Send the Correction Back

The correction has been posted. The account owner sends Tess a link and writes that a great many people have seen it.

Tess opens the customer message that began her morning. There is no reply yet. She still does not know whether that customer has seen the correction. The large number beneath the new post cannot answer her small question.

This is a gap between publishing an answer and returning an answer. A correction can be clear, well supported and widely praised while missing a person whose decision is already changing. We should care about the quality of the words. We should also care where they go.

A publisher may find this harder to measure. The original post may have been sent through groups and private messages that the publisher cannot inspect. A person may have copied the image without its link. Some readers may have left the platform. These limits should be stated. They make complete reach difficult to establish; they do not make the question unimportant.

Twelve people is not twelve repairs

Return to the invented meeting from our source map. It is separate from Tess’s story. We now give its reports small, fixed audiences so that we can count exactly what happens in the example.

Sixteen different people encountered at least one original report. Twelve different people encountered at least one correction. Eight people are in both groups. Another eight saw an original report but no correction. Four correction recipients had not encountered an original report.

What proportion of the original audience encountered a correction? Eight out of sixteen: half.

What proportion of the correction audience had encountered an original report? Eight out of twelve: two-thirds.

Both answers are right. They answer different questions. The first concerns coverage of the original audience. The second concerns the composition of the correction audience. Substituting one for the other would turn a 50 percent answer into a roughly 67 percent answer without reaching one additional person.

It would be worse to divide all twelve correction recipients by the sixteen original recipients and announce that three-quarters of the original audience was reached. Four of those twelve are new people. That calculation would count them as though they belonged to the group we are trying to follow.

The example uses invented audiences, not measured people. Exposure is stipulated in the input. It does not mean that a person read closely, understood, remembered or changed their opinion. Even the correct fraction of one-half establishes only the exposure we put into the model. It cannot establish that half the harm was repaired.

The model also contains repeated encounters. Its original-report lists hold 37 report-person memberships, but only sixteen distinct people. Someone who appears on three lists remains one person when we count individuals. Likewise, sixteen correction memberships reduce to twelve distinct people. Adding page audiences without accounting for overlap would exaggerate how many different people we reached.

In a real dispute, these identities and overlaps may be unknown. We should not manufacture precision to fill that gap. A publisher could report how many corrections were sent, how many pages were updated and which decisionmakers acknowledged receipt. These are narrower measures. Their honesty makes them more useful than a confident estimate built from incompatible counts.

An old study, a current question

Andrew Guess, Brendan Nyhan and Jason Reifler linked survey responses with desktop and laptop browsing by 2,525 Americans during October 7 to November 14, 2016. Among 111 people who visited misleading articles that could be matched to fact-checks, three also visited the matching correction. Their 2020 paper does not comprehensively capture mobile, News Feed or offline exposure. The result concerns that historical sample and measured browsing, not a universal rate of correction failure. It does show why the existence of a fact-check and the original audience’s encounter with it require separate evidence. ¹

Our hypothetical calculation asks a related but different question. Because every audience is supplied, it can make the denominator problem completely visible. The study examined actual recorded browsing with important limits. Neither gives a publisher permission to assume that the people who need an update will find it unaided.

Suppose Tess’s correction page is accurate and easy to read. A customer who searches for it may find the answer. A customer who never knew there was a dispute to revisit may not search. Posting a correction creates a place for the answer. Delivery gives a person a reason and an opportunity to go there.

That is why the route of the original matters. If the accusation was sent by email to a project committee, a public post somewhere else is an incomplete response. The committee should receive the update through a channel it uses, with enough information to recognize the earlier claim. If a newspaper article supplied the allegation, the article itself should make its changed status visible. If a private recommendation carried the story, the sender can correct that recommendation to its recipient.

The appropriate route follows the actual act. We do not need to expose a private dispute to a wider audience merely to demonstrate effort. Sometimes a quiet correction to the right person does more of the required work than a public statement to thousands of strangers.

Give the correction a destination

A workable response starts with a short list of known places where the claim may still guide action. The list might include the originating page, major copies, a customer, an editor, a manager and a particular record. It should distinguish confirmed recipients from guesses. We cannot call somebody reached just because a message could have appeared in their feed.

For each destination, ask what must change. A page may need a visible notice. A private message may need a reply in the same thread. A decision file may need its evidential basis revised. A person who cancelled a booking may need the chance to reconsider. The word correction covers these acts too loosely unless we describe them.

Then assign the work. Tess should not have to discover that five sympathetic people each assumed somebody else would contact the editor. One person can make the contact, another can check the resulting page, and Tess can decide how much of her own attention to give it. Helping means taking a defined part of the burden, with her agreement, rather than sending advice that creates more work for her.

A record of status can be simple. Requested differs from acknowledged. Acknowledged differs from changed. Changed differs from checked. These distinctions can fit in a few words beside each destination. They prevent a friendly reply from being counted as completed repair when the old sentence remains on the page.

The check should be proportionate. If an account owner says that a public post has been corrected, someone can open that post and see what it now says. If a recipient says that they have forwarded the correction, that statement is evidence of their report, not direct access to every recipient’s response. We can accept reasonable confirmation without pretending to know more than we do.

This balance matters. A repair process can become intrusive if it demands screenshots of private conversations from everyone in a chain. People who were not responsible for the error may have their own privacy to protect. Ask for what is needed to establish the action, and do not collect personal material merely because a dispute has made everyone anxious.

Do not reward the wrong number

Imagine two response plans in our small model. The first reaches eight original recipients and nobody new. The second still reaches only those same eight original recipients but also reaches four people who never saw the accusation.

The second correction audience is larger. Coverage of the original sixteen remains one-half. The new recipients might help distribute the correction later, but that later help is not yet in the calculation. A report should not claim it in advance.

Now imagine a third plan that reaches all sixteen original recipients and nobody else. Original-audience exposure becomes sixteen out of sixteen, or 100 percent. This is a genuine increase in the outcome we chose to measure. It still says nothing about how many recipients understood the correction or changed their minds. We have closed one specified gap.

These are extensions of the invented example, not estimates of a real strategy’s effect. Their purpose is to test a reporting habit. Which number would make the team look busier? Which number would show that more of the relevant people received an answer? If the reward follows the first number alone, attention can drift away from the second.

A serious correction report can include several measures without collapsing them. The number of known pages updated describes coverage of the visible record. The number of acknowledged notices describes contact. Reconsidered decisions describe a further action. None can stand in for the whole experience of the person harmed.

We should also avoid defining success so narrowly that nothing counts. A corrected customer decision is worth recording even when the wider audience is unknown. A misleading search snippet replaced by an accurate one is a concrete change, though not proof that every earlier reader returned. Partial repair deserves recognition with its limits intact.

What should be said

Delivery cannot rescue an empty correction. The recipient needs to know what changed, why it changed and which earlier claim is affected.

A useful message might identify the image, state that the quotation has been withdrawn because no source supports its attribution, and link to the publisher’s explanation. It should avoid announcing that Tess has been cleared of every possible criticism. That broader claim would require broader evidence. The purpose is to repair the specific unsupported statement.

If the publisher has established a definite misidentification, it should say so. If the claim remains unverified, that is the status to report. An account owner should not use “unverified” to blur a proven error, or “false” to turn a missing source into proof of the opposite. The wording should follow what has actually been established.

An apology can accompany the information. It may matter greatly to Tess. But the audience also needs the factual change. A regretful tone without a clear correction gives readers too much room to retain the earlier account.

The recipient’s next action should be just as clear. Please attach this update to the post you shared. Please review the decision that used this quotation. Please tell the people to whom you sent it. These requests have an object and a destination. They do not ask a person to promise that nobody will ever remember the error again.

If a recipient declines, record the refusal accurately. There may be reasons worth understanding. They may lack control over the surviving copy, need more evidence, or dispute the correction itself. Further action should respond to that reason. Repeating a demand more loudly is not always the step that answers it.

Back to the person waiting

The original customer finally replies to Tess. He has read the account owner’s withdrawal. He says that he had taken the picture and quotation together as proof of an interview. He will send the correction to his neighbor, who sent him the image.

Tess asks about the kitchen appointment. They choose another time.

That is a specific repair. It does not restore the lost morning, reveal the first creator of the image, or settle every future search result. It does mean that a decision made under the pressure of an unsupported claim has been reconsidered.

An institution reporting this result should say exactly that. A person living through it can allow it to matter. We do not have to declare the whole internet corrected before we recognize that an answer arrived where it was needed.

Source notes

The audience arithmetic is an original hypothetical construction. The baseline, sensitivity scenarios and reproducible set calculations appear in the methods appendix and analysis files.