「经济学人」The lie-detectors

How to distinguish between weapons-grade disinformation and everyday codswallop

Correcting the gigabytes of digital gibberish that circulate at high speed online is a never-ending task. YouTube removed more than half a million channels last year for broadcasting misinformation. Facebook and Instagram deleted 27m falsehoods about covid-19 at the height of the pandemic. The doughty fact-checking organisations that try to keep the internet honest face more claims than they can handle. How should they prioritise?

The question is growing more pressing as fact-checking resources become scarcer. Meta, which built probably the world’s largest network of fact-checkers for its social networks, announced in January that it would start to replace professional sleuths with volunteers. America’s government is dismantling USAID, which had channelled funds to fact-checking organisations.

So checkers came up with a new approach: forecasting which claims are most dangerous and thus which most deserve to be put under the microscope. Researchers from the University of Westminster and fact-checkers from Full Fact, Africa Check and the AFP news agency developed a triaging system to sort weapons-grade misinformation from everyday nonsense.

One test of a false claim is whether enough people will believe it for it to cause any harm. To swing an election with misinformation, you need to persuade many people; to kill someone with fake medicine you need to convince only one. Another test is whether those believing a false claim have the capacity to act on it. People may be misled about the genesis of covid-19, for example, but whether they think it came from a market or a lab is unlikely to change their behaviour. The researchers estimated that, of the false claims in their sample, 57% were unlikely to contribute to any specific real-world effect.

Of the remaining, potentially consequential falsehoods, the checkers considered whether the consequence would be “direct”—such as persuading people not to take a vaccine—or “cumulative”, contributing to a false narrative about immigration, say. The claims were roughly evenly split. “Cumulative” harm is harder to assess, says Peter Cunliffe-Jones of the University of Westminster, but large data sets make it possible to see how often a claim is repeated, and thus when a narrative is forming.

Triaging may help overworked fact-checkers to focus on the most dangerous false claims. But harm is not everything. Karl Malakunas of AFP points out that one of his organisation’s most-read fact-checks concerns a photograph of an elephant carrying a lion cub in its trunk (fake, alas). It seems most unlikely that anyone fooled by it would seek a pachyderm playmate. But correcting viral falsehoods matters for digital literacy, Mr Malakunas says.

Time devoted to selecting which dodgy claims to check is probably well spent. It takes five minutes to triage a claim, whereas carrying out a thorough check takes five to six hours. The fact-checking world needs to get more systematic in its approach, says Mr Cunliffe-Jones. “If this community is going to learn anything from Meta…it’s that data is the future.”

This article appeared in the International section of the print edition under the headline “The lie-detectors” (May 15th 2025)

我没有译文，仅供拓展阅读

LearnAndRecord

2015年2月8日

2025年5月19日

第3754天

每天持续行动学外语