Hours after the 2013 Boston Marathon bombing, a photograph of a missing child circulated on Twitter and Facebook with the caption identifying him as a victim. The child was Missing-Persons case Sunil Tripathi β but he was not at the marathon. The image had been repurposed from an entirely different missing-person campaign. His family saw their son's face attached to an atrocity he had nothing to do with. Reverse image search, had anyone applied it immediately, would have traced the photo's origin in minutes.
Photographs carry an automatic presumption of truth. When a compelling image accompanies a claim, readers process the claim faster, believe it more strongly, and share it more readily β a phenomenon researchers call the photo-as-proof heuristic. Misinformation actors exploit this by attaching real photographs to false contexts: a flood image from 2011 labeled as 2023, a protest from one country labeled as another, or a celebrity photograph with fabricated captions.
Three categories of visual misinformation dominate fact-checker caseloads: recycled images (real photos, wrong date or location), manipulated images (cropped, colorized, or composited), and synthetic images (AI-generated content). Lesson 3 covers AI-generated media specifically; here we focus on the first two, which remain far more common in the wild.
Reverse image search allows you to submit an image (or image URL) and retrieve pages where that image β or visually similar images β appears. Three engines dominate professional verification work:
The SIFT method β Stop, Investigate the source, Find better coverage, Trace claims β maps directly onto visual verification. When you encounter a compelling image:
During the catastrophic 2019β20 Australian bushfires, a photograph labeled "Australia on fire" went globally viral. Fact-checkers at AAP FactCheck and AFP traced it via TinEye to a NASA satellite image of California wildfires from 2018. The misattribution did not diminish the real severity of the Australian crisis β but it demonstrated how emotional urgency short-circuits verification instincts even among experienced news consumers.
Geolocation is the process of determining where a photograph or video was taken using visual cues in the image itself. It requires no special software β only systematic observation and access to mapping tools. The Bellingcat investigative collective has published detailed geolocation guides and demonstrated the method in dozens of conflict-zone investigations, including tracking Russian military hardware movements in Ukraine by matching road signs, bridge structures, and tree lines to satellite imagery.
Key visual anchors for geolocation: road signs and house numbers, distinctive architecture or landmarks, shadow length and direction (which can be used with sun-position calculators to verify the time of day and season), vegetation species (which narrow geographic region), and military or emergency vehicle markings.
A social media post claims a photograph shows flooding in Jakarta, Indonesia, from "last week." You suspect the image may be recycled. Your AI lab partner will walk you through the verification workflow β ask it anything about how to investigate this image, what tools to use, what to look for, and how to document your findings.
Stanford researchers asked three groups β historians, professional fact-checkers, and undergraduate students β to evaluate the credibility of websites. The historians and students carefully read within the websites they were given, scrutinizing design, "About" pages, and citations. The professional fact-checkers did something entirely different: they immediately opened new browser tabs to search for what others said about each source. The fact-checkers were faster and dramatically more accurate. The researchers named this behavior lateral reading.
Lateral reading means leaving a website almost immediately and searching for what credible third parties say about it β rather than reading the website itself more carefully. This sounds obvious, but it runs against the instinct most people develop in school, where careful close reading is rewarded. Online, close reading plays into the hands of sophisticated misinformation sites that invest heavily in appearing credible from within.
The practical workflow: when you encounter an unfamiliar source making a significant claim, open a new tab, type the source name into a search engine, and scan the first few results. Wikipedia, journalism watchdogs, academic critiques, and news stories about the organization appear quickly. You are not trying to find the "truth" about the claim yet β you are establishing whether the source is trustworthy enough to read further.
Over 300 fact-checking organizations operate globally, indexed by the Duke Reporters' Lab. The most cited in English-language verification work:
Fact-check verdicts are more nuanced than simple true/false. PolitiFact's "Half True" rating means a claim contains accurate elements but omits critical context that would change its meaning. AFP's "Misleading" rating often covers technically accurate statements weaponized through selective framing. Understanding rating systems prevents the mistake of treating a "Mostly True" verdict as full endorsement.
A critical caveat: absence of a fact-check is not evidence of truth. Fact-checking organizations are small and must triage. The vast majority of circulating misinformation has never been reviewed. A claim that hasn't been checked is not thereby verified β it simply hasn't been checked.
When Facebook launched its third-party fact-checking program in 2016 after criticism over misinformation during the US election, it required partner organizations to be certified IFCN signatories. The IFCN Code requires: a commitment to nonpartisanship, transparent sourcing, transparent funding, transparent methodology, and an open corrections policy. Checking whether a fact-checking outlet is an IFCN signatory is itself a quick lateral-reading move that filters out politically motivated "fact-checkers" operating under that label.
Professional fact-checkers maintain bookmarked sets of starting points. A functional personal toolkit includes: a reverse image search extension (InVID/WeVerify), three fact-checking bookmarks (Snopes, PolitiFact or AFP, and a science-specific checker like FullFact for health claims), a media bias reference (AllSides or Ad Fontes Media Bias Chart), and an archived search tool (Wayback Machine / archive.org) for retrieving deleted content.
The Wayback Machine deserves special mention: when a source deletes a page after publishing misinformation, the archived version often remains. Fact-checkers routinely use it to document claims that have since been removed, establishing a permanent record.
You encounter a health article from a website called "NaturalHealthInsider.net" claiming a government agency suppressed a study on vaccine side effects. You've never heard of this outlet. Your AI lab partner will help you apply lateral reading and other source evaluation techniques to assess its credibility.
On March 22, 2023, a series of photorealistic images depicting former President Donald Trump being arrested by police officers spread across Twitter, Telegram, and news aggregators. The images were created by Eliot Higgins of Bellingcat using Midjourney v5, posted with a label identifying them as AI-generated β and then extensively recirculated without that label by other accounts. Multiple news outlets ran queries to their newsrooms asking for authentication. The images had never represented a real event, yet they provoked genuine public confusion about whether an arrest had occurred.
AI image generation has outpaced detection technology. In 2023, researchers at the University of California, Berkeley, and elsewhere demonstrated that leading AI detectors β including those embedded in major platforms β had false-positive rates that made them unreliable for operational fact-checking. A human rights organization using an AI detector could incorrectly flag a genuine atrocity photograph as synthetic, with serious real-world consequences.
The honest assessment from the fact-checking community: no single AI detection tool is reliable enough to use as a sole verification method. Detection tools are useful as a triage signal β raising a flag for closer investigation β but not as a final verdict. The verification workflow must combine technical signals with contextual analysis.
Current AI image generators leave characteristic artifacts that trained eyes can learn to spot β though this list will become obsolete as models improve. As of 2024, common signals include:
Days before Slovakia's September 2023 parliamentary election, an audio clip circulated on Facebook appearing to be a recording of Progressive Slovakia leader Michal Ε imeΔka discussing how to rig the election by buying votes from the Roma community. AFP Fact Check and DFRLab analyzed the audio and identified characteristics consistent with AI voice synthesis β unnatural prosody, background noise inconsistencies, and phrasing atypical of Ε imeΔka's documented speaking style. The clip could not be definitively proven synthetic with available tools, but multiple signals pointed strongly in that direction. Ε imeΔka's party narrowly lost the election. The case became a reference point in EU discussions about AI regulation ahead of elections.
Because technical detection is unreliable, experienced fact-checkers lean heavily on contextual signals. The key questions for any suspicious image or video: Does corroborating evidence exist from independent sources? Is the claim being amplified by accounts with patterns of sharing synthetic content? Does the content appear at a suspiciously convenient political moment? Can the claimed event be verified through other media?
The First Draft organization, before its 2023 closure, documented a principle they called verification through triangulation: if an event happened, multiple independent witnesses with different devices and angles will have captured it. A single perfect image of a major event with no corroborating footage is a red flag regardless of what detection tools say.
A 40-second video is going viral on social media showing what appears to be a prominent politician making an inflammatory statement at a press conference. No major news outlets have reported on the event. The account that posted it was created three months ago and has 12,000 followers. Ask your AI lab partner how to approach this investigation β covering detection tools, contextual red flags, and how to document your findings if you suspect it's synthetic.
In 2018, the Senate Intelligence Committee released reports documenting the Internet Research Agency's (IRA) social media operations during the 2016 US election cycle. The Oxford Internet Institute's Computational Propaganda Project and the firm New Knowledge analyzed IRA content and found that the operation did not primarily originate viral claims β it identified organic grievances and amplified them. One documented technique: IRA accounts would create content on fringe platforms, then amplify it through a network of accounts to push it into mainstream feeds, where authentic users would pick it up and spread it further, erasing the trail back to its origin.
Understanding where a claim originated reveals whether it emerged from an organized influence campaign, a single bad-faith actor, honest error, or a satire account whose label was stripped. Each of these requires a different response. A claim originating from a known state-sponsored influence operation warrants different public messaging than one that started as genuine misinformation from a confused grandmother β even if the claim text is identical.
Claim tracing also helps fact-checkers prioritize. When the same claim appears simultaneously across hundreds of accounts that share behavioral characteristics β posting at similar times, using similar language, linking to the same obscure sources β this is a signal of coordinated inauthentic behavior, not organic spread.
The Stanford Internet Observatory's 2020 report on the "Plandemic" video documented a classic coordinated spread pattern. The video was uploaded on May 4, 2020. Within hours, hundreds of accounts with no prior connection began posting it simultaneously with nearly identical framing language β a behavior inconsistent with organic discovery. The SIO found evidence of pre-coordinated distribution planning in private Facebook groups and Telegram channels before the video's public launch.
Key signatures of coordinated inauthentic behavior: synchronized posting times, shared templated language, accounts created in batches (similar registration dates, similar naming patterns), cross-platform simultaneous deployment, and amplification before organic engagement metrics could plausibly have developed.
EU DisinfoLab documented a 15-year operation running 265 fake media outlets in 65 countries, all traced back to a network of front organizations with links to Indian PR firm Srivastava Group. The investigation used WHOIS records, network analysis, and open-source domain tracing to map the operation. Fake NGO websites created recycled UN reports to make their content appear to have been cited at the UN. The operation had been running since 2005, demonstrating that even long-running sophisticated campaigns leave traceable infrastructure when investigators have the right tools.
When confronted with a viral claim, the full toolkit from this module provides a layered approach. Start broad and narrow: check fact-checkers first (fastest, most efficient), then trace the claim's origin using search operators and date filtering, then assess the source via lateral reading, then verify any images or video using reverse search and geolocation, then examine the amplification network if coordinated spread is suspected.
No single tool is sufficient. The power of the verification toolkit is its redundancy β multiple independent methods pointing to the same conclusion is the gold standard. One tool flagging suspicious behavior is a signal; three independent methods converging on the same finding approaches a defensible verdict.
A claim appears on X/Twitter that a major bank is about to fail, sourced to an obscure website registered 6 weeks ago. Within 4 hours it has been posted by 400+ accounts, many of which have posted about financial collapse topics repeatedly this month. Your AI lab partner will help you investigate: how to trace the claim's origin, assess the domain, analyze the amplification network, and determine whether this shows signs of a coordinated operation.