Book summary “Calling Bullshit”: summary

Introduction

You live in a world saturated with misleading data, manipulated charts, and polished misinformation. Modern digital platforms flood your feeds with deceptive claims designed to capture your attention and exploit your cognitive biases. This quantitative nonsense masquerades as rigorous science, leaving you vulnerable to manipulation in your daily life.

Carl T. Bergstrom and Jevin West argue that you can combat this deception by developing systematic skepticism. You will learn to recognize when numbers lie, question hidden assumptions, and evaluate opaque models. The book equips you to spot bullshit across digital media, statistical claims, data visualizations, and scientific studies.

• How do engagement-maximizing algorithms spread misinformation?
• What distinguishes genuine causation from spurious statistical correlation?
• How can you expose flaws in big data and data visualizations?

1. Bullshit Everywhere

Misinformation and misleading rhetoric pervade modern digital environments, driven by the attention economy and structural asymmetries between creating nonsense and refuting it. You encounter deceptive claims daily across news platforms, corporate statements, and social networks.

The energy required to refute bullshit is vastly greater than the energy needed to produce it.

Alberto Brandolini coined this principle in 2014. It explains why bullshitters hold a massive structural advantage over truth-tellers. It is simpler and cheaper to generate nonsense than to clean it up. Concurrency and speed allow false narratives to outpace corrections.

The Spread of Bullshit

Social media environments facilitate the rapid proliferation of rumors and fake news during crises. False rumors reach more people and persist even after fact-checkers debunk them.

During the 2013 terrorist bombing of the Boston Marathon, a false story about a young girl killed in the blast was shared by ninety-two thousand people. This occurred despite two thousand tweets refuting the false rumor and rapid debunking by Snopes.

Animal Bluffing and Human Deception

Mantis shrimp perform empty threat displays while molting because the danger of leaving their hole is too high. This is merely an evolved reflex rather than calculated deceit. Mantis shrimp perform threat displays even when molting and unable to punch, because fleeing their crevice exposes them to predators on the reef. Their hammer-like claws strike at fifty miles per hour when intact.

Sophisticated bullshitting requires a theory of mind, meaning the ability to put oneself in the mark's place. Corvids such as ravens and crows demonstrate advanced cognitive abilities by caching food to recover later. Ravens tested in experiments with peepholes generalize from their own experience of watching through screens. They recognize that an unseen bird could be observing them and engage in a fake-cache by pretending to stash a food item while keeping it in their beak.

Human deception often relies on paltering and weasel words. Paltering occurs when someone deliberately leads others to draw wrong conclusions using statements that are technically not untrue. Bill Clinton claimed to Jim Lehrer that there is no sexual relationship with Monica Lewinsky, defending himself later by pointing to the present-tense verb is. Advertisers, politicians, and managers use hedged language, passive voice, and euphemisms to avoid culpability. A 2019 NBC News report revealed global manufacturers used child labor in raw material mines where families earned forty cents a day. A Fiat Chrysler spokesperson described this practice as engaging in collaborative action across the value chain.

Andrew Wakefield published a discredited study in 1998 in The Lancet linking MMR vaccines to autism. Ten co-authors retracted the interpretations section in 2004, and The Lancet fully retracted the paper in 2010 after Wakefield was found guilty of professional misconduct. British Medical Journal editor Fiona Godlee declared the study a fraud in 2011. Despite a meta-analysis of 1.3 million children showing no relation between vaccination and autism, the antivax movement persisted and contributed to a record high 41000 measles cases reported in Europe in the first six months of 2018.

2. The Impact of Digital Media on Information

The rise of the Internet and smartphones did not eliminate misinformation. Instead, these technologies accelerated a bullshit pandemic by lowering the cost of publishing and replacing subscription-based news models with click-driven advertising economies. You now navigate platforms where publishers must capture immediate attention rather than build long-term relationships. This shift created an arms race of sensationalized headlines, fluff, and viral content that crowds out thoughtful reporting.

The rise of the internet and commercial social media platforms accelerated a pandemic of misinformation by replacing journalistic filters with engagement-maximizing algorithms and identity signaling.

Steve Rayson analyzed 100 million articles from 2017 and found that the most widely shared Facebook headlines promise emotional experiences rather than facts. The most common phrase in these forward reference headlines was will make you.

Social media communication is primarily driven by social bonding and identity signaling rather than the objective transmission of information. You might share false stories simply to affirm tribal affiliations. Judith Donath notes that sharing content online serves to signal political allegiances and cultural values to a peer group. When your primary goal is to demonstrate belonging or express hatred for an opponent, the literal truth becomes completely irrelevant.

A user shares an absurd conspiracy theory about airplane contrails not to prove a scientific claim, but to signal membership in a group that distrusts the liberal establishment.

Algorithms and Echo Chambers

Commercial social media algorithms are designed to maximize user engagement and screen time rather than provide accurate information. These systems inevitably drive you toward increasingly extreme content and conspiracy theories. Algorithms analyze your data, geolocation, and network connections to feed you content that aligns with your sociopolitical orientations while suppressing alternative viewpoints. Because outrage and radical viewpoints are highly effective at retaining attention, platforms like YouTube systematically recommend more extreme videos over time.

Jevin Bergstrom and his six-year-old son watched real-time footage from the International Space Station, located 254 miles above spherical Earth where the station orbits, and YouTube's sidebar immediately filled with recommendations for flat-earth videos.

The rapid spread of misinformation and disinformation via messaging apps and social media platforms can trigger violent real-world tragedies and destabilize international geopolitics. Platforms like WhatsApp, which connects 1.5 billion WhatsApp users worldwide, allow unverified rumors and fake videos to travel faster than official corrections. A 2018 study found that 2.6 percent of US news articles were false. During the 2016 US election, teenagers in Macedonia created 140 popular fake news websites.

In India, fake WhatsApp videos warning of child kidnappings led angry mobs to brutally beat and kill innocent travelers, forcing authorities to shut down the internet.

Bots and Counterfeit People

Modern digital networks face a massive wave of human counterfeiting where automated bots, click farms, and AI-generated deepfakes mimic real individuals to manipulate public sentiment. Advanced algorithms can generate photorealistic faces of nonexistent people from scratch and synthesize convincing audio or video through adversarial machine learning. These synthetic personas and bot networks can overwhelm democratic consultations and trick media outlets into broadcasting fabricated controversies. Facebook deleted nearly three billion fake Facebook accounts in 2018.

The FCC received 21.7 million public comments regarding net neutrality, but nearly 10 million comments were sent using stolen identities. Millions of submissions were traced to automated bots, stolen identities, and Russian email addresses submitting duplicate boilerplate text.

3. Defining Bullshit and Black Boxes

Bullshit uses language, statistics, or data presentations to persuade or impress through a blatant disregard for truth or logical coherence. Drawing on philosophical concepts from Harry Frankfurt and G.A. Cohen, bullshit differs from an outright lie because the creator exhibits complete indifference to whether the claim is true or false. Bullshitters deploy rhetorical veils, jargon, and superfluous details either to actively mislead or simply to fill space and feign authority.

Bullshit functions like a sociological black box, using technical jargon or opaque statistics to shield weak claims from critical examination.

Bruno Latour noted that specialized techniques are accepted by communities to facilitate work, but bullshitters exploit them to make claims unassailable to lay readers. A friend trying to prove that cat people earn higher salaries than dog people backs up the claim by mentioning statistical terms like ANCOVA and log transformations, creating an opaque black box. Rather than struggling to unpack the internal mechanics, you can successfully evaluate a claim by inspecting its raw inputs and final outputs.

Spotting Bullshit Without Technical Expertise

You do not need technical expertise in complex algorithms or statistics to call bullshit. You can debunk flawed claims by carefully evaluating whether the input data are unbiased, reasonable, and relevant.

Researchers claimed an AI algorithm could detect criminality from facial headshots with 90 percent accuracy. Critics discovered the training sets used professional smiling photos for noncriminals and unsmiling police ID photos for criminals, turning a smile detector into a fake crime predictor based on labile facial expressions rather than physiognomy. The study analyzed 1,800 photos of Chinese men, repeating old errors of physiognomy through automated systems.

• Check if the data that go into the analysis are unbiased, reasonable, and relevant to the problem at hand.
• Check if the results pass basic plausibility checks.
• Check if the results actually support whatever conclusions are drawn.

4. Causality and Statistical Associations

Human brains naturally seek patterns. This tendency leads you to assume that if two things are correlated, one must cause the other.

Correlation does not imply causation, and you must carefully distinguish between genuine causal mechanisms and spurious correlations driven by confounding factors or common causes.

Core Principles

Observational data can hide alternative explanations. A study found a correlation between self-esteem and having had a first kiss by college, but it remains unclear whether self-esteem causes kissing, kissing causes self-esteem, or both are driven by a third factor.

Scientists use a correlation coefficient on a scale from 1 to minus 1. The correlation coefficient between college football program budget rankings and success rankings from 2006 to 2015 is 0.78.

Conflating probabilistic causes with necessary or sufficient causes is a rhetorical trick. Mike Pence argued against tobacco regulations by stating that smoking does not kill because two out of three smokers do not die from smoking-related illnesses and nine out of ten do not get lung cancer.

Data and Media Pitfalls

Data dredging involves comparing numerous trends against each other until finding patterns that match closely by chance. Tyler Vigen found a tight statistical correlation between the age of Miss America and the number of murders caused by hot vapors, as well as a correlation between anticoagulant deaths and US sociology degrees from 100 data series allowing nearly 10,000 pairs to be compared.

The popular press frequently twists observational studies into unwarranted causal claims. MarketWatch published a story on a Zillow housing price report with the causal headline Another Adverse Effect of High Home Prices: Fewer Babies, even though the original report explicitly noted the correlation could be driven by confounding factors. Out of 50 research studies shared most often on Facebook and Twitter regarding health factors, only 15 did a decent job of demonstrating cause and effect, and just 2 met the highest standards for demonstrating causality.

5. Numbers and Nonsense

Numbers carry an aura of scientific authority and exactness. You might wrongly believe that data never lies and measurements speak for themselves, while words are recognized as human constructs used to blur the truth.

Numbers carry an unearned aura of objectivity, making them prime vehicles for bullshit when presented without context, distorted by percentages, or gamed through metrics.

Correct numbers can still be misleading if they lack appropriate context or fail to allow for fair, meaningful comparisons. Breitbart published a headline claiming 2,139 DREAM Act recipients had been accused of crimes, omitting the context that this represents only about 0.3 percent of the roughly 700,000 to 800,000 people who held DACA status.

Percentages can easily obscure reality by making large values look small, small values look large, or confusing changes in numerators and denominators. Google VP Ben Gomes stated that only around 0.25 percent of their daily search traffic returns offensive or misleading content, which sounds small but actually amounts to well over thirteen million queries a day out of 5.5 billion daily search queries on Google.

Distilling Numbers

Mathiness involves formulas and equations that mimic mathematical rigor while completely disregarding logical coherence, formal units, and factual accuracy. Authors sometimes invent equations to impress audiences, using algebraic symbols to represent ill-defined, unmeasurable qualities without validating dimensional analysis or functional forms. Cliff Arnall created the Blue Monday formula to calculate the saddest day of the year, adding incompatible quantities like weather and debt and raising time units to powers.

Goodhart's Law and Zombie Statistics

When a quantitative measure becomes a target for evaluation or reward, people will alter their behavior to game the system, thereby destroying the metric's usefulness. Employees and administrators under pressure to look good will find ways to manipulate scores or meet quotas, which corrupts the underlying social process or administrative indicator. During the French colonial period in Hanoi, a bounty on rat tails caused residents to cut off tails and release the rats to breed, and entrepreneurs even started raising rats in captivity to harvest tails.

Zombie statistics are numbers that are cited out of context, outdated, or completely fabricated, yet persist indefinitely because they are endlessly repeated. Without tracking down original sources and examining exact contexts, people often propagate unsupported claims that mutate significantly over time as editors and writers pass them along without checking facts. Arthur Jago traced the popular trope that 50 percent of scientific articles are never read back through decades of citations to find it originated from a misinterpreted statistic about papers going uncited for four years.

6. Selection Bias and Hidden Pitfalls

Selection bias occurs when the individuals sampled for a study differ systematically from the target population eligible for the study.

Drawing valid statistical inferences requires looking at a random sample of the population. When the sampling mechanism depends on variables related to the study question, the results become skewed and fail to generalize. Carl and his father noticed that skiers at Solitude constantly claimed it was the best resort in the world, ignoring the fact that anyone who disliked Solitude would be skiing elsewhere. Insurance company advertisements claiming massive average savings for customers who switch are driven entirely by selection bias. GEICO and Allstate advertise that customers who switched saved an average of over $500 or $498 per year, even though competing companies make nearly identical claims.

The Hidden Cause of Murphy's Law

The average class size experienced by students is significantly larger than the average class size offered by an institution because large classes serve disproportionately more students. Marquette University offers courses across various bins resulting in a mean class size of 26, but the experienced mean class size for a student is approximately 51. The friendship paradox states that the vast majority of people have fewer friends than their friends have on average. Researchers analyzing 69 billion friendships among 720 million Facebook users found that 93 percent of users have fewer friends than their friends have.

Disarming Selection Bias

Observational studies on workplace wellness programs falsely attribute better health outcomes to the programs due to self-selection bias. The University of Illinois at Urbana-Champaign randomized employees into treatment and control groups for a wellness program and found no significant effect on medical costs or physical activity, despite observational correlations showing participants were healthier.

7. Data Visualization and Visual Deception

Bad charts manipulate axes, violate proportional ink principles, and employ inappropriate visual frameworks to manufacture false conclusions.

Axes and Scales

You can reverse an apparent trend by inverting a graph's vertical axis. A Reuters data visualization covering 22 years of Florida homicides after Stand Your Ground laws placed zero at the top instead of the bottom. This made a rapid increase in murders look like a sharp decline.

Truncating the vertical axis of a bar chart visually exaggerates differences. A Maclean's magazine graph on social trust used truncated values of 35, 45, and 50 on the vertical axes.

Independent axes can manufacture fake correlations. A discredited research paper linked autism prevalence with MMR vaccine coverage by using a scale of 0 to 0.6 percent for autism and 86 percent to 95 percent for MMR coverage.

A Wall Street Journal graph on President Obama's tax plan used variable bin widths. It used narrow bins for low incomes and wide bins for middle incomes, totaling $5.65 trillion of total taxable income from all individual taxpayers in 2008, to make the middle class look like the primary tax base between $50,000 and $200,000.

Visual Distractions and Proportional Ink

Dataviz ducks use excessive decoration that overwhelms the data. Examples include USA Today Daily Snapshots featuring lipstick tubes as bar charts.

Glass slippers force data into unsuitable visual frameworks to borrow undeserved authority. Copycat periodic tables and subway maps represent corporate structures or literary plots.

Violating the principle of proportional ink leads to distorted perceptions. A Tennessee Department of Labor chart showed a ratio of 1.08 times for 2014 nonfarm jobs compared to 2010. Truncation caused the 2014 bar to use 2.7 times the amount of ink compared to the 2010 bar.

• Ensure the size or area of a shaded region representing a numerical value is directly proportional to that value.

Three-dimensional charts distort proportions. An Ontario polling pie chart represented the Ontario NDP wedge with 35 percent vote share, taking up 47 percent apparent surface area. It represented the Ontario PC wedge with 40 percent vote share, taking up 32 percent apparent surface area.

8. Calling Bullshit on Big Data

Modern artificial intelligence and machine learning technologies are built upon decades-old logical circuits called perceptrons. Invented by psychologist Frank Rosenblatt on July 8, 1958, a perceptron is a simple circuit that takes numerical inputs and outputs a zero or one. The first perceptron used an IBM 704 computer costing two million dollars to simulate its operations.

Artificial intelligence and big data models are bound by the garbage in, garbage out principle, making them prone to overfitting, proxy usage, and spurious predictions.

On December 28, 2013, The New York Times published a story on neural networks using the same superlatives found in its original 1958 article. Recent advances in artificial intelligence rely more on massive increases in computing power and data than on new algorithmic approaches. In supervised machine learning, an algorithm receives training data and labels to generate a program that labels unfamiliar test data.

Artificial Intelligence Hype

Overfitting occurs when a machine learning model treats random noise in training data as meaningful patterns, ruining its performance on new data. Complicated models memorize idiosyncratic errors and scanning artifacts, causing accuracy to drop significantly on test data. Simpler models often generalize much better.

The US postal system processes half a billion pieces of mail every single day. Its handwriting recognition system correctly interprets addresses 98 percent of the time using the MNIST database containing seventy thousand labeled images of handwritten digits. Human address experts in Salt Lake City process up to 1,800 addresses per hour when machines encounter illegible text.

Model Generalization and Black Box Opacity

Researchers Yilun Wang and Michal Kosinski collected nearly eight thousand men's photos and nearly seven thousand women's photos to train a deep neural network on sexual orientation. In early September 2017, outlets like The Economist and The Guardian published stories claiming neural networks could infer sexual orientation from facial photos. This research ignored confounding factors like grooming and photo self-selection while failing to prove structural facial differences.

Deep learning models are opaque black boxes that frequently latch onto background correlates or proxy variables instead of core causal features. A husky and wolf classification algorithm achieved a 70 percent success rate by relying on whether snow appeared in the background. Google Flu Trends relied on 45 key search queries to nowcast outbreaks, but predictions missed the mark by a factor of two after Google introduced autocomplete functionality in 2008, forcing project cancellation in 2014.

9. The Susceptibility of Science

Scientific literature is vulnerable to statistical fallacies, p-hacking, publication bias, and predatory publishing, requiring careful evaluation of study legitimacy.

Philosopher Philip Kitcher distinguishes between epistemically pure scientists seeking only truth and epistemically sullied scientists who act from human motivations such as earning money, winning promotions, and gaining peer recognition.

Statistical Fallacies and P-Hacking

Confusing the probability of data given an innocent hypothesis with the probability of the hypothesis given the data leads to severe statistical errors. People frequently mix up $P(\text{match}\mid\text{innocent})$ with $P(\text{innocent}\mid\text{match})$. A prosecutor argues there is a one-in-ten-million chance of a random fingerprint match, but across a database of fifty million people, this yields about five innocent matches for every one guilty match.

A p-value measures the probability of observing data as extreme as recorded assuming the null hypothesis is true, not the probability that the hypothesis itself is false. Carl guessing 19 out of 52 playing card suits correctly yields a p-value of 0.043, meaning there is a 4.3 percent chance of getting this result by random guessing alone.

Researchers can manipulate statistical tests and variables until they achieve a statistically significant result, rendering standard thresholds unreliable. Data collection involves numerous degrees of freedom, such as selecting specific demographics, metrics, or timeframes. Researchers Leif Nelson, Joseph Simmons, and Uri Simonsohn p-hacked an experiment to demonstrate that listening to the Beatles song 'When I'm Sixty-Four' statistically significantly makes people younger.

To combat these analytical pitfalls, you should follow specific practices:

• Specify all methodological choices before looking at the data.
• Test the one hypothesis committed to in advance.

Publication Bias and Predatory Publishers

Publication bias creates a distorted scientific record because journals overwhelmingly publish positive results while negative results remain hidden. Erick Turner found that while published literature showed a 94 percent success rate for antidepressants, FDA-registered trials including unpublished and recast data showed a much lower actual success rate.

Open-access publishing models have enabled predatory publishers to flood the literature with low-quality, unvetted articles for a fee. Scientific editor John McCool submitted a fake case report summarizing the plot of a Seinfeld episode involving 'uromycitisis poisoning,' which was quickly accepted by an open-access journal demanding a fee of $799.

To navigate the published literature, you should use reliable verification tools:

• Check the paper on the publisher's website or in the PubMed database for retractions or corrections.
• Evaluate the reputation of the journal and publisher using citation indices or impact factors.

10. Spotting and Refuting Bullshit

You can effectively identify and dismantle misinformation by questioning sources, applying Fermi estimations, using counterexamples, and engaging in disciplined online fact-checking.

To evaluate any piece of information, you should always ask who is telling you, how they know it, and what they are trying to sell you. Tracing the background of the source reveals expertise or hidden motives.

Lifestyle brand Goop interviewed shamanic energy practitioner Colleen McCann, who claims crystals retain all past information, while her website sells eighty-five dollar chakra-healing crystal sets.

• Ask: Who is telling me this?
• Ask: How does he or she know it?
• Ask: What is this person trying to sell me?

Fermi estimation lets you break numerical claims down into rough components and check them to the nearest power of ten. National Geographic claimed nine billion tons of plastic waste enter the ocean every year, which would mean over one ton per person on Earth, vastly exceeding total historical global plastic production of eight billion tons.

• Break down a number into components you can estimate.
• Estimate to the nearest power of ten or order of magnitude.
• Use known baseline figures to calculate an approximate total and check if it makes sense.

Refuting Bullshit Effectively

Calling bullshit is a performative utterance in which you publicly repudiate something objectionable, such as lies, trickery, or injustice. Because it is a performative act, you must use it with prudence and correctness rather than carelessness.

Nothing is as devastating to specious arguments as a simple counterexample that demonstrates a premise does not imply the claimed conclusion. An immunologist at a Santa Fe Institute workshop asked "But what about trees?" when a physicist argued that long-lived multicellular organisms must have adaptive immune systems, successfully demolishing the model since ancient bristlecone pines lack those features.

• Keep it simple by focusing on core points without getting bogged down in tangential technicalities.
• Take it offline to private spaces rather than calling people out in public forums or social media feeds.
• Find common ground by establishing a shared perspective before introducing skeptical ideas.
• Don't overemphasize the myth to avoid increasing its familiarity and stickiness.
• Fill in knowledge gaps with alternative explanations to provide a complete narrative.

You must be correct by doing background research before speaking. Be charitable by assuming honest mistakes rather than malice. Admit fault swiftly and graciously when you make a mistake. Be clear by constructing understandable arguments. Be pertinent by ensuring your objection undermines the core claim rather than a trivial technicality.

10 Key Ideas

1. Brandolini's principle states that the energy required to refute bullshit is vastly greater than the energy needed to produce it.
2. Commercial social media algorithms are designed to maximize user engagement and screen time rather than provide accurate information, which inevitably drives users toward increasingly extreme content and echo chambers.
3. You do not need technical expertise in complex algorithms or statistics to call bullshit, as you can debunk flawed claims by carefully evaluating whether the input data are unbiased, reasonable, and relevant.
4. Correlation does not imply causation, and you must carefully distinguish between genuine causal mechanisms and spurious correlations driven by confounding factors or common causes.
5. When a quantitative measure becomes a target for evaluation or reward under Goodhart's law, people will alter their behavior to game the system, thereby destroying the metric's usefulness.
6. Selection bias occurs when the individuals sampled for a study differ systematically from the target population eligible for the study, skewing results and preventing generalization.
7. Bad charts manipulate axes, violate proportional ink principles, and employ inappropriate visual frameworks to manufacture false conclusions.
8. Artificial intelligence and big data models are bound by the garbage in, garbage out principle, making them prone to overfitting, proxy usage, and spurious predictions.
9. Scientific literature is vulnerable to statistical fallacies, p-hacking, publication bias, and predatory publishing, requiring careful evaluation of study legitimacy.
10. You can effectively identify and dismantle misinformation by questioning sources, applying Fermi estimations, using counterexamples, and engaging in disciplined online fact-checking with prudence and correctness.