Corrections nobody reads: the continued influence effect

Series: How minds are moved · Fake news, fact-checks and how to check for yourself · 7 of 7

On Saturday 7 November, 30 news sites in 13 languages ran a headline saying that thousands of Syrian troops were going to Yemen and Saudi Arabia. Within hours Damascus denied the Yemen part, and two days later its foreign minister denied it at a press conference. My news engine LSN News kept every headline, so I could ask a simple question: how many of the sites that ran the claim later ran the denial? The answer was 4 of 30. This post is about that gap: why corrections reach a fraction of the original audience, and why, even when they do reach us, the first version keeps working in our heads.

Contents · 9 sections

TL;DR #

• In my data, only 4 of 30 outlets that headlined the Syrian troops claim later headlined the official denial.
• A retracted fraud accusation in Seattle: the original videos were viewed millions of times; the correction reached 42 news sites, 34 of them running one syndicated piece.
• Psychologists call the second problem the continued influence effect: even people who read and remember a correction keep using the false detail when they reason.
• Corrections work better when they give an alternative explanation, are repeated, and lead with the fact rather than the myth.
• The “backfire effect”, where correcting makes belief stronger, turns out to be rare. Corrections help; they just help less than the original hurt.

Case 1: a claim, a denial, and who carried which #

On 7 November AFP reported, citing several Yemeni officials, that thousands of Syrian troops were set to deploy to Saudi Arabia and Yemen, with some already on the ground. The first version in my data appeared at 14:00 UTC in an Arabic paper. By evening it was in English, French, German, Dutch, Norwegian, Slovak, Vietnamese, Swahili and more.

The denial came fast. At 16:45 UTC the first headlines carried a Syrian diplomatic source, quoted by the state agency SANA, saying military agreements with Saudi Arabia covered training and expertise, not deployment to Yemen. On 9 November Foreign Minister Asaad al-Shaibani said: “There is no deployment of troops to Yemen nor is there a plan for such deployment in the near future.” On 12 November Yemen’s chargé d’affaires in Damascus said the reports “do not reflect the reality on the ground”.

What exactly is true is still disputed, and I am not judging it. AP reported that Syria was preparing to send around 600 special forces to Saudi Arabia near the Yemeni border; Al Jazeera reported a Syrian official confirming troops sent to Saudi Arabia for training; The New York Times reported reconnaissance teams. The “into Yemen” part is what both governments denied. That makes it a good test: a contested claim, and an official denial that any careful reader would want to see next to it.

Here is what the headlines show:

• The claim story on LSN News grew to 87 articles in 18 languages; only 7 of its 87 headlines mentioned the denial, 5 of them Greek.
• The main denial story reached 49 articles in 8 languages, a little over half the claim’s reach.
• I took the 30 outlets that ran the claim as a headline on 7 and 8 November and looked for any later headline from them on the subject. 4 ran the denial. 2 more ran the narrower AP version. The other 24 never returned to it in a headline.

Greek outlets were the exception worth copying: five of them put the claim and the denial, or “conflicting statements”, in the same headline (“…Damascus denies”). The reader got both at once.

A caveat: some sites may have updated their original article in place, which I cannot see. But a reader who saw the first headline on Saturday would not see an edit made on Monday.

Case 2: a retraction, and who heard it #

In October, the American YouTuber Nick Shirley published two videos accusing a small family-run Asian grocery in Seattle of food-stamp fraud. According to The Seattle Times, the videos were viewed millions of times, and the store’s managers said staff faced a wave of abusive calls.

On the night of 11 November he retracted. “I have not uncovered any evidence that [the store] itself engaged in EBT fraud,” he said in a short video, and took the originals down.

In LSN News, the correction reached 42 news sites. 34 of those ran one identical syndicated article from a national desk that feeds local TV stations; the main story for the retraction shows 34 outlets but only 4 independent voices. The other eight were a Seattle TV station, The Seattle Times’ report reprinted by a regional paper, and six national news and opinion sites.

Notice the mismatch. The accusation travelled through his own channels to his own audience. The correction travelled mostly through news sites, many of which that audience does not read. Even a full, honest retraction by the author does not reach the same people.

I cannot measure his video views myself, so I cannot give you a ratio. But the pattern is the one researchers keep finding.

What the research says about reach #

In 2018 three MIT researchers published in Science the largest study of its kind: about 126,000 rumour cascades on Twitter from 2006 to 2017, shared by some 3 million people. False stories spread farther and faster than true ones in every category, and falsehoods were 70% more likely to be retweeted than the truth. They also found that bots spread true and false news at about the same rate: people, not robots, made the difference. Novelty was a key reason. A correction is rarely novel.

Craig Silverman’s 2015 report for Columbia’s Tow Center followed more than 1,500 news articles about over 100 online rumours. In interviews at the time his summary was blunt: corrections are less interesting than the claim, are shared less, and often do not reach the people who saw the original.

Why corrections that do arrive still fail #

Reach is the first problem. The second is stranger.

In 1994 Hollyn Johnson and Colleen Seifert ran experiments with a simple story about a warehouse fire. Early in the story, readers learned that volatile materials, paint and gas cylinders, had been stored carelessly in a closet. Later, some readers were told plainly: the closet was actually empty. Then everyone answered questions like “what caused the black smoke?”

The people who got the correction remembered it. And they still used the paint and gas to explain the fire. Johnson and Seifert named this the continued influence effect.

A large 2012 review led by Stephan Lewandowsky, in Psychological Science in the Public Interest, summarised two decades of follow-up work: retractions “rarely, if ever” eliminate reliance on misinformation, even when people believe, understand and later remember the retraction. The explanation they favour is that we build a mental model of events. Remove a key piece and the model has a hole. The false piece fills it better than nothing.

The same review gives a real-world example from the government side. Before the 2003 invasion of Iraq, US officials said there was no doubt Saddam Hussein had weapons of mass destruction. None were found, the official bipartisan position changed, and corrections were published. Yet years later some 20 to 30% of Americans believed such weapons had actually been discovered.

It is not only governments and influencers. In October 2023 The New York Times published an editors’ note saying its early coverage of the blast at al-Ahli hospital in Gaza had “relied too heavily on claims by Hamas” without making clear they could not be verified. Nobody can measure how many of the people who saw the first headline ever saw the note.

What does help #

The good news is that corrections are not useless. A 2019 study by Thomas Wood and Ethan Porter tested 52 contested issues on more than 10,100 people and found no case where a correction made belief in a false claim stronger. The feared “backfire effect” is rare. Corrections move people towards the facts; they just do not move them all the way back.

The research points to a few things that make corrections work better:

• Give an alternative explanation. “The closet was empty” leaves a hole; “the fire was caused by an electrical fault” fills it.
• Lead with the fact, mention the myth once, explain why it was wrong, and repeat the fact.
• Repeat the correction. One correction against many repetitions of the claim loses.
• Warn people in advance that misleading claims of this kind are circulating.
• Correct in the same place, and in the same form, as the original. A headline needs a headline.

How to notice it #

• When you remember a striking claim, ask yourself: did I ever see how it ended?
• Search the claim again a few days later, adding words such as “denied”, “retracted”, “correction”.
• Be suspicious of your own certainty about details you learned from a single early report.
• Notice when a story you shared turns out wrong, and share the correction to the same people.
• Prefer sources that correct visibly, in the headline, rather than silently in the text.

Frequently asked questions #

What is the continued influence effect? #

The tendency to keep relying on information after it has been corrected, even when you remember the correction. It was named by Johnson and Seifert in 1994.

Do corrections and fact-checks work? #

Partly. They reduce belief in false claims, and backfire is rare, but they rarely undo the first impression completely, and they usually reach fewer people.

Why do corrections reach fewer people? #

They are less new and less emotional than the original claim, they come later, and they often travel through different channels from the original.

How can I correct someone without making things worse? #

Lead with the accurate fact, offer an alternative explanation, mention the false claim only once, and use the same channel where it was shared.

Did the outlets in the Syrian case do anything wrong? #

Not necessarily. They reported AFP’s sourced claim. What my data shows is that most did not headline the denial afterwards, so their readers saw only half the story.

The lesson #

A claim and its correction are not a fair fight. The claim arrives first, it is new, and it fills a gap in our picture of the world. The correction arrives later, to fewer people, and asks us to leave a hole. That is why the habit that matters most is not spotting the lie on day one, but going back on day three to see how the story ended, and then telling the people you told the first time.