Two researchers trained small AI models on 96 invented languages, and the models picked up some grammar patterns as easily as human languages tend to. On action nominal constructions, though, the AI’s preferences split away from what real languages do, and that gap is the most interesting part of the study.
The paper, posted to arXiv by Iskar Deng and Shane Steinert-Threlkeld of the University of Washington, asks an old question with new tools: do languages look the way they do because some patterns are simply easier to learn?
Key takeaways
- The study built 96 synthetic languages and trained GPT-2-small models from scratch on each, with three models per language.
- On basic clause word order, the models’ learning matched patterns found in the world’s languages, often scoring near 100% accuracy.
- On complement clauses and action nominal constructions, model preferences diverged from real-world distributions.
- The models learned some combinations that are rare or entirely absent in human languages.
- The authors argue that learnability alone doesn’t explain how human languages are distributed. Communicative needs matter too.
What are action nominal constructions?
An action nominal construction (ANC) is a phrase that packs an event into a noun. Take “the enemy destroyed the city”. It becomes “the enemy’s destruction of the city”. The same event and the same participants are there, but in noun form.
The tricky part is how a language marks the participants once the verb turns into a noun. Some languages keep sentence-style marking. Others treat the doer, the thing affected, or both as possessors, like the “‘s” in “the enemy’s”.
Linguist Maria Koptjevskaja-Tamm mapped these strategies across 168 languages in the World Atlas of Language Structures (WALS). Her counts show how unevenly the options are spread:
| ANC type | How participants are marked | Languages in WALS sample | Example language |
|---|---|---|---|
| Sentential | Same marking as a normal clause | 25 | Godoberi |
| Possessive-Accusative | Doer is a possessor; object keeps clause marking | 29 | Meadow Mari |
| Ergative-Possessive | Object (and intransitive subject) is a possessor; doer is pushed to the side | 21 | Russian |
| Double-Possessive | Everyone is a possessor | 7 | Tukang Besi |
| Mixed | Several patterns at once | 14 | English |
| No action nominals | No such construction | 42 | Gagauz |
Typologists, the linguists who compare how grammar works across languages, want to know why the distribution looks like this. That’s the question the AI study set out to test.
How do you test grammar with fake languages?
You can’t run a controlled experiment on Russian or Yoruba, because real languages differ in thousands of ways at once. So the team used what they call a grammar-engineered method. They took the meaning of real sentences from the 100-million-word BabyLM corpus and re-expressed it through computational grammars, one per invented language.
Each language changed five settings:
- Clause word order (3 options, such as SVO, the English subject-verb-object order)
- Noun-phrase word order (2 options)
- Alignment, meaning how subjects and objects are case-marked (2 options)
- Complement system: whether embedded clauses keep a full verb (“balancing”) or use a reduced form (“deranking”) (2 options)
- ANC strategy: sentential, possessive-accusative, ergative-possessive or nominal (4 options)
Multiply those together and you get 96 languages. Think of it like a car designer’s test track. Instead of comparing a Toyota to a Volvo, you swap one part at a time and see what changes.
Each language got roughly 4.1 million generated sentences. The models were then scored on 34 grammatical phenomena using minimal pairs, which are two sentences that differ in one detail where only one is grammatical. Average accuracy across everything was 88.4%.
Where did the models match real languages?
The basics. Clause word order, the position of subject, verb and object, was learned almost perfectly, often near 100%. Combinations that are common in the world’s languages, like SVO with nominative-accusative marking (the English-style system), performed well across the board.
That fits earlier work. In “Mission: Impossible Language Models” (ACL 2024), Julie Kallini and colleagues showed GPT-2 struggles with languages that break human-like rules. A 2025 follow-up, “Anything Goes?”, extended that test to 12 languages and found the split between possible and impossible languages was real but not perfectly clean.
Where did language models and action nominal constructions diverge?
Once the grammar got richer, the parallels faded. Models found the verb form inside an ANC easy (91.9%) and marking the object easy too (90.1%). Marking the doer was much harder, at 68.7%. Marking an intransitive subject was the weakest of all, at 62.9%.
The big surprise came from rare combinations. SVO paired with deranked complements appears in just one language in the comparison sample, yet the models handled it well. Deranking combined with the nominal ANC strategy appears in zero sample languages, and the models still scored 89.7% on ANC-related tests.
In other words, if ease of learning were the whole story, you’d expect that combination to show up in human languages. It doesn’t.
Why does this matter beyond linguistics?
Critics including Noam Chomsky have argued that neural language models tell us little about human language because they can learn almost anything. This paper offers a nuanced answer. Models do share some human biases, but not all of them.
The authors suggest the gaps come from communicative and usage pressures that a lab setup can’t reproduce: what speakers need to say, how often, and to whom. My take: that’s a healthy result. It warns against treating a chatbot’s grammar sense as a stand-in for a child’s, a point worth remembering whenever someone claims AI “understands” language. If you’re curious how these models are reaching everyday users, our look at how AI is transforming small businesses covers the practical side.
What happens next
The paper (arXiv:2610.22468, submitted on 15 October 2026) is a 41-page preprint, so it hasn’t been peer reviewed yet. Its method builds on the University of Washington’s grammar-engineering tradition, including work adding action nominals to the LinGO Grammar Matrix.
The obvious next steps are bigger models, other architectures, and training setups that add some communication pressure, such as agents that must actually be understood. Each would show whether the divergences are a quirk of GPT-2-small or something deeper.
FAQ
What is an action nominal construction in simple terms?
It’s a noun phrase that describes an event, like “the enemy’s destruction of the city” instead of “the enemy destroyed the city”. Languages differ in how they mark who did what inside it.
Which AI model did the study use?
GPT-2-small, trained from scratch on each of 96 synthetic languages, with three models per language and about 4.1 million sentences each.
Does this prove AI learns language like humans?
No. The models matched human languages on basic word order but diverged on complex structures, and they even learned patterns no human language in the sample uses.
Want to see the data yourself? Read the full paper on arXiv, then compare its results with the WALS maps.
Leave a comment