November 25, 2025
In 1869, a French civil engineer named Charles Joseph Minard published what Edward Tufte would later call “the best statistical graphic ever drawn.” The famous visualization traced Napoleon's catastrophic 1812 Russian campaign, with a thick tan line representing the 422,000 soldiers who marched toward Moscow shrinking to a thin black thread of 10,000 survivors limping home. What made Minard's graphic revolutionary wasn't just its elegant brutality, it was its honesty about what could and couldn't be known. Minard carefully noted his sources and acknowledged the limitations of his casualty figures, embedding uncertainty into the very fabric of his masterpiece.
Today's scientists have inherited this duty of epistemic humility, and they express it through one of the most ubiquitous yet least understood features of modern research: the error bar. Those little vertical lines that sprout from data points like whiskers on a graph have become the visual vocabulary of doubt in an age that demands certainty. Yet for all their prevalence in scientific papers, medical studies, and policy reports, error bars remain deeply misunderstood, not just by the public, but often by the very researchers who deploy them.
The modern error bar emerged from the statistical revolution of the early twentieth century, when mathematicians like Karl Pearson and Ronald Fisher were developing the conceptual machinery that would transform science from a collection of observations into a rigorous system of inference. Fisher, working at the Rothamsted agricultural research station in the 1920s, pioneered the analysis of variance and developed many of the statistical tests still used today. His work established the framework for expressing not just what researchers observed, but how confident they could be in their observations.
But here's where things get surprising: error bars don't actually tell you what most people think they tell you. When you see a bar graph with error bars in a newspaper article about a new drug trial or climate study, you might reasonably assume those bars represent the range where the true value probably lies. You would be wrong. Or at least, you would often be wrong. Error bars can represent standard deviations, standard errors, confidence intervals, or credible intervals, and each tells a fundamentally different story. Standard deviation describes the spread of the data itself. Standard error describes the uncertainty about the mean. A 95% confidence interval, meanwhile, doesn't mean there's a 95% chance the true value falls within those bounds (a common misconception that drives statisticians to despair) but rather that if you repeated the experiment infinitely, 95% of the calculated intervals would contain the true value.
This confusion isn't merely academic. A 2005 study published in Psychological Methods found that even professional researchers frequently misinterpret error bars on graphs. The psychologist Sarah Belia and colleagues showed graphs to scientists and asked them to make inferences about statistical significance. The results were sobering: researchers consistently overestimated their ability to judge whether differences were statistically meaningful by eye. The error bar, that symbol of rigor, had become a Rorschach test.
The deeper surprise is that error bars may be teaching us the wrong lesson about uncertainty altogether. They create what might be called an aesthetic of precision by offering a visual suggestion that we have successfully bounded the unknown. But the universe doesn't actually come with error bars attached. What we're really seeing is the uncertainty that stems from sampling, from measurement, from the particular statistical model we've chosen to impose on nature's chaos. Crucially, error bars say nothing about systematic errors, about whether we've asked the right question, or whether our entire theoretical framework might be subtly wrong.
Consider the replication crisis currently roiling psychology, medicine, and other fields. Studies fail to replicate not because someone calculated the error bars incorrectly, but because the error bars were calculated on noisy data from small samples, or because researchers selected results that looked significant, or because publication bias means we see only the studies where error bars didn't overlap. The error bar is impeccable in its mathematics, but powerless against these deeper problems.
The rise of artificial intelligence has brought these tensions into sharp relief, revealing both the power and the peril of our statistical certainties. Modern AI systems, particularly the large language models and image generators that have captured the public imagination, don't typically come with error bars at all. When ChatGPT answers a question or DALL-E generates an image, there's no confidence interval, no visual whisker indicating, “I'm 73% certain about this token.” Instead, these systems project an unsettling confidence, hallucinating facts with the same smooth assurance they use to report genuine information. The absence of error bars in AI outputs represents a regression to a pre-statistical age, where authority spoke without acknowledging doubt.
Yet paradoxically, AI is also transforming how we think about uncertainty in scientific research itself. Machine learning models trained on vast datasets can now generate predictions with error estimates that dwarf traditional statistical approaches in complexity. In drug discovery, neural networks predict how molecules will bind to proteins, outputting not just predictions but confidence scores based on the model's uncertainty about its own weights and architecture. In climate science, ensemble models run thousands of simulations to generate probability distributions that get distilled into error bars on temperature projections. These aren't Fisher's error bars anymore but rather the error bars of computational brute force, where uncertainty is estimated by exploring vast possibility spaces rather than calculated from narrow mathematical formulas.
The AI revolution has also exposed a troubling truth about error bars in an age of big data: sometimes we have so much data that our error bars shrink to invisibility, creating an illusion of certainty even when our models are fundamentally misspecified. Training a neural network on millions of images produces predictions with tiny error bars, but those bars say nothing about whether the model has learned genuine patterns or merely latched onto spurious correlations in the training data. An AI system might predict recidivism rates with impressively tight confidence intervals while encoding historical biases that no error bar can reveal. The error bar becomes a fig leaf, covering systematic uncertainty with the appearance of statistical rigor.
Perhaps the most profound irony is that as error bars have become more sophisticated (bootstrapped, Bayesian, adjusted for multiple comparisons), they may have paradoxically increased our false sense of certainty. The physicist Richard Feynman warned about this in his famous 1974 Caltech commencement address about “cargo cult science,” noting that the first principle of science is “you must not fool yourself—and you are the easiest person to fool.” Error bars can become a performance of rigor that substitutes for genuine self-skepticism.
Yet for all their limitations, we cannot do without them. Error bars represent science's crucial acknowledgment that knowledge is provisional, that measurements are imperfect, that we see through a glass, darkly. They are visual reminders that certainty is not a prerequisite for action, and that understanding the boundaries of what we know is itself a form of knowledge.
The challenge for this century is to cultivate what might be called error bar literacy, which is not just the technical ability to calculate them correctly, but the wisdom to interpret them appropriately. We need to teach students and journalists and policymakers not just that error bars exist, but what they mean and what they don't. We need visualizations that communicate uncertainty more honestly, that resist the seductive pull toward false precision.
Charles Minard understood this in 1869, even before modern statistics existed. His graphic about Napoleon's army didn't pretend to perfect knowledge. It simply tried to show, as honestly as possible, what could be known and with what degree of certainty. The error bar is our attempt to do the same. It is a small mark with an enormous burden: to remind us, in every scientific chart and graph, that knowledge is hard-won, that certainty is rare, and that acknowledging what we don't know is the beginning of wisdom.
– Geoffrey W. Smith
First Five
First Five is our curated list of articles, studies, and publications for the month.
1/ More exercise, fewer heartbeats
“New research from Australia overturns the old idea that exercise “uses up” heartbeats. It shows that fitter people actually use fewer total heartbeats each day thanks to their lower resting heart rates, even when accounting for workouts. Athletes’ hearts beat around 10% less daily, saving over 11,000 beats per day compared to sedentary people. This efficiency not only signals cardiovascular fitness but also translates to longer life expectancy and lower disease risk.”
Read more here >
2/ More music, less dementia
“Older adults who regularly listen to or play music appear to have significantly lower risks of dementia and cognitive decline. The data suggests that musical engagement could be a powerful, enjoyable tool for supporting cognitive resilience in aging.”
Read more here >
3/ More coffee, less A-Fib
“In an unexpected twist, scientists found that a daily coffee may protect against atrial fibrillation. New research finds that daily coffee drinking may cut A-Fib risk by nearly 40%, defying decades of medical caution. Scientists discovered that caffeine’s effects on activity, blood pressure, and inflammation could all contribute to a healthier heart rhythm. The DECAF clinical trial’s findings suggest coffee could be not only safe but beneficial for people with A-Fib.”
Read more here >
4/ More cocoa extract, less aging
“Daily cocoa extract supplements reduced key inflammation markers in older adults, pointing to a role in protecting the heart. The findings reinforce the value of flavanol-rich, plant-based foods for healthier aging.”
Read more here >
5/ Less light, less inflammation
“Boston researchers linked nighttime light exposure to greater stress-related brain activity and inflamed arteries, signaling a higher risk of heart disease. The study suggests that artificial light at night disrupts normal stress responses, leading to chronic inflammation. Experts call for reducing unnecessary light in cities and homes to protect cardiovascular health.”
Read more here >
To subscribe to Engineering Biology by Jacob Oppenheim, and receive newly published articles via email, please enter your email address below.
