An artificial intelligence model reveals the hidden decisions underlying human reading, showing how people choose which words to focus on, which to skip, and which to revisit. The study’s findings were published in the journal Nature Human Behaviour.
Reading appears to be a smooth movement across a line of text, yet your eyes pause on certain words, skip others, and return to phrases seen moments earlier. Four researchers now view each of these steps as a decision regarding where the next gaze will yield the greatest benefit. Based on this single premise, their program operates in a manner strikingly similar to human reading.
The model was developed by a team from Aalto University in Finland, in collaboration with colleagues from the City University of Hong Kong and the National University of Singapore. They created a computer program rather than a device for measuring human brain activity; its ability to explain the reading process is gauged by how closely its behavior aligns with observed human behavior.
For decades, two distinct lines of research have described reading, yet they have evolved largely independently of one another. On one hand, researchers have explained eye movements by treating each saccade—or rapid jump—as a byproduct of word recognition speed.
On the other hand, researchers have explained how meaning accumulates—sentence by sentence—coalescing into something that can be summarized later. Neither approach, however, explained the link between eye movements and comprehension.
Yunpeng Bai, Xiaofu Jin, Shengdong Zhao, and Antti Oulasvirta start from a simple premise: attention, memory, and time are finite resources, forcing the reader to constantly decide which words warrant attention and which do not. “Reading seems easy, but your brain is constantly deciding where to look, what to skip, and when to backtrack—allocating attention like a budget to maximize comprehension,” says Zhao, a professor at the City University of Hong Kong.
The program operates on three levels simultaneously. At the lowest level, it selects which letters within a word to fixate on, since a single glance captures only part of the word. Above that, a mechanism decides which word to read next or whether to skip it. At the highest level, it determines which sentences to read and which to reread.
No rules were hard-coded at these levels to forbid skipping short words or rereading confusing ones. Instead, the team trained each level using trial and error on millions of texts, selecting the strategies that yielded the best comprehension with the least time and effort.
This training method is known as reinforcement learning—the same approach used to teach robots how to move.
“We placed the model in a world containing millions of texts. Then, using AI-based reinforcement learning, we trained it to optimize eye movements so that it truly understood what it was reading,” said Oulasvirta.
The team tested the program against eight datasets of human reading patterns—most of which had been collected by other laboratories using eye-tracking equipment over the past few decades.
Humans spent longer looking at long words, and the program did the same. In both cases, common, predictable words were skipped more frequently, and less time was spent on those that weren’t skipped. Both humans and the program frequently backtracked to complex words or words that could be read in two different ways.
For every additional letter in a word, a human spent approximately 14 milliseconds longer looking at it; in the simulation, this figure was about 21 milliseconds. Thus, the program correctly identifies the direction of each effect but is only approximate regarding its magnitude.
One of the less significant results is particularly striking. When people first focus on a word, they do not aim for the middle; their gaze lands between the beginning and the center. The program did exactly the same thing, even though it was never told where the target was.
In one experiment conducted by the team, 39 adults read short texts on a screen while an eye-tracking device recorded their gaze 1,200 times per second.
Thirty-two people were included in the text comprehension analysis, and 28 in the eye-movement analysis. The average age of the participants was 24. The team wanted to determine what a reader sacrifices when time is limited.
Each participant read for 30, 60, or 90 seconds. This was followed by a 20-second arithmetic task designed to interrupt them.
…memorizing the material, followed by a free-recall test and five multiple-choice questions. Due to time constraints, people read faster, skipped more words, backtracked less frequently, and retained less information. The program underwent similar changes.
Based on 15 aggregated metrics spanning three time periods, the program’s performance matched that of the human participants with a correlation coefficient of 0.9999.
The team also created versions without constraints; these versions no longer resembled standard reading programs. When the model could retain all the information it had read, it answered 85.2% of the multiple-choice questions correctly, compared to 71.6% for the human participants.
It also performed better on the free-recall test and virtually eliminated backtracking. There was no need to reread anything, as nothing had been forgotten.
The team created other, “shortsighted” versions that considered only the immediate next step. Users of these versions would read the beginning of the text and stop there, progressing at speeds of 34 and 42 words per minute, whereas others managed 176 words per minute.
Reading that mimics human behavior requires both a memory capable of perfect retention and some sense of the value of each page. The authors explicitly state that the model describes readers in general rather than you personally; they did not tailor it to any specific individual or verify whether the results matched that reader’s actual habits.
Secondly, the evidence is quite limited. The team conducted a single study under time constraints using one type of short text and university-aged participants; furthermore, the comprehension tests assessed only recall and quiz performance. The model also predicts that reading patterns for people with dyslexia should look different: longer pauses and more regressions—even on words the reader might otherwise guess.
Researchers aim to develop reading support tools based on these findings. One concept involves text that adapts to the individual and the specific situation, similar to the smart glasses and lightweight headsets currently in development.
“We want to assist users in situations requiring quick reactions—for instance, by generating text that helps drivers without distracting them,” said Oulasvirta.
About the Author
