Cognitive Load and Information Processing in Financial Markets: Theory and Evidence from Disclosure Complexity
This paper challenges the traditional view of disclosure complexity as a static document property by proposing a reader-task-interface framework, demonstrating through historical data and large-scale experiments that cognitive load in financial markets is dynamically determined by the interaction between content, access representation, and AI-driven transformation technologies rather than the document alone.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the world of finance, companies are required to publish detailed reports about their money, their debts, and their future plans. For decades, the assumption was that the difficulty of understanding these reports depended entirely on the document itself. If a report was long, filled with complex sentences, or packed with dense tables, it was considered "hard to read," and that difficulty was thought to be a fixed property of the paper. This idea shaped how researchers studied investor behavior, assuming that a complicated report would confuse everyone equally, regardless of who was reading it or what tools they used. However, the way information is delivered has changed. Companies now submit these reports in two forms at once: a traditional document meant for human eyes, and a structured digital file designed for computers to read instantly. This shift raises a new question: does the difficulty of a report stay the same when a human reads it versus when a machine processes it?
A team of researchers set out to answer this by treating the difficulty of a financial report not as a single number, but as a result of three things working together: the reader, the task, and the interface. They wanted to see if the old rules still applied in an age where artificial intelligence and automated software are increasingly the ones reading these documents. To do this, they looked at historical data from the US Securities and Exchange Commission, analyzed thousands of modern filings, and ran a massive experiment where they asked computer models to answer questions about company reports using different methods. Their work reveals that the complexity of a financial report is not a stable trait of the document. Instead, it changes depending on who is asking the question and how they are asking it.
The researchers began by looking back at a specific moment in history when the rules for filing changed. Around 2009, large companies were required to start submitting their financial data in a special, machine-readable format called XBRL, alongside their standard reports. The team reconstructed the traffic logs from the government website to see how people and computers reacted to this change. They found that the rule did not act like a simple on-off switch. While the requirement did increase the use of the machine-readable files, it did not cause a massive drop in people reading the standard documents. More importantly, the data suggested that the two formats served different purposes. The machine-readable files were accessed more often by automated systems, while the human-readable documents remained the primary source for direct reading. This indicated that the "burden" of reading a report was shifting. For a human, the burden was still about reading long pages of text. For a machine, the burden was about finding the right numbers inside a structured file.
To understand how these two worlds interact today, the researchers examined over 200,000 modern filings. They discovered a surprising pattern: the companies that produced the longest, most complex reports for humans were often the same companies that provided the richest, most detailed data for machines. In the past, researchers might have assumed that a long report was simply "hard" for everyone. The new data showed that a report could be very difficult for a person to read because of its length and vocabulary, yet be extremely easy for a computer to extract specific facts from. This means that the old way of measuring complexity—counting words or sentence length—no longer tells the whole story. A report can be "hard" for a human and "easy" for a machine at the same time. The researchers also found that larger companies, which have more complex businesses, tended to have both higher human reading burdens and higher machine accessibility. This suggests that the gap between what is easy for a person and what is easy for a computer is not random; it is a structural feature of the modern financial system.
The most revealing part of the study was a controlled experiment where the researchers tested how well artificial intelligence could answer questions about these reports. They took 432 specific questions about company finances, such as "What was the net income?" or "What is the current ratio?" and asked six different AI models to answer them. They tested these questions under twelve different conditions. In some cases, the AI was given the full, messy human document. In others, it was given a perfectly organized list of just the facts it needed. In some cases, the AI had to find the facts itself, while in others, the facts were handed to it directly.
The results were clear and specific. When the AI was given the full human document and had to find the numbers itself, it got the answers right only about 47 percent of the time. However, when the researchers gave the AI the exact same text but organized it so the relevant facts were easy to find, the accuracy jumped to nearly 100 percent. This proved that the main problem for the AI was not that it couldn't understand the language or the math. The problem was that it couldn't find the right numbers in the middle of thousands of words. The researchers also tested whether the special machine-readable format (XBRL) was inherently better than plain text. They found that once the AI was given the exact same facts in both formats, the difference in accuracy was almost zero. The advantage of the machine-readable format came not from the format itself, but from the fact that it made it easier to locate the specific numbers needed.
This finding challenges the idea that simply switching to a machine-readable format solves all problems. The researchers showed that the real bottleneck is "evidence localization"—the ability to find the right piece of information in a large document. If a computer can find the number, it can calculate the answer correctly, whether the number came from a structured file or a plain text document. The study also tested whether the AI was just memorizing the answers from its training data. They changed the numbers in the reports to new, made-up values and asked the AI to answer again. The AI still got the answers right, proving that it was actually reading the document and doing the work, rather than just recalling old facts.
The study concludes that we can no longer treat the complexity of a financial report as a single, fixed score. A report's difficulty depends on the reader, the task, and the tools used to access it. For a human investor, a long report with complex language is a heavy burden. For a machine, the burden is finding the right data point in a sea of text. The researchers found that the gap between human difficulty and machine ease is widening, and that the tools we use to access information matter more than the document itself. This means that as we move further into an era where artificial intelligence helps us make financial decisions, we need to stop thinking of reports as static objects and start thinking of them as dynamic inputs that change based on how they are processed. The difficulty is not in the paper; it is in the path between the paper and the person or machine trying to understand it.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.