Home Blog Page 6

What Should You Remove Before Uploading a Document to AI?

0
Blank documents pass through a glass privacy gate while identity symbols remain in a separate vault.

An AI tool may need a paragraph to improve your wording. It probably does not need the passport number, signature, private address and entire email thread attached to that paragraph.

The useful habit: provide the smallest amount of information that can support the task. Treat an upload as sharing information with a service, rather than as a purely local edit.

First decide whether the upload is appropriate

Check your organisation’s rules and the service’s actual data controls. An approved workplace environment can differ from a personal account. Retention, access, model-improvement settings and connected tools vary. A paid subscription alone does not establish permission to upload someone else’s confidential material.

If you cannot establish that the service is suitable, use a fictional example or ask for a general structure instead. You can request a complaint-letter outline without uploading the original case file.

Separate the task from the identity

Information Possible treatment when unnecessary
Names and contact details Replace consistently with Person A, Organisation B and placeholders
Identity and account numbers Remove entirely rather than expose full identifiers
Passwords, recovery codes and API keys Keep out of the upload
Exact dates, locations and amounts Generalise if precision is irrelevant to the task
Signatures, faces and document images Exclude or properly redact when not needed
Private comments and revision history Inspect and remove from the sharing copy

For a fictional editing task, “Person A worked at Organisation B from 2021 to 2024” may preserve the sentence’s structure without exposing the original employer and identity. Keep the replacement key locally if you need to restore names later.

Look beyond the main page

Check headers, footers, filenames, embedded images, spreadsheet tabs, comments, tracked changes and document properties. A pasted screenshot may reveal account names or notifications outside the area you intended to discuss.

A black rectangle placed over text is not necessarily a real redaction. The underlying text may remain selectable or extractable. Use an appropriate redaction method, then inspect the exported sharing copy. Do not upload the unredacted original merely to ask an AI tool to remove its secrets.

Removing names may not make a file anonymous

A rare role, precise location and distinctive event can identify someone together. Consider a fictional personnel note describing “the only night-shift supervisor at a named branch on a particular date.” Replacing the person’s name may do little to prevent recognition.

Ask whether the combination of remaining details can point back to a person. Use broader categories or a synthetic example when the task allows. Pseudonyms reduce direct identification but are not a guarantee of anonymity.

Preserve the information the task needs

Removing too much can also produce a misleading answer. If you want help checking arithmetic, preserve the relevant quantities and units. If you want help comparing contract clauses, removing the conditions that change their meaning can make the comparison unreliable.

The solution is to narrow the task. Ask for wording feedback on one sanitised section, or ask what questions to take to a qualified professional. Avoid presenting a stripped-down file as a complete basis for a consequential decision.

A practical pre-upload workflow

  1. State the task. Write one sentence describing what help you need.
  2. Create a separate sharing copy. Keep the original protected and unchanged.
  3. Remove unnecessary details. Use consistent placeholders and inspect hidden content.
  4. Review the final file. Open that exact copy and check text, images and properties.
  5. Check the destination. Verify the service, account, sharing scope and applicable workplace policy.
  6. Review the output. Check for invented details and restore private information locally only where appropriate.

If sensitive information was uploaded accidentally

Stop further sharing. Follow the provider’s deletion and support process and your organisation’s incident procedure where applicable. If an exposed item is a credential, arrange its revocation or rotation through the relevant service; deleting a chat is not a substitute for invalidating a secret.

Note to keep: AI document privacy begins before the upload. A useful prompt explains the task without automatically including every detail in the source file.

Related: check an AI answer before relying on it and understand your digital footprint.

Sources

Your Digital Footprint: What You Leave Behind Without Posting

0
Glowing glass footprints containing photos, parcels and location symbols lead toward an archive beside the sea.

You can avoid posting photographs and still leave a digital footprint. Browsing a shop, opening an account, using a map or sending a file can create information about your activity.

The useful habit: think beyond “What have I posted?” Ask “What information does this action create, who receives it and how long might it remain?”

Three layers to recognise

Layer Example Question
Information you provide A delivery address or account profile Is every requested field necessary?
Activity recorded around your use Pages visited or sign-in times What activity controls and retention settings exist?
Information inferred from patterns Possible interests suggested by repeated visits Can you inspect or limit personalisation?

A digital footprint is not necessarily public. A provider may hold records that other users cannot see. Conversely, a public profile may expose details that you assumed were visible only to friends.

Browsing creates records without a public post

Websites and apps can use cookies, pixels and other technologies to record activity. Some support ordinary functions such as keeping you signed in; others support analytics or advertising. The exact behaviour depends on the service, browser settings and available privacy controls.

Imagine looking at hiking equipment over several evenings. You have not announced a holiday, but a service may infer an interest in outdoor travel. That inference can be wrong. A category assigned to you is not a reliable statement about your intentions.

Files can say more than their visible content

A photo may contain metadata such as capture time, device information or location. A document may contain author details, comments or revision information. What survives sharing depends on the file, export process and platform.

Before sharing a sensitive file, inspect its properties and the actual exported copy. Do not assume every platform removes location information. Also inspect the visible content: a street sign, reflected screen or boarding document can reveal information even after metadata is removed.

Other people contribute too

Tagged photographs, group posts, public directories and forwarded messages can add to what is discoverable about you. Ask contacts before publishing their private information and use available tag-review controls for your own accounts.

For example, a picture of a celebration might inadvertently expose an address on an envelope. Cropping or retaking that photograph before sharing is easier than trying to retrieve every copy afterward.

Run a focused privacy review

  1. Check the public view. Search for your name and review what your profiles expose when you are not signed in. Results can vary by location and search engine.
  2. Review active accounts. Remove unnecessary profile details and inspect audience, activity and advertising settings.
  3. Review device access. Check app permissions and browser site permissions.
  4. Inspect older sharing. Look at public folders, old posts, tagged photographs and documents shared by link.
  5. Retire unused services deliberately. Save anything you need, cancel relevant subscriptions separately, and use the service’s account-deletion process where appropriate.

Understand what deletion changes

Deleting browser history affects that browser’s local record; it does not automatically delete records held by a website. Removing a public post may reduce its visibility without retrieving screenshots or copies that other people already made.

Deletion and retention options differ by provider and jurisdiction. Read the service’s current process instead of assuming a single button clears everything. Where available, use account export tools to understand your stored information before deciding what to remove.

Make the review manageable

You do not have to erase your online life. Start with high-value information: home address, personal contact details, precise location, sensitive files and accounts that could unlock other services.

A practical monthly task is to review one account and one shared folder. A practical daily habit is to pause before submitting optional details. Small reductions in unnecessary collection can be more sustainable than a large cleanup that you never repeat.

Note to keep: a private account, a private browser window and a clean public profile solve different problems. Our private browsing and VPN comparison explains another part of that picture.

Sources

Private Browsing, VPNs and Anonymity: What Each One Does

0
Three miniature glass, frosted and enclosed tunnels illustrate different layers of online privacy.

A private window and a VPN protect different parts of your online activity. Combining them can be useful, but neither makes every action anonymous.

Start here: private from whom? Another person using your browser, the network operator and the website you visit are three different observers.

Private browsing mainly changes what the browser keeps

In Chrome Incognito, the browser does not retain the session’s browsing history and site data after the session ends. That helps when you do not want a search to remain in the ordinary browser history on that device.

It does not make you invisible to websites or organisations operating your network. Downloads remain on the device, and bookmarks you create can remain. Closing one tab may not end a private session if other private windows are still open.

Imagine planning a surprise gift on a shared home computer. A private session can reduce traces in that browser. It does not remove an order confirmation sent to your shared email account or hide the purchase from a joint shopping account.

A VPN changes the network route

A VPN typically creates an encrypted connection between your device and a VPN server. Traffic routed through that server reaches websites from the server’s network address rather than your ordinary public IP address.

This changes who you trust. The local network sees a connection to the VPN; the VPN operator handles the routed traffic. HTTPS still matters for protecting communication between your browser and the destination website.

Coverage depends on the product and configuration. A browser-only proxy may cover that browser rather than every app. Split tunnelling, exclusions and disconnections can also affect which traffic uses a VPN. Read the actual scope instead of relying on a shield icon.

Compare the questions each tool answers

Your goal What to consider
Avoid ordinary browser history on a shared device Private browsing, closing the session, and downloaded files
Change the network route and public IP seen by websites A suitable VPN and its traffic coverage
Keep a website from knowing your account identity Logging in still identifies the account, with or without a VPN
Protect information sent to a website HTTPS and the website’s own data practices
Stop a malicious download or phishing message Careful verification, updates and appropriate security controls

An IP address is only one clue

Websites can recognise account logins and may use cookies or other technical signals. Information you submit can identify you directly. If you enter your name and delivery address, hiding your usual IP address does not erase those details.

Similarly, a VPN cannot make a post anonymous if the post itself reveals your identity. Privacy depends on the whole activity, including accounts, files, permissions and people who receive the information.

Use a simple three-part check

  1. Name the observer. Are you concerned about another household member, a network administrator or a website?
  2. Name the information. Is it your browser history, IP address, account name or the content you upload?
  3. Match the protection. Check what the tool actually covers and what remains outside it.

For example, signing in to work email through a VPN does not prevent the employer’s email system from recording activity in that account. Private browsing on a managed device also does not override device-management or monitoring controls.

Questions worth asking before installing a VPN

Who operates it? What data does it collect? Which applications does it cover? What happens when the connection drops? Are its claims supported by clear documentation and relevant independent scrutiny? A free price or an expensive subscription is not, by itself, evidence of trustworthy handling.

You do not need to buy a tool before defining the problem. Sometimes the useful action is closing a private session, reviewing account sharing or restricting an app’s access. See our app permissions guide for that separate layer.

Note to keep: privacy tools have boundaries. Understanding those boundaries is more useful than expecting a single setting to hide everything.

Sources

What App Permissions Actually Allow—and How to Review Them

0
A ceramic app building has separate doors for a camera, microphone and location garden, illustrating selective permissions.

A permission prompt is easy to dismiss when you are trying to finish something. Weeks later, the access you granted may still be available even though you no longer use the feature.

The useful habit: connect each permission to a specific job. Ask what the app needs, when it needs it and whether a narrower option would work.

Permission is access to a capability

Camera access enables camera features. Microphone access supports audio capture. Location access can help with navigation or nearby results. Contacts and photo access concern different information again. A permission is not a promise that the app will use that capability only in ways you personally expect.

Permissions also do not describe every kind of data handling. Information you type, upload or send to an online service may be processed under that service’s settings and policies. Denying location permission does not make your internet connection anonymous.

Match access to the task

Feature you want Question to ask
Scan a document Does the app need the camera only while scanning?
Attach one photograph Can you select that image instead of exposing a larger library?
Find nearby shops Would approximate location or a typed area be sufficient?
Record a voice message Is microphone access connected to that recording action?
Navigate during a journey Does the feature need access while the screen is off?

These are decision questions, not universal settings. Available choices depend on the operating system, version and permission category. Some legitimate features need continuing access; others work with a one-time grant.

Review by app and by permission

Reviewing one app answers “What can this app access?” Reviewing a category answers “Which apps can access my microphone or location?” Both views are useful because an old app may be easy to forget.

On Android, use Settings to find the app’s Permissions page or search for Permission manager. Supported devices also provide a Privacy dashboard showing recent permission use. Menu names vary by manufacturer and software version.

On iPhone, review the relevant categories in Settings under Privacy & Security. Apple’s App Privacy Report, where available and enabled, provides information about permission use and network activity. Consult the current official instructions for your device.

Make one deliberate change at a time

Choose an app you rarely use. Identify a permission with no clear role in the features you need. Reduce or deny it, then try the relevant feature again. If the app asks later, you can make the decision with the task in front of you.

For example, imagine a fictional travel app that you installed for one holiday. You still read saved bookings, but no longer need nearby attractions. Reviewing its location access is more useful than reflexively allowing every request because the app was once helpful.

A feature failing after a permission change does not automatically prove the app is malicious. Equally, a feature working without access is a useful sign that the broader permission may not be necessary for your use.

Read activity without jumping to conclusions

A record of camera or microphone access is a reason to understand the activity. It is not, by itself, proof of spying. Compare the timing with calls, recordings or other features you used. If the explanation remains unclear, review the developer’s documentation and consider removing an app you do not trust.

Some permissions sit outside the ordinary camera-and-location list. Accessibility services, notification access, device administration and special file access can have substantial implications. Read the operating system’s description before enabling them; a generic “required for performance” message is not a useful explanation.

Revoking access is not retroactive deletion

Changing a permission controls future access through that permission. It does not necessarily delete information previously uploaded to a provider. Check the app’s account controls and data-deletion process separately. Removing an app may also leave an online account active.

A ten-minute review

  1. List the apps you no longer recognise or use.
  2. Review camera, microphone, location, contacts and photo access.
  3. Choose narrower access where it still supports the feature.
  4. Inspect special access requests individually.
  5. Remove unused apps after checking whether important local information needs saving.
  6. Repeat after installing a major new app or changing phones.

Note to keep: permission choices work best when they reflect today’s needs rather than every feature you might someday use.

Related: prepare for a lost phone and understand cloud sync versus backup.

Sources

What Does “Scientifically Proven” Actually Mean?

0
Three glass testing chambers examine matching crystals in different conditions, with light converging into a prism.

“Scientifically proven” sounds like the end of a discussion. On a product page, in a headline or beneath a viral video, it can suggest that someone has already done all the checking for you.

But the phrase alone tells you very little. What was tested? Compared with what? In which people or conditions? And how closely does the advertised promise match the result?

The useful habit: replace “Is it proven?” with “How strong is the evidence for this exact claim?” Science can support very confident conclusions, while still describing their limits and remaining open to better evidence.

Start by making the claim specific

A broad promise is difficult to evaluate. “Improves learning” might mean a higher score immediately after a lesson, better recall a month later, or an improvement in everyday performance. These are different outcomes.

Imagine a fictional study of a study-planning app. Forty university students use it for two weeks and report feeling more organised. That result would not, by itself, establish that the app raises examination grades for all students.

Write down: the intervention, the comparison, the participants, the measured outcome and the duration. If the claim leaves these out, find the underlying study before accepting the slogan.

Different studies answer different questions

Laboratory experiments can explore mechanisms. Observational studies can identify patterns in real-world data. Randomised trials can help test whether an intervention makes a difference by assigning participants to comparison groups by chance. Systematic reviews assess a defined body of research; some also combine numerical findings in a meta-analysis.

These labels are starting points. The design must fit the question, and the execution matters. A review does not automatically repair weak studies. A trial with substantial missing data or an unsuitable comparison can still leave important doubts.

Evidence described Useful question to ask
A laboratory finding Has the relevant real-world outcome been tested?
An association in a survey Could other differences explain the pattern?
A randomised trial Was the comparison fair, and were the outcomes relevant?
A research review How were studies selected and their limitations assessed?

For the distinction between a pattern and an explanation, see our guide to correlation and causation.

Read the size of the result

A result can be statistically significant without being large enough to matter in practice. Statistical tests help assess observations relative to a model and sampling variation; a significance label is not a measurement of usefulness or a guarantee that a claim is true.

Suppose a fictional typing tool improves average speed from 50 words per minute to 51. The gain is one word per minute, or 2%. Calling it a “scientifically demonstrated improvement” would still leave you needing to know the uncertainty, the cost and whether that difference matters for your task.

Check: How big was the change in understandable units? What range of results remains compatible with the evidence? Were benefits accompanied by disadvantages?

Ask who and what the evidence applies to

A finding has a setting. A test on adults using one version of a tool does not automatically establish the same effect in children, in a different language or after a major redesign.

In our fictional app example, imagine the volunteers already enjoy planning their studies. Their experience may be informative, but it does not settle how students who dislike planning would respond. Equally, a very large sample does not automatically correct a biased recruitment method.

Look for the population, setting, duration and exact product or procedure studied. A claim about an ingredient, component or earlier version needs a further argument before it becomes a claim about the finished product you are considering.

One study is part of a larger picture

Confidence generally grows when different investigators and appropriate methods produce compatible findings. Repeating work with new data can reveal whether a result holds up and where its boundaries lie.

Disagreement does not automatically mean fraud or that science is useless. Studies may examine different groups, use different measurements or contain different limitations. A single unsuccessful replication does not conclusively refute a finding, just as one successful replication does not settle every question.

Check: Is the claim based on an isolated result, or does it reflect the wider evidence? Does a recent review explain why studies agree or disagree?

Peer review and funding are clues to examine

Peer review provides expert scrutiny before publication in many journals. It is a useful quality-control step, but it cannot guarantee that every method, calculation and conclusion is correct. A preprint may be useful early research, while still awaiting that journal review process.

Also read funding and conflict-of-interest disclosures. Commercial funding does not automatically make a result false; independent funding does not make it flawless. The practical questions are whether the methods are transparent, the comparison is fair and the reporting presents the full relevant results.

A five-minute claim check

  1. Locate the source. Ask for the study, not just a testimonial or a screenshot of a headline.
  2. Match the promise. Compare the advertised outcome with what researchers actually measured.
  3. Read the methods. Identify participants, comparison, duration and important sources of bias.
  4. Read the numbers. Note the effect size, uncertainty, missing data and relevant disadvantages.
  5. Check the wider evidence. Look for independent work and a careful review of the topic.

A useful note might read: “One short study suggests better self-reported organisation among university volunteers. It did not measure long-term grades.” That is more informative than repeating “proven” or dismissing the research outright.

For claims circulating through social media, use the same source-tracing steps in our online claim checklist.

Common questions

Does “clinically tested” mean something works?

The phrase says little about the outcome on its own. Ask what the test found, how it was designed and whether the evidence supports the specific benefit being claimed.

Should I wait for absolute certainty?

Decisions often involve uncertainty. Match the strength of evidence you require to the consequences of being wrong. Trying a free planning template and changing a medical treatment involve very different stakes; discuss treatment changes with a qualified clinician.

Can strong scientific conclusions change?

Yes. New evidence may refine a conclusion or reveal limits. That possibility does not make all claims equally credible. Assess the quality and consistency of the evidence available.

Sources and further reading

The examples in this article are fictional. The explanation draws on these primary educational and professional sources:

How Charts Mislead Without Using False Numbers

0
Small stone columns cast exaggerated shadows beside a magnifying glass, illustrating misleading visual comparisons.

A chart can contain perfectly accurate numbers and still leave you with the wrong impression. A small change can look enormous. A short recovery can look like a lasting boom. Two unrelated lines can appear to move together.

The useful habit: read the scale, time period and measurement before accepting the visual story. Misleading charts can result from poor design as well as deliberate persuasion; a suspicious presentation alone does not establish intent.

1. A cropped bar chart can exaggerate a small difference

Imagine two fictional monthly scores: 80 and 84. The second score is 4 points higher, a 5% increase relative to 80.

What you compare First score Second score
Actual value 80 84
Visible bar length if the axis begins at 78 2 units 6 units

On that cropped chart, one bar would be three times the visible length of the other. The labels could be correct while the picture suggests a much larger proportional difference.

Check: Where do the bars begin? The Office for National Statistics recommends a zero baseline for bar charts because length represents quantity. A dot plot can show small differences without implying that truncated lengths represent the complete values.

2. A line chart does not always need to start at zero

The rule changes with the chart type. A line chart can legitimately use a narrower vertical range to make small changes visible. Think of a measurement fluctuating between 98 and 102: a zero-based chart might hide the variation you need to examine.

The question is whether the scale is clearly labelled and appropriate to the question. Read the actual size of the change before describing it as a surge or collapse. Also check whether two charts being compared use the same scale.

Note to keep: “Non-zero axis” is a reason to inspect a line chart, not an automatic verdict that it is dishonest.

3. The starting date can change the story

Consider a fictional index that starts a year at 100, reaches 130 in October and finishes at 120 in December. Across the year it rises 20%. From October to December it falls about 7.7%.

Both descriptions are mathematically valid. They answer different questions. A chart showing only the final months could support a gloomy headline while hiding the longer rise; a chart starting at a temporary low could make an ordinary recovery look exceptional.

Check: Why does the chart begin on this date? Look for a longer series and consider whether seasonal patterns or an unusual starting point affect the interpretation. A longer period is useful context, though it does not make a relevant short-term change disappear.

4. Similar-looking lines may use different scales

A chart with a left-hand axis and a right-hand axis can show two measurements with different units. Independently adjusting those scales can make their movements look closely aligned.

Matching heights do not mean matching quantities. Matching shapes do not establish that one variable causes the other. Read each axis separately and ask whether the comparison has a sensible explanation. Separate, clearly labelled charts may be easier to interpret.

Related note: why correlation does not establish causation.

5. Bigger pictures can multiply the apparent difference

Suppose an infographic represents one quantity with a small circle and another with a circle twice as wide. If the circle’s height also doubles, its area becomes four times as large. That is a poor area representation of a quantity that merely doubled.

Decorative icons, perspective and three-dimensional shapes can make size comparisons harder. Ask whether the quantity is represented by height, area or volume. When the picture is ambiguous, use the labelled values or the accompanying data table.

6. Counts and rates answer different questions

Imagine two fictional towns. Town A records 100 incidents among 100,000 residents; Town B records 40 among 20,000 residents. A count chart puts Town A higher. A population-adjusted comparison gives 100 incidents per 100,000 residents in A and 200 in B.

The counts help describe the total workload. The rates help compare incidents relative to population. Neither measure answers every question, and meaningful comparisons also need consistent definitions and reporting periods.

Check: What is the denominator? Is the chart showing totals, percentages, per-person figures or something else? For percentage claims, also distinguish percentages from percentage points.

7. Read what the chart leaves out

  • Time spacing: Equal gaps on a time axis should represent equal intervals. Missing months need to be apparent.
  • Scale type: On a logarithmic axis, equal steps represent equal ratios rather than equal numerical additions. That can be useful, but it changes how you read the slope.
  • Measurement: Cumulative totals and new events per month tell different stories. A cumulative line can keep rising even as monthly additions slow.
  • Uncertainty: A precise-looking dot may be an estimate. Look for intervals and methodological notes before treating small differences as settled facts.
  • Coverage: Missing groups, changed definitions and incomplete reporting can limit a comparison even when every plotted value is correct.

A quick checklist before you share a chart

  1. What exactly is being measured, and in which units?
  2. Where do the axes start, and are their scales consistent?
  3. Why was this time period selected?
  4. Are the figures counts, rates, percentages or cumulative totals?
  5. Are important groups, dates or uncertainty missing?
  6. Can you find the original source and underlying figures?
  7. Does the headline claim more than the chart actually demonstrates?

If the source cannot be traced, pause before sharing. Our online claim-checking guide explains how to follow a claim back to its evidence.

Common questions

Does a misleading chart mean the underlying data is fake?

No. The problem may be the selection, scale or presentation. Check the source data separately from the visual design.

Should I reject every chart with two vertical axes?

No. Read both units and scales carefully. Such charts can be difficult to interpret, and an apparent relationship needs evidence beyond visual alignment.

What is the simplest way to check an alarming chart?

Write down its first and last values, the units and the dates. Calculate the change yourself, then look for a broader time window and the original data.

Sources and further reading

This explainer uses original fictional examples. Chart-design guidance was checked against the following primary sources:

Correlation vs Causation: Why Two Things Moving Together Proves Less Than You Think

0
Two miniature trains on separate parallel tracks beneath shared clockwork, illustrating a pattern without one train pulling the other.

Correlation describes a pattern between measurements. Causation means that changing one thing produces a change in another. Two things can move together because one influences the other, because something else influences both, or because the apparent connection is misleading.

The difference matters whenever someone turns “these things are linked” into “do this and you will get that result.” A pattern can be useful evidence without being a complete explanation.

The question that separates them

Imagine a shop introduces a new window display and sales rise the following week. The owner says the display caused the increase.

That may be true. But perhaps the same week included a holiday, a local event or a discount. A before-and-after comparison records the change; it does not automatically isolate what produced it.

The causal question is: “What would sales have been during that same period without the new display?” You cannot directly observe both versions of the same week, so a good investigation needs a credible comparison.

Three explanations to consider before celebrating

Possible explanation Question to ask
One thing influences the other. What evidence shows the proposed cause changed the outcome?
A shared factor influences both. What else changed at the same time?
The pattern reflects chance or selection. Does it remain when we examine more suitable data?

The shop example is fictional. Its purpose is to keep several explanations open until the evidence distinguishes them.

A shared cause can make unrelated actions look connected

Suppose a café records more cold-drink sales and more outdoor-seat use on warm days. Buying a cold drink is not necessarily what makes people sit outside. Temperature can influence both choices.

A factor affecting the relationship you are studying may be called a confounder. Simply listing one is not proof that it explains the whole pattern, but it tells you what needs investigation.

The Australian Bureau of Statistics explains the distinction between correlation and cause and effect. It cautions that a correlation alone cannot establish that one measurement caused another to change.

The direction can run the other way

Imagine a company notices that its busiest teams send the most messages. It decides that sending more messages will make every team busier and more productive.

Perhaps the workload creates the messages: more tasks require more coordination. Increasing messages without increasing useful work may simply create interruptions.

Ask what happened first and what mechanism could connect the two. Timing is necessary for a proposed cause to produce a later outcome, but timing alone is not sufficient evidence. “After” does not automatically mean “because of.”

The dataset can make a relationship look stronger than it is

Look at how the observations were selected. A report that includes only successful customers may miss everyone who tried the same approach and saw no benefit. A short period can also hide a longer trend or a seasonal cycle.

For example, a shop’s weekly sales and advertising spending might both rise during a festive season. A graph of those weeks can look persuasive while leaving the season itself out of the story.

Ask whether the pattern survives different reasonable time periods, whether unusual observations dominate it, and whether the measurements mean the same thing across the comparison.

Why an impressive graph cannot finish the argument

A scatter plot can reveal patterns, unusual observations and possible relationships. NIST’s statistical handbook makes clear that a scatter plot cannot by itself prove cause and effect.

A correlation coefficient is also a summary, not a mechanism. Common measures such as Pearson’s correlation describe linear association. A value close to zero does not rule out every possible curved relationship.

Read the axes and examine the underlying observations where possible. A chart can help you ask better questions; it cannot supply missing study design.

What makes causal evidence stronger?

A well-designed experiment deliberately changes an input and compares outcomes. Random assignment can help make comparison groups similar on average, reducing systematic differences in who receives the change.

JMP’s statistics guide discusses controlled experiments and randomisation as ways to investigate causal relationships. The design, measurement and execution still matter; the label “experiment” is not a guarantee.

Sometimes experiments are impractical or unethical. Researchers can use observational designs and other evidence to investigate causal explanations, but their assumptions and limitations need scrutiny. Correlation alone being insufficient does not mean causation is unknowable.

A practical business example

Return to the shop display. A useful small investigation could compare a planned change across comparable locations or periods while recording prices, promotions and customer traffic. If feasible, randomise which eligible locations receive the display first.

Decide in advance what outcome matters: total sales, profit, conversion rate or something else. A display that increases revenue while sharply reducing margin may not deliver the benefit the owner wanted.

A small trial gives evidence for the conditions tested. Record uncertainty, check for other changes and avoid generalising one encouraging week to every future location.

Five questions for the next “X causes Y” headline

  1. Was this an observation, an experiment or a review of several studies?
  2. Did the proposed cause come before the outcome?
  3. Could another factor influence both?
  4. What comparison supports the claim?
  5. How far does the evidence justify applying the result?

If the evidence only establishes an association, use wording such as “was associated with.” If a study provides causal evidence, describe its actual scope rather than stretching its conclusion.

Common questions

Does correlation mean the relationship is useless?

No. It can help identify patterns, support prediction and guide further investigation. Prediction and explaining what will happen after an intervention are different tasks.

Does a strong correlation prove causation?

No. A strong pattern can still reflect a shared influence or another explanation.

Can a real cause have a weak overall correlation?

Yes. Other influences, measurement problems or the shape of the relationship can obscure a simple overall pattern.

Related: Use our root-cause investigation guide to test explanations, and percentage-points explainer to keep numerical comparisons clear.

A Real Photo Can Still Tell a False Story: Understanding Context

0
Conceptual illustration of a framed red boat image with a wider view revealing a tabletop model basin and stage light.

A photograph can be genuine while the story attached to it is false. The image may show a real scene, but the caption can supply the wrong place, date, identity or explanation. Checking whether a picture looks edited answers only part of the question.

Before treating a photo as evidence, separate what you can actually see from what the surrounding post asks you to believe.

Two claims are hiding inside one post

Imagine a photograph of a red boat on water. You can see the boat, reflections and part of its surroundings. A caption says it shows a rescue operation in a particular city this morning.

The pixels alone do not establish that city, that morning or that purpose. The caption adds those claims. They need evidence of their own.

This is an invented example. The featured image is a conceptual illustration of a model scene and a tight frame, rather than a photograph of a real news event.

What you can observe What still needs checking
A boat on water Where the scene is located.
A visible moment When it happened.
People or objects in the frame Who they are and what they are doing.
A selected view What happened outside the frame or before and after it.

Four ways a real photo can mislead

1. An old photograph receives a new date

A picture from a previous storm may circulate during a new one. It can be a genuine record of damage without being evidence of today’s conditions. Check the event date as well as the date on the post.

2. The location changes in the caption

A scene from one city may be described as another. Familiar architecture or weather can make a caption seem plausible, but resemblance is not identification.

3. A crop removes a decisive detail

A tight frame may hide a sign, a barrier, a stage or the rest of a queue. Cropping is normal in photography and is not automatically dishonest. The question is whether the crop supports an interpretation that the wider scene contradicts.

4. The image is used to claim more than it shows

A photograph of a crowded entrance does not establish the total attendance at an event. A single expression does not reliably explain a person’s feelings or motives. A still image records a selected moment, not a complete investigation.

Start with the post’s precise claim

Write down the asserted place, time, event and source. Then ask which details are visible and which come only from the caption. This prevents the caption from quietly becoming part of your visual observation.

If a post says “today,” translate that into an actual date. Screenshots can circulate long after the original post, making relative words such as “yesterday” especially confusing.

Search for earlier appearances

Reverse image search uses an image to find matching or visually similar material. It can uncover earlier publication and different captions. AFP’s image-search guide explains how this can expose old images and images reused outside their original context.

Use the whole image first. If a border, overlaid text or collage dominates the results, try searching the relevant picture area separately. Compare distinctive details to confirm you found the same image, rather than another similar scene.

Open the result pages. A thumbnail or search snippet can be incomplete. Look for the photographer, agency, original caption and surrounding report, and check whether later corrections change the account.

Understand what “first seen” means

Google’s About this image guidance says the feature may show when Google first encountered a similar version and other pages using similar images. Availability varies by region.

That is a clue about the image’s online history. It is not necessarily the date the camera captured it or the first time anyone published it.

Similarly, a search that finds no match does not prove an image is new, real, fabricated or AI-generated. Search engines do not contain every image. Treat a failed search as an unresolved lead.

Use several clues that can be checked independently

A useful investigation might combine an original credited caption, reliable reporting of the event and recognisable public features that agree with the claimed location. One vague resemblance should not carry the whole conclusion.

Metadata can add information, but missing metadata does not prove deception. Files can lose it during sharing, and metadata can be changed. Google also cautions that image-credit and digital-source metadata may be modified. Seek evidence that corroborates it.

Keep the investigation proportionate. Avoid identifying private individuals or exposing sensitive locations simply to settle a social-media argument.

A worked example: the “today” caption

Suppose a post claims a waterfront photo shows an event this morning. A reverse search leads to a reliable report containing the same photograph from two years earlier.

  1. Compare distinctive objects and framing to establish that it is the same photograph.
  2. Read the older report and its image credit.
  3. Check that the older publication date belongs to the actual page and image use.
  4. Explain the narrow finding: this photo was already published earlier, so it cannot by itself document this morning’s event.

That finding does not prove nothing happened today. It challenges the photograph’s use as current evidence. Keep the correction as precise as the evidence allows.

A short checklist before sharing

  • What does the image show without its caption?
  • Who supplied the place, date and explanation?
  • Can I find an earlier or wider version?
  • Does the original caption support the current claim?
  • Which parts remain uncertain?

If important details remain unresolved, leave the image unshared or explain the uncertainty without presenting the claim as confirmed.

Common questions

Does a genuine photograph prove the caption is true?

No. Authentic image content and accurate context are separate questions.

Does an earlier publication prove a current event is false?

It can show that the image is not new. It does not, on its own, settle every claim about the current event.

Can I verify an image just by looking for visual glitches?

No. A misleading caption may need no editing at all. Trace the image’s history and check the attached claims.

Related: Use our online claim-checking guide for the wider verification process, and AI answer checklist when an automated tool supplies the explanation.

How to Check an Online Claim Before You Share It

0
Blank paper messages pass through a magnifying-lens inspection gate in a floating miniature workshop.

Before sharing an online claim, identify exactly what it says, trace it to its source, and check whether independent evidence supports it in the stated context. If the important details remain unclear, leave it unshared.

A message can arrive from someone you trust and still contain an error. Your friend may have forwarded it with good intentions. The useful question is what evidence supports the message, rather than whether the sender meant well.

Pause when the message demands an immediate reaction

“Share before it gets deleted” and “everyone must know” create pressure to act quickly. An urgent tone does not prove a claim is false, but it is a reason to slow down and inspect it.

The United Nations’ information-integrity guidance encourages people to pause before sharing and consult more than one news source. For everyday use, make that pause a specific action: open the source before opening the share menu.

Write the claim as a sentence you can check

A long post may mix facts, opinions, predictions and instructions. Separate them. “A new rule starts tomorrow” is a factual claim. “This rule is unfair” is an opinion. “It will ruin every business” is a prediction requiring a different kind of evidence.

Write down who supposedly did what, where and when. If those details are missing, the claim may not yet be precise enough to verify.

Message says Checkable question
All services are closed tomorrow. Which services, in which location, on what date, according to which announcement?
This photo proves it happened today. Where and when was the image first published, and what event does it actually show?
A study proves this always works. Which study, what did it measure, and how strong was its conclusion?

Follow the message back to its origin

A screenshot of a headline is a clue, not the complete report. Search for a distinctive phrase, find the original article or announcement, and read beyond the headline. Check whether the screenshot omits a date, qualification or correction.

For a service closure, look for the organisation responsible for the service. For a research claim, look for the original study and informed interpretation of its methods. For an unfolding event, distinguish what an eyewitness reports from what anyone has independently established.

Do not treat a logo pasted onto an image as proof of authorship. Where possible, navigate independently to the organisation’s established website or account and find the announcement there.

Leave the page to learn about the source

Reading a website’s own About page can help, but it is not enough. Search for what credible sources say about the publisher, author or organisation. Look for relevant expertise, a visible corrections process and evidence behind the specific claim.

Poynter’s MediaWise guidance describes lateral reading: moving beyond the source itself to investigate its credibility. This is useful when an unfamiliar site looks professional but gives you little reason to trust its reporting.

A familiar publisher can also make mistakes. Evaluate the individual claim rather than accepting or rejecting everything from one outlet automatically.

Check the date, place and missing context

A genuine announcement from last year may be recirculated as new. A real image from another country may be attached to a local story. A quotation can be accurate while the surrounding discussion changes what it means.

Inspect the publication date and any update date. Then check the date of the event itself. These are not always the same. If a post says “tomorrow,” translate that into a calendar date before relying on it.

For an image, reverse image search can help locate earlier appearances. An earlier match is useful evidence of prior circulation. No match does not prove an image is new, authentic or accurately captioned.

Count independent reporting, not copied headlines

Ten sites may repeat one unverified post. Follow their links and attributions. Do they have separate reporting or do they all lead back to the same anonymous message?

Look for evidence that adds something: an original document, a named accountable source, direct reporting or a clear explanation of verification. Repetition alone cannot turn an unsupported statement into a supported one.

A worked example: the forwarded closure notice

Imagine receiving an undated image saying all public libraries in a city will close tomorrow. This is an invented scenario.

  1. Identify the city and the exact day the message refers to.
  2. Search independently for the library service’s official announcement.
  3. Check whether the notice covers every branch or only one building.
  4. Look for a newer update or correction.
  5. If you still cannot establish the closure, do not present it as confirmed.

A useful reply might be: “I haven’t found a current notice confirming this. Please check the library’s official service updates before travelling.” You can help people without declaring an unsupported claim true or false.

Choose an honest outcome

  • Supported: Share the original reliable source with the relevant context.
  • Misleading: Explain the specific missing detail or mismatch.
  • Unresolved: Say what remains unknown and avoid amplifying it.

If you already forwarded an incorrect message, correct it where you shared it. Keep the correction clear, link to sound evidence and avoid repeating the sensational claim unnecessarily. The aim is to repair understanding, not embarrass the original sender.

Common questions

Does a trusted friend make a claim reliable?

No. Trust in the person and evidence for the message are separate questions.

Does “not verified” mean false?

No. It means you have not established it. Preserve that distinction when explaining your conclusion.

Should I share a warning just in case?

Check first. An inaccurate warning can create confusion or direct people away from useful action. For an immediate safety concern, use the relevant official guidance and reporting channels.

Related: Read how to check an AI answer for verifying generated responses and percentage versus percentage points for checking numerical claims.