Science communication explains research evidence accurately and accessibly, so readers can understand what is known, what is uncertain and which decisions the evidence can support. Good practice joins source checking, plain language, clear statistics, honest visuals and respect for the audience—not just a memorable science headline.
How do you explain scientific uncertainty without confusing people? What is the difference between a relative risk and an absolute risk? This Gold Standard guide answers those questions across 25 practical chapters, four worked communication examples, an editorial checklist and connected eduKate science and education articles.
Choose how you want to learn
Find and frame the question · Chapters 1–5 · Follow this route →
Establish the evidence · Chapters 6–11 · Follow this route →
Make information understandable · Chapters 12–17 · Follow this route →
Work through real communication choices · Chapters 18–21 · Follow this route →
Teach, test and continue · Chapters 22–25 · Follow this route →
Complete contents · Specialist directory · Gold Standard master library
Index of 25 chapters
Find and frame the question
- 1. Science communication begins with the audience’s decision
- 2. Explain the answer immediately, then earn the conclusion
- 3. Separate scientific evidence, interpretation, policy and values
- 4. Choose one main message before choosing a channel
- 5. The first two paragraphs are a public promise
Establish the evidence
- 6. Trace a claim back to its original study
- 7. Correlation, causation and confounding
- 8. Describe what researchers actually measured
- 9. Uncertainty should be informative rather than theatrical
- 10. Confidence intervals and variability are not interchangeable
- 11. Tell the absolute number before celebrating a percentage
Make information understandable
- 12. Denominators, sampling frames and missing people
- 13. Graphs need units, scales and honest comparisons
- 14. Images are evidence only when provenance survives
- 15. Explain risk without generating unnecessary fear
- 16. Make technical vocabulary a tool, not a gatekeeper
- 17. A story can illuminate an idea but cannot replace the data
Work through real communication choices
- 18. Worked example 1: the headline that halves the risk
- 19. Worked example 2: separate measurement from inference
- 20. Worked example 3: a beautiful graph with an invisible baseline
- 21. Worked example 4: two audiences, one evidence base
Teach, test and continue
- 22. Pre-bunk misinformation without repeating it as fact
- 23. Communicating across languages, cultures and educational levels
- 24. Evaluate whether readers learned or could act
- 25. Frequently asked questions and the connected reading library
The evidence-to-reader chain
| Step | What to establish | What to avoid |
|---|---|---|
| Question | Audience, decision and main claim | Answers that arrive too late |
| Evidence | Original measurements, methods and limits | Converting inference into observed fact |
| Message | Plain language, denominator, context and useful visual | Jargon, vague percentages and unlabeled axes |
| Feedback | Reader comprehension and ability to apply the idea | Treating likes as evidence of understanding |
1. Science communication begins with the audience’s decision
Science communication makes scientific evidence accessible without making it less accurate. It is not the job of decorating findings with simpler words or persuading everyone to like science. It begins by asking what readers need to understand, which decision or question is actually before them, and what the evidence can responsibly answer.
A parent deciding whether a claim about tutoring methods is reliable does not need the same message as a laboratory specialist assessing instrument calibration. Both deserve honesty, but their initial questions, vocabulary and useful level of detail differ.
A quick audience diagnosis before writing
Imagine two people reading about the same air-quality study. A school principal wants to know whether an outdoor activity should change; a scientist wants the sampling intervals, detector corrections and uncertainty estimates. Presenting only the methods to the principal may fail to answer an urgent question. Giving only a practical conclusion to the scientist may obscure whether the claim is sound. The evidence need not change, but the order and emphasis should.
Before drafting, write the intended audience in one sentence, describe the decision they face and identify what a mistaken interpretation could cost them. Then state the short answer, the critical qualification and the next reliable source. This audience diagnosis is not manipulation. It is a way to make the same truth usable by readers with different immediate needs.
← Chapter index · Next chapter →
2. Explain the answer immediately, then earn the conclusion
People often arrive at a page with one urgent question. An effective first paragraph supplies a careful answer, names the central condition and tells the reader where to go for evidence. Burying the answer under institutional background loses both readers and clarity.
Structure the explanation in layers: direct answer; what scientists measured; what the measurements mean; limits and competing explanations; and next action or related reading. Readers should be able to stop early without being misled.
The five-sentence test
A good opening can often be audited through five short sentences. First: answer the actual question. Second: say what kind of evidence supports it. Third: name the main limitation. Fourth: explain what the result changes for the reader. Fifth: offer the path to deeper detail or official action. Not every article must mechanically follow this order, but all five jobs should be visible.
Try it with a school science project. ‘Our sensor detected higher particles on the busy road’ is an observation. ‘Traffic caused the difference’ is a causal inference requiring a stronger comparison. The first sentence can be useful without pretending to prove the second. Students learn that brevity and caution are compatible.
3. Separate scientific evidence, interpretation, policy and values
A measurement can tell us what happened under specific conditions. Interpretation proposes what explains it. Policy involves priorities, constraints and trade-offs. Values affect which outcomes people choose to protect or pursue.
These are related yet distinct. Scientific results can inform a decision without uniquely determining the most appropriate political or personal action. Good communication labels where evidence ends and a judgment begins.
One study, three levels of conclusion
Suppose an experimental study reports that an intervention changed an outcome within a controlled setting. A narrow conclusion says that the effect was observed under those specified conditions. A broader inference asks whether it extends to other populations and settings. A policy recommendation adds costs, equity, alternatives and public priorities. These three statements are not equivalent.
Science communicators should label which level appears in each paragraph. This practice is especially important with public health, education and environmental research because audiences may assume a research result automatically establishes a universal rule. Good writing keeps the result strong where supported without exaggerating its reach.
4. Choose one main message before choosing a channel
An article, classroom lesson, short video, podcast, infographic and public announcement have different strengths. The suitable medium depends on whether readers must compare numbers, recognise a warning, learn a process or ask questions.
A vivid animation can be memorable but poor at conveying the uncertainty of a numerical estimate. A long research report can document methods exceptionally well but may be unsuitable for an urgent alert. Match the job, channel and attention available.
When a chart is better than a story
Use a chart when relationships, distributions or changes over time carry the meaning. Use a short story when the reader needs to understand a human problem or the sequence of a decision. Use a table for precise comparisons, and use a diagram when a process is central. A decorative photograph is rarely a substitute for any of these.
Ask readers to solve the task without the caption. If they cannot identify which group is larger, which units are shown or whether the differences matter, the visual needs repair. It is also reasonable to place a short sentence beside an unfamiliar plot: ‘Read the horizontal axis as time and the vertical axis as measured concentration.’ Clarity is part of data integrity.
5. The first two paragraphs are a public promise
Write the main question and its practical answer in plain language early. In the next paragraph name what is known, what remains uncertain and where the deeper discussion begins. This gives both busy readers and researchers an honest map.
Do not stuff the introduction with search terms, exaggerated guarantees or brand claims. Relevance should come from actually answering the question rather than repeating the same phrase in several forms.
6. Trace a claim back to its original study
Before writing ‘science proves’, ask who produced the underlying evidence, what was measured, how large the sample was, which population was studied and whether the result is observational, experimental, model-based or speculative.
Press releases, social posts and summaries can point toward a study but may omit important qualifiers. Return to the primary source and, where appropriate, compare independent replication or systematic reviews. A claim becomes stronger through evidence, not through repetition.
Trace the ‘according to scientists’ phrase
An article may say ‘according to scientists’ without naming a paper, measuring instrument or research team. That phrase gives readers no meaningful path to verify the statement. A better source trail identifies the actual study, date, methods and relevant limitations. If the source is a preprint, conference abstract or press announcement, label it accurately rather than calling it peer-reviewed by implication.
Check whether the data were collected at one site, across multiple populations or in a simulation. If a finding is controversial, describe the credible competing explanations and what newer evidence would settle them. Repetition by many secondary articles does not make an original weak measurement stronger.
7. Correlation, causation and confounding
If two quantities change together, the association may be causal, confounded by another factor, reversed in direction or coincidental. A diagram showing a correlation does not settle the mechanism.
Consider a story where students attending extra lessons score better. Differences in earlier attainment, motivation, resources and selection may contribute. Clear reporting states which alternative explanations were investigated instead of announcing that the lessons alone caused the outcome.
Three causal claims that sound alike but are not
‘People who exercise more report better sleep’ describes an association. ‘An assigned exercise programme improved reported sleep in a particular trial’ describes a result under an intervention. ‘Exercise improves sleep for everyone’ is an unsupported universal statement. The change in wording carries a major change in the claimed scope of knowledge.
Teach students to underline the causal verb—causes, prevents, improves, reduces—and ask what research design justifies it. A controlled experiment may support a particular causal interpretation, yet adherence, subgroup variation, chance, measurement error and study conditions still matter. The correct replacement for weak evidence is a narrower claim, not an empty adjective such as ‘interesting’.
8. Describe what researchers actually measured
Outcomes can be defined in many ways: examination scores, self-reported confidence, hospital admissions, laboratory signals or modelled risk. Tell readers which variable was measured and the timescale involved.
A proxy is sometimes helpful, but it is not the outcome itself. A rise in app engagement is not automatic proof of learning, and a molecular marker does not necessarily imply clinical benefit. Rename the result precisely before discussing meaning.
9. Uncertainty should be informative rather than theatrical
Uncertainty can come from sampling, measurement, model assumptions, future variability or missing evidence. Some uncertainties shrink with more data; others reflect real variation that cannot be eliminated.
Do not present ‘scientists are unsure’ as if every established result were in doubt. Equally, avoid hiding meaningful uncertainty to make a headline more attractive. Explain what is well supported and where genuine questions remain.
Uncertainty with consequences, not just error bars
The phrase ‘there is uncertainty’ helps little unless readers know what varies and whether that variation could change the decision. A temperature forecast with a two-degree uncertainty band has a different meaning for a warm outdoor picnic than for a sensitive scientific instrument near its operating limit.
A useful explanation distinguishes uncertainty in the quantity estimated, variability among individuals and uncertainty about the underlying mechanism. Where possible, show a plausible range and identify which assumptions move its endpoints. When uncertainty is irrelevant to the main decision, mention it proportionately; when it could reverse the conclusion, put it near the headline.
10. Confidence intervals and variability are not interchangeable
A standard deviation describes dispersion of observations under a defined calculation. A confidence interval characterises uncertainty in an estimated parameter according to its statistical procedure and assumptions. They answer different questions.
Reporting ‘±12’ without saying what it represents is not transparent communication. Show the estimate, denominator, interval type, unit and definition. The eduKate case study on a mislabeled confidence interval provides a useful diagnostic.
A confidence interval does not mean 95% of people lie inside it
Suppose a report gives an estimated mean and a 95% confidence interval. That interval arises from a statistical procedure designed, under its assumptions, to achieve a stated coverage property over repeated datasets. It is not a statement that 95% of raw measurements or individuals fall in that numeric interval.
If the question is how varied the individual observations are, use appropriate spread statistics or a distribution. If the question is uncertainty in the estimated mean or effect, a confidence interval may be appropriate. A communicator must identify which question the graphic is answering before giving a reassuring-sounding percentage.
11. Tell the absolute number before celebrating a percentage
A relative change can sound dramatic while representing a small absolute difference. If an invented event rate changes from 2 in 1,000 to 1 in 1,000, the relative reduction is 50%, but the absolute reduction is 1 in 1,000, or 0.1 percentage points.
Both calculations are mathematically correct. Presenting only ‘50% lower risk’ without the baseline prevents readers from judging scale. Whenever possible, use consistent groups, time windows and denominators.
Explain a ratio and a base rate together
Consider an invented survey where a rare outcome rises from 1 per 10,000 to 2 per 10,000. The relative increase is 100%, which is correct arithmetic, but the absolute increase is one additional event per 10,000. Neither statistic is meaningless. They illuminate different aspects of a change.
Present both when the difference matters to a decision. Use matching time periods and populations, and state whether counts are observed or modelled. A relative percentage detached from its baseline can exaggerate a small change; an absolute number without the relative comparison can hide a meaningful pattern when starting rates differ.
12. Denominators, sampling frames and missing people
A survey of 400 responses can represent a narrow self-selected group rather than a whole country. The denominator is not only the number in a graph; it includes who could enter the study, who responded and who was omitted.
Ask whether excluded observations, nonresponse, measurement thresholds and subgroup differences could change the interpretation. More decimal places do not repair an unrepresentative sample.
Selection bias can survive a large number of responses
Ten thousand responses sound impressive, but a voluntary online poll may systematically omit those without internet access or those less invested in the topic. The sample’s size describes how many answers were collected, not whether the underlying process represented the population of interest.
Explain the recruiting method and the intended population in ordinary language. A sentence such as ‘among visitors who chose to answer our website survey’ is more accurate than ‘the country believes’. When subgroup differences matter, describe who was included before extrapolating findings into universal advice.
13. Graphs need units, scales and honest comparisons
Graphs are arguments about quantities. A bar chart with a truncated baseline can exaggerate differences; an axis mixing percentages and raw counts can create confusion. A decorative ‘before and after’ photo can imply causality it cannot establish.
Make labels, units, baseline, time period, number of observations and uncertainty easy to find. Readers should be able to reconstruct the claim from the visual without trusting the caption.
A graph audit readers can do in one minute
Read the title, both axes and their units. Find the plotted time period and baseline. Check whether the graph shows counts, percentages, rates or model estimates. Look for uncertainty bands, missing categories and the sample size. Then restate the claimed conclusion without adjectives such as ‘massive’ or ‘tiny’.
If the chart’s visual emphasis disagrees with your restatement, investigate its design. For instance, a narrow vertical axis can make a two-point change appear dramatic. A graph might still legitimately use a truncated scale for precision, but this choice should be plainly marked so viewers know what difference they are being invited to notice.
14. Images are evidence only when provenance survives
Scientific images can be microscope outputs, telescope composites, diagrams or artist’s impressions. Each carries different informational meaning. A false-colour image uses a colour map to show selected measurements; it is not automatically a natural-colour photograph.
Give the origin, date where relevant, instrument or model, processing choices and what the image cannot verify. An illustration should be labelled as an illustration, and a stock image should not masquerade as a documented case.
Real photograph, measurement image and illustration
A journalist photographing a flooded street supplies visual documentation of a particular place and time, subject to provenance. A satellite composite may combine instrument channels with assigned colours to highlight a physical feature. A clean conceptual drawing may explain a mechanism without depicting any actual event. All three can be useful, but they should not share an ambiguous ‘photo of’ caption.
Ask: who produced this image, what is shown, which transformations were made, when was it recorded and which claim does it support? Where a scene is reconstructed or AI-generated, say so. A beautiful illustration is not evidence that the depicted event happened, while a genuine photograph is not proof of a population-wide trend.
15. Explain risk without generating unnecessary fear
Risk communication should identify the event, plausible severity, likelihood or uncertainty, exposure route and actions supported by evidence. The purpose is informed preparation, not fear.
A warning about a hazard should also say who is affected, what is not implied and where official updates can be checked. Avoid universal statements such as ‘dangerous to everyone’ when the evidence concerns a specific population or setting.
Public alerts should translate into specific safe actions
An emergency notice needs a clear hazard, affected place, time window, official source and practical action. General phrases such as ‘major risk in your area’ can create fear without giving people enough information to respond. Technical details may belong behind a link, but the immediate action must remain accurate.
Risk estimates can change as observations arrive. Distinguish ‘possible’, ‘likely’ and ‘confirmed’ according to the issuing authority’s actual meaning; do not invent probabilities. When the situation is not urgent, give people time to compare evidence and trade-offs. A good public message is usable at the moment the reader receives it.
16. Make technical vocabulary a tool, not a gatekeeper
Some specialist terms are necessary. Explain them at the point of use and keep the correct scientific meaning. For example, ‘positive test’ may mean a diagnostic criterion was met, not necessarily that a person has a particular condition with certainty.
Define terms precisely once, use them consistently and provide a short glossary for readers. Oversimplification becomes harmful when the new phrase changes the claim, while needless jargon discourages questions.
Make unfamiliar terms memorable without distorting them
Define a technical term beside the first time it affects the explanation. A confidence interval, for example, cannot be replaced by the loose phrase ‘almost sure range’ because that changes its interpretation. Better to use one plain sentence explaining the estimating procedure and then link to a fuller statistical guide.
A short analogy can help if its boundaries are explained. Calling a cell membrane a ‘gatekeeper’ conveys selective exchange but may suggest conscious decisions. A good analogy includes the sentence ‘The likeness stops here.’ Readers learn a durable mechanism when they can explain both what the comparison illuminates and what it leaves out.
17. A story can illuminate an idea but cannot replace the data
Stories provide people, places and stakes. An anecdote can reveal why a research problem matters. It does not give a reliable estimate of frequency, causation or average effect without a suitable study.
Use an illustrative character to open the door, then clearly separate their example from representative evidence. Label invented or composite cases as teaching examples, never as observed field reports.
Anecdotes are entry points, not denominators
A student’s story about improving after a new study routine can inspire a useful article. It cannot determine how many students would improve or whether the routine was the cause. Likewise, one striking picture of a damaged building cannot measure the risk facing every building in a city.
Use human stories to explain why a question matters, then connect to systematic evidence for general claims. When no reliable population evidence exists, say what remains unknown. Audiences can appreciate a genuine anecdote without being asked to confuse it with a controlled study.
18. Worked example 1: the headline that halves the risk
Invented teaching dataset: group A had 20 events among 10,000 people, and group B had 10 events among 10,000. A 50% relative reduction is correct because the event count per equal denominator halved.
But the absolute difference is 10 events per 10,000 people, or 0.1 percentage points. The responsible headline gives both: ‘From 20 in 10,000 to 10 in 10,000 in this study,’ with the study design and uncertainty described.
19. Worked example 2: separate measurement from inference
Invented classroom experiment: a lamp causes a sensor voltage to rise from 1.2 V to 1.8 V. The observation is a 0.6 V increase in sensor output. Without a calibration and controlled comparison, the data do not justify ‘light intensity increased by 50%’ or ‘the room became safer.’
The repair is simple: name the instrument and units, describe what was changed, check calibration and determine whether the sensor output corresponds linearly to the claimed quantity.
20. Worked example 3: a beautiful graph with an invisible baseline
Invented scores: one group averages 84 and another 86 out of 100. A chart beginning its vertical axis at 83 makes the second bar look three times the first bar’s height above the baseline.
For ordinary bar charts, a zero baseline better reflects magnitude. Where a truncated scale is justified for another type of plot, disclose it visibly. Also report the sample size and uncertainty before claiming a meaningful difference.
21. Worked example 4: two audiences, one evidence base
Invented air-quality project: a research team measures particle concentrations at a set of outdoor sensors. A technical report should explain sensor accuracy, calibration, siting and temporal coverage. A neighbourhood notice should say which areas were monitored, what was found and what residents can reasonably do.
Both outputs share the same evidence and limitations. Simplifying the presentation may remove detail; it must not change a local measurement into a national conclusion.
22. Pre-bunk misinformation without repeating it as fact
Some misleading claims exploit fear, identity, confusion between association and causation, or an out-of-context graphic. Effective corrections supply the accurate account, explain the critical reasoning step and provide a trustworthy path to check the source.
Repeatedly displaying a false claim in a sensational headline can amplify it. Lead with the verified information, and distinguish between genuine scientific debate and unsupported assertions.
Correct false information with an explanatory replacement
When addressing a misleading graph or claim, focus on the accurate information and the specific broken reasoning. If the error is an unstated denominator, supply the denominator. If the error is correlation described as causation, explain the missing research design. If the error is an old result portrayed as current, give the study’s date and the relevant subsequent evidence.
Corrections should avoid unnecessarily amplifying provocative slogans. A respectful explanation often works better than mocking the people who believed a claim, because embarrassment can make honest questions harder to ask. The aim is transferable reasoning, not a public scorecard of who was wrong.
23. Communicating across languages, cultures and educational levels
An accessible explanation is more than translating nouns. Familiar units, local examples, cultural context, literacy levels and the possibility of asking questions influence whether a message can be understood and used.
An international vocabulary guide may need examples in different languages, but its scientific terms and numerical meaning must remain consistent. Invite readers from the intended audience to test the message before distribution.
24. Evaluate whether readers learned or could act
Page views and likes are measures of exposure or engagement, not direct evidence of comprehension. To test communication, ask readers to describe the main claim, restate its limitations and apply it in a changed situation.
The CDC Clear Communication Index and the National Academies’ research agenda both emphasise systematic thinking about audiences and communication quality. Pair quality checks with real audience feedback rather than assuming a polished design proves effectiveness.
Measure understanding with a changed-context question
Suppose readers are shown a claim that a treatment reduces an event rate from 4% to 2%. After explaining absolute and relative difference, give them a second case: an event rate changing from 0.4% to 0.2%. Can they recognise that the relative reduction remains 50% while the absolute reduction shrinks to 0.2 percentage points?
This test goes beyond asking whether readers liked an article. It reveals whether they can apply the idea with new numbers and notice when a persuasive headline hides scale. Science communication succeeds when readers can reason independently after the author leaves the room.
25. Frequently asked questions and the connected reading library
Should every article include a technical method section? It should provide access to sources and assumptions, even if the main reader-facing answer is short. Can an AI summary count as a primary scientific source? No. Its claims need checking against the underlying evidence.
Is uncertainty a sign science failed? No. Explicit limits are part of responsible inference. How do educators teach better science communication? Ask learners to take one claim through a source check, numerical rewrite, visual audit and evidence-based explanation.
Connected Gold Standard specialist directory
Evidence and thinking
- The Gold Standard of Evidence-Based Thinking
- The Gold Standard of Analytical Thinking
- The Gold Standard of Biostatistics
- Gold Standard Case Study 209 | Standard Deviation or Confidence Interval?
Science and measurement
- The Gold Standard of Signal Processing
- The Gold Standard of Astronomy
- The Gold Standard of Microscopy
- The Gold Standard of Bioinformatics
- Gold Standard Case Study 210 | Compostable Cups and a Rejected Waste Process
Learning and public life
- The Gold Standard of Education | Learning Transfers into Life
- How Education Works | Science After a School Move
- The Gold Standard of School Citizen Science Data Contributions
- How Singapore Education Works | Sustainability and Global Decisions
Primary practice standards and sources
- National Academies — Communicating Science Effectively: A Research Agenda
- National Academies — Challenges in communicating uncertainty
- CDC — The Clear Communication Index
- CDC — How to use the Clear Communication Index
- CDC — Test messages with the intended audience
Before publishing, test these five questions
- Is the main answer visible in the first paragraph?
- Can readers identify the original evidence and its limits?
- Are denominators, units, axes and risk measures labeled?
- Can the intended audience repeat the correct conclusion?
- Does the material offer a useful next step without overstating the science?
Return to The Gold Standard or explore Evidence-Based Thinking to continue checking how a claim becomes knowledge.
For the larger connected reading journey, follow How X Works and The Gold Standard of Education. Good explanation is transferable: it helps readers identify the claim, evidence, uncertainty and next sensible action in unfamiliar subjects.