Public pension trustees and administrators routinely encounter hundreds of pages of investment reports, actuarial analyses, legal opinions, policies and supporting documents. Some passages are easy to absorb. Others require a second—or third—reading before their meaning becomes clear.

New research suggests that going back over difficult material is not simply a sign that a reader has lost focus. It is part of how the human mind detects and corrects an initial misunderstanding.

Researchers from New York University and the University of Massachusetts Amherst used eye-tracking technology to study 368 adults as they read syntactically challenging sentences. The researchers then compared the participants’ eye movements with predictability estimates generated by more than 400 language-model configurations.

The study, published in the Proceedings of the National Academy of Sciences, finds that humans and large language models process language similarly during the early stages of reading. Both use context to anticipate what word is likely to come next.

The difference appears when a sentence becomes more complicated or leads the reader toward an incorrect interpretation. Humans often stop and move their eyes backward to reconsider earlier words. The language models’ next-word predictions did not fully account for that rereading and reinterpretation.

Researchers estimate that about 20% of human eye movements during reading are backward. Futurity recently reported on the findings, which researchers describe in the PNAS study.

What the research means for pension boards

The study does not examine pension board packets or AI document-summary tools. Still, its findings provide a useful lens for trustees and administrators working with complex governance documents.

For trustees, rereading may represent the moment when meaningful analysis begins. A reader recognizes that a sentence does not fit the surrounding information, revisits an assumption and reconstructs the meaning. That process can be particularly important when reviewing actuarial assumptions, investment recommendations, proposed policy changes or legal guidance.

A concise AI-generated summary may identify major topics and help a trustee navigate a lengthy document. It may define unfamiliar terminology, create a list of questions, or point readers toward sections that deserve closer attention.

But a summary should remain an entry point—not a substitute for the underlying material. AI may omit a qualification, flatten competing arguments, or fail to recognize why a particular phrase changes the meaning of a recommendation. A chatbot also does not carry a trustee’s fiduciary responsibility.

Trustees using an AI-assisted summary should be able to trace each material point back to the original document, review the surrounding context, and question anything that appears unclear or inconsistent. Systems should also follow their data-security policies and avoid placing confidential, privileged, or personally identifiable information into unapproved AI tools.

A lesson for administrators, too

The research also reinforces the importance of how board materials are written and organized. Complexity sometimes comes from the subject matter, but it can also come from long sentences, undefined terminology, buried qualifications or unclear decision points.

Administrators can make packets easier to process by:

• Providing a short executive summary that identifies the issue, recommended action and supporting rationale.
• Using descriptive headings and separating background information from material requiring board action.
• Defining actuarial, legal and investment terminology when it first appears.
• Breaking long, multi-clause sentences into shorter statements.
• Clearly identifying risks, alternatives and assumptions.
• Adding page references that allow trustees to verify summary statements against source material.
• Providing board materials early enough to allow time for careful reading and follow-up questions.

AI tools may eventually help systems identify unusually complex passages or produce preliminary reading guides. The new research, however, underscores why the human reader remains essential.

For pension boards, the backward glance is not a flaw in the reading process. It may be the point at which information becomes understanding—and understanding becomes sound governance.

Want a concise yet in-depth explanation? Press play above for an easy-to-follow TEXPERS Deep Dive podcast discussion of the research and what it means for public pension trustees and administrators.

Editor’s Note: Generative artificial intelligence was used to assist in creating the banner graphic accompanying this blog post.

About the Author: Allen Jones is the director of communications and event marketing for TEXPERS, where he leads editorial strategy, member communications, conference marketing, and digital engagement initiatives for Texas public employee retirement systems. He began his journalism career in 1998 and has worked in journalism and communications for more than 25 years.