Real data arranged to mislead beats outright lies

1

Bullshit is not lying

Bullshit aims to impress through the look of rigor and does not care about truth, so contrary facts cannot refute it the way they refute a lie.

A liar knows the truth and works to hide it. A bullshitter wants to persuade or impress and is simply indifferent to whether the claim holds. That gap blocks the usual remedy: stacking up contrary facts does nothing when facts were never the point.

Quantitative bullshit rides cultural deference to numbers and science. A spreadsheet feels more credible than an opinion. Phrases like “studies show” and charts with a veneer of method shut down pushback even when the underlying work is thin.

Data is cheap to produce and easy to share, so this strain floods social media, news, and workplace decks faster than corrections can catch it. Pattern recognition—not another pile of facts—becomes the practical defense.

In practice: A company boasts a 40% improvement without stating the baseline, so readers picture a large absolute gain when the real change may be trivial.

2

Correlations sold as causes

Most public causal claims are unsupported correlations dressed as causation because a cause-and-effect story satisfies and shares better than “we don’t know.”

A clean mechanism feels complete. Once one appears, checking often stops even when the same numbers fit confounds or reverse causation just as well. Ice cream sales rise with drowning deaths; summer heat drives both, yet the causal upgrade still tempts.

Confirmation bias locks the story in. People stop asking what else could produce the pattern. Simpson’s paradox sharpens the trap: a trend in the aggregate can reverse when the data is split into subgroups.

The authors want a hard default. Treat every correlation-based claim as causally neutral until a randomized experiment or a strong natural experiment rules the alternatives out. The question to keep asking is what other story fits the same data equally well.

In practice: UC Berkeley graduate admissions in the 1970s looked biased against women in the totals; department-level data showed women had applied more to the most competitive programs, reversing the apparent effect.

3

True numbers, false stories

Selectively chosen statistics—mean instead of median, relative risk without absolute numbers, or a flattering time window—let a technically true figure tell a false story.

Summary choices steer hard. The mean follows outliers, so average income in an unequal population looks richer than what most people actually earn. The median would sit closer to lived experience.

Relative risk manufactures alarm or comfort on demand. Reporting a doubled risk sounds dramatic; leaving out the starting absolute rate hides whether the change matters in real life. Selective start and end dates do the same job for trends in finance and climate arguments alike.

Narrow fact-checking fails here because the number is true. What is needed is a second-order check: true compared with what, chosen by whom, and for what purpose. The move is curation, not fabrication, which is why it thrives in respectable reports and headlines.

In practice: A drug that “doubles” a risk sounds scary, yet if the baseline is one in a million, two in a million is still negligible.

4

Charts that quietly distort

Small design choices—truncated axes, dual scales, or unequal ink—distort what viewers see even when every plotted number is accurate.

Viewers treat graphs as neutral machines. They are not. Axis range, color, chart type, and aggregation level are all choices, and each can be tuned to enlarge or shrink an effect.

A bar chart that starts near the top of the scale turns a minor gap into a dramatic cliff. Dual-axis charts place unrelated series on separately scaled axes so the lines cross or align on purpose and hint at a relationship that is not there. Shapes that violate proportional ink make visual area lie about magnitude.

A verbal claim invites scrutiny. A chart discourages it. The mechanical habit is to read the axes first, notice what the frame excludes, and redraw the same points with a zero baseline before trusting the first impression.

In practice: A bar chart starting at 90 instead of 0 can make a 2% difference look like a 300% leap.

5

The data you never see

Conclusions drawn only from surviving, visible, or published cases are systematically biased because the missing data is almost never random.

The sample you can observe is often entangled with the outcome you care about. That is selection bias, and it needs no deliberate liar to do damage.

Publication works the same way. Journals and media reward positive, exciting findings; null results stay invisible. The public record then overstates how often effects hold up.

Incentives—novelty, drama, career rewards—do much of the filtering. Skepticism has to reach past the presented table and ask what was excluded and why the exclusion rule itself may carry the real signal.

In practice: Abraham Wald saw that WWII bombers returning with holes showed where a plane could survive damage; armor belonged where the returning planes were clean, because hits there never made it home. The 2015 Open Science Collaboration later reproduced only about 36% of 100 psychology studies it re-tested.

6

Algorithms that cannot care

Machine learning can generate bullshit at industrial scale because it optimizes patterns in training data with no mechanism for caring whether outputs are true, fair, or meaningful.

Automated systems inherit the biases of their inputs and then add a halo. “The algorithm decided” sounds more objective than a person deciding, even when proxies quietly carry the old skew.

Recommendation engines optimize pure engagement. Outrage and novelty hold attention better than accuracy, so the loop favors the extreme. Nothing in the objective function checks whether a pattern is meaningful or merely convenient.

An algorithm literally cannot care about truth; caring requires intent, and the system has only a target to maximize. Scale and speed then outrun any human’s ability to inspect each output. The result deserves at least the skepticism given a human claim—and often more.

In practice: COMPAS risk scores used in U.S. courts held no explicit race variable, yet ProPublica’s 2016 investigation found Black defendants flagged high-risk at nearly twice the rate of white defendants relative to actual reoffense.

7

Why debunking often fails

Corrections frequently backfire or stall because they are boring, leave a narrative gap, and travel slower than the charged claim they try to kill.

People keep relying on a false claim in later reasoning even after they see it corrected. The original filled a story hole; a bare “that’s false” does not replace it. That is the continued influence effect.

Virality is lopsided. Sensational or novel claims spread farther and faster than accurate ones because emotion and surprise drive sharing more than precision does. Platforms built to maximize engagement lean the same direction, so individual fact-checks fight the grain of the system.

The practical response is a replacement causal story, delivered early and repeated, plus hard choices about where scarce effort still has a chance. Some claims become effectively unkillable once they saturate the network.

In practice: An MIT team’s 2018 Twitter study found false news traveled significantly farther, faster, and more broadly than true news.

8

Habit beats expertise

Calling bullshit is a practiced checklist, not innate brilliance, so anyone can improve at it regardless of technical background.

Subject-matter expertise is weak armor. Bullshit aims at blind spots outside whatever you know best. A statistician can still swallow a biased study; a scientist can still buy a cooked financial chart.

What protects people is a short list run every time, in every domain: check the source’s incentive, check plausibility against known base rates, check what the frame leaves out, check axes and units, and check whether simpler explanations were ruled out.

The list has to become reflex rather than a special effort—the way experienced checkers flinch at round numbers, too-neat correlations, and single-study miracles. Each person who verifies before sharing weakens the paths bullshit travels through ordinary trust.

In practice: Apply the same five checks hardest to the claim that flatters what you already believed; those pass unchallenged most often.