What a story can substantiate — and why it is not enough, by itself, for a general rule

What we know: A case report, a testimony or a personal experience can be evidence of a particular event and a source of a hypothesis. By themselves they usually do not suffice to estimate frequency, causality or an effect in a population. The experiments and meta-analyses cited below show that a narrative presentation can influence perceived risk, attitudes and intentions.^[5][6][7]

What we do not know: From these studies one cannot determine how strongly one particular story will affect every Czech reader, nor whether it was the cause of a particular change in vaccination coverage or of a court decision. For an individual claim it is necessary to check its own sources, population and denominator.

A table from a large dataset and an interview with one person supply different kinds of information. A story can truthfully describe a particular experience, but without a denominator and a comparison it does not determine frequency or cause in a population. This boundary is especially important where a claim affects a health decision, a punishment or public money.

1. A witness and a denominator are not the same

Imagine an editorial meeting about a health report. The editor-in-chief wants a three-minute interview with a mother whose child ran a fever after vaccination: “People will understand that.” The data journalist puts a coverage graph and intervals on the table. The public-health doctor says that one case is not incidence. The campaign marketer objects that without a face nobody will read the statistics to the end. The lawyer recalls that a court has already believed a strong story without DNA. The reader, who is not in the room, will by evening remember only the sentence that could be told.

At first glance this is a dispute about emotion versus figures. In fact each person is describing a different object. The editor-in-chief measures attention. The data journalist measures a population. The doctor measures mechanism and frequency. The marketer measures the intention to get vaccinated. The lawyer measures the procedural defensibility of a punishment. The reader measures what will remain in memory after the phone is closed.

A newsroom may legitimately look for an understandable illustration, while a doctor or a court needs evidence matching the seriousness of the conclusion. The problem arises when attention or memorability is passed off as the validity of a claim. Persuasiveness is not the same as evidence, a case report is not a general rule, and correcting one fact need not remove the whole narrative.

The first task is to determine whether the story is meant to illustrate an experience, generate a hypothesis, or support a general claim. Each of these uses requires a different degree of independent evidence.

A story can substantiate an individual experience. Without a denominator, however, it does not measure frequency, and without a comparison it usually does not prove a cause.

— Jiný Kontext

2. What can and cannot be inferred from one testimony

When the screen shows “an acquaintance got vaccinated and fell ill”, it is easy to get the impression that the sentence “proved a connection”. The result, however, arises in a chain. Each layer may raise validity, lower it or change the type of error. Two identical stories therefore need not mean the same thing. The same story may behave differently in a family, in an infographic and in a courtroom.

Five layers of one testimony

1. The event: what happened to one person, at one time, with one sequence.
2. Selection: who chose the story, who is silent, who has no camera.
3. The causal bridge: the leap from “afterwards” to “because of that”.
4. Spread: how many times the sentence is repeated, how vividly, with what affect.
5. Use: a hypothesis for research, advice to a friend, a headline, or a basis for punishment and public policy.

The event may be true and the bridge still false. The selection may be honest and silent evidence still missing. Spread may cite an existing study and still fail to carry the sentence attached to it. An honest comparison must state not only the story, but also sample size, the definition of success, the period, the population and whether it is meant to be a hypothesis or a rule.

This is not an academic dispute about words. The cut of the layers determines where to turn when something fails. If the story does not know the denominator, incidence is missing. If it does not know those to whom the same thing happened without a punchline, silent evidence is missing. If a fact-check corrects a figure and people remember the scene, layers 4 and 5 failed, not necessarily layer 1. The sentence “people believe nonsense” is the start of a diagnosis, not its end.

Illustrative scheme: The same anecdote can, in a different use — as a hypothesis, as a headline, as evidence before a court — create a different harm. The truth of a scene is not the truth of a rule.

3. The cost of error depends on how the story is used

The same evidential error has a different cost in an advertising slogan, a health decision and criminal proceedings. A headline can be corrected, but the consequences of changed health behaviour or of an unjust judgment are far harder to put right. The more serious the decision, the less its basis can rest on the vividness of the telling alone.

Daniel Kahneman in the book Thinking, Fast and Slow (Myšlení rychlé a pomalé, 2011) describes how a coherent story produces a feeling of certainty even when it rests on a narrow slice. That is not an insult to the reader. It is a description of the machine we have in our heads. Risk does not grow only with the falsity of a scene. It grows when the fluency of the telling tempts one into undue trust.^[1]

This also changes the question. We do not ask only which text is more moving. We ask which will fail in a way that our process will catch. A newsroom can tolerate ten unusable human stories because it will recognise one well as an illustration. A court, a doctor and a ministry cannot tolerate a single hidden substitution of a case report for a rule.

In assessment, therefore, it matters whether an error can be recognised, put right, and how serious its impact will be. For a decision a drier figure with a clear denominator can be more reliable than a widely shared scene without a comparison.

4. A story describes a case, not a frequency

Stories are needed because they replace part of the abstraction with experience that can be passed on. The problem arises when one scene starts to stand for a whole population. Nassim Nicholas Taleb in the book The Black Swan (Černá labuť, 2007) called this vulnerability the narrative fallacy: an after-the-fact telling that attributes to chance and to a slice an order the data did not have. It is one book and one thesis, not evidence that every telling lies.^[3]

A case report in medicine can pose a question, describe an unusual case and propose further measurement. It cannot, however, by itself carry a population conclusion from twelve children. Likewise a company anecdote about “one brave decision” does not show how many comparable firms failed with the same step.

Five items without which “I know it from my own experience” is only advertising certainty

• Sample size and the method of selection, not only the vividness of the scene.
• The denominator: from how many attempts this punchline came.
• The period, the population and the definition of success or harm.
• Whether it is a hypothesis, an illustration, or a general claim.
• What independent primary source the story would have to carry if it were to hold as a rule.

For buying a conclusion, therefore, a small honest denominator matters more than someone else’s overall excitement. Thirty real cases with a record of failures can say more than one viral paragraph. A public story helps to choose a question. The decision must be made by data that know the denominator.

5. System 1 fills the gap before System 2 has time to count

Kahneman in Thinking, Fast and Slow distinguishes two modes. System 1 is fast, associative, and constructs a coherent story from what is at hand. He calls the principle WYSIATI — What You See Is All There Is: what you see is all there is. System 2 is slow, effortful, counts, checks premises, looks for an alternative. The first can act when there is no time. The second can count when there is a reason and calm.^[1]

When we meet an unknown phenomenon, System 1 constructs a bridge: X happened because Y. Once the story fits and is easy to read, cognitive ease arises. Ease feels like truth. That is not evidence that System 2 does not exist. It is evidence that it does not start by itself, merely because a child’s health or a firm’s collapse is at stake.

It does not follow that people are stupid. It follows that the story machine is the default mode and statistics is a mode that has to be switched on. A newsroom that relies on this machine is not taking a “human approach”. It is choosing a layer. A court that relies on it is not doing justice to a scene. It is taking the risk that the scene will override what is not in the file.

6. Availability, representativeness and what a story leaves out

Amos Tversky and Daniel Kahneman, in the article Judgment under Uncertainty: Heuristics and Biases, published in Science in 1974, described the shortcuts by which people estimate probability when they lack time or data. Availability: a phenomenon we recall easily we take to be more frequent. Representativeness: we judge probability by similarity to a stereotype, not by the base rate in the population. The illustrations in the article are laboratory and betting as well as clinical analogies; the authors write explicitly that the shortcuts are useful and at the same time a source of systematic errors.^[2]

An air crash on television is available. Vascular deaths in the silence of statistics are not. Lichtenstein, Slovic, Fischhoff, Layman and Combs showed in 1978 that people overestimate causes of death that are often written about, and underestimate the quiet ones. That is one experimental set on perceived frequency, not Czech mortality tables. It does, however, state the same mechanism: memory is not frequency.^[20]

A third shortcut attaches to the story from another side: silent evidence. We hear about a person who got rich on a risky bet. We do not hear of thousands who were ruined by the same step, because they have no stage. Taleb in The Black Swan links this slice to the narrative fallacy. A story is always a cut. A cut in principle hides what did not fit into the scene.^[3]

Anyone who merges these three shortcuts into one insult “people do not understand statistics” loses the diagnosis. Availability is treated differently from representativeness. Silent evidence is not treated by another moving hero. It is treated by a denominator.

7. Affect is a feeling of risk. It is not a probability

Paul Slovic and colleagues, in work on the affect heuristic (among other things the book The Feeling of Risk, 2010, building on earlier articles), distinguished risk as analysis and risk as feeling. When an object is liked, benefits seem high and risks low. When it arouses fear, the reverse. Figures can be in the room and still lose to the tone of a voice.^[4][22]

Zebregs, van den Putte, Neijens and de Graaf, in a meta-analysis in Health Communication (2015), compared statistical and narrative evidence by type of outcome. Statistics had a stronger relationship to beliefs and attitudes, narrative to intentions.^[6] These are different outputs, not a universal victory of one form. A campaign about frequency therefore needs different evidence from a message aimed at action.

Freling and colleagues in Organizational Behavior and Human Decision Processes (2020) assembled a broader database of 61 papers on anecdotal versus statistical evidence. Overall, statistics persuaded more. At high emotional involvement — health, a serious threat, an impact on oneself — the ranking reversed and the anecdote strengthened. This is a moderation, not evidence that statistics are useless. It is also 61 papers, of which 50 were published and 11 unpublished, not one Czech cohort.^[7]

Figures inform. A story triggers a feeling from which an estimate is made. Anyone who merges the two will wonder why a graph “did not work”. It did not work because it was not aimed at the layer that affect won in the evening.

8. Two stories can outweigh statistics that are in the room

Cornelia Betsch, Corina Ulshöfer, Frank Renkewitz and Tilmann Betsch in Medical Decision Making (2011) gave participants a statistical figure on the risk of vaccine adverse effects and at the same time a discussion forum with a varying number of negative narratives. In the first experiment they varied the number of stories reporting an adverse effect (1, 2, or 4 out of 10). Willingness to vaccinate fell with the number of narratives; perceived risk mediated this. In the second experiment statistical risk was 20 %, or 40 %, and the narratives 2 out of 10, or 4 out of 10. The number of stories moved the estimate more strongly than the statistical figure. More emotional texts raised perceived risk more; the “richness” of the description by itself did not.^[5]

This is not a study with “99.9% safety”. It is a laboratory forum with given ratios. That is precisely why it is useful: it shows a mechanism, not Czech vaccination coverage. In the study several negative tellings reduced the weight of statistical information in participants’ judgment. An editorial selection of comments can similarly change the impression of risk, but it is not automatically an experiment, and this article does not from it, without measurement, determine the size of the effect.

A pleasant style can conceal the same old problem

A well-told case feels like understanding. It is not automatically that. In health or judicial work it is necessary to check especially whether the telling has mistaken temporal sequence for a mechanism, omitted the denominator, or smoothed two different occurrences into a fluent but non-existent rule. The quality of a scene lowers the cost of reading. It can raise the cost of detecting an error, because the answer does not look suspicious.

9. Twelve children are not enough for a general rule

In 1998 The Lancet published a paper by Andrew Wakefield and colleagues on twelve children, ileal lymphoid hyperplasia and pervasive developmental disorder in temporal association with MMR vaccination. A sample of size twelve is not a cohort or an estimate of incidence. It is a case series that can pose a question, but does not by itself determine a population association. In February 2010 the editors of The Lancet retracted the article. Retraction is an act of the journal about a particular publication, not a guarantee that its reading will disappear from public memory.^[8][9]

Subsequent large studies did not confirm an association of MMR and autism; an Institute of Medicine review did not support this hypothesis.^[17] That is a different level of evidence from a series of twelve cases. The article here does not claim that this particular publication caused a particular fall in vaccination coverage or a measles outbreak; such a causal chain would require separate data.

It does not follow that every case report is a fraud. It follows that a size of 12 cannot carry a claim about millions, and that retracting a study corrects the literature faster than it corrects memory. The check belongs on the “because of that” bridge, not only on the existence of children who really existed.

10. Vaccination coverage is not the same as an anecdote from a forum

Czech HPV figures belong in a different column from Wakefield, but they illustrate the same vulnerability: vivid fears can move coverage even when tumour incidence does not speak in a graph. From public health insurance, HPV vaccination has been covered for girls since 2012. According to ÚZIS data, as summarised in 2023 in a presentation for the health committee, coverage of 13-year-old girls in 2012 75.6 % and in 2016 fell to 63.1 %; in 2018 64.1 %, in 2021 69.2 %. For boys, covered later, in 2017 15.8 % and in 2021 42.5 %. The Ministry of Health in 2025 reported for girls born in 2010 coverage of 79.0 % and for boys 57.4 % in the reimbursement interval — a different cohort, a different year, not to be merged with the series of 13-year-olds 2012–2021.^[10][18]

The fall from 75.6 % to 63.1 % is a fall in coverage of one cohort, not “hundreds of unnecessary deaths a year”. Incidence of cervical cancer and mortality have their own series in the National Cancer Registry and the screening programme; from a year-on-year drop in vaccination of 13-year-olds one cannot by arithmetic manufacture an annual number of deaths. WHO uses an elimination target of 90–70–90. That is a health and political target, not a Czech figure from 2016.^[10][18][21]

An anecdote from a forum can be one of the pressures on a fall. It is not the only one: reimbursement, availability, trust in institutions, covid years. Anyone who from the graph 75.6 → 63.1 makes a clean proof that “disinformation kills hundreds” has built a causal bridge that the data in this form do not carry. Anyone who from the same graph makes proof that stories “can do nothing” ignores that coverage is behaviour and behaviour is a layer that narrative can move.

11. A fact-check corrects a claim, not automatically memory

Against untruths, newsrooms and projects have arisen that compare a claim with a source. In Czechia Demagog.cz, Manipulátoři.cz, Ověřovna iROZHLAS. Lucas Graves in the book Deciding What's True (2016) describes American fact-checking as a genre with a method, not as a magic of truth. The method can say: this sentence has no support. It cannot by itself say: this scene has disappeared from identities.^[12][16]

Nyhan and Reifler in 2010 in Political Behavior described that a correction can, on some political items, strengthen the original error — the so-called backfire. Later work limited this. Wood and Porter (2019) in five experiments with more than 10 thousand participants and 52 items did not find backfire as a common phenomenon. The Debunking Handbook 2020 by Lewandowsky, Cook and colleagues summarises the consensus: a correction usually reduces the error; backfire is not a reason not to correct. Persistence of error is explained more by repetition from elites, decay of the effect over time and the so-called continued influence: people in their judgment still lean on withdrawn information even when they “know” it as untrue. Johnson and Seifert described this as early as 1994; Ecker and colleagues in reviews hold it as the main problem, not as spectacular backfire.^{[13][14][15][19]}

A fact-check is needed, but it need not by itself remove the memory trace of the original story. Control moves from the authority of the medium to whether the claim is actually supported by the source. A correction is usually more effective when, besides a denial, it offers a documented explanation of the mechanism and does not leave only an empty place.^[13][15]

12. Silent evidence: a grandmother and those we do not hear

An illustration, not a clinical study. A sentence circulates on the network: a grandmother ate a spoon of turmeric daily and lived to 98 without a doctor. The post has tens of thousands of shares. The reader buys a supplement because the scene is concrete, cheap and has a happy ending. Beside it there does not run a list of people who did not eat turmeric and lived to the same age, nor of those who ate it and died in their fifties. That is not because the list would not exist as a logical possibility. It is because it has no punchline.

Here representativeness and silent evidence fail at once. One long-lived woman can be a true individual fate, but she does not by herself prove an effect of curcumin on longevity. That would require a comparison with a denominator and control of other factors. Anyone who inserts a made-up p and ten thousand patients into the illustration makes the same error the text describes: invents a precise figure so that the scene looks like a study.

The judicial analogy is more serious, because the consequences of a judgment are hard to put right. Innocence Project data on DNA exonerations in the USA show that in its set mistaken eyewitness identifications and false confessions played a significant role.^[11] This is not Czech criminal procedure and the percentages cannot be transferred to all judgments. They do, however, support a narrow warning: the persuasiveness of a testimony or a confession does not replace an independently reviewable chain of evidence.

13. Citations do not remove a hallucination of meaning

Before hypertext the reader checked whether the author had invented a fact. With links one checks something more complicated: whether the source exists, whether it is primary, whether it speaks to the same question and whether it really carries the sentence to which it is attached. A citation is navigation. It is not automatic evidence.

One answer can combine three true sources into an untrue conclusion. The first describes a worldwide population, the second Czech girls, the third a different period. Each link works. The synthesis is methodologically wrong. At other times a system — human or machine — uses a high-quality secondary text even though the primary document is available. Or it cites a page that has since changed.

A four-step check of a story that wants to be evidence

1. Existence: the link works and leads to the document described.
2. Authority: the source is primary for the given claim, or clearly admitted as secondary.
3. Support: the particular passage carries the particular sentence.
4. Scope: population, period, definition, unit and whether it is a hypothesis or a rule all fit.

The most reliable working procedure separates searching from writing. First a table of claims, sources and limits. Only then a headline. For every figure the denominator is written down. If a source cannot be opened, a claim is not promoted to a fact merely because someone told it nicely.

A research or news system is good when it shortens the path to a check, not when it replaces the check. For a low-risk anecdote a scene with several links may suffice. For a health, legal or publicly significant conclusion the last step must be taken by a person who understands the source and bears responsibility for its use.

14. The price of sharing is not the price of truth

The attention market uses several price languages at once. One is reach. The second is correctness. The third is the reversibility of harm. The fourth is who pays for the correction. A direct comparison “this worked because it went viral” without these columns deceives.

A story can cheaply bring attention, but in a health or judicial decision an error can be very costly. This article does not, without separate data, determine whether a particular telling caused a fall in vaccination, an outbreak of disease or another population consequence. A more demanding analysis pays off where it reduces the risk of an unfounded causal conclusion with a high impact.

The real price of an accepted story

Attention + time for checking + corrections + harm from an untrue rule + the price of silent cases that the scene erased.

The result is cheap only when it can carry a denominator and can be used for a decision that can partly be reversed. Otherwise we have bought fluency and paid for layer 5.

15. What to choose when you decide about health or punishment

A reasonable choice does not start with whether the text is “human”, but with the kind of claim and the control mechanism. The following signpost is not a ranking of genres. It is a first pilot. For each row it makes sense to take two candidates — a scene and a denominator — and compare them on the same question.

For Czech use a separate test is necessary. An English experiment will not measure whether a Czech headline preserves the difference between 13-year-old girls 2012 and the cohort born 2010. A test set should contain diacritics, a local institution, a table with a denominator and a text in which the correct answer is “this is a hypothesis, not a rule”.

Some newsrooms in the end need several genres. One for the scene, a second for the register, a third for the correction. That is not a failure of voice. It is the same decision as using a different tool for a column and a different one for an epidemiological report. The condition is that the division is understandable and that the scene does not pass into a rule unnoticed.

What needs to be tested before a scene leaves the newsroom

Three tests suffice. The first is the denominator: from how many cases did this punchline come? The second is selection: who would have taken the same step and had no camera? The third is the bridge: what would have to change in the world for the story not to hold as a rule? A product that answers a minute earlier with a moving sentence need not be better if an hour later nobody can find out why one scene was made into a law.

16. A short test of a story that will say more than an impression

A choice can be made without a six-month committee and without an impulse from the comments. A short, disciplined audit suffices. Its aim is not to prove that stories are bad. The aim is to find situations in which a scene fails as evidence, and to see whether we can catch the decision.

Fourteen days from impression to usability

1. Days 1–2: Write down whether the text is meant to be a hypothesis, an illustration, or a rule.
2. Days 3–4: Prepare the same inputs: a primary source, a denominator, a period, a population.
3. Days 5–7: Verify each decisive claim at least three times. Keep a register of errors: anecdote, bridge, silent evidence, affect.
4. Days 8–10: Check citations in four steps. Retracting a study is not the same as erasing a scene.
5. Days 11–12: Count the price of the accepted conclusion including reversibility: a headline, a vaccine, a punishment.
6. Day 13: Stress tests: two negative comments beside statistics, a grandmother without a denominator, twelve children as “evidence”.
7. Day 14: Decide according to the matrix, write down what the scene does not measure, and the date when the claim will be reopened.

The assessor should if possible not know whether the text is meant to move them. For health they follow the denominator and the mechanism. For justice a chain of traces. For a campaign, whether the narrative complements statistics or overrides them. For a correction, what remains after a week. Every failure is classified.

Conditions must be comparable. If one material may be a scene and another must be a register, we are testing genres — which can be right, but we must name it so. If we want to test evidential strength, we must unify the denominator.

The pilot should also have exclusion conditions. For example: a single non-existent figure in a decisive sentence, a mix-up of cohorts, use of a retracted study as if it still held, or a punishment resting on the coherence of a telling without a trace. High sharing must not override an error whose impact the organisation cannot accept.

The last output is not only “a story, or a table”. It is an operational rule: what the scene is used for, what it is not, which data must lie beside it, what the source must carry and when the decision will be reopened. Media, studies and networks change. Without a review date even a good fact-check turns into a superstition that the correction was enough.

17. When an error becomes part of identity

Stories in a feed, a campaign and a television report can join an individual claim to a group identity. In such a situation a correction no longer competes only with one fact, but with a coherent scene that a person has remembered.

That is why the competition is not only for the best fact. It is a competition for the working context of memory. For which scene becomes “our” story. The more comfortable the coherence that arises, the higher the price of leaving it can be. Corrections are hard to export into identity. People stop distinguishing their own experience from a selection someone assembled for them.

A telling can show understandably why statistics matter. A newsroom, an office and a court must, however, label an illustration, add a denominator, separate a hypothesis from a general conclusion and, when correcting, offer a documented explanation.

The strength of evidence is not measured by reach, the length of a quotation or the number of links. A claim must match the size and selection of the sample, the method used and the seriousness of the decision it is meant to support.

The moment a story only illustrates a table, we are choosing a voice. The moment it reads health, punishment and public money, we are also choosing a structure of responsibility. And then the question is no longer: “Which story is the strongest?”

Let us ask: How many silent cases did we not see, who chose the scene, and what evidence would be needed for an individual experience to become a general rule?

— Jiný Kontext