An Error Taxonomy for AI-Copilot Assisted Spreadsheet and Analytics Work

For decades, the spreadsheet has been the quiet engine of the modern world, the digital ledger where businesses calculate profits, track inventory, and plan budgets. It is a tool so ubiquitous that most people use it without thinking about how it works, yet researchers have long known it is fragile. Studies going back to the 1990s have shown that the vast majority of these spreadsheets contain at least one serious mistake, often hidden in a single formula that looks correct but calculates the wrong number. For years, the assumption was simple: these errors happen because a human made a slip of the hand or a mistake in logic, like typing a wrong number or dragging a formula to the wrong cell. The solution, in the eyes of experts, was always human review—having another person check the work to catch the slip before it caused damage.

But the landscape of how these spreadsheets are built has shifted dramatically in the last few years. Since around 2023, a new kind of helper has appeared inside these tools: an artificial intelligence assistant that can write formulas and build complex reports just by listening to a human's request in plain language. This technology promises to speed up work, but it introduces a new uncertainty. We know these AI systems make mistakes, sometimes getting the math wrong even when the result looks plausible. The critical question that has gone unanswered is whether a human reviewer can spot an AI mistake as easily as they spot a human one. If an AI generates a number that is slightly wrong but looks reasonable, will a tired employee notice it, or will they trust the polished output and let a dangerous error slip through?

A new study proposes a way to answer this question by treating the human reviewer not just as a safety net, but as a subject of the experiment itself. The researchers, working from the National University of Bangladesh and East West University, designed a controlled comparison to see if the source of an error changes how likely it is to be caught. They do not simply ask computers to solve math problems and see how often they fail; instead, they propose a scenario where real working analysts build spreadsheets in two different ways. One group would build their work manually, typing out every formula themselves. The other group would use an AI assistant to generate the same formulas based on their instructions. Both groups would be given identical tasks, such as creating a budget report or summarizing sales data, ensuring that the difficulty of the work is exactly the same for everyone.

Once the spreadsheets are built, the researchers plan to perform a rigorous audit. They will examine every file to find errors, classifying them into specific types. They will look for "loud" errors, which are obvious mistakes like a broken link or a total that clearly does not add up, and "silent" errors, which are trickier. A silent error is one where the computer produces a number that looks perfectly normal but is actually wrong because it used the wrong data or misunderstood a rule. The researchers hypothesize that while AI assistants may be good at avoiding obvious typos, they frequently produce these silent errors. They often get the structure of a formula right but misinterpret the subtle relationships between different parts of the data, resulting in a plausible-looking but incorrect answer.

The most significant part of the study is the second layer of testing, which focuses on the human reviewers. A separate panel of analysts, who do not know which spreadsheets are built by humans and which by AI, will review the files under normal, time-pressured conditions. They will not be told that errors are present; they will simply be asked to check the work as they would in a real job. The researchers want to see if the reviewers are more likely to miss an error if it comes from an AI. The hypothesis driving this part of the study is that humans suffer from a kind of trust bias when looking at automated output. Because the AI-generated work often looks clean and professional, a reviewer might feel more confident in it and spend less time scrutinizing the details, leading them to overlook mistakes that they would have easily spotted if they knew a human had made them.

The study does not yet have the final numbers, as it outlines the method for a controlled experiment rather than reporting the completed results. However, the design is built to test a specific and worrying possibility: that the errors introduced by AI are harder to catch than errors made by humans, even when the mistakes are of the same severity. If the study confirms this, it suggests that the current way companies check their work is insufficient. It implies that simply having a human look at the screen is not enough if that human is subconsciously trusting the machine more than a person. The researchers argue that if AI errors are indeed harder to detect, organizations adopting these tools might be increasing their risk of making bad decisions, even if the AI is technically faster or more accurate on paper.

This research aims to bridge a gap between two fields that have rarely spoken to each other. On one side, there is decades of research on spreadsheet errors, which has established a clear map of how humans make mistakes. On the other side, there is a flood of new data showing how AI models fail at these same tasks. What has been missing is a direct comparison that measures the human reaction to both. By using the same classification system for errors made by humans and by machines, and by measuring how often those errors are found by a third party, the study seeks to provide a clear, evidence-based answer. It moves beyond asking "how often does the AI get it wrong?" to ask the more practical question: "how often does the AI get away with it?"

The implications of this work extend far beyond the spreadsheet itself. It touches on a broader issue of how humans interact with intelligent machines in high-stakes environments. If the study finds that we are less vigilant when reviewing AI work, it suggests that the tools we use to make decisions may be creating a blind spot in our own judgment. The researchers plan to measure not just whether errors are found, but also how confident the reviewers feel when they say a file is correct. This "calibration" is key; if reviewers are just as confident in AI-generated work as they are in human work, but are actually less accurate in their judgment, it points to a fundamental shift in how we need to verify information.

Ultimately, this paper is a proposal for a necessary experiment in a world that is changing faster than our safety habits. It recognizes that the old rules of checking work may no longer apply when the work is being written by a machine. By setting up a fair, controlled test, the authors hope to give businesses and analysts a clear picture of the risks they face. They are not saying that AI is useless or that it should be banned; rather, they are suggesting that we need to understand the specific nature of its failures before we trust it with our most important numbers. The goal is to ensure that as we hand over more of the writing to machines, we do not lose the ability to see the truth when it is hidden in plain sight.