AI language models can develop a genuine mathematical understanding of the real world that closely mirrors how humans judge whether events are commonplace, unlikely, impossible, or nonsensical. A new study from Brown University reveals that large language models (LLMs) encode something like the causal constraints of reality itself, opening new doors for building more trustworthy and interpretable AI systems.

What Does It Mean for AI to "Understand" the Real World?

Researchers at Brown University designed an experiment to peer inside the mathematical workings of several AI language models and determine whether they truly grasp the difference between plausible and implausible scenarios. The team tested models by feeding them sentences describing events of varying plausibility, then examining the internal mathematical states the models generated in response.

The researchers used an approach called mechanistic interpretability, which functions like neuroscience for artificial intelligence. Rather than treating AI models as black boxes, mechanistic interpretability seeks to reverse-engineer what the model is doing when exposed to particular inputs, essentially understanding what is encoded in the "brain state" of the machine.

"Mechanistic interpretability can be appropriately characterized as something like neuroscience for AI systems. It seeks to reverse-engineer what the model is doing when exposed to a particular input. You could kind of think about it as understanding what is encoded in the 'brain state' of the machine," explained Michael Lepori, a Ph.D. candidate at Brown who led the work.

Michael Lepori, Ph.D. Candidate at Brown University

How Did Researchers Test AI Understanding?

The Brown team created test sentences across four categories to measure how well language models distinguish between different types of events. The experiment included commonplace scenarios like "Someone cooled a drink with ice," improbable events like "Someone cooled a drink with snow," impossible scenarios like "Someone cooled a drink with fire," and nonsensical statements like "Someone cooled a drink with yesterday".

For each input, the researchers examined the resulting mathematical patterns, or vectors, generated inside the AI model. By comparing the differences in these "brain states" across sentence pairs from different categories, they could determine whether, and how well, the models internally differentiate between plausibility levels. The experiments were repeated across several different open-source language models to ensure the findings were not specific to one system.

• Models Tested: Open AI's GPT-2, Meta's Llama 3.2, and Google's Gemma 2 to achieve a model-agnostic understanding of how these systems distinguish between event categories
• Accuracy Rate: Models of sufficient size developed distinct mathematical patterns that could distinguish between even the most similar categories, like improbable versus impossible events, with roughly 85% accuracy
• Emergence Threshold: These vectors started to emerge in models with more than 2 billion parameters, which is relatively small compared to today's trillion-plus-parameter models

Do AI Models Capture Human Uncertainty About Plausibility?

One of the most striking findings was that the mathematical vectors revealed by the study reflected human uncertainty about which category a statement might fall into. Consider the statement "Someone cleaned the floor with a hat." When people encounter this, they may genuinely disagree about whether it represents something impossible or merely unlikely.

The researchers analyzed the vectors to see how ambiguous the AI systems thought these borderline statements were, then compared that assessment with survey results from human participants. The results were remarkable: in cases where roughly 50% of people said a statement was impossible and 50% said it was improbable, the models were assigning approximately 50% probability as well.

"What we show is that the models actually capture that human uncertainty pretty well. In cases where, say, 50% of people said a statement was impossible and 50% said it was improbable, the models were assigning roughly 50% probability as well," noted Michael Lepori.

Michael Lepori, Ph.D. Candidate at Brown University

How Can This Research Improve AI Systems?

Understanding what AI models know and how they came to know it is crucial for developing smarter, more trustworthy systems. Mechanistic interpretability studies like this one provide a window into the internal logic of language models, revealing whether they have genuinely learned causal relationships about the world or are simply pattern-matching based on training data.

This research has practical implications for AI safety and reliability. If we can understand how models encode real-world constraints, we can better predict when they might fail, identify potential biases in their reasoning, and build systems that are more transparent and accountable. The findings suggest that modern AI language models do indeed develop an understanding of the real world that is reflective of human understanding, which is a significant step toward more interpretable and trustworthy AI.

The study was presented at the International Conference on Learning Representations in Rio de Janeiro in April 2026. The research was led by Michael Lepori and advised by Ellie Pavlick, a professor of computer science, and Thomas Serre, a professor of cognitive and psychological sciences, both faculty affiliates of Brown's Carney Institute for Brain Science.