Picture this: It’s 2001, and Enron is riding high as one of America’ds most innovative companies. Behind the scenes, executives are making billion-dollar decisions based on spreadsheets. Complex financial models that determine the fate of investments, employee pensions, and shareholder wealth. Surely these critical documents are error-free, right?
Not quite. When researchers later analyzed over 15,000 real Enron spreadsheets, they discovered formula errors scattered throughout the corpus. While the Enron case revealed specific instances of errors, it was just the tip of an iceberg that research would later reveal spans every industry, every document type, and every level of human expertise.
While we obsess over AI hallucinations and their potential for error, there’s a more uncomfortable truth we need to face: humans have been quietly making mistakes in documents at rates that would make even the most error-prone AI model look impressive.
The Spreadsheet Disaster Hidden in Plain Sight
Let’s start with the numbers that should keep CFOs awake at night. A comprehensive review of spreadsheet audit studies found that approximately 94% of operational spreadsheets contain errors. Not 9.4%. Not “a few problem cases.” Ninety-four percent.
The average cell error rate hovers around 5.2%, meaning that in a typical business spreadsheet with 1,000 cells, you can expect about 50 cells to contain mistakes. Some audits found error rates exceeding 6%, and in certain models, over 22% of unique formulas were flagged as incorrect.
Think about what this means for your organization. That quarterly revenue model? The budget projections your board approved? The risk calculations determining your insurance premiums? There’s a 94% chance they contain at least one error, and a meaningful chance that 1 in 20 cells is simply wrong.
The researchers put it bluntly: spreadsheet errors are “pervasive and can have severe impacts.” This isn’t about typos in someone’s personal expense tracker—these are errors in documents that drive million-dollar decisions.
When Doctors and Data Don’t Mix
Healthcare, with its rigorous training and life-or-death stakes, should surely perform better. The reality is more sobering.
An analysis of 1,485 healthcare data breaches between 2015 and 2020 revealed that human error was the leading cause of security incidents. Unintentional insider errors—like misdirected emails, lost devices, or falling for phishing scams—exposed over twice as many patient records per incident as malicious external attacks.
Let that sink in: the well-meaning doctor accidentally sending patient records to the wrong email address causes more data exposure than sophisticated hackers trying to break into the system.
The numbers are stark. In many analyses, over half of all healthcare data breaches involve human behavior rather than system failures. Phishing attacks—which exploit human psychology rather than technical vulnerabilities—were the single biggest cause of records compromised.
Even within medical records themselves, human documentation errors directly harm patients. Studies link over 60% of diagnostic malpractice claims to faulty electronic health record documentation. The very systems designed to improve patient safety are undermined by the humans using them.
The Peer Review Paradox
Science represents humanity’s most rigorous attempt at accuracy. Papers undergo months of peer review, with experts scrutinizing every claim and calculation. Surely this process catches human errors?
The evidence suggests otherwise. Researchers used automated tools to scan over 250,000 p-values in published psychology papers, checking for consistency between reported statistics and actual calculations. The results were striking:
- Approximately 50% of papers contained at least one statistical inconsistency or error
- 12.5% contained “gross” errors—mistakes large enough to potentially change a study’s conclusions
- Error rates showed no improvement over the nearly three-decade study period
These aren’t obscure papers in predatory journals. These are peer-reviewed studies in established psychology publications, representing the supposed gold standard of scientific rigor.
The Gene Name Catastrophe
Perhaps the most telling example of human error comes from genomics research. Scientists studying genetics often use Excel spreadsheets to manage gene lists. Excel, being helpful, automatically converts text that looks like dates into actual dates. A gene named “SEPT2” becomes “Sep-2.” “MARCH1” becomes “1-Mar.”
When researchers analyzed supplemental Excel files from genomics papers published between 2005 and 2015, they found that 19.6% of papers with gene lists contained these autocorrect errors. Over 700 papers had published corrupted gene data.
Here’s the kicker: after this problem was widely publicized in 2016, you’d expect scientists to adapt. Instead, a follow-up study found the error rate had increased to 30.9% by 2021. Despite warnings, training, and public awareness, nearly one-third of genetics papers continued to publish corrupted data. Eventually scientists rename human genes to stop Microsoft Excel from misreading them as dates.
This perfectly illustrates a key point about human error: knowledge of the problem doesn’t automatically solve it. Even highly trained experts, aware of specific pitfalls, continue making the same mistakes.
The PowerPoint That Killed
Sometimes human errors in documents have consequences beyond financial loss or research confusion. On February 1, 2003, the Space Shuttle Columbia broke apart during re-entry, killing all seven crew members.
The investigation revealed that engineers had identified potential damage from foam debris during launch. Their analysis was technically sound, but their PowerPoint presentation to NASA leadership drastically understated the risks. The Columbia Accident Investigation Board specifically criticized the reliance on “terse slides” that obscured critical information.
A document formatting decision—how to present technical analysis in slides rather than detailed reports—contributed to a catastrophic failure of risk assessment. Seven people died, in part, because humans struggle to accurately communicate complex information in simple formats.
Why Humans Keep Making Mistakes
Understanding these error rates requires examining why humans struggle with document accuracy:
Cognitive Limitations: Our brains excel at pattern recognition and creative problem-solving but struggle with sustained attention to detail. Error rates increase with fatigue, stress, and multitasking—conditions that define modern work environments.
Tool Mismatches: Many professionals use general-purpose tools (like Excel) for specialized tasks. Excel’s automatic formatting features that help in general business contexts actively corrupt scientific data. Users often don’t realize the tools are working against their accuracy goals.
Process Gaps: Organizations rarely implement systematic error-checking. The psychology paper study found that statistical errors persisted despite peer review, suggesting that even rigorous quality control processes have blind spots.
Overconfidence: Expertise can paradoxically increase error rates when experts become overconfident and skip verification steps that novices would take.
The AI Error Reality Check
This brings us to AI hallucinations—the current boogeyman of artificial intelligence. Yes, AI models sometimes generate false information with high confidence. Yes, this poses real risks that require mitigation strategies.
But let’s maintain perspective. The research shows that humans operating in their areas of expertise, with established quality control processes, still produce error rates up to 94% depending on the task and context.
Early evidence suggests that in specific domains, AI systems can achieve competitive or better error rates than humans. In radiology, for example, studies show human error rates of 3-6% on general examinations and 10-14% on specialized tasks, while some AI diagnostic systems approach these levels or better. In legal document review, AI systems can achieve 2-4% error rates compared to human paralegal error rates of 10-20%.
However, AI errors present unique challenges that differ qualitatively from human mistakes:
Systematic vs Random Patterns: AI models tend to make systematic errors that are consistent and reproducible, while human errors are often more random. This creates a trade-off: systematic errors can be easier to identify once detected but may be harder to initially notice precisely because they’re consistent.
Auditability Challenges: While AI outputs can theoretically be logged and tracked, research shows that auditing AI systems remains extremely challenging, particularly for large-capacity models commonly used in practice. The promise of easy AI auditability has not yet been realized in practice.
Reproducibility Issues: The AI field itself faces significant reproducibility challenges, with only 33-50% of AI studies being successfully reproduced. This suggests that claims about AI reliability and improvability should be viewed cautiously.
Context Sensitivity: Like humans, AI systems show sensitivity to how questions are framed and exhibit biases toward familiar information, creating error patterns that may be less predictable than initially assumed.
The Path Forward
This isn’t an argument for replacing humans with AI systems everywhere. Rather, it’s a call for realistic expectations and strategic deployment.
Instead of asking “How can we eliminate AI errors?” we should ask “How can we achieve better overall accuracy than current human-AI systems?” The answer likely involves:
Hybrid approaches where AI handles tasks with high human error rates (data entry, calculation verification, initial drafts) while humans focus on tasks requiring judgment, creativity, and contextual understanding.
Error detection systems that use AI to audit human work and humans to audit AI work, creating multiple layers of verification.
Honest accounting of error rates across human and AI systems, moving beyond anecdotal concerns to systematic measurement and improvement.
The spreadsheet audit findings suggest that we’ve been living with a 94% error rate in critical business documents while worrying about AI accuracy. The genetics research shows that even after identifying specific error patterns, human performance often gets worse rather than better.
Maybe it’s time to stop treating AI hallucinations as a uniquely disqualifying flaw and start treating them as another type of error to manage—one that might actually be more manageable than the human errors we’ve been tolerating all along.
The real question isn’t whether AI makes mistakes. It’s whether AI plus proper verification systems can make fewer mistakes than the error-prone humans currently creating our documents. Based on the evidence, that’s not just possible—it might be inevitable.
Sources
- Finding Errors in the Enron Spreadsheet Corpus – Schmitz & Jannach (2016)
- A Critical Review of the Literature on Spreadsheet Errors – Powell et al. (2008)
- Human Factors in Electronic Health Records Cybersecurity Breach: An Exploratory Analysis – Yeo & Banfield (2022)
- Human Error in Data Breaches of Electronic Health Records: A Systematic Literature Review – Alvarado & Triantis (2024)
- The Prevalence of Statistical Reporting Errors in Psychology – Nuijten et al. (2016)
- Gene Name Errors Are Widespread in the Scientific Literature – Ziemann et al. (2016)
- Gene Name Errors: Lessons Not Learned – Abeysooriya et al. (2021)
- The Boeing PowerPoint Slide – Columbia Accident Investigation Board analysis
- Should Artificial Intelligence Have Lower Acceptable Error Rates Than Humans? – Larsen et al. (2023)
- The Range of Error Accepted in AI Utilization vs. Human Error in the Legal Sector – ELLE Law Review (2023)
- AI vs Human Intelligence: Accuracy and Error Rates – Future Savant (2023)
- Strange Errors and Familiar Flaws: How AI and Human Mistakes Compare – UNU Campus Computing Centre (2025)
- The Unreasonable Effectiveness of Open Science in AI: A Replication Study – Gundersen et al. (2024)
- AI Auditing: The Broken Bus on the Road to AI Accountability – Mökander et al. (2024)
- Under Manipulations, Are Some AI Models Harder to Audit? – Godinot et al. (2024)