Literary Insights

“If you liked the book, you can purchase it using the links in the description below. By buying through these links, you contribute to the blog without paying any extra, as we receive a small commission. This helps us bring more quality content to you!”

BOOK LINK:

CLICK HERE

Here is a summary of the preface:

The world is filled with too much bullshit in both politics and other areas like science and business. It pollutes discussion and undermines trust in information.
There are two main types of bullshit - “old-school” bullshit that uses fancy rhetoric and language, and “new-school” bullshit that uses quantitative language and data to seem rigorous but can be misleading.
The authors teach courses on data analysis and critical thinking. They wrote this book to help readers detect both types of bullshit without needing advanced technical skills, by applying basic logic and fact-checking.
They aim to help all citizens resist propaganda and manipulation so democracy can thrive. Detecting bullshit should not take political sides, but rather empower people of all views to critically examine information from all sources.
The book will provide examples of bullshit and techniques for questioning dubious quantitative claims or detecting flaws in data visualizations, statistics, studies, and causal arguments. The goal is to educate citizens and help politics be grounded in facts rather than misinformation.
Bullshit has its origins in deception which has existed in the animal kingdom for hundreds of millions of years. Certain marine crustaceans like mantis shrimp engage in bluffing behavior, waving their claws during molting when they are vulnerable even though they cannot punch full force. However, this is an instinctual behavior rather than sophisticated deception.
Corvids like ravens and crows are among the most intelligent non-human animals and show signs of having a theory of mind. They are cautious about caching food where others can see due to risk of theft. When being observed, they will quickly cache food or fake-cache by pretending to stash food but keeping it concealed. Faking caching only qualifies as bullshitting if the corvid understands the impression it creates in the observer’s mind, indicating they have a theory of mind about what others can see/know.
Ravens have been shown to exhibit deception when caching food, hiding treats from other ravens by considering what those ravens can and cannot see. This suggests ravens have a basic “theory of mind” of understanding what others know.
Humans take deception to another level with our rich language and ability to model how messages will affect others. Language allows us to convey a vast range of ideas and manipulate beliefs.
“Bullshit” is common in human communication partly because we all try to promote ourselves, and our language gives others handles to manipulate us.
“Paltering” uses technical truths to mislead without outright lying. Implicature, what language implies rather than literally states, provides wiggle room for misleading claims with plausible deniability.
Weasel words and passive language in advertising, politics and corporate communications diffuse responsibility to avoid accountability, though some uses obscure human suffering.
While some bullshit aims to deceive, others are indifferent to truth, focusing on self-regarding signals about the communicator rather than objective claims.
People often tell stories and spread information to create impressions of themselves and convey things about who they want to be seen as. This can lead to a lot of “bullshit” being produced and spread.
It is much easier and less work to produce and spread bullshit than it is to refute and clean up bullshit. This is known as Brandolini’s principle.
Andrew Wakefield’s now retracted and discredited 1998 study linking vaccines to autism had huge consequences and spread a lot of misinformation, despite being thoroughly debunked with evidence. The vaccine-autism link belief persists due to parental fears and its simplicity.
Wakefield’s hoax has had serious public health impacts, with vaccine rates dropping and diseases resurging. Debunking complex bullshit takes a huge effort compared to creating and spreading it initially.
Bullshit and falsehoods spread much faster than the truth can catch up, as captured by Jonathan Swift’s saying. Social media provides an environment where rumors and bullshit can spread widely and be analyzed at scale.
Chapter 2 discusses how changes in media and technology have contributed to the proliferation of misinformation.
If smartphones had enabled easy fact-checking, it was predicted this would end the spread of bullshit. However, people do not typically fact-check claims on their phones.
Instead, smartphones have become another vehicle for spreading bullshit as false information goes largely unchecked.
The rise of the Internet changed the way people acquire and share information. Things like news feeds, social media platforms, and constant information availability via mobile devices have created an environment where misinformation can rapidly spread.
These changes in media environment provide “fertilizer” for the proliferation of distractions, misinformation, bullshit and fake news, making it difficult to correct false claims. The chapter examines how and why this shift occurred.

So in summary, the chapter looks at how new media technologies like smartphones and social media, rather than ending misinformation, have enabled its faster spread by changing how people consume and share information. This new environment promotes the proliferation of false and misleading claims.

The internet has greatly lowered the cost of producing and sharing information, allowing much more content to be published from a wider range of sources. This has democratized information but also led to more misinformation.
Content is increasingly driven by what gets clicks and spreads on social media, rather than intellectual depth or factual accuracy. Fluff, clickbait, and tabloid-style content have proliferated at the expense of serious journalism.
News coverage has become more partisan as amateur publishers without journalistic standards can reach large audiences. It’s harder for people to discern reliable sources.
The sheer quantity of information available online through search and social media feeds makes it difficult for people to filter out low-quality content and focus on important, well-researched information. People are overwhelmed by a firehose of information.
Traditional news organizations were supported by subscriptions and focused on maintaining quality over the long run. On the internet, the business model rewards quantity of clicks over accuracy or depth of reporting. This shift has further incentivized fluff over substance.

So in summary, while the internet has democratized information, it has also enabled the spread of misinformation due to changes in how content is produced, shared, and financially incentivized on digital platforms. Quality is increasingly lost amid a deluge of low-effort, attention-grabbing but shallow content.

Headlines on social media often appeal to emotions and avoid conveying facts to drive more clicks and shares. Emotional experiences like “will make you cry” perform better than headlines about facts.
Traditional headlines clearly conveyed a story’s essence, but click-driven media incentivizes headlines that don’t reveal too much to pique curiosity.
Partisanship has increased in mainstream media as cable news channels and online publishers deliver slanted or hyperpartisan content, since it is shared and engaged with more on social media.
Algorithms personalize feeds based on preferences and activity, amplifying agreeing views and suppressing dissent, further isolating people in separate information bubbles and tribal epistemologies not based on facts. This fragmented partisan media landscape makes nationwide discussion more difficult.
Social media facilitates the spread of misinformation and disinformation. Fact-checking takes time so publishers rush to be first, cutting fact-checking in the process.
A 2018 study found about 2.6% of US news articles were false, potentially reaching millions daily. While some false stories are harmless, others can have serious consequences.
Rapid internet expansion in India and elsewhere increases susceptibility to false rumors. On WhatsApp, fake kidnapping videos spread fear and led to mob attacks killing dozens.
Even educated people can be fooled. A fake news story led Pakistan’s defense minister to threaten nuclear retaliation against Israel on Twitter.
Political propaganda spreads more easily on social media since people are more likely to believe stories from trusted connections. Russian propaganda reached 126 million Americans on Facebook.
The “firehose strategy” confuses people by broadcasting high volumes of contradictory stories, exhausting critical thinking abilities.
Governments use social media for surveillance and manipulation of public opinion. However, most fake news is created for profit, not propaganda, with some sites making thousands monthly from advertising.
During the 2016 US election, a fake news story that Pope Francis endorsed Donald Trump went viral on Facebook, garnering almost 1 million engagements - more than a top NY Times article. The pope denounced “coprophagia” or obsession with scandals in the media.
Fake accounts and bots are a modern form of counterfeiting that undermines trust in information. An estimated half of internet traffic is non-human. The FCC net neutrality debate saw massive flooding of fake comments.
Sophisticated “deepfakes” using machine learning can generate realistic fake images, videos and audio that are hard to distinguish from real ones. This complicates verifying the truth.
However, society can still discern truth by cross-checking multiple independent sources and considering context. While technology and regulation aim to tackle misinformation, an arms race is likely and censorship concerns remain. Individual discernment remains important.
The passage discusses the nature of bullshit and how it differs from lies. Bullshit aims to persuade or impress through rhetorical techniques rather than convey the truth. It can involve exaggerated details, meaningless language, or disregard for factual accuracy.
Bullshit is characterized by the speaker aiming to manipulate the listener rather than communicate truthfully. The speaker may conceal this through rhetorical flourishes or overwhelming the audience with information.
One example given is Freud describing how he gave a lecture without proper preparation but spoke confidently using cocaine, trying to impress the audience rather than convey understanding.
The sociologist Bruno Latour’s work is discussed, how authors may appear authoritative not just by being correct but through citation lists, jargon, and treating scientific claims as “black boxes” that are difficult for readers to scrutinize.
Effective bullshit can act like black boxes, shielding claims from fact-checking through complexity or appealing to established authority rather than objective truth. An example of an easily disputable claim about cat and dog people’s salaries is given.
The passage discusses an example of someone making unfounded claims about a link between pet ownership (cats vs dogs) and personality/career success. They start embellishing the claims with made-up details to make it sound more credible.
This is described as a form of “bullshit” that functions like a “black box” - it shields the original claim from scrutiny by adding complex-sounding filler that is difficult for the listener to verify or disprove.
The passage then considers what would happen if someone cited an actual research study making this claim. For most people, the statistical analysis in the study would act as another opaque “black box” that is difficult to unpack without expertise.
However, the key point is that you usually don’t need to understand the technical details (open the black box) to call out problems. Oftentimes the issues lie in biases in the data used or implausible results, which don’t require deep expertise to identify.
As an example, the passage analyzes a research paper claiming to use AI to detect “criminality” from facial photos. It questions the data and biases used to train the algorithms, not the algorithms themselves. This illustrates how to identify problems without understanding the technical details.
The passage discusses an association found in a study between higher self-esteem and having had a first kiss before college. However, the study only showed an association and did not prove that self-esteem causes kissing or vice versa.
Merely finding an association between two things does not demonstrate that one causes the other. There could be other factors influencing both variables.
People often erroneously infer causation when they find two things are correlated or associated, but more evidence is needed to establish causation.
Establishing causation versus just association is an important part of rigorously understanding relationships between variables and avoiding bullshit claims. The passage uses this study as an example to introduce the topic of distinguishing correlation from causation.
It discusses how associations work and how scientists think about linear correlations, but notes just finding a correlation does not prove one thing causes the other. This distinction between association and causation is a common source of bullshit claims.
Understanding linear correlations involves plotting two variables (like height and weight) on a scatter plot and seeing if they appear correlated (points form a slanted line).
Correlation strength is measured by the correlation coefficient from -1 to 1, with 1 being perfect positive correlation and -1 being perfect negative correlation.
Correlation only indicates association, not necessarily causation. More data and analysis is needed to determine causation.
Sports examples show correlation between factors like team spending and win rates. The direction of causality is unclear - success may drive spending or vice versa.
Media often wrongly claims causation based only on correlations found in studies. More rigorous analysis is needed to establish cause and effect relationships rather than just associations.
Diagrams can be used to depict correlations, potential causal relationships, and feedback loops between variables. But the true nature of the relationships often remains unclear without further evidence.

The key point is that correlation only indicates association, not causation. Claims of causality require stronger evidence and reasoning beyond just identifying a correlational relationship between two variables. Media coverage often overstates causality suggested by correlations alone.

A Zillow study found a correlation between high home prices and lower birth rates among women aged 25-29, but cautioned that correlation does not prove causation.
Popular media reports on the study suggested a causal relationship, implying high home prices were causing lower birth rates.
Correlation does not imply causation. There are many alternative explanations for the observed trends.
Similar issues arise when studies find correlations in health/medical contexts but media reports suggest causation, e.g. between exercise and cancer risk.
Prescriptive claims based solely on correlations are problematic and premature. More evidence of causation is needed before making recommendations.
Even scientific papers sometimes use causal language when only correlations have been observed, which can mislead.
A classic example is Ronald Fisher hypothesizing cancer causes smoking based on correlations, though he was wrong.
A study found drinking beer from pitchers correlated with drinking more, but banning pitchers based on this alone assumed causation without evidence.

So in summary, the key issue is media and others inferring causation from correlations in a premature or unwarranted way, when more research is needed to establish causative relationships. Correlation does not equal causation.

Studies have found an association between poor sleep and Alzheimer’s disease, but the direction of causality is unclear - poor sleep could cause Alzheimer’s or vice versa.
Causality can be implied subtly through grammar (subjunctive mood implies causality more than indicative) and data visualization (which variable is placed on the x or y axis).
The marshmallow test found kids who delayed gratification had later success, but replication found this was due to socioeconomic factors, not delayed gratification causing success. Wealthy parents produced kids who could wait and be successful, not the waiting itself.
Correlations can be spurious or “illusory” when two things regularly co-occur by chance with no meaningful causal connection. Examples are the age of Miss America correlating with murders by steam/hot objects. These correlations tell us nothing about how the world works. We must be cautious about inferring causation from correlation.

Here are the main points:

Spurious correlations arise when unrelated data series are compared over time. Minor similarities can emerge due to chance rather than any causal relationship.
Tyler Vigen finds many examples of spurious correlations by systematically comparing trends across different data sets covering things like deaths, sociology degrees awarded, etc. Even seemingly strong correlations mean little and do not imply causality.
Researchers need to be careful of data dredging, where too many comparisons are made and correlations emerge by chance rather than reflecting real relationships. Large data sets require accounting for this risk.
Simple upward or downward trends over time are likely to correlate by chance when paired, even without causation between the variables.
Experiments, especially those involving randomization, are needed to distinguish correlations from causal relationships. Manipulating potentially causal variables allows establishing causality rather than just association.

So in summary, the key points are how spurious correlations often arise by chance when unrelated variables are compared, and that experiments are a stronger approach than observational studies alone for determining causal relationships.

Numbers are widely used to quantify and analyze the world, but they can be misused to spread misinformation if not properly understood and presented.
Some numbers come from exact counts or measurements, but many estimates are based on sampling a population due to the infeasibility of counting or measuring everything.
Sampling introduces potential sources of error like sampling error from chance fluctuations, measurement error, and selection bias if the sample is not truly representative of the overall population.
It’s important to consider the context, methodology, and potential biases when interpreting numbers and estimates rather than taking them at face value. Representativeness of the sample and potential for manipulation must be evaluated to avoid being misled.
While numbers feel objective, they can still be used to spread bullshit if not accompanied by sufficient context about how they were derived and potential limitations or biases in the methodology.

So in summary, the passage cautions that numbers are not as objective as they seem and must be carefully evaluated and placed in context to avoid being misled, as sampling and other factors can introduce errors or biases.

Summary statistics and percentages can be used to mislead if presented without proper context. Politicians may cite an average tax cut amount to make a tax plan sound beneficial when it actually only helps the wealthy. Similarly, percentages are commonly used in marketing to persuade consumers without providing all relevant details.

Speed and other quantities like whale populations are sometimes inferred indirectly rather than directly measured, introducing possible errors. Quantities can be miscounted, samples may not be representative, and inference procedures can be flawed.

When communicating numbers, it’s important to choose appropriate summary statistics, provide necessary context, and present data in a way that allows for meaningful comparisons. Simply citing numbers alone does not remove them from context, as the choices in representation set an implicit framing. Honest and transparent reporting of quantitative information requires judges discretion in presentation for clarity and to avoid misleading the audience.

The passage discusses how percentages, numbers, and relative risks can be used in ways that obscure meaningful comparisons or fail to provide proper context.
It gives examples like a hot cocoa package claiming to be 99.9% caffeine-free, which sounds low risk but doesn’t tell you how it compares to coffee. Or a Breitbart article citing crimes by DACA recipients without comparing the rate to citizens.
Percentages can make large numbers seem small (Google said 0.25% of searches return inaccurate results) or small numbers seem large (bitcoin lost 34% value but it was over a 52% drop from the high).
Reporting percentage changes requires specifying what the change is relative to.
Relative risks are better but still need context, like a study finding low drinking increased health risks by 0.5% but failing to note the original risk was close to 0%.

The key point is that statistics like numbers, percentages and relative risks can be used in misleading ways unless properly framed with relevant comparisons and full context about what the numbers actually mean. Proper perspective is important to interpret risks and draw meaningful conclusions.

This passage summarizes several key points about percentages and how they can be misleading if not interpreted carefully:

The difference between percentage points and percentage changes. The same numerical change can be expressed as a small percentage point increase or a large percentage change, depending on how it’s framed.
Context is important. What may seem like a small reduction needs to be viewed in the overall context, like the influenza vaccine reducing cases by 1% but cutting the rate in half.
Changing denominators can obscure what’s happening to the numerator. Like incarceration rates for African Americans decreasing as a percentage even as their numbers increased in reality.
Percentage changes don’t necessarily cancel out. A 10% rise followed by a 10% fall won’t leave you where you started due to the changing base.
It can be misleading to assign percentage contributions when totals are changing in different directions, like rideshares increasing while taxis decreased overall.
This discusses Goodhart’s Law - that quantitative metrics can inadvertently influence the behaviors they intend to measure, like the unintended consequences of a rat bounty program in colonial Vietnam.

The passage discusses the concept of “mathiness” - the use of mathematical formulas and expressions that have the appearance of rigor but disregard logical coherence and formal mathematical principles.

Two examples are provided:

The “VMMC Quality Equation” relating factors like appropriateness, outcomes, service, and waste to quality. However, the quantities are not well-defined or measurable, and the relationship could be expressed just as well without equations.
The “Trust Equation” relating factors like credibility, reliability, authenticity and self-interest to trust. Again, the relationships are plausible but not uniquely captured by this particular equation.

The key point is that while these equations express basic qualitative truths, there are usually many mathematically equivalent ways to do so. By using equations without sufficient justification or definition of terms, these examples demonstrate a “disregard for logic or factual accuracy” akin to “bullshit” - they trade on the appearance of rigor without substance. The purpose seems to be persuasion rather than accurate representation.

The passage defines “mathiness” as the use of mathematical expressions in this way, prioritizing impression over coherence or evidence. It is a form of “bullshit” that employs mathematical notation rather than plain language.

This passage discusses some specific implications and issues with expressing trust, credibility, and relationships quantitatively through mathematical equations. Some key points:

Trust approaches infinity as perceived self-interest approaches zero in the trust equation, but infinite trust in a random coin flip doesn’t make sense in reality.
The trust equation treats increases in reliability, credibility, authenticity equally, but trust could be high even if one factor is low or zero.

-Dimensional analysis (tracking units) often doesn’t make sense for equations created just for “mathiness.” Happiness and sadness can’t realistically be measured the same way as variables in the equations presented.

Zombie statistics, like the claim that 50% of scientific papers are never read, are often cited out of context or were made up originally but spread widely. Tracking the original source of such statistics is difficult. The 50% paper figure is inaccurate for various reasons outlined.

In summary, the passage critiques using equations to quantify complex human relationships and attributes when the equations don’t reflect reality and lack proper units or context. It warns about unfounded statistics spreading without verifying their original sources.

The chapter discusses the concept of selection bias, where the samples used for analysis are not truly representative of the overall population.
It gives the example of skiers at the Solitude ski resort in Utah consistently saying it’s the best place to ski, even better than more renowned resorts like Alta and Snowbird. However, the people one talks to there are not a random sample - they chose to ski at Solitude over the other options.
Selection effects can obscure the logic and mislead interpretations when analyzing data. The data fed into statistical analyses and models comes from samples that may not accurately represent the whole group.
Important to recognize when samples are not truly random and may be skewed or biased in some way. This is a major source of confusion and misunderstanding when interpreting data analysis results.
The chapter introduces selection bias and non-representative sampling as a type of “black box” problem, where flawed or biased data inputs can undermine and conceal flaws in statistical analyses.

The key points are:

When conducting research on a population, it’s typically not feasible to study the entire group directly. Researchers instead take samples to make inferences about the broader population.
For the inferences to be valid, the sample must be random and representative of the population. Otherwise, selection bias can occur if the sample is not truly random.
Selection bias arises when the studied sample systematically differs from the target population in a way that impacts the question being asked. This can lead to misleading conclusions if extrapolated to the full population.
Examples given include surveys that obtain non-random samples, like only surveying students on a sunny day or people with certain attributes like gun show attendees. Insurance company claims of average savings also involve selection bias as only people with most to gain switch.
For research to be useful, the results must be carefully evaluated for how broadly they can be generalized beyond the specific studied sample/population.
Internet users are more likely to be young males who drive more miles recklessly. But one insurance company found a huge difference in claim rates between people who provided email addresses vs those who didn’t.
Upon further investigation, they realized the data was flawed. People only provided email addresses when filing a claim, so those with emails must have filed claims. It didn’t mean those with emails filed more claims.
Selection effects like this are common and can mislead analyses if not accounted for. For example, people only visit psychiatrists if they have too much anxiety, not too little, so studies only see one side.
Class size statistics can also be misleading. Universities may report average class sizes of 18, but students experience much larger classes due to a few very large classes enrolling many students. This distorts perceptions.
The friendship paradox theory states most people’s friends have more friends than they do. This seems illogical but occurs due to a few very social people inflating average friend counts. Studies on Facebook data confirm over 90% of people experience this paradox.
The passage discusses a phenomenon called Berkson’s paradox, which can explain why some people feel like attractive people tend to be jerks.
If you plot all potential partners based on attractiveness and niceness, these traits would likely be uncorrelated.
However, when considering who you are willing to date, you select for both attractiveness and niceness. This creates a selection bias.
Specifically, among those you would date, there is now a modest negative correlation - very attractive people are less likely to be nice, and very nice people are less likely to be very attractive.
Similarly, not everyone is willing to date you. Some attractive people have better options. This further restricts your potential dating pool.
Berkson’s paradox essentially arises because of selection biases in who we choose to date and who chooses to date us. It provides an alternative explanation to the idea that attractive people can afford to be jerks.

In summary, the passage uses Berkson’s paradox and ideas of selection bias to explain why some have the perception that attractive people tend to be worse partners, even if attractiveness and personality may not actually be correlated.

The graph showing musicians dying at younger ages in rap/hip-hop than other genres is misleading due to right-censoring of the data.
Right-censoring occurs when individuals still alive at the end of a study are excluded from analysis, biasing results toward shorter lifespans.
For newer genres like rap, most artists are still alive today so only those who died young are included. For older genres, data includes those who lived into old age.
This makes it appear rap artists die much younger, when in reality the data is skewed due to right-censoring and the genres’ different times being established.
It’s a form of selection bias, as the rap/hip-hop samples only include those with shortest lifespans rather than a true random sample across all lifespans.
In reality, there is no evidence these newer genres inherently lead to younger deaths when accounting for the genres’ different eras.

So in summary, the graph is misleading due to right-censoring of data biasing results for newer music genres that are still establishing longer track records over time.

The article discusses the problem of selection bias in studies of workplace wellness programs. Previous studies found benefits, but they did not use randomized designs and likely suffered from selection bias.
A randomized controlled trial at the University of Illinois found that simply offering a wellness program had no effect on costs, absenteeism, or health behaviors. This challenged previous observational studies.
However, when just looking at employees who chose to participate versus not, large differences emerged - suggesting a selection effect. Healthier employees self-selected into the program.
The randomized design was key, as it allocated some employees to a control group not offered the program. This allowed researchers to disentangle the effects of self-selection from any program impacts.
Previous non-randomized studies likely found benefits because healthier employees opted into wellness programs to begin with. The program itself may not have caused the observed differences. Randomization is important to minimize selection biases in evaluations.
The passage discusses the history and evolution of data visualization, from its early uses in the 18th century by scientists like William Playfair and Johann Heinrich Lambert, to its limited adoption in popular media through the 20th century.
Simple charts like bar charts, line graphs and pie charts were occasionally published in newspapers and magazines in the early 20th century, but showing relationships between multiple variables was rare.
The use of data visualization increased in the 1980s as digital plotting software became more widely available. However, graphics still mainly showed only one or two variables, like how a price changed over time.
Today, newspapers like the New York Times employ data visualization experts who create sophisticated, interactive visualizations exploring complex relationships within datasets. Well-designed data visualizations can provide deeper insights while promoting quantitative literacy.

So in summary, the passage traces the history of data visualization from its scientific origins to its proliferation and increasing sophistication in modern media, enabled by advances in digital technologies.

The passage criticizes the increasing prevalence of “data visualization ducks” - visualizations that prioritize form over function by using clever or eye-catching graphical elements that make it harder to understand the underlying data.
Examples are given of charts shaped like lips, ice cream cones, forks, etc. from sources like USA Today. These distract from the actual data comparison task.
While not outright misleading like “bullshit”, ducks shade in that direction by prioritizing attention-grabbing over clear communication.
“Glass slippers” are an even worse form that takes one type of data and presents it in the form of another visualization type, like periodic tables of many business/marketing topics.
The periodic table is a successful data visualization because its structure truly reflects atomic structure patterns. But marketing/business periodic tables arbitrarily adopt this format without meaningful theoretical mapping.
In summary, the passage criticizes trendy or clever visualizations that prioritize form over function, making the data harder rather than easier to understand.

The passage discusses how certain types of visualizations like periodic tables, subway maps, Venn diagrams, and labeled schematics are often misused as metaphors when displaying data that does not actually fit those structures.

Some examples given where the visualization form was appropriately used include subway maps of music genres and human anatomy. However, many other examples are critiqued for shoehorning unrelated data into the visual structure in a misleading way just for aesthetic purposes.

Specific issues called out include Venn diagrams that don’t logically represent set relationships, labeled diagrams that don’t meaningfully connect labels to image parts, and axes on charts being truncated or lacking labels to exaggerate differences in the data.

The key point is that visualizations work best when the structure genuinely reflects relationships in the data, rather than just borrowing a familiar visual form as a surface-level metaphor when it is not a natural fit for the information being conveyed. Appropriateness of form to content is important for clear and honest communication through data visualization.

Bar graphs should extend to zero on the vertical axis so that the lengths of the bars accurately represent the proportional differences between the quantities. Some of the bars in the first example graph are misleading because they do not extend all the way to zero.
Line graphs do not need to include zero on the vertical axis because they are showing changes over time rather than absolute magnitudes. Including zero can sometimes make small changes appear smaller.
Graphs can be intentionally misleading if they manipulate scales or axis ranges in a way that obscures important information or makes relationships appear stronger than they really are. Examples discussed include using different scales for the two axes, not including zero when it is needed, choosing time ranges that hide parts of the story, and using uneven bin sizes in bar charts.
Graphical choices should be consistent with telling the intended story as accurately as possible. Manipulating aspects like scales, ranges, or bin sizes in a way that distorts the interpretation is misleading. Readers need to carefully examine all aspects of a graph to identify potential bias or spin.

Here are the key points about how a designer can tell different stories by choosing variable bin widths:

The designer took the same tax data but grouped it into different bins/intervals to tell three different narratives.
By widening or narrowing the bin widths, the designer was able to define different income groups (e.g. poor, middle class) and craft stories about taxing each group.
Narrow bins showed more granular differences, while wider bins obscured details and grouped more people together, shaping how the data was interpreted.
Choosing the bin structure allowed the designer to selectively highlight or downplay certain parts of the data distribution to fit different narrative goals, even if the underlying numbers remained the same.
Readers need to be aware that binning choices can influence the story told by the data visualization and potentially mislead if not transparently explained. The same data can support various interpretations depending on bin definition.

Here is a summary of the key points from the page:

3D graphs are often used unnecessarily when the data only has one independent variable. In these cases, a standard 2D graph would be better.
3D graphs can be misleading as they use perspective which distorts the relative sizes of elements. Bars on the left will appear shorter than equivalently valued bars on the right.
3D pie charts in particular are problematic as front wedges appear larger than rear wedges, distorting the true percentages.
Graphs should be designed to accurately represent the underlying data, not tell a story the designer wants to portray. Subtle design choices like axis ranges can influence the narrative.
Questions to ask include whether a 3D element adds useful information or just impresses the viewer, and whether the graph respects the principle of proportional ink (the amount of ink used corresponds accurately to the data values).
In general, simpler 2D graphs are preferable in most cases as they avoid many of the misleading effects of 3D visualization.

So in summary, the key message is to be skeptical of unnecessary 3D effects in graphs and check if they accurately represent the raw numbers or distort the story being told. Simpler 2D designs are often better.

In 1958, an article in the New York Times described a new “embryo of an electronic computer” called the perceptron, which was envisioned as being able to walk, talk, see, write, reproduce itself and be conscious. This was the work of Frank Rosenblatt at Cornell.
Rosenblatt described perceptrons, simple logical circuits modeled after biological neurons, as the building blocks for creating intelligent machines. He made bold predictions about perceptrons being able to recognize faces, translate speech, assemble other perceptrons, and potentially gain consciousness.
Many of Rosenblatt’s predictions proved true. Modern neural networks and deep learning use the same basic perceptron architecture. Systems can now do things like facial recognition, machine translation, and more.
However, the key to machine learning’s capabilities is vast amounts of data to train algorithms, not fundamentally new approaches. Garbage in results in garbage out. The hype around AI potential often exceeds the reality of current capabilities.
The focus should be more on issues like algorithmic bias and privacy threats from existing applications, rather than hypothetical human-level or superintelligent machines of the future. Understanding data and systems used today is more important than futuristic what-ifs.

Here are the key points from the passage:

The US Postal Service processes half a billion pieces of mail daily using machine learning systems to read addresses. Their handwriting recognition system correctly interprets addresses 98% of the time.
The remaining 2% that can’t be read by machines go to a facility in Salt Lake City where human experts interpret them at high speeds.
Reading addresses is well-suited to machine learning as it involves classifying digits, which computers can do by treating images as grids of pixels and using labeled training data to recognize patterns.
The Postal Service uses the MNIST dataset of 70,000 labeled handwritten digits to train their systems through algorithms that identify key patterns without overfitting to noise in the data.
Training data is critical—garbage in leads to garbage out. While the Postal Service’s data is high quality, sampling bias can still impact performance if the training data isn’t fully representative. Overall, machine learning has helped automate address reading at huge scale for the Postal Service.
An AI study from Stanford claimed it could detect someone’s sexual orientation from their facial photo alone, with accuracy above chance. This garnered media attention focused on ethical implications.
However, the study made two questionable inferences that should have been interrogated more:

It claimed the AI could see subtle facial features humans couldn’t. But humans were at an unfair disadvantage - untrained vs a highly trained model, and bad at probabilistic decision making. A simpler explanation is humans made poor judgments.
It linked facial differences to the prenatal hormone theory of sexual orientation. But extraordinary claims require extraordinary evidence, and the study provided almost no evidence, just AI inferences on self-selected photos. Many confounding factors could explain facial variations beyond prenatal hormones. Stronger evidence like 3D lab measurements was needed.

In general, when an AI study makes bold claims, we should scrutinize the training data and outputs more than just taking the results and inferences at face value.
It is difficult to understand how machine learning algorithms make decisions, as the rules they learn are often opaque to humans. Opening the “black box” of these algorithms can provide more transparency.
Researchers have started examining what specific features machines focus on to better understand limitations and potential biases. One study found an algorithm distinguishing wolves from huskies was using background snow rather than morphological features.
Algorithms can pick up on auxiliary or indirect cues in data that don’t actually relate to the task, like a portable X-ray machine watermark. This limits generalizability.
Machines tend to perpetuate existing human biases in data if not properly accounted for. Algorithms have shown racial biases in criminal risk assessment and lending. Gender biases have also emerged in some hiring algorithms.
There are calls for algorithmic accountability and transparency to address impact on humans. However, even revealing all details may not explain why decisions are made, due to interpretability issues with complex models.
Biases can be difficult to remove as correlated indirect factors may still proxy for protected attributes like gender or race, requiring careful algorithm design and testing.

In summary, understanding how machines make opaque decisions is an important challenge, and examining what they focus on can help address potential biases and limitations before high-stakes application.

Here are some key points about gender-skewed professional organizations, hobbies, or memberships from the passage:

Professional organizations can sometimes have skewed gender representation, with disproportionate numbers of one gender dominating membership. For example, a women’s professional organization would skew female in representation.
Certain hobbies can also have skewed gender participation, with one gender being more heavily represented than the other. For example, a hobby stereotypically thought of as more female-dominated or male-dominated could exhibit skewed gender representation.
Membership or involvement in such groups provides insight into social influences and norms around expected gender roles and interests. Certain fields, activities, or organizations may face barriers that discourage one gender from joining.
Looking at skewed representation can help understand perceived gender appropriateness of different types of work or leisure activities according to prevalent social constructs of femininity and masculinity. It reflects implicit or unconscious biases surrounding “women’s work” vs. “men’s work.”

So in summary, gender-skewed professional organizations, hobbies or memberships refer to groups where the participation or leadership is heavily dominated by one gender over the other due to social or cultural factors influencing expected gender norms.

Science is a human invention that allows us to understand the universe beyond human scales of time and space. However, science is still a human practice and relies on human cognition, so it is not an absolute pathway to truth.
Scientists are motivated by both curiosity and scientific discovery, as well as more “sullied” motivations like status, recognition, promotion and career advancement.
Scientists build their reputations by publishing novel, relevant findings in peer-reviewed scientific journals. Publication in top journals carries the most prestige.
The peer review process aims to uphold standards of novelty, completeness, correctness, and appropriate scale for scientific papers. However, flaws can still occur in the system.
Scientists compete for priority and recognition of discoveries. Findings can be overturned if not replicated by other researchers. High-profile cases like cold fusion claims show how flawed findings can still enter the scientific literature initially. Overall, science self-corrects through replication and skepticism, but is not flawless.
Researchers claimed to have achieved “cold fusion” by generating excess heat from deuterium molecules in palladium metal, which would provide a new source of clean energy. However, many labs tried and failed to replicate their findings. Within a year, the scientific community concluded that cold fusion was not real.
Science can change when current discoveries disprove long-held assumptions. For example, epigenetics showed that nongenetic inheritance plays a role alongside genetics in determining traits, altering views on inheritance through genes alone.
In the early 2000s, replication problems began occurring more frequently across fields. Fraud cannot fully explain failed replications, raising doubts about the reliability of some published results.
P-values are commonly used to determine if results reflect real patterns or random noise. But how they are interpreted, particularly in legal cases, requires careful consideration to avoid logical fallacies. The defense attorney in the example story illustrates how the prosecutor’s argument about extremely low probabilities was misleading given the actual application.

In summary, the passage discusses the evolution of scientific understanding, increasing concerns about reproducibility, and proper interpretation of statistical significance tests to avoid incorrect inferences.

P-values are commonly misunderstood and misrepresented. They indicate the probability of observing results at least as extreme as what was actually observed, given that the null hypothesis is true. They do not directly indicate the probability that the null hypothesis is false or the alternative hypothesis is true.
Confusing a p-value with the probability that the null hypothesis is false is known as the “prosecutor’s fallacy.” It parallels how a prosecutor may incorrectly imply that a fingerprint match means the defendant is very likely guilty.
Calculating the probability of the alternative hypothesis directly is difficult because it depends on prior beliefs, which people often disagree on. However, p-values only consider the null hypothesis.
As research spreads beyond scientific circles, p-values are often inaccurately described, such as implying a 99% certainty of a discovery rather than just a 1% p-value.
Proper interpretation of p-values is important for scientific reasoning and avoiding “bullshit” claims not supported by evidence. Low p-values do not necessarily make improbable alternative hypotheses suddenly likely.

So in summary, it focuses on the correct vs. incorrect interpretation and representation of p-values, which is a common source of confusion and potentially misleading communication in both science and law.

Researchers and journals are less interested in presenting “negative” results that show no statistically significant relationship or effect compared to “positive” results that do show a statistically significant relationship.
Negative results can feel boring or inconclusive since they don’t provide new insights. They may also suggest experimental failure or incompetence.
It’s easy to generate false hypotheses, so just finding something is true among many things tested isn’t that meaningful. Significant positive findings are more interesting.
Researchers have incentives to p-hack their data or analyses to find statistical significance since that’s what gets published. This undermines scientific integrity.
With many hypotheses tested, some false ones will appear statistically significant by chance even with no true effect. Publishing only positives creates a biased sample that overrepresents false relationships.
This “file drawer effect” means p-values alone no longer serve as a good measure of statistical support since what gets published depends on being below an arbitrary cutoff of 0.05 rather than actual evidence strength.
Ioannidis argued this bias towards publishing only positives means most published research findings may actually be false due to these selection biases and pressures, not accurate representations of reality.
Ioannidis argues that most published research findings may be false due to issues like publication bias, where negative results often go unpublished. His analogies compare this to interpreting medical tests.
Interpreting a positive test result as definitive proof of a condition is prone to error if the condition has a low prevalence in the population. This is known as the base rate fallacy.
An example is testing for Lyme disease, which has a low infection rate. Even with a 95% accurate test, a positive result only indicates a 2% chance of actually having the disease.
Testing common conditions like H. pylori is less prone to this error since the base rate is higher (20% prevalence). Most positive results would correctly identify infections.
Ioannidis assumes researchers test highly unlikely hypotheses, but scientists likely test hypotheses with reasonable chance of being true, like 10-75%. This makes publication bias less of a problem.
Studies of published vs. registered clinical trials show negative results are much less likely to be published. Published results often overstate findings by selectively publishing and spinning negative results positively.
This “iceberg effect” makes the true picture obscured if relying only on published literature without considering unpublished results. More work is needed to quantify the extent of the problem.
Public distrust in evolution and climate change can be attributed to deliberate misinformation campaigns by industries like tobacco and oil. Religious groups also attempt to sow doubts about evolution.
Science reporting contributes to misunderstandings by amplifying preliminary results without noting limitations or failing to report on null findings that don’t support the initial results. This gives the public the impression that science can’t decide on issues.
Individual studies should not be taken as definitive facts but as contributions to ongoing debates, as interpreting results requires weighing evidence across multiple studies. However, media often portrays individual studies as facts rather than arguments.
Scientists are also at fault, like overstating a 1980 study on opioid addiction risks due to uncritical citations. Reporting is also selective, focusing on surprising results and omitting contradictory evidence.
Press releases and media reports sometimes overstate or misinterpret scientific findings, like overstating changes to an astronaut’s DNA after space travel. Corrections often receive less attention than initial misleading reports.
Jonathan Swift’s saying highlights how false information can spread more quickly than the truth when correcting it.
The market for academic journals is different than typical consumer goods markets, as authors provide their work to publishers who then charge fees without compensating authors.
Open access publishing has advantages of free access but also risks lower quality if journals accept papers just for the fees rather than quality.
So-called “predatory publishers” have arisen that provide little peer review but charge fees, polluting the literature with unreliable studies to make money.
Predatory journals attract submissions by spamming researchers and sometimes hiding fees. Their lack of rigor allows publication of fringe or deliberately deceptive studies.
No single authority validates “scientific” journals, and online publication lowered barriers for scammers.
While peer review helps quality, it does not guarantee correctness, as reviewers cannot verify all aspects of a study. Mistakes can still occur even in top journals. There is no foolproof way to know if any single study is fully valid.

So in summary, the passage discusses how false information in science can spread due to the rise of predatory open-access journals seeking fees over quality, taking advantage of the inability of peer review to catch all errors.

Evaluating the legitimacy of published scientific papers is important but difficult. The best one can usually do is determine if a paper was conducted in good faith, using appropriate methodologies, and is taken seriously by the relevant scientific community.
Key factors to consider include the journal it was published in (higher impact journals are generally better), the publisher’s reputation, and whether the paper’s claims are appropriately ambitious given the journal’s prestige level. Extraordinary claims in less prestigious journals warrant more skepticism.
Checking if a paper has been retracted or corrected is also advised, as retractions are uncommon but do occur when data cannot be validated. This can be done on the publisher’s website or PubMed database.
While problems like bias, misconduct and false positives exist, science continues to work well due to factors like hypotheses being reasonably likely to be correct a priori, cumulative progress building on previous findings, and contradictory results casting doubt on prior claims becoming more publishable over time. Empirically, science has been very successful at understanding the natural world.

Here are the key points summarized from the passages:

Systems for facial recognition and other biometrics suffer from high error rates. One system tested by London police had a 0.1% error rate for correct identification of true negatives, but only identified 8 out of 22 actual suspects correctly, resulting in a much higher 64% error rate among positive results.
P-values are often misinterpreted as the probability that a hypothesis is true or false. A low p-value only means the data is unlikely if the null hypothesis is true, not that an alternative hypothesis is definitely true. Studies with surprising or unlikely results still require replication before strong conclusions can be drawn.
P-hacking refers to selective reporting and analysis practices that make statistically significant results more likely by chance. This can include testing multiple hypotheses without correction, continuing data collection until reaching significance, and selectively reporting variables that support the desired result.
Through deliberately p-hacking two implausible hypotheses, researchers showed they could claim children’s music makes people feel older and that listening to “When I’m Sixty-Four” literally makes people younger, both with statistical significance. This illustrated how easily p-hacking can produce misleading significant results.
Selection and publication bias can distort the published literature, as studies with non-significant or negative results are less likely to be published. This means the literature may present an exaggerated view of certain effects.
When evaluating new claims or studies, it’s important to consider the source and potential biases or agendas. Questions like who is presenting the information, how they obtained it, and what they may want you to believe can help identify bullshit and misleading claims. Rigorous fact-checking, looking for independent replication, and applying healthy skepticism to all claims regardless of political alignment are important critical thinking skills.
The passage discusses how headlines, lists, and comparisons can sometimes be misleading or unfair if they ignore important context or definitions.
It gives two examples - a headline comparing germs on airport security trays to toilet seats, without noting they looked at different types of germs, and a list of “most dangerous cities” that may be skewed by how city boundaries are drawn.
For the cities list, it hypothesizes that crime rates could depend on how tightly or loosely the city boundary encircles the urban core versus suburbs. Tighter boundaries may result in higher reported crime rates.
It suggests this hypothesis could be tested by comparing city size to surrounding metropolitan area size, predicting tighter cities would have higher average crime rates. Understanding how city boundaries are defined is important for interpretting such lists and comparisons.
In general, the passage advocates carefully considering the metrics, comparisons, and context involved in headlines, lists and rankings to avoid potentially unfair or misleading interpretations that ignore important factors. Additional data may be needed to fully evaluate the claims.
The passage compares the violent crime rate in different cities to the percentage of the metro area that is contained within the city limits. It finds that cities with narrower boundaries tending having higher crime rates, while cities with more expansive boundaries have lower crime rates.
This suggests that how a city draws its boundaries can influence perceptions of how safe or dangerous it is, since more narrow boundaries only include the urban core while more expansive ones incorporate suburbs.
Ranked lists of cities need to ensure the entities being compared are directly comparable, taking into account factors like boundaries.
If a claim seems too extreme to be true, it probably merits further investigation. Examples given are a supposed 40% drop in international student applications due to Trump policies, and reports of 9 billion tons of plastic waste entering oceans annually.
To assess extraordinary claims, it helps to think in orders of magnitude - breaking numbers down into rough estimates of their component parts to see if they seem plausible. This can help identify statistical or factual errors without detailed research.

In summary, the passage discusses how city boundary definitions can influence crime statistics, cautions that sensational claims should be fact-checked, and recommends thinking in orders of magnitude as a tool for skepticism.

The passage talks about using “back-of-the-envelope” or “Fermi estimations” to spot bullshit claims without precise calculations. This involves making reasonable assumptions and order-of-magnitude estimates.
It provides examples of using this approach to estimate the number of John Smiths in the UK, the potential sea level rise from cliffs eroding, and the scale of food stamp fraud claims in the US.
The estimates show the original dubious claims were off by factors of 10 or more, demonstrating the value of ballpark thinking to identify nonsense.
One section discusses avoiding confirmation bias when evaluating claims. Confirmation bias is the tendency to accept information that confirms our existing beliefs without critical scrutiny. This makes us more susceptible to bullshit.
Being aware of confirmation bias in ourselves is important for objectively assessing the reliability of extreme or feel-good claims, according to the author. Back-of-envelope thinking can also help overcome reliance on confirmation.

So in summary, the passage promotes the use of approximate reasoning and order-of-magnitude thinking as a tool for recognizing implausible claims, and cautions against letting confirmation bias influence our skepticism.

Here is a summary of key points from the passage about spotting bullshit online:

Corroborate and triangulate claims by searching for reports of the same claim from multiple reliable sources. Be suspicious of surprising claims only reported by unknown or unreliable sources.
Pay attention to where information comes from. Unsourced claims should be treated like unidentified candy - with skepticism.
Dig back to the original source of a story, not just a headline or tweet. Read full reports and trace claims back to primary reports/data to verify facts.
Use reverse image search to verify pictures and videos are not taken out of context or manipulated in some way by tracing where online an image first appeared.

The passage emphasizes the importance of individual vigilance in verifying claims seen online before sharing to avoid spreading misinformation. It provides concrete techniques like source verification, corroboration, tracing claims back to origins, and image searches to fact check suspicious information.

The chapter introduces the concept of “calling bullshit” as a way to refute and repudiate things like bullshit, lies, trickery, and injustice.
Calling bullshit is defined as a performative utterance - an utterance that performs an action through the act of speaking. Examples given are speech acts like dubbing someone a knight or swearing an oath.
Performative utterances are characterized as statements rather than questions, usually using “I” as the subject in the present tense, like “I resign” rather than past or future tenses.
The chapter suggests that while detecting bullshit is important, a solution to the ongoing problem of bullshit also requires actively calling it out through performative utterances that repudiate and demand better. This moves beyond just spotting bullshit to also refuting and shining a light on it.

So in summary, the chapter introduces the concept of “calling bullshit” as a way to actively refute problematic information through performative language, not just detect it, as part of addressing the proliferation of bullshit. It draws on the idea of performative utterances from philosophy of language.

The passage discusses calling out bullshit or false claims, which is known as a performative utterance since it directly refutes or rejects an idea rather than just reporting skepticism.
Calling bullshit responsibly is important to avoid antagonizing others while still protecting the community from misinformation. It’s best targeted at ideas, not people.
Two strategies for effectively calling out bullshit are presented: reductio ad absurdum and being memorable.
Reductio ad absurdum involves showing how an argument leads to an absurd conclusion, like extrapolating a linear model of track times to predict negative finishing times in the future.
Being memorable involves critique that stays in the mind, like labeling a study that found brain activity in a dead salmon to question assumptions in fMRI research methodology.
The key is to dismantle claims respectfully using logic and humor rather than personal attacks, in order to persuasively dispel false ideas while maintaining relationships.
A study showed photos of people in social contexts depicting emotions like inclusive or exclusive to a salmon to see if it could determine the emotions. Despite the salmon being dead, brain regions associated with emotion processing showed increased activity.
This highlighted the risk of false positives in fMRI studies by showing activation in a dead subject. The authors humorously concluded either they discovered amazing post-mortem fish cognition or had issues with their statistical approach.
The salmon study drew attention to problems with fMRI studies in a memorable way using humor. Humor can be effective for spreading ideas through informal discussion even when not required for logical arguments.
Providing counterexamples is another effective way to critique claims or arguments. For the traffic analysis example, a simple counterexample questioned the premise by noting trees are long-living multicellular organisms but lack the assumed immune system features.
Analogies can also help reframe claims that initially seem reasonable to encourage critical thinking. The baseball analogy highlighted how measuring only travel times ignored increased trips from infrastructure work, analogous to only measuring batting average ignoring a pitcher’s role.
Accurate data can be displayed misleadingly through things like inappropriate scale ranges, as evident in examples with climate change and iPhone sales graphs. The most effective refutation is to redraw the graph correctly to reveal the true picture.
A null model can help determine if an observed pattern could occur without the proposed underlying process. The example shows how using a null model of no age-related performance decline still produces the pattern of decreasing world records with age due solely to smaller sample sizes of older athletes. This undermines using the data as evidence for senescence.
Myths that are intertwined with identities are harder to debunk, as rejecting the myth threatens a person’s sense of self. It’s better to decouple issues from identities and find common ground first. Private discussions are also preferable to public call-outs.
When debunking, keep things simple, fill in knowledge gaps with alternative explanations, don’t overemphasize the original myth, and find opportunities for common ground. Maintaining humility and correctness are also important principles.
When calling out incorrect information, make sure to thoroughly research the facts yourself first before asserting anything. Double and triple check sources.
Approach rebuttals charitably by not assuming malice or incompetence on the part of the original speaker. Consider the possibility of honest mistakes.
If you make a mistake in your own arguments, admit fault gracefully instead of doubling down. Owning up to errors maintains credibility.
Present rebuttals clearly and understandably instead of through an overload of unorganized facts or jargon. Effective refutation takes work to communicate persuasively.
Focus criticisms on the argument itself, not the person making the argument. Attack the ideas, not the individual.
Ensure any corrections are pertinent to the actual discussion at hand. Avoid irrelevant “well actually” corrections that don’t meaningfully address the core topic.
Consider your audience and whether confronting a claim will convince anyone or simply be confrontational. Pick battles wisely.

The overall message is that effectively rebutting incorrect information requires thoroughness, charity, honesty, clarity, relevance and judgment - not just pointing out errors for their own sake. The goal should be advancing truthful understanding, not personal aggrandizement.

Confirmation bias can make us overconfident in our beliefs, and humility is important to counter this. Self-reflection on how we arrive at our views is a mark of mature thinking.
Calling out bullshit is a moral imperative given the prevalence of misinformation in the world today. Both innocuous and harmful forms of bullshit exist.
While laws and technology can help address the rise of misinformation, each individual needs to be more vigilant, thoughtful and careful about sharing information. We should point out bullshit when we see it.
The authors thank their wives, friends, colleagues, students and others who provided feedback and helped refine their ideas over many years of developing and teaching the concept of “Calling Bullshit.” They acknowledge the important editorial and production support from their publisher in bringing these ideas to a wider audience in book form.

So in summary, it outlines why addressing misinformation is important, acknowledges influences on developing this work, and thanks those who helped provide feedback and publish the final book.

Here is a summary of the selected sources:

Several sources discuss the spread of misinformation online, particularly through social media platforms. Facebook reported removing over 1.5 billion fake accounts between 2017-2018. Misinformation can spread quickly through social networks and expose users to untrustworthy websites.
Studies analyzed fake comments submitted to the FCC regarding net neutrality repeal, finding many inaccuracies and duplicates. New York investigated indicating fake identities were used.
Features of online media like clickbait headlines, outrage-inducing content, and algorithmic recommendations can drive traffic to misinformation and extreme content. YouTube and social media may push users to radical content.
Propagandistic “firehose of falsehood” techniques from Russia intentionally exhaust critical thinking abilities. Modern propaganda aims to make truth itself unclear.
Historical sources discussed coping with information overload as early as the 1500s due to the printing press. One scribe criticized the new technology in 1474 for allowing untrue things to be easily spread.
Partisan outlet bias and “tribal epistemology” have undermined shared understandings of facts and truth in public debates. Profits from ads and data collection may incentivize some spread of misinformation online.

Here are brief summaries of the key articles:

Silverman et al. (2016) found that hyperpartisan Facebook pages were publishing false or misleading information at a high rate, undermining civic discourse.
Somaiya (2019) discusses how “junk news” sites produce low-cost, partisan clickbait content that spreads widely on social media.
Sonnad (2017) describes how a bot was used to submit over a million fake comments to the FCC opposing net neutrality, manipulating the public debate.
A 2018 study found that 70% of Facebook users only read the headline of science stories before commenting on them.
Subramanian (2017) reported on Macedonian teenagers who created false news websites for profit during the 2016 US election.
Tufekci (2018) argued that YouTube’s recommendation algorithm can unintentionally expose users to radicalizing and extremist content.
Wiseman (2018) found that hyper-partisan Facebook content continues to perform much better than less partisan posts.
The Wrap (2017) covered a completely fake pro-Trump Twitter account that was actually created by Russian trolls to influence US politics.

The summaries focus on how these articles document problems with misleading information, “junk news”, bots, Facebook algorithms, and foreign influence campaigns undermining public discourse and debate.

Here are summaries of several of the sources:

“Selection Bias” by Hernán et al. (2004) defines selection bias as the error introduced into an estimate of the effect of a treatment or exposure due to the mechanisms of selection of subjects into the study. It discusses different types of selection bias.
Carroll (2018) argues that workplace wellness programs don’t actually generate savings or reduce healthcare costs as claimed, noting that studies finding positive results suffer from various forms of selection bias.
Jones et al. (2018) analyze data from a large workplace wellness program in Illinois and find no evidence that the program lowered healthcare costs or generated a positive return on investment. They note methodological flaws like selection bias in previous studies finding positive results.
Kenny and Asher (2016) analyze life expectancy and causes of death in popular musicians across genres like rock, pop, country, and hip hop. They find hip hop artists have significantly lower life expectancies than other genres, which they hypothesize could be due to selection biases in the personal attributes and lifestyle risks of those drawn to different musical genres.
Ugander et al. (2013) analyzes friendship networks on Facebook and discusses selection bias, noting that friends tend to be similar in attributes like age, gender, education and geographic location, biasing estimates of attributes determined by friend selection. It develops methods to correct for selection biases.

Here are summaries of the selected sources:

Arlow, “The Anatomy of the Facebook Social Graph.” 2011. arXiv: 1111.4503.

Analyzes the structure and properties of Facebook’s social graph using a large dataset of Facebook profiles and relationships. Finds that the graph exhibits small-world and scale-free properties.

“U.S. Survey Research: Collecting Survey Data.” Pew Research Center. December 2019.

Describes Pew Research Center’s methods for collecting survey data in the U.S., including telephone, mail, and online surveys. Details considerations around sampling methodology, question wording, and interview procedures.

Alden, “Statistics Can Be Misleading.” 2008.

Discusses how statistics can be manipulated or presented in misleading ways. Provides examples like only showing a subset of data or changing the scale of graphs. Advocates for full transparency around data and methodology.

Angwin et al., “Machine Bias.” ProPublica. May 23, 2016.

Investigates whether commonly used computational decision-making systems like COMPAS exhibit racial bias. Finds some risk assessments were overwhelmingly likely to incorrectly label black defendants as future criminals. Highlights challenges of bias in algorithms.

In summary, the sources cover topics like analyzing Facebook’s social network structure, methods for collecting survey data, how statistics can be misleading if not properly presented, and identifying racial bias in computational risk assessment systems.

Here is a summary of the key points from the readings:

Several articles discuss how to identify unreliable or exaggerated claims, such as those from unreliable news sources, pseudoscience, or corporate PR/marketing. Red flags include lack of evidence, uncertainty expressed as fact, logical fallacies, motive to mislead, and implausibility.
Studies found exaggerated claims in health science press releases and news coverage compared to original studies. News often misinterprets statistical significance.
Fact-checking sites have pointed out false and misleading claims made by some news sources about topics like food stamp fraud levels, crime rates, foreign student enrollment. Follow-up corrections are sometimes issued.
Concepts like “bullshit detection” and “guesstimation” aim to improve critical assessment of claims by considering sources and limitations rather than assuming perfect accuracy. Information spreads widely but verification is important.
Some products and services make medical or wellness claims not supported by evidence, like certain crystal therapies or expensive airport security tray disinfection tools. Marketing language aims to impress rather than inform.
Data on name frequencies in the UK showed implausible claims made about the author’s name being highly unusual or trendy were false. Sources knowing little details can mislead.

So in summary, the readings highlighted ways unreliable, exaggerated and false claims can spread, and approaches for carefully evaluating evidence and considering limitations/motivations to identify “bullshit”. Fact-checking and follow-up corrections were also discussed.

Here is a summary of the tweet:

Benjamin Wittes wonders whether he would eat hypothetical candy if it were on the street, or give it to kids and friends. He poses this rhetorical question to highlight how one should view and evaluate information found online, just as one would carefully consider eating unknown food left unattended on the street. The tweet prompts the reader to think critically about trusting or spreading unverified claims, rumors and statements just because they are accessible on platforms like Twitter.

#book-summary

Calling Bullshit - Jevin D. West & Carl T. Bergstrom