In the months before the 2016 US presidential election, a story spread across Facebook claiming that Pope Francis had endorsed Donald Trump. The story came from a site called WTOE 5 News — which disclosed in tiny footer text that it was a "fantasy news website." It was shared nearly a million times. Most people who shared it never left the page to check whether WTOE 5 News was real. That single click they didn't take cost them the truth.
Researchers at the Stanford History Education Group spent years studying how people evaluate online information. They compared high school students, college students, professional historians — and working fact-checkers at organizations like PolitiFact and Snopes. The fact-checkers were dramatically faster and more accurate than everyone else.
The surprising part: the historians had deep subject expertise, but expertise wasn't the winning factor. What separated the fact-checkers was a single habit called lateral reading — leaving the page you're evaluating almost immediately, and opening new tabs to find out what other sources say about the source.
Regular readers tend to read vertically — scrolling down through a page, examining its design, reading its "About" section, evaluating what it says about itself. Fact-checkers do the opposite. They read laterally — they leave the site within seconds and search for what the wider web says about it. They check the source before trusting the claim.
In 2019, Stanford researchers found that fact-checkers were able to accurately assess the credibility of a website in under 30 seconds using lateral reading — while historians spent up to 10 minutes reading a site and still reached incorrect conclusions. Speed and accuracy both improved when users left the page.
You can apply lateral reading yourself. Here is the exact process professional fact-checkers use:
In 2017, NewsGuard researchers documented sites that described themselves as "objective," "non-partisan," or "community news" in their About sections — while consistently publishing fabricated content. The site American News (ampnews.us) described itself as "your trusted source" while Snopes had already documented dozens of false stories it published. Self-description is not evidence of credibility.
The same principle applies to AI-generated content. When an AI tool produces a confident, well-formatted article, the polished presentation is not evidence of accuracy. It is a formatting decision, not a fact-checking decision. Lateral reading bypasses presentation entirely and asks: what do independent observers say?
Lateral reading works because it shifts your question from "Does this look trustworthy?" to "Is this trusted by people with no stake in the answer?" Design, tone, and confidence are easy to fake. Independent corroboration is hard to fake.
Your AI lab partner will give you a fictional article headline and source name. Walk through the lateral reading process: identify what you'd search, what you'd look for, and how you'd decide whether to trust the source. Your partner will coach you through each step.
When Hurricane Harvey struck Houston in August 2017, a photograph circulated on Twitter showing a shark allegedly swimming on a flooded Texas freeway. The image was real — but it was taken in 2011 during a shopping mall flood in Brazil, thousands of miles away. It was shared hundreds of thousands of times with the Houston caption. A five-second reverse image search would have surfaced the original 2011 post.
Recycled images exploit a cognitive shortcut: we process photographs as evidence of what they show. When a dramatic image accompanies a headline about a current event, the brain connects them automatically. Misinformation producers know this — they routinely search for striking old photographs and re-caption them during breaking news events when fact-checkers are overwhelmed and emotion is high.
First Draft News documented over 40 cases of recycled photographs during the 2015 European refugee crisis alone. Many of the most-shared images — including a photograph of a drowned child described as a Syrian refugee — were real photographs but misattributed in time, location, or context. The photograph was real. The caption was wrong.
Google Images, TinEye, and Bing Visual Search all allow you to upload an image — or paste an image URL — and find every other place on the web where that image appears. The oldest indexed occurrence is typically the original. If a photo captioned "March 2024" first appeared in a database entry from 2017, the caption is wrong.
In 2022, fact-checkers at Reuters used reverse image search to debunk photographs circulated during the Russian invasion of Ukraine that actually showed events in Syria (2013), Gaza (2021), and Iraq (2016). The photographs were real war images — the misattribution was the lie.
In 2020, fabricated screenshots of tweets attributed to public figures spread widely on platforms including Twitter and Facebook. The screenshots showed politicians and celebrities saying things they never said — created using browser developer tools or tweet-generator websites that anyone can access freely.
The BBC's Reality Check team documented a fabricated screenshot attributed to a UK politician during the 2019 election. The tweet template was created at a site called "Freetweetgenerator.com" in under two minutes. Verification required going to the actual Twitter account and checking the archive — the tweet had never existed.
For screenshots: search the exact quoted text in quotes on the platform itself. If the original post doesn't exist, a wayback archive doesn't show it, and no journalists cite it — it's almost certainly fabricated.
AI image generators (Midjourney, DALL-E, Stable Diffusion) can now produce photorealistic images of events that never happened. Standard reverse image search may not find AI-generated images because they have no prior indexed occurrence. Use AI image detection tools like Hive Moderation, AI or Not, or Google's SynthID where available. Look for unnatural finger counts, lighting inconsistencies, and background distortions — common AI artifacts as of 2024.
Your AI partner will describe realistic image verification scenarios. For each scenario, identify what tool you'd use, what you'd look for, and whether the described evidence suggests the image is authentic, recycled, or fabricated.
In April 2020, a widely shared post claimed: "The CDC just admitted masks don't work." The CDC had, in fact, updated its guidance — twice — during the pandemic's first weeks. The claim latched onto a real event (guidance changed) and added a false conclusion (therefore masks don't work). FactCheck.org documented the post reaching millions of Facebook shares before platforms applied labels. The fact at the center was real. The interpretation wrapped around it was fabricated.
Professional fact-checkers at AP, Reuters, and AFP are trained to decompose every claim into its component parts before beginning verification. They distinguish between:
In the CDC mask example: the anchor fact was "CDC updated guidance." The unstated bridge was "updated guidance means previous guidance was wrong." The false conclusion was "therefore masks don't work." Each step away from the anchor fact introduced additional error.
When you encounter a viral claim, ask these questions in order:
The "97% of scientists agree" claim (climate): This stat is real — it derives from multiple meta-analyses of peer-reviewed climate literature, including Cook et al. (2013). But a decomposed version appeared claiming "97% is based on just 77 scientists" — the anchor fact being a paper by Doran & Zimmerman (2009) that surveyed 77 active respondents, one of many studies that reached similar conclusions. The real anchor was made to sound flimsy by hiding the broader consensus evidence.
The "crime in immigrant communities" pattern (2017–2019): PolitiFact documented multiple claims stating "immigrants commit more crimes" accompanied by real crime statistics. The anchor facts were real crime numbers — but the bridge (comparison to crime rates in non-immigrant populations, adjusted for income, age, and urbanization) was always absent. Without the bridge, the conclusion was unsupported. Real data, missing context, false implication.
The single most common misinformation structure documented by Reuters Fact Check is a real statistic presented without the denominator, comparison group, or time frame needed to understand it. A number that sounds alarming in isolation often becomes unremarkable — or reversed — when context is restored. Always ask: compared to what? Over what period? Among which population?
AI systems are particularly prone to producing claims that have a true anchor but a false bridge. In 2023, journalists at the Chicago Sun-Times used an AI tool that generated a summer reading list containing real author names but invented book titles. The anchor (real authors exist) was true. The bridge (these authors wrote these specific books) was fabricated. Without decomposing each element of the claim, readers had no way to know which parts were real.
When evaluating AI-generated content: verify the anchor facts separately from the conclusions. Don't treat the whole output as one unit — treat it as a list of individual claims, each requiring independent verification.
Your AI partner will give you viral-style misleading claims. For each one, identify: (1) the anchor fact, (2) the unstated bridge, and (3) the implied conclusion — then explain where and why the reasoning breaks down.
Between 2014 and 2018, BuzzFeed News built one of the most respected verification desks in digital media. Their reporters documented a public workflow — later published as an open-source guide — that allowed individual journalists to verify a social media claim in under three minutes. The workflow was adopted by AFP, the BBC, and the International Fact-Checking Network. It was fast because it was systematic: the same steps, every time, in the same order.
Human judgment about credibility is unreliable under stress, time pressure, and emotional engagement — exactly the conditions that surround viral misinformation. Research published in Science (Vosoughi, Roy, Aral, 2018) found that false news spreads six times faster than true news on Twitter, in part because false news tends to be more novel and emotionally arousing. When a claim makes you feel something strongly, that feeling reduces your accuracy at detecting false information.
A written workflow counteracts this by creating procedural distance. You follow the steps regardless of how you feel about the claim. Professional fact-checkers at organizations like Africa Check, Full Fact (UK), and Chequeado (Argentina) all use written checklists specifically because checklists reduce the influence of prior belief on verification decisions.
This workflow synthesizes the lateral reading, image verification, and claim decomposition methods from this module:
These signals don't prove misinformation, but each one alone warrants applying the full workflow before sharing:
AI-generated articles, summaries, and social posts require one additional step in the workflow: treating each factual claim as a separate unit. An AI summary of five facts may have three correct and two fabricated — and the correct ones make the fabricated ones feel trustworthy. Reuters' 2023 internal audit of AI tools used in newsrooms documented this pattern: AI-generated copy often had accuracy rates of 70–85% by sentence, meaning 1 in 7 to 1 in 4 sentences contained a checkable error.
For AI content specifically, the workflow adds: (6a) identify every proper noun, statistic, and named event in the AI output and verify each independently against a primary source. This sounds laborious but typically takes under two minutes per paragraph.
Research by the Reuters Institute for the Study of Journalism found that regular fact-checkers do not apply the full workflow to every piece of content they read — they apply it selectively when red flags trigger it. The goal is not to fact-check everything, but to recognize which content warrants deeper investigation, and to run the workflow reliably when those signals appear.
Your AI lab partner will present you with a realistic piece of viral content — a social media post, an article excerpt, or an AI-generated paragraph. Walk through all six steps of the verification workflow and explain your reasoning at each step. Your partner will give feedback on your process.