How Eye-Tracking and AI Reveal Predictive Human Reading

Psycholinguists are combining eye-tracking technology with artificial intelligence to understand how the human brain processes written language in real time. This article explores how researchers use reading corpus models and AI-driven surprisal theory to test whether humans probabilistically predict upcoming words during sentence comprehension. By comparing human eye movements with the mathematical probability distributions generated by language models, scientists can accurately measure the cognitive effort required to process unexpected words and confirm the predictive nature of the brain.

Understanding Surprisal Theory

At the core of this research is surprisal theory, which states that the cognitive effort required to process a word is directly proportional to how unexpected that word is in a given context. In information theory, "surprisal" is calculated as the negative logarithm of a word's probability of appearing next, given the preceding words.

If a sentence begins with "The baker took the bread out of the...", the word "oven" has a high probability and low surprisal. Conversely, the word "freezer" has a low probability and high surprisal. Psycholinguists use this framework to test if the human brain constantly calculates these probabilities while reading.

Eye-Tracking Reading Corpora

To measure human cognitive effort during reading, researchers rely on eye-tracking reading corpora. These are large datasets recorded while human participants read natural texts, such as newspaper articles or novels, while a high-speed camera tracks their gaze.

When we read, our eyes do not move smoothly; they make rapid jumps called saccades and pause on words in moments called fixations. Eye-tracking captures three primary metrics that reflect processing difficulty:

• First-Fixation Duration: The length of the very first pause on a word, indicating initial recognition effort.
• Gaze Duration: The sum of all fixations on a word before the eyes move to another word, reflecting deeper integration into the sentence.
• Regressions: Instances where the reader’s eyes jump backward to earlier words to resolve confusion.

Longer fixation times and frequent regressions directly indicate higher cognitive load.

Artificial Intelligence as a Probabilistic Model

To test if eye movements align with surprisal theory, psycholinguists need a way to calculate the exact probability of every word in the reading corpus. This is where artificial intelligence, specifically large language models (LLMs), enters the equation.

AI language models are trained on vast amounts of text to predict the next word in a sequence. For any given point in a sentence, the AI generates a probability distribution across its entire vocabulary for what the next word might be. Researchers feed the exact texts from the eye-tracking corpora into these AI models to extract the precise surprisal value for every single word.

Testing Probabilistic Prediction

Once researchers have both the human eye-tracking data and the AI-generated surprisal values, they use statistical models to compare them. They control for baseline factors that naturally affect reading speed, such as word length and how frequently a word appears in the language overall.

The results consistently show a strong, linear relationship between AI surprisal and human fixation durations. When an AI model assigns a low probability (high surprisal) to a word, human readers reliably spend extra milliseconds looking at that exact word.

Furthermore, because this effect is visible in the earliest eye-tracking metrics (like first-fixation duration), it proves that the brain is not merely reacting to words after seeing them. Instead, the brain actively pre-activates and predicts upcoming words using a probabilistic system remarkably similar to artificial intelligence. When the actual word violates the brain's statistical prediction, the eyes pause longer to allow the brain to update its understanding of the sentence.