In 2016, researchers at the Stanford History Education Group tested how well students, professors, and professional fact-checkers evaluated online sources. The results were striking: professional fact-checkers outperformed both groups — not because they read articles more carefully, but because they did the opposite. They clicked away almost immediately and searched for what other sources said about the site they were evaluating.
The researchers named this "lateral reading." Novices read vertically — top to bottom through one page. Experts read laterally — across multiple sources about the source itself.
Lateral reading means opening new browser tabs to search for information about a source before — or instead of — reading the source itself. You are not asking "what does this site claim?" You are asking "who runs this site, and are they credible?"
When professional fact-checkers at PolitiFact, Snopes, or AFP Fact Check encounter an unfamiliar website, they type the site's name into Google within seconds. They look for news stories, academic references, or Wikipedia entries that tell them whether the outlet has a track record of accuracy or a history of spreading misinformation.
The Stanford study, published in 2019 in the Proceedings of the National Academy of Sciences, gave the same set of web sources to three groups: high school students, college students, and professional fact-checkers. Fact-checkers correctly identified the reliability of sources in a fraction of the time the students took — and they did it by leaving the page almost immediately.
One typical fact-checker comment recorded in the study: "I'm going to Google this site right away. I don't want to read their About page — they wrote that themselves." That instinct — distrust the source's own self-description — is the core of lateral reading.
AI tools like ChatGPT cannot perform lateral reading in real time — they don't browse the web during conversation (unless given a specific tool). They also cannot tell you whether a website has a history of publishing false stories. That's a judgment that requires live, lateral search — and it's something humans need to do themselves.
A source that is credible on one topic may be unreliable on another. The fact-checker's goal is not to permanently label a source but to understand its track record for the type of claim being made right now.
You've just encountered a headline on a website called "NaturalHealthAlert.net" claiming that "Scientists Confirm Vitamin C Cures COVID-19." Walk through the lateral reading process with the AI assistant below. Ask it how you would evaluate this source before deciding whether to believe the claim.
During the 2012–2015 period of the Syrian civil war, photographs of refugees, bombed buildings, and displaced children circulated globally. Researchers at the BBC's User Generated Content Hub and BuzzFeed News documented dozens of cases in which the same photograph was reused in stories about entirely different events — floods in Pakistan, earthquakes in Haiti, protests in Ukraine. The images were real; the captions were false.
The tool that allowed investigators to expose each case was the same: reverse image search. By uploading an image to Google Images or TinEye, they could find every previous time that image had appeared online — often years before the story claiming to show it.
Standard search engines let you search by text. Reverse image search lets you search by the image itself. You upload a photo — or paste its URL — and the search engine finds every indexed page where that image (or a visually similar image) has appeared.
The two most powerful tools are:
During Hurricane Harvey in 2017, a photograph circulated on Twitter showing a dog appearing to swim through floodwater past a person's feet. It was captioned as a Harvey rescue scene. Snopes and AFP used reverse image search and traced the image to a 2011 flood in the Philippines — six years before Harvey. The image was real; the captioning was fabricated.
This pattern — real photograph, false context — is one of the most common forms of image-based misinformation. Reverse image search is the primary tool for detecting it.
When you find an image in reverse search, always sort results by "oldest first" using TinEye's filter. An image that appeared in 2015 cannot be from an event that happened in 2023. The date gap is your most decisive clue.
Reverse image search finds copies of existing images. It cannot reliably catch AI-generated images that were never previously published anywhere. Investigators now use additional tools — including Hive Moderation, Illuminarty, and Google's SynthID system — to detect synthetic images. But these tools have limitations. The core fact-checking skill remains: when an image seems designed to provoke strong emotion, apply extra scrutiny before assuming it is real.
AFP Fact Check's published methodology states: "No image can be accepted as evidence of a specific event without reverse image verification." This is now standard practice at major wire services and broadcast news outlets worldwide.
You've seen a viral tweet showing a dramatic photo of flooded streets, captioned "Flooding in Miami today — this is what climate change looks like." Walk through how you would verify or debunk this image using reverse image search techniques.
Mike Caulfield, a researcher at the University of Washington's Center for an Informed Public, spent years studying why traditional media literacy education was failing. Students were taught to ask "Is this source biased?" but that question proved too vague to be actionable in real-time online browsing.
In 2019 he published SIFT — a four-move framework that translates the instincts of professional fact-checkers into a systematic routine. SIFT stands for: Stop · Investigate the source · Find better coverage · Trace claims to their origin. It has since been adopted by universities, libraries, and newsrooms across the English-speaking world.
In 2021, a widely shared claim asserted that "the COVID-19 vaccine causes infertility in 97% of recipients." Tracing the claim revealed a chain: viral tweets cited a YouTube video, which cited a German article, which cited a letter by Wolfgang Wodarg and Michael Yeadon submitted to the European Medicines Agency. The letter itself contained a speculative hypothesis, not a finding — and the EMA had rejected the petition. The "97%" figure appeared nowhere in any primary document.
SIFT's fourth move — tracing claims to their actual origin — is the only way to expose this kind of citation laundering, where a speculation gets progressively inflated as it passes through layers of secondary sources.
Citation laundering is the process by which a speculative claim in a fringe document gets cited by a blog, then a website, then a social media post, with each step making it sound more certain and authoritative. SIFT's "Trace" move is designed specifically to reverse this process.
Move 3 — "Find better coverage" — is often misunderstood. It does not mean finding a source you personally agree with. It means finding the most authoritative, primary, or expert source on the topic. For scientific claims, that means the actual peer-reviewed paper. For legal claims, that means the actual court document. For government statistics, that means the actual agency data, not a journalist's summary of it.
The Reuters Fact Check team, in its published methodology, uses identical language: "We always seek the primary document — court filing, government release, scientific paper — before evaluating a claim based on media reports about it."
AI assistants can help with some SIFT moves. They can summarize topics (supporting Move 3), explain scientific concepts, and help you understand jargon in primary documents. But AI cannot reliably perform Move 2 (investigating sources in real time) or Move 4 (tracing a live viral claim through its actual citation chain online). Those moves require live web access and the kind of skeptical, iterative searching that human fact-checkers do.
SIFT works because it converts vague advice ("be critical") into specific actions with specific tools. Each move has a defined goal and a defined stopping point. That specificity is why it works where earlier frameworks failed.
A friend shares a Facebook post claiming "New study proves coffee causes cancer — scientists demand warning labels." The post links to a wellness blog, not a scientific journal. Practice applying all four SIFT moves to this scenario with the AI assistant.
In 2019, EU DisinfoLab published a report called "Indian Chronicles" — one of the largest investigations into coordinated inauthentic behavior ever documented. Researchers discovered a network of over 265 fake news sites across 65 countries that appeared to be independent local media outlets but were all connected to a single Indian PR company.
The key discovery method was not analyzing the articles themselves — it was examining the WHOIS registration records of the websites. Multiple "independent" outlets shared identical registration details, the same IP addresses, and the same hosting infrastructure. The metadata told the story that the content carefully concealed.
Metadata is data about data. Every digital file — a photograph, a Word document, a PDF, a web page — carries metadata that records how, when, and sometimes where it was created. This data is often invisible to casual readers but accessible to investigators using freely available tools.
WHOIS is a public database of domain registration information. To query it, you can use whois.domaintools.com or simply type "whois [domain name]" at lookup.icann.org. Key things to examine:
In 2020, researchers investigating health misinformation found that several websites presenting as legitimate medical news outlets had previously been general content farms or even spam sites before rebranding. The Wayback Machine preserved their earlier versions. A site that was selling discount shoes in 2018 and is now publishing COVID-19 treatment advice in 2021 carries a different level of credibility than a site with a continuous, consistent editorial history.
AI tools cannot query live WHOIS databases, read current Wayback Machine snapshots, or extract EXIF data from images you haven't shared. These are manual steps requiring direct tool access. Knowing they exist — and how to use them — is a human skill that complements what AI can help you do.
Professional investigators at Bellingcat, EU DisinfoLab, and DFRLab follow a consistent principle: the content of a piece of misinformation is often carefully crafted to deceive — but the infrastructure used to publish and distribute it often is not. Look at the container, not just the content.
You've discovered a website called "AmericaFirstHealthNews.com" publishing alarming stories about vaccine side effects. The site's "About" page says it was "founded in 2003 by concerned doctors." Walk through with the AI how you would use WHOIS, the Wayback Machine, and image metadata to verify or challenge this claim.