Capítulo 1
The Art of Detecting Deception in the Information Age
In a world increasingly saturated with misinformation, Carl Bergstrom and Jevin West's "Calling Bullshit" arrives as an essential survival guide. This book has quickly become required reading at universities nationwide, with over 70 institutions adopting its principles into their curricula. When Bill Gates included it on his 2021 summer reading list, he noted that "critical thinking has never been more important-or more challenging." The book's cultural impact extends beyond academia; its techniques have been embraced by journalists at outlets from The New York Times to ProPublica as essential professional tools. What makes this work particularly compelling is how it transforms complex statistical concepts into accessible strategies anyone can use. As we navigate an era where, according to MIT research, false news spreads six times faster than truth on social media, Bergstrom and West offer something increasingly rare: practical hope for reclaiming our information ecosystem.
Capítulo 2
Bullshit: Ancient Practice, Modern Problem
Deception isn't uniquely human. Ravens demonstrate sophisticated deception by pretending to cache food in one location while secretly storing it elsewhere when they believe they're being watched-approaching what humans do when spreading misinformation online. But human deception operates on an entirely different scale thanks to our complex language and theory of mind, allowing us to manipulate others' beliefs in countless ways.
We've developed linguistic techniques that exploit the gap between literal meaning and implication-a practice called "paltering." Bill Clinton's famous "there is no sexual relationship" defense exemplifies this technique, technically true in the present tense while deliberately misleading about the past. Corporate communications excel at this approach, using passive voice and euphemisms to diffuse responsibility. When Fiat Chrysler faced allegations about child labor in their supply chain, they responded with vague talk of "collaborative action with global stakeholders" rather than addressing the four-year-olds working in crude mines for pennies.
The digital landscape has supercharged bullshit's spread. As Jonathan Swift noted in 1710, "falsehood flies, and truth comes limping after it." Modern research confirms this principle: Facebook studies show false rumors consistently outperform true ones, even after being debunked. During the Boston Marathon bombing, a false story about an eight-year-old Sandy Hook survivor being killed received 92,000 shares while corrections garnered only 2,000. This creates a substantial disadvantage for truth-tellers, who consistently find themselves outpaced by those spreading misinformation.
The problem isn't that we lack access to facts-smartphones give us unprecedented information access-but that we rarely use this technology to verify claims. Instead, our media ecosystem has evolved to prioritize engagement over accuracy, creating an environment where bullshit thrives.
Capítulo 3
How Media Ecosystems Fuel Misinformation
Information revolutions have always triggered concerns about quality. In 1474, scribe Filippo de Strata lamented how the printing press would flood the market with cheap, trivial content. The internet has accelerated this democratization dramatically, allowing anyone to publish globally without cost. While this brings marginalized voices into conversation, it also enables amateur writers with poor journalistic standards to reach mass audiences.
The economics of online media fundamentally differ from traditional subscription models. Where newspapers built long-term relationships with readers who valued quality information, the internet runs on clicks generating advertising revenue. Headlines no longer aim to inform but to provoke emotional responses with phrases like "will make you" or "melt your heart." The unvarnished truth simply cannot compete in this marketplace where straight information is less valuable than emotional appeal.
Political polarization compounds these problems. Media outlets now cater to specific political perspectives, creating echo chambers that deepen ideological divides. Hyperpartisan content thrives on social media because sharing such content isn't primarily about conveying information but signaling tribal affiliation. Algorithms worsen this problem by personalizing content feeds to maximize engagement rather than inform, creating vicious cycles that limit exposure to diverse viewpoints.
By 2017, Facebook admitted 126 million US users had been exposed to Russian propaganda designed to deepen ideological divides on emotionally charged issues. This "firehose strategy" aims not to convince people of specific falsehoods but to overwhelm with contradictory stories until audiences become disoriented and unable to distinguish truth from fiction. As chess grandmaster Garry Kasparov noted, "The point of modern propaganda isn't only to misinform or push an agenda. It is to exhaust your critical thinking, to annihilate truth."
The problem extends beyond deliberate disinformation. Nearly half of internet traffic comes from bots, with Facebook deleting almost three billion fake accounts in 2018 alone. This undermines democracy when fake voices drown out real ones, as seen during the FCC's net neutrality comment period where millions of comments were fraudulent-including half a million submitted simultaneously in a single second.
Capítulo 4
Understanding Bullshit's Nature
What exactly constitutes bullshit? Philosopher Harry Frankfurt defined it as what people create when trying to impress or persuade without any concern for truth. It's the high school essay written without reading the book or the Silicon Valley pitch filled with meaningless jargon. The key distinction is that while liars know and care about the truth (they're just hiding it), bullshitters simply don't care whether what they're saying is true or false-their only concern is achieving their persuasive goal.
Sigmund Freud perfectly illustrated this concept in an 1884 letter to his fiancee about a lecture he gave: "I told about my discoveries in brain anatomy, all very difficult things that the audience certainly didn't understand, but all that matters is that they get the impression that I understand it."
Effective bullshitters shield their claims from scrutiny by creating rhetorical "black boxes"-complex-sounding explanations that intimidate questioners. When someone makes a simple claim like "cat people earn higher salaries than dog people," it's easy to question. But when they add layers of jargon, supposed TED talks, and personality types, refuting them requires substantial work.
The beauty of spotting bullshit is that you rarely need to open these black boxes-instead, examine what goes in (the data) and what comes out (the results). If the data are flawed or the results implausible, the specific technical details don't matter. This approach is especially important today when quantitative evidence carries undeserved weight despite how easily numbers can be manipulated.
Consider the 2016 paper claiming an algorithm could identify criminals from facial features with 90% accuracy. By examining the training data rather than the algorithm itself, we can spot the fatal flaw: the "criminal" photos came from government IDs where subjects weren't smiling, while "non-criminal" photos came from professional headshots where people typically smile. The algorithm wasn't detecting criminality-it was detecting smiles.
Capítulo 5
The Confusion of Correlation and Causation
We frequently mistake correlation for causation, leading to widespread misunderstanding. Correlation simply measures association between variables on a scale from -1 to 1, while causation indicates that one thing actually produces an effect on another. Though philosophers debate what causation truly is, we primarily care about it for practical purposes-to make things happen or prevent problems.
Media transforms correlational findings into causal headlines because prescriptive advice sells better than nuanced correlations. Health publications declare "Exercise Cuts Cancer Risk" based on correlational studies. Even scientific articles make similar mistakes, using causal language ("protects against") despite only showing correlation.
The famous marshmallow test perfectly illustrates this problem. The original research showed children who could delay gratification at age four had better outcomes later in life. While researchers carefully noted this as correlation, popular media transformed it into causation, with outlets like Fast Company claiming delayed gratification could transform performance "from mediocre to top-notch." However, follow-up research with better controls found socioeconomic status was the common cause of both delayed gratification ability and later success.
Some correlations exist purely by chance and reveal nothing meaningful about how the world works. Tyler Vigen's collection of spurious correlations humorously demonstrates this, showing strong correlations between unrelated variables like Miss America's age and murders committed with hot objects. These arise from data dredging-comparing numerous data sets until chance alignments appear. Simple trends over time are especially vulnerable; any two increasing quantities will correlate positively without causal connection.
When discussing causality, we must distinguish between probabilistic causes (A increases the chance of B), sufficient causes (if A happens, B always happens), and necessary causes (unless A happens, B can't happen). These distinctions are often misused to deny causal relationships. Mike Pence once claimed "smoking doesn't kill" because only one-third of smokers die from smoking-related illness-conflating sufficient cause with probabilistic cause. Similarly, arguing smoking doesn't cause lung cancer because some non-smokers get it conflates necessary cause with probabilistic cause.
Capítulo 6
Numbers as Vehicles for Deception
We live in a thoroughly quantified world where everything is counted, measured, and analyzed. While numbers appear objective and scientific, they're actually ideal vehicles for bullshit-they feel precise but are easily manipulated to tell whatever story one desires. Unlike words, which we recognize as subjective human constructs, numbers seem to come directly from nature, giving them an unearned authority.
Numbers come from different sources with varying reliability. Some are exact counts or direct measurements. Others are estimates derived from samples, which introduce potential sampling error. More problematic are measurement biases (like men exaggerating their height) and sampling biases (like estimating average height by measuring basketball players). Summary statistics can be particularly misleading, as when politicians report mean tax savings that benefit only the wealthy while the median family receives nothing.
Numbers need proper context to tell an honest story. Scotland loses 440,000 barrels of whiskey annually to evaporation-the "angels' share." This can be expressed per distillery (Macallan loses 220,000 LPA yearly), as percentage of starting volume (Pappy Van Winkle loses 58% during aging), or as percentage of final volume (138% of final volume). Without appropriate context, numbers become meaningless numerosity.
Percentages can mislead in multiple ways, especially when comparing changes. A 1 percentage point decrease in flu incidence from 2% to 1% actually represents a 50% reduction in cases. Changing denominators create further confusion, as with incarceration statistics where African American jail populations increased by 13% from 2000-2005, yet their percentage of total inmates decreased from 41.3% to 38.9% because white incarceration grew even faster at 27%.
When measurements become targets, they cease to be effective measurements-a principle known as Goodhart's Law. In colonial Hanoi, officials offered bounties for rat tails to control the rodent population, but people began breeding rats and cutting off tails while leaving the animals alive to reproduce. Similarly, college rankings drive perverse behaviors-schools manipulate acceptance rates by recruiting unlikely applicants, cap class sizes just below measurement thresholds, and employ tricks to boost reported SAT scores.
Zombie statistics-numbers that refuse to die despite being outdated, decontextualized, or entirely fabricated-spread because numbers appear rigorous simply by virtue of being quantitative. The common claim that "50 percent of scientific articles are never read" exemplifies this phenomenon. When investigated, this statistic traced back to papers from 1990-1991 that actually claimed 50 percent of papers go uncited after four years-not unread, a crucial distinction. Even this corrected claim was flawed: it only measured citations within a limited database and counted all journal items including obituaries.
Capítulo 7
The Hidden Influence of Selection Bias
Statistical analysis relies on examining samples to make broader inferences, but what you observe depends critically on where and how you look. Selection bias occurs when sampled individuals differ systematically from the eligible population. For instance, surveying class attendance among students present on a sunny Friday afternoon yields unreliable results-the diligent students who showed up aren't representative of the entire class.
Insurance companies exploit this when advertising that "customers who switched saved $500"-of course they did, as people only switch when they'll save money! Different insurers use different algorithms, so each attracts customers who benefit most from their specific approach. Similarly, a psychiatrist rarely sees patients with too little anxiety-they simply don't seek treatment.
Seemingly contradictory statistics can both be true due to selection effects. In Portugal, 60% of families have only one child, yet 60% of children have siblings-because multi-child families contribute disproportionately to the child population. Similarly, universities boast about small average class sizes while students complain about large classes. Both are right: if a department offers many small classes and a few large ones, administrators correctly report a low average class size, while the "experienced mean class size" for students is much higher since large classes serve more students.
Berkson's paradox explains seemingly negative correlations created through selection bias. When Google found programming contest winners performed worse as employees, this wasn't because contests teach bad habits, but because their hiring process already selected for programming skill, creating an artificial negative correlation. Similarly, the complaint that "hot guys are jerks" emerges when dating pools are restricted to those above certain thresholds of both attractiveness and niceness.
Data censoring creates selection bias when initially random samples become non-random. The viral graph showing rap musicians dying young while jazz performers lived longer suffered from right-censoring-excluding those still alive. Since rap is relatively new, most performers are still alive; only those who died prematurely appear in the data. It's not that rap stars die young; it's that rap stars who have died must have died young, because the genre hasn't existed long enough for performers to die of old age.
Capítulo 8
The Deceptive Power of Data Visualization
Data visualizations are essential because humans struggle to process raw data, but they can easily mislead without careful examination. Throughout the United States, Stand Your Ground laws permit civilians to use deadly force when threatened with serious harm. A Reuters data visualization of Florida homicides appeared to show murders dropping after the 2005 Stand Your Ground law-but the vertical axis was inverted! What seemed like a decrease was actually a sharp increase.
Data visualization "ducks" occur when ornament overwhelms purpose in graphics-named after buildings like the Big Duck of Flanders, NY where form dominates function. USA Today pioneered this approach with their Daily Snapshots feature, presenting simple information with thematically-related visuals (like lipstick tubes as bar charts). These designs sacrifice clarity for novelty, using only a fraction of space for actual data while making comparisons difficult.
Even more problematic are "glass slippers"-visualizations that shoehorn data into visual forms designed for entirely different purposes. Examples include countless "periodic tables" of unrelated concepts (marketing, typefaces, cryptocurrencies) that mimic Mendeleev's chemical arrangement without its underlying logic. Similarly, designers create "subway maps" for concepts lacking the sequential and spatial relationships that make actual transit maps useful.
Data visualizations can mislead through axis manipulation. Watch for truncated vertical scales that don't reach zero, disproportionate bar lengths, and inappropriate scaling choices. While bar charts should always include zero to avoid distorting magnitude comparisons, line graphs properly focus on changes and may exclude zero. Be wary of dual-axis plots where scales are manipulated to create false correlations, cherry-picked time frames that hide context, and uneven horizontal axis intervals that distort trends.
When shaded regions represent numerical values, their areas should be directly proportional to those values-the principle of proportional ink. Violations occur when truncated axes make bars visually disproportionate to their values, as in Tennessee's jobs chart where a 1.08x increase appears as 2.7x more ink. Other violations include book sales bars with titles below zero, filled line charts with truncated axes, and donut charts where outer rings misleadingly appear larger than inner ones despite representing smaller values.
Three-dimensional visualizations create numerous problems despite their superficial visual appeal. While 3D charts can legitimately display data with two independent variables, they become pure bullshit when representing single-variable data. The unnecessary third dimension makes basic comparisons difficult, as demonstrated with birth rate data that becomes instantly clearer when presented as a simple 2D chart.
Capítulo 9
The Limitations of Artificial Intelligence
The hype around artificial intelligence follows a recurring pattern of breathless optimism dating back to the 1958 invention of the perceptron by Frank Rosenblatt. His simple neural circuit design spawned grandiose predictions about machines that would think like humans, recognize faces, translate speech, and even attain consciousness. Remarkably similar claims continue today, with the same technology (albeit with vastly improved hardware) generating the same superlatives in modern news articles.
Unlike traditional programming where you write code that processes data to produce output, machine learning inverts this process. You provide training data with correct labels, and a learning algorithm generates a program that can then classify new test data. The critical insight is that data quality determines everything-no algorithm, however sophisticated, can overcome poor training data. This principle, known as GIGO (garbage in, garbage out), means anyone can evaluate AI claims by examining the input data rather than needing to understand the complex algorithms themselves.
In 2017, media outlets uncritically reported Stanford researchers had developed AI capable of detecting sexual orientation from facial photographs. The researchers claimed their algorithm detected subtle facial features imperceptible to humans and that differences supported prenatal hormone theory. However, the study had major flaws. The human-machine comparison was unfair, as humans received no training while the algorithm processed thousands of images. The researchers failed to provide strong evidence for facial structure differences, relying on self-selected dating site photos rather than controlled 3D measurements.
Understanding how AI algorithms make decisions is extremely difficult, even for computer scientists who create them. This opacity presents a major challenge, as machines create their own decision-making rules that often make little sense to humans. In one revealing case, an algorithm distinguishing wolves from huskies wasn't analyzing facial features but simply detecting snow in wolf photos. Similarly, medical algorithms meant to detect pneumonia from X-rays were actually identifying the word "PORTABLE" on images, which correlated with sicker patients who needed bedside imaging.
The curse of dimensionality plagues machine learning systems as they incorporate more variables. While Google Flu Trends used 45 search queries and medical systems might analyze thousands of genes, adding variables requires exponentially more training data to distinguish true predictive capacity from chance correlations. Without sufficient data, algorithms might find spurious connections-like using Yankees' win-loss records to predict stock market performance-that fail when applied to new situations.
Capítulo 10
Science's Vulnerability to Bullshit
Science represents humanity's greatest invention, allowing us to transcend our evolved limitations and understand phenomena across vastly different temporal and spatial scales. However, science isn't an unerring path to ultimate reality but rather a collection of institutions, norms, and traditions developed through trial and error. While science excels at self-correction through organized skepticism, it isn't immune to bullshit-both accidental and deliberate.
To understand scientific reasoning, we must grasp the concept of p-values. A p-value tells us how likely an observed pattern could arise by chance alone. By convention, p<0.05 indicates statistical significance, but this threshold creates problems. Researchers have numerous degrees of freedom in data analysis-which elections to analyze, which demographics to include, which specific drugs count as painkillers-allowing them to mine data until finding something significant.
Faced with promising but non-significant results after months of work, many researchers engage in p-hacking: collecting more data until reaching significance, trying different statistical tests, or analyzing subgroups separately to find significance. This creates a severe publication bias where the roughly 5% of false hypotheses that appear significant by chance get published, while the 95% of negative results remain in researchers' "file drawers." This selection bias means published papers represent a skewed sample of all experiments conducted, making p-values unreliable measures of statistical support.
Public skepticism toward science stems partly from deliberate campaigns to manufacture uncertainty, but scientists and science reporters share blame by amplifying publication bias. News outlets eagerly report potential breakthroughs without indicating their preliminary nature, rarely following up when studies are disproven. The endless parade of contradictory health stories leaves the public feeling jerked around and increasingly cynical about science.
Popular science writing often misrepresents science as a "collecting process" where each paper represents a definitive fact, rather than one argument in an ongoing conversation. Scientists understand this complexity, weighing evidence across multiple studies, but the popular press rarely presents science this way. "Cafeteria science" compounds the problem, as writers cherry-pick studies to tell compelling stories, while selection bias ensures that surprising studies receive disproportionate coverage.
Despite these problems, the institution of science remains fundamentally effective. Scientists typically test hypotheses with reasonable chances of being correct, making most positive findings true positives. Science's cumulative nature means false results eventually fail when others build upon them, revealing errors. Most importantly, science simply works-it allows us to understand the physical world beyond our evolved senses and create technologies that would seem magical to previous generations.
Capítulo 11
Practical Tools for Bullshit Detection
Cultivating proper habits of mind is crucial for avoiding deception in an environment saturated with bullshit. Though becoming adept at spotting bullshit requires lifelong practice, a few simple techniques can significantly improve our defenses.
First, question the source of information. Journalists apply three essential questions: Who is telling me this? How do they know it? What are they trying to sell me? While we naturally apply this skepticism when dealing with used-car salesmen, we must extend this vigilance to social media, news, and health advice.
Second, beware of unfair comparisons. Headlines like "Airport Security Trays Carry More Germs Than Toilets!" create shock value through misleading comparisons. The study only measured respiratory viruses, which naturally accumulate on frequently touched surfaces like trays rather than toilet seats. Similarly, "America's Most Dangerous Cities" rankings often reflect arbitrary political boundaries rather than actual danger.
Third, if it seems too good or too bad to be true, it probably is. When NBC tweeted that "International student applications are down nearly 40 percent" following Trump's travel ban, the claim warranted skepticism for its implausible magnitude. Digging to the source revealed crucial context: applications decreased at 39% of universities but increased at 35%-essentially statistical noise rather than a meaningful trend.
Fourth, think in orders of magnitude. Many numerical claims can be debunked through simple order-of-magnitude thinking. When National Geographic claimed "9 Billion Tons of Plastic Waste End Up in the Ocean Every Year," a quick reality check reveals this would require each person on Earth to contribute over a ton annually-clearly impossible when total historical plastic production is only about eight billion tons.
Fifth, avoid confirmation bias. Our tendency to accept information that confirms our existing beliefs makes us vulnerable to misinformation. When a viral tweet suggested recommendation letters used dramatically different language for male versus female candidates, many people readily shared it because it matched their understanding of gender bias. However, checking the original research revealed the graphic merely illustrated the study's hypothesis, not its findings.
Finally, consider multiple hypotheses. Bullshit often appears as incorrect explanations for true events. When Reuters tweeted "Walt Disney shares down 2.5 percent after ABC cancels 'Roseanne' show," they implied causation where none existed-the stock drop actually occurred before the cancellation announcement and reflected broader market trends.
Capítulo 12
Effectively Refuting Misinformation
Spotting bullshit is a private activity, but calling bullshit is a public one that can protect entire communities. One powerful refutation strategy is reductio ad absurdum-showing how an opponent's assumptions lead to ridiculous conclusions. Biostatistician Ken Rice demonstrated a flawed model's weakness by extending it further to show it would predict negative sprint times by 2636, an obviously impossible result.
Finding counterexamples can immediately dismantle specious arguments. At a Santa Fe Institute workshop, a physicist presented a mathematical model claiming that long-lived multicellular organisms must have specific immune system features to survive. An immunologist responded with a simple question: "But what about trees?" Trees are long-lived multicellular organisms without the immune characteristics the physicist claimed were necessary, instantly invalidating his argument.
Analogies help recontextualize claims by drawing parallels to situations audiences intuitively understand. To counter vaccine skeptics, comparing vaccination to seatbelts helps people apply familiar risk assessment reasoning to unfamiliar medical contexts.
Redrawing misleading graphs is one of the most effective ways to refute visual bullshit. The Washington Post redrew the National Review's climate change chart to show the dramatic temperature increase hidden by the original's inappropriate scale. Similarly, Quartz redrew Apple CEO Tim Cook's impressive-looking graph of cumulative iPhone sales to reveal that quarterly sales had actually been declining-information completely obscured by the cumulative format.
When calling bullshit, accuracy is paramount. Thoroughly research facts and double-check them before making claims. Consider alternative explanations: you might be wrong, the person might be incompetent rather than malicious, or they might have simply made an honest mistake. Focus criticism on arguments rather than people.
Humility matters. When you make mistakes, own them swiftly and graciously. The common internet practice of doubling down on errors wastes everyone's time and damages productive discussion.
Effective refutation requires clarity. A disorganized flood of facts won't convince anyone to change their beliefs. Arguments need to be understandable, persuasive, and jargon-free-which often demands more effort than spotting the original bullshit.
Above all, remember to "think more, share less" to keep our information environments clean. In a world drowning in bullshit, this might be our most important civic responsibility.
