When transformers learn "impossible" languages, what do they learn?

When Transformers Learn "Impossible" Languages: A Deeper Look at Inductive Bias

A new preprint (arXiv:2606.30815) challenges recent claims that transformer language models exhibit a "human-like" bias toward natural languages over artificial, "impossible" languages that humans cannot acquire. The authors argue that previous studies conflated sample efficiency with fundamental learnability, and that transformers may not actually prefer human languages in the way earlier research suggested.

What the Research Actually Shows

The core finding is nuanced: when transformers are trained on both natural and impossible languages, differences in performance may stem from statistical properties of the training data rather than any innate preference for human-like linguistic structures. The paper suggests that transformers can learn impossible languages just as well—or nearly as well—as natural ones, provided they receive sufficient data. The apparent bias reported in prior work likely reflects differences in sample efficiency, not an architectural predisposition toward human language.

This distinction matters because it separates what a model can learn from what it learns quickly. Transformers, being statistical pattern matchers, will exploit whatever regularities exist in the data, regardless of whether those regularities align with human linguistic universals.

Why This Matters for AI Research

The debate over whether LLMs have an inductive bias toward human language touches on fundamental questions about the nature of language acquisition. If transformers genuinely preferred natural languages, it would suggest a deeper connection between neural network architectures and human cognitive biases—potentially validating certain Chomskyan claims about universal grammar.

This paper pushes back on that narrative. Instead, it reinforces the view that transformers are powerful general-purpose pattern learners. Their success on human language does not require any special linguistic endowment; it simply reflects the fact that natural languages contain rich statistical structure that these models are designed to exploit.

For AI safety and alignment research, this has implications: if models can learn "impossible" languages just as well, they may also internalize non-human-like reasoning patterns or reward structures that diverge from human values, even when trained on human data.

Implications for AI Practitioners

• Data quality over architecture: The findings underscore that model behavior is primarily driven by training data distribution, not architectural biases. Practitioners should focus more on data curation than on searching for "human-like" architectures.

• Sample efficiency as a metric: The difference between natural and impossible languages appears to be one of sample efficiency, not ultimate learnability. This suggests that benchmarks comparing model performance across language types should control for data quantity.

• Interpretability caution: Claims about models having "human-like" biases should be scrutinized. Transformers may simply be better at exploiting the statistical regularities present in human language data, not preferring it for deeper reasons.

Key Takeaways

• Transformers can learn "impossible" languages as effectively as natural languages, given sufficient data—the apparent bias is about sample efficiency, not fundamental learnability.
• Prior claims of human-like inductive bias in LLMs may be overstated, reinforcing the view that these models are general-purpose pattern matchers.
• For practitioners, this means data distribution and quantity matter more than architectural choices for shaping model behavior.
• AI safety researchers should note that models can internalize non-human-like patterns, even when trained on human data, which has implications for alignment.