Key Takeaways
Judge claims by their allegiance to truth, not their polish or precision.
Truth is the dividing line. The authors use “bullshit” for communication designed to impress or persuade without caring whether it is true, coherent, or informative. Unlike a straightforward lie, it may contain accurate details. Its trick is to wrap weak reasoning in jargon, confident rhetoric, or impressive-looking statistics, making listeners feel they lack the expertise to question it. A corporate statement about “supply-chain collaboration,” for instance, can obscure the human cost of child labor without making a directly false claim.
Look past the wrapping. The authors call technically truthful statements used to mislead paltering. The key test is not merely whether each sentence can be defended literally, but what impression it is meant to create. Ask what evidence supports the claim, what is being left out, and whether the language clarifies the issue or makes scrutiny harder.
The distinction between falsehood and indifference to truth is useful because it catches more than fabricated facts. A speaker can mislead through omission, ambiguity, or theatrical expertise while keeping individual statements technically defensible. This resembles a problem in law and professional ethics: compliance with the letter of a rule does not guarantee compliance with its purpose. Still, suspicion should not become a shortcut for dismissing unfamiliar language. Technical terms can be essential in specialized work. The practical question is whether the speaker can translate them into testable claims and explain what evidence would count against the conclusion.
Trace who profits when a claim outruns its correction.
Speed favors the rumor. The Internet lowers the cost of publishing, but it also weakens the filters that once came with professional reporting. Click-based outlets compete for attention, and emotional headlines often outperform careful explanations. Social platforms add another motive: sharing a story can signal group identity, even when the sharer has not checked whether it is true.
Corrections face a head start. After the 2013 Boston Marathon bombing, a false story claimed that a child who survived the Sandy Hook school shooting had died in the race. More than 92,000 people shared it, while over 2,000 tweets tried to correct it. The authors call the imbalance between producing and correcting falsehood Brandolini’s principle. Deliberate disinformation can go further, flooding audiences with contradictory claims until people lose confidence that truth can be found at all.
This account makes misinformation a problem of infrastructure, not merely of gullible individuals. A platform optimized for engagement can reward outrage without taking a position on truth, much as an advertising system rewards attention without caring whether the attention is healthy. That perspective points toward practical interventions: slower sharing, visible sourcing, and incentives for corrections to reach the same audience as the original claim. Yet no distribution system can eliminate the human appeal of stories that confirm identity or offer a simple explanation for fear. Better tools help, but social trust and habits of verification remain necessary.
Treat an association as a clue, not permission to prescribe a cause.
Correlation leaves the story open. When two things vary together, either may influence the other, both may be caused by something else, or the pattern may be accidental. The authors describe a study linking preschoolers’ ability to wait for a second marshmallow with later academic success. A later, larger study found a much smaller relationship and pointed to family socioeconomic status as a common cause: it can affect both a child’s sense of security and educational opportunities.
Ask what would separate causes. Students who order pitchers may drink more because of the container, or because they already intend to drink more. A randomized experiment can help distinguish these explanations by assigning treatments rather than observing people who chose them. When experiments are impractical, competing explanations and background factors still need to be considered before making medical or policy recommendations.
Causal reasoning is difficult partly because a plausible story feels like an explanation before it has been tested. A causal diagram helps slow that leap by making assumptions visible: What points toward what, and what common factor could produce both observations? This is closely related to counterfactual thinking, which asks what would have happened to the same person or group under a different condition. Randomization can approximate that comparison, but it is not always ethical or feasible. Observational evidence can still matter; its strength depends on how seriously alternative causes have been investigated.
Never trust a percentage until you know its baseline and denominator.
Numbers need a comparison. A raw total can sound alarming or reassuring depending on what it is divided by. The authors contrast a report of 2,139 DACA recipients accused or convicted of crimes with the larger eligible population: the figure was about 0.3 percent, and comparison with relevant rates among US citizens changed the impression again. The percentage did not settle the question, but it made the missing context visible.
Absolute risk matters too. A study reported that one alcoholic drink a day increased the annual risk of certain health problems by 0.5 percent. Because the baseline risk was already below 1 percent, the increase amounted to about four additional cases per 100,000 people. A relative change can sound dramatic while the absolute change is small. Check the group, time period, starting value, and denominator before interpreting a numerical claim.
This is a lesson in measurement as much as arithmetic. A percentage is not a neutral translation of a count; its denominator defines the comparison, and choosing that denominator shapes the story. Risk communication therefore benefits from showing both relative and absolute changes, especially when readers must decide what to do. There is a limit, however: a small individual risk can still matter when millions of people are exposed, or when consequences are severe. Context should clarify scale, not minimize harm. The aim is to make magnitude legible without letting framing do the reasoning for the reader.
Expect people to game any score that controls their rewards.
Targets change behavior. Goodhart’s law describes what happens when a measure becomes a goal: people adapt to the measure, so it stops reflecting the quality it was meant to track. In colonial Hanoi, officials paid residents for rat tails. People began cutting off tails and releasing the rats, while others bred rats to collect the bounty. The count rose, but the rat problem remained.
Institutions do it too. Once college rankings rewarded small classes, universities found ways to improve the reported metric, including creating classes just below the ranking’s size threshold. The score could rise without improving the average student’s classroom experience. A measure can still be useful, but its designers should ask how people will respond when money, status, or access depends on it. Look for gaming, unintended incentives, and outcomes that the metric leaves out.
Metrics are not just mirrors; once attached to consequences, they become part of the system being measured. This creates a design problem familiar in management and public policy: people respond to incentives, often in ways that are individually rational and collectively damaging. One safeguard is to use multiple indicators, audit for unexpected behavior, and keep the underlying goal visible. More measurement is not automatically better, since each added target can invite a new form of gaming. The hardest question is not whether a measure is imperfect, but whether the institution can notice and correct the distortions it creates.
Ask who is missing before trusting what a sample seems to show.
The observed group may be selected. Selection bias occurs when the people included in an analysis differ systematically from the population the conclusion is about. At a ski resort, asking skiers where they most enjoy skiing naturally favors that resort: people who dislike it may have gone elsewhere. Insurance companies can truthfully advertise large average savings among customers who switched, because people most likely to switch are those offered unusually favorable rates.
The observer matters too. Large classes contain more students than small classes, so students experience a larger average class than the average class listing suggests. In one example, the average class had 50 students, but the average student attended a class of about 140. Before accepting a finding, ask how participants entered the sample, who dropped out, and whether the selection process is related to the result being measured.
Selection effects are especially treacherous because the data can be collected and analyzed correctly while still answering the wrong question. The missing cases are often invisible, which makes this error harder to detect than an obvious arithmetic mistake. This is why sampling is a matter of causal structure, not just sample size: a huge group can remain systematically unrepresentative. The useful habit is to imagine the path into the data set. Who had to act, qualify, respond, survive, or be noticed to appear there? That question often reveals the bias before advanced statistical analysis is needed.
Read the axes before believing the picture.
Design choices carry meaning. A graph can use accurate data and still encourage a false impression. A Reuters graphic on Florida homicides inverted its vertical axis, placing zero at the top. The designer said the intention was to depict deaths in negative terms, not to deceive, but the result could make rising murders look like a decline. In other charts, a truncated axis exaggerates small differences, or a changing time scale hides a trend.
Check how visual space works. The authors’ principle of proportional ink says that when shaded area represents a number, the area should change in proportion to that number. Three-dimensional bars, decorative chart shapes, and mismatched scales can violate that rule. Ask what each axis measures, whether intervals are equal, and whether the visual size of an element fairly represents its value.
Graphs are persuasive because perception works quickly: viewers notice slope, size, and contrast before they inspect labels. That makes visualization a form of argument, not mere packaging. A well-designed chart can reveal patterns that a table conceals, but a poor one can exploit the same cognitive shortcuts. The best defense is often to redraw the figure with a clearer scale or simpler form, rather than debate the designer’s intentions. At the same time, not every unconventional chart is deceptive. The test is whether its visual encoding preserves the comparison the data are supposed to support.
Audit an algorithm’s training data before admiring its accuracy.
Models learn the examples provided. Machine-learning systems infer patterns from training data and labels, then apply those patterns to new cases. A study claimed that facial images could reveal criminality with high accuracy. But the supposed noncriminals appeared in professional portraits, while the convicted criminals appeared in official identification photos. The algorithm could have learned to detect smiles, not criminal tendencies.
Test beyond the training setting. A wolf classifier learned to associate snow in images with wolves, while a medical imaging model used the word “portable” on an X-ray as a clue for pneumonia. Those patterns may work in the original data but fail elsewhere. Ask who supplied and labeled the examples, whether they represent the people and settings where the system will be used, and whether its apparent accuracy holds on genuinely new cases.
Algorithmic accuracy is not the same as valid knowledge. A system can predict a label reliably by exploiting a proxy that is irrelevant, unfair, or unstable outside its training environment. This is a form of measurement failure hidden behind technical sophistication: the model optimizes the task it was given, not the broader human purpose its users may assume. Independent testing across settings and groups is therefore essential, as is responsibility for decisions made with algorithmic tools. Transparency alone may not solve the problem if even the designers cannot readily explain a complex model’s reasoning.
Treat each scientific paper as one argument, not the final verdict.
Science is powerful, not infallible. Scientific claims can be challenged and corrected, but the incentives surrounding research also create errors. A p-value describes how surprising observed data would be under a specified null hypothesis. It does not give the probability that the preferred explanation is true. Confusing those two questions can make weak evidence sound conclusive.
Publication filters distort the record. Researchers and journals often favor positive, statistically significant findings. In a review of antidepressant trials, 51 of 74 registered studies were published, and nearly all published reports appeared positive, while many negative or uncertain results stayed hidden or were presented more favorably. Replication attempts have also failed to reproduce substantial shares of influential findings in several fields. These problems do not make science useless. They make converging evidence, replication, and corrections more trustworthy than one exciting result.
The most productive view of science is neither reverence nor cynicism. Scientific knowledge is an institutional achievement, produced by people with ordinary motives and checked through shared methods that can expose error. That distinction matters: a flawed study is not proof that the whole enterprise has failed, but institutional self-correction is not automatic either. Publication incentives, selective reporting, and press coverage can leave the public with a distorted picture of what researchers know. Reading a result in relation to the broader body of evidence is a form of calibrated trust, not a refusal to trust expertise.
Refute the claim clearly, carefully, and without humiliating its speaker.
Make the correction useful. The authors demonstrate several ways to test a claim: extend its logic until it produces an absurd result, find one counterexample, compare it with a familiar situation, redraw a misleading graph, or build a simple null model showing that the pattern could arise without the proposed explanation. A model predicting that sprinters will eventually run a 100-meter race in negative time is a memorable warning against careless extrapolation.
Protect credibility and relationships. Check the facts, criticize the argument rather than the person, and admit mistakes promptly. In a personal conversation, a private correction and a shared point of concern may work better than a public takedown. Avoid repeating a myth more than needed, and offer a plausible alternative explanation so the audience is not left with an unanswered question. Choose battles where someone is willing to engage.
Effective correction is a communication problem as well as an evidentiary one. A technically airtight rebuttal can fail if it threatens someone’s identity or turns an audience into spectators for a public defeat. The emphasis on clarity, relevance, and charity helps separate truth-seeking from status competition. There is also a real tension: private correction protects dignity, while public misinformation may require a public response to protect others. A useful decision rule is to match the response to the harm and audience, while keeping the standard of evidence constant. The goal is not to display cleverness, but to improve what people can reasonably believe.
Analysis
The book’s central move is to treat bullshit as a failure of allegiance, not simply a catalog of false statements. That makes its target broad: literal truth can still be deployed to mislead, and rigorous-looking statistics can conceal indifference to whether a conclusion follows. The chapters then build a practical ladder of skepticism. First inspect incentives and rhetoric. Then examine causal claims, numerical context, sample construction, visual encoding, and the data behind algorithms. Finally, apply the same scrutiny to science without confusing fallibility with failure.
Its most powerful teaching device is the gap between a claim’s polished surface and the ordinary mechanism underneath it. The criminality algorithm detects smiles because its image sets differ; the marshmallow result partly reflects family circumstances; the college class-size statistic changes depending on whether classes or students are counted. These examples show that technical complexity is often not the main obstacle. A sensible question about how the data were produced can be more revealing than expertise in the machinery that processed them.
The book also treats skepticism as civic practice. Digital systems make sharing cheap, correction expensive, and attention commercially valuable. A reader who merely recognizes errors privately does little to slow their circulation. Yet public correction can itself become performance, rewarding the debunker for humiliation rather than accuracy. The authors’ closing ethic is therefore essential, not decorative: verify first, address the argument, and preserve room for honest revision.
One tension deserves attention. The memorable phrase “call bullshit” can invite overconfidence, especially when a claim is unfamiliar or technically difficult. The authors counter this by urging charitable interpretation and acknowledging that science works through imperfect institutions. Their strongest model is neither automatic trust nor reflexive suspicion. It is disciplined provisional judgment: identify what evidence would change the conclusion, compare plausible explanations, and calibrate confidence to the quality of the evidence. This approach makes skepticism a method for improving shared understanding, rather than a badge of superiority.
Review Summary
Calling Bullshit receives mostly positive reviews for its accessible approach to critical thinking and spotting misinformation. Readers appreciate its timely content, practical examples, and humor. Many find it useful for navigating today's information-heavy world. Some criticize its density or repetitiveness, while others praise its thoroughness. The book is often recommended as essential reading for students and adults alike. Reviewers highlight its insights on data manipulation, scientific publishing, and social media. Overall, it's seen as a valuable guide for developing skepticism in a data-driven society.
People Also Read
Glossary
Bullshit
Persuasion indifferent to truthCommunication intended to impress or persuade while disregarding whether its content is true, logically coherent, or genuinely informative. It may use false statements, but it can also rely on technically accurate claims, jargon, detail, or quantitative presentation to create a misleading impression or discourage scrutiny.
Paltering
Misleading with literal truthsThe deliberate use of statements that are literally true to lead another person toward a false or unsupported conclusion. Paltering exploits the gap between what a sentence strictly says and what listeners reasonably infer from it, allowing the speaker to mislead while claiming not to have lied.
Brandolini's principle
Refutation costs more effortThe observation that producing a false or misleading claim generally takes much less effort than carefully disproving it. The imbalance is especially visible online, where a rumor can be copied widely in moments, while checking sources, reconstructing context, and reaching the same audience with a correction takes substantially more work.
Selection bias
Unrepresentative cases distort findingsA distortion that occurs when the people or cases included in a study differ systematically from the broader group the study aims to describe. The selection process can be connected to the outcome being measured, making a sample misleading even when its members report accurately and the calculations are correct.
Berkson's paradox
Selection creates false relationshipsA statistical pattern in which selecting cases based on two characteristics can create an association between those characteristics, even when they are unrelated in the wider population. For example, choosing potential partners who meet a threshold for both attractiveness and niceness can make those traits appear negatively related within the selected dating pool.
Goodhart's law
Targets corrupt useful measuresThe principle that a measure can stop representing its intended goal once people are rewarded for improving that measure. Because people adapt to incentives, a ranking or score can be raised through behaviors that satisfy its rules without improving the underlying outcome, as when schools reshape classes to score better on a class-size metric.
Mathiness
Math-like form without rigorThe use of equations or mathematical notation to give an argument an appearance of precision without defining measurable quantities or justifying the relationships shown. A formula that combines vague ideas such as trust, authenticity, or happiness can imply exact arithmetic even when the concepts have no shared units or defensible mathematical form.
Fermi estimation
Quick order-of-magnitude estimateA method for checking whether a numerical claim is plausible by breaking it into simpler quantities that can be estimated from ordinary knowledge. The estimates can be rough. Their purpose is often to determine whether a claim is off by a factor of ten or more, not to produce a precise final answer.
Overfitting
Learning noise instead of patternA machine-learning failure in which a model matches the training examples too closely, including their accidental quirks, and performs worse on new data. A model that memorizes handwritten characters can appear perfect on examples it has seen, yet fail on unfamiliar handwriting because it has not learned a pattern that generalizes.
P-hacking
Searching analyses for significanceThe use of undisclosed flexibility in data collection, outcome choice, or statistical analysis to obtain a statistically significant result. Trying multiple approaches and reporting only the one that crosses a significance threshold can make chance patterns look like strong evidence, particularly when readers do not know how many alternatives were tested.
File drawer effect
Unpublished negative findingsThe bias created when studies with positive or statistically significant findings are more likely to be published than studies with null or negative results. The visible research literature then overrepresents apparent discoveries, while unreported experiments remain unseen, making weak or nonexistent effects seem more consistent than they are.
Prosecutor's fallacy
Confusing inverse probabilitiesThe error of treating the probability of evidence under innocence as if it were the probability of innocence given that evidence. In a fingerprint search, a rare false-match rate does not by itself establish that a matched person is guilty. The number of possible innocent matches and the base rate of guilt also matter.
Null model
Pattern without proposed causeA deliberately simple model used to show that an observed pattern could appear even when the proposed process is absent. If a simulation with no age-related decline still produces slower record times among older runners because fewer older runners compete, then the observed trend alone cannot establish senescence.
Proportional ink
Visual area matches numerical valueA rule for graphics that use shaded regions to represent numbers: the area of each region should change in proportion to the value it depicts. If a chart uses area, volume, or decorative shape in ways that enlarge some values disproportionately, the visual impression can exaggerate or conceal differences in the data.
Firehose strategy
Overwhelm with contradictory claimsA form of propaganda that spreads large quantities of information rapidly and across multiple channels, without requiring consistency among the claims. Its purpose can be less to persuade an audience of one specific story than to exhaust attention, create confusion, and weaken confidence that reliable truth can be distinguished from falsehood.
FAQ
What's Calling Bullshit: The Art of Skepticism in a Data-Driven World about?
- Focus on Misinformation: The book explores how misinformation spreads in a data-driven world, particularly through social media and data visualization.
- Critical Thinking Skills: Authors Carl T. Bergstrom and Jevin D. West aim to equip readers with tools to critically analyze quantitative arguments and recognize manipulation.
- Democracy and Bullshit: It argues that detecting and refuting misinformation is crucial for the survival of liberal democracy, especially in the age of fake news.
Why should I read Calling Bullshit: The Art of Skepticism in a Data-Driven World?
- Empowerment Through Knowledge: The book empowers readers to think critically about daily information, helping to discern fact from fiction.
- Engaging and Accessible: Complex ideas are presented in an engaging manner, using humor and relatable examples to illustrate key concepts.
- Civic Responsibility: It emphasizes the importance of being an informed citizen, relevant for anyone concerned about public discourse.
What are the key takeaways of Calling Bullshit: The Art of Skepticism in a Data-Driven World?
- Understanding Bullshit: Bullshit is defined as presentations intended to persuade or impress by distracting or overwhelming with a disregard for truth.
- Causation vs. Correlation: The book warns against assuming causation from correlation without sufficient evidence.
- Data Visualization Awareness: Readers learn to critically evaluate data visualizations, recognizing misleading design choices.
What are the best quotes from Calling Bullshit: The Art of Skepticism in a Data-Driven World and what do they mean?
- Energy to Refute Bullshit: "The amount of energy needed to refute bullshit is an order of magnitude bigger than [that needed] to produce it." This highlights the challenge of combating misinformation.
- Bullshit Defined: "Bullshit involves language, statistical figures, data graphics, and other forms of presentation intended to persuade or impress an audience." It emphasizes manipulation over falsehoods.
- Goodhart's Law: "When a measure becomes a target, it ceases to be a good measure." This reflects how metrics can be gamed when they become the focus of evaluation.
How does Calling Bullshit: The Art of Skepticism in a Data-Driven World address data visualization?
- Critical Evaluation of Graphs: The authors discuss how data visualizations can mislead by manipulating axes or using inappropriate scales.
- Importance of Context: Numbers must be presented in context to be meaningful, as raw figures can obscure important comparisons.
- Design Choices Matter: The design of a graph can significantly influence data perception, urging vigilance about graphical representations.
What is the difference between correlation and causation in Calling Bullshit: The Art of Skepticism in a Data-Driven World?
- Definition of Correlation: Correlation is a statistical relationship between two variables, but it does not imply causation.
- Causation Explained: Causation means one event is the result of another, directly affecting the other variable.
- Common Misunderstanding: The book warns against assuming causation from correlation, emphasizing the need for rigorous evidence.
What specific methods does Calling Bullshit: The Art of Skepticism in a Data-Driven World suggest for spotting misinformation?
- Corroborate and Triangulate: Check multiple sources to verify surprising claims, treating uncorroborated information with skepticism.
- Use Reverse Image Lookup: Verify image authenticity using reverse image search tools to identify manipulated images.
- Consider Multiple Hypotheses: Evaluate claims by considering various explanations, avoiding confirmation bias.
What is the principle of proportional ink mentioned in Calling Bullshit: The Art of Skepticism in a Data-Driven World?
- Definition of Proportional Ink: The size of shaded areas in a graph should be proportional to the values they represent.
- Importance in Data Visualization: Violating this principle can lead to misleading interpretations of data.
- Application in Graphs: Examples from sports and economic data show how improper scaling can mislead audiences.
What is the concept of "selection bias" in Calling Bullshit: The Art of Skepticism in a Data-Driven World?
- Definition of Selection Bias: It occurs when a study's sample is not representative of the population, leading to skewed results.
- Impact on Research: Selection bias can distort findings, as seen in surveys and studies with unrepresentative samples.
- Real-World Implications: Understanding selection bias is crucial for accurately interpreting research and trends.
How does Calling Bullshit: The Art of Skepticism in a Data-Driven World relate to democracy?
- Critical Electorate: A functioning democracy relies on a critically thinking electorate that can discern truth from misinformation.
- Impact of Misinformation: Misinformation can undermine public trust and distort democratic processes.
- Civic Responsibility: The book calls for individuals to take responsibility for their information consumption, advocating for informed citizenry.
What is the significance of the "prosecutor's fallacy" discussed in Calling Bullshit: The Art of Skepticism in a Data-Driven World?
- Understanding Conditional Probabilities: It illustrates confusion between the likelihood of a match given innocence and innocence given a match.
- Implications for Scientific Claims: Parallels are drawn with p-values in research, highlighting potential misinterpretations.
- Critical Thinking Application: Recognizing this fallacy encourages careful consideration of statistical claims.
How can I apply the lessons from Calling Bullshit: The Art of Skepticism in a Data-Driven World in my daily life?
- Develop Critical Habits: Question the sources of information, considering motivations and evidence.
- Practice Fermi Estimation: Use rough estimates to evaluate the plausibility of numerical claims.
- Engage in Discussions: Consider multiple perspectives and hypotheses to foster open dialogue and informed conclusions.
Download PDF
Download EPUB
.epub digital book format is ideal for reading ebooks on phones, tablets, and e-readers.