Aviv University, with Stanford, report that models can build prediction skills from data they generate themselves. The method gives no knowledge of the real world, but may make later learning more efficient.

Researchers at Tel Aviv University, working with Stanford University, have reported a way for AI models to learn from data they generate themselves, with no human-made text, images or recordings. Ynet and the Jerusalem Post, along with Xinhua and The Media Line, have covered the work, which was posted to arXiv in early November. It is an early research result and not a product.
What has been reported
- The approach, described in the coverage as Zero-Data Pretraining, uses two models that train each other. One generates sequences of symbols and the other learns to predict the next symbol.
- The training material was generated entirely inside the system.
- The researchers tested the models on language, images, sounds, music, computer code and DNA sequences, and reported that prediction ability improved as more computing power was used in training.
- The method does not give a model factual knowledge about the world. According to the coverage, that still requires exposure to real data.
- The researchers suggest it could help future systems learn more efficiently from real data afterwards.
Why it matters
Today's large models are trained on enormous amounts of human-written and human-made material, and coverage notes that developers are approaching the limits of the high-quality data available. Research that reduces dependence on that data touches questions that matter to many people: whose writing is used, and how much a system must read before it is useful.
It also adds to a steady stream of Israeli university research feeding the global AI field. The headlines are often about startups and acquisitions, but the groundwork is laid in labs like this one.
What it means for families, schools and frum professionals
- It is a reminder to separate research from products. A paper on arXiv is not yet something a child or a business will use, and early results are often refined or revised.
- Learning to predict patterns is not the same as knowing facts. A model trained this way would still need real material to be useful, and it can still be wrong.
- For teachers, it is a useful classroom example of how much of what AI appears to know comes from the data it has seen.
Putting the result in context
It is worth being plain about how far this result goes. According to the coverage, the models learned to predict patterns in material that the system made for itself. They were not taught what is true about the world, and they were not shown to be better assistants. What the researchers report is that the skill of predicting what comes next can be built without human-made data, and that it grew with the computing power used during training.
The coverage also notes that the practical hope is a modest one: that a model given this kind of head start might learn more efficiently once it meets real data. Whether that holds at the scale of today's largest systems is a question for further work and for other researchers to test. A preprint on arXiv has not been through peer review, and readers should treat it as a claim awaiting scrutiny.
Why the source of data matters to our community
Many families and educators ask where an AI system's knowledge comes from, especially when the subject is Torah, Hebrew or Jewish history. A model's answers reflect the material it was trained on. If that material is thin, mistaken or biased, the answers can be too, and the model will often sound equally confident either way. That is why we encourage checking any AI answer against the source, and bringing real questions of halacha to a rav.
Research like this does not change that advice. Even if future models need less human-made data to learn how to learn, they will still need real material to know anything about the world, including our texts.
What it changes for someone using Navōn
Nothing today. This is research, and Navōn's features do not depend on it. As with all AI, answers can be wrong, and a halachic question belongs with a rav.
Common questions
What did Tel Aviv University researchers develop?
They report a training method in which two models generate and predict symbol sequences, so a model builds prediction skills without human-generated data.
Can AI really learn without data?
Not about the real world. According to the coverage, the method builds pattern-prediction skills, but factual knowledge still needs real data.
Has this been peer reviewed?
The coverage says the work was posted to arXiv, which is a preprint server. It should be treated as early research.
Does this affect Navōn?
No. It is a research result and does not change how Navōn works today.
Sources
- Ynet: Israeli breakthrough could transform AI: Models can learn from data they generate themselves (November 2026)
- Xinhua: Study finds AI can learn without real-world data (November 4, 2026)
- The Media Line: Israeli Researchers Develop Method That Lets AI Teach Itself Without Human-Generated Data (November 2026)
This story was researched and written by Navōn.
Inside Navōn, News writes an article on any topic from sources you choose, checked against your family's or school's standards.
Published by Navōn. How these guides are written and checked.
Read next
- Kosher AI: what it means and how to evaluate it
A complete guide to what people mean by kosher AI, why no certification exists, and the specific properties to check before a tool enters your home or classroom.
- Israeli AI startup Two Delta raises a seed round to speed up AI models
Calcalist reports that Two Delta, a small Israeli AI company, raised a seed round backed by Bill Ackman and Eric Schmidt. What was reported, and what wasn't.
- Israeli AI chip startup NeuReality bought by Unconventional AI
A US chip startup led by Naveen Rao is buying Israel's NeuReality and keeping an Israeli development centre. What Reuters, Calcalist and Globes reported.