Fact-Checking Tools and OSINT: Reading the Results

Fact-checking tools return results that look like answers: a match, a rating, a timestamp. Each one answers a narrower question than it appears to. A reverse image search shows where a copy appears, a published check shows what an organization concluded, and a metadata field shows what a file records about itself. Concern about misleading content is widespread, with the Reuters Institute’s Digital News Report 2026 finding that concern about fake news stood at 62% on average across the markets it surveyed, up four percentage points on the previous year.

Moving from a tool result to a justified conclusion requires another step, which is interpreting the evidence, and that step is the subject of this guide. It is written for journalists, researchers, students and everyday readers, and it relies on open-source intelligence (OSINT), meaning the evaluation of publicly available information to answer a defined question. The guide explains how to choose a tool, interpret its output and decide what to check next.

In this article, we examine how to match a claim to the right tool, what fact-checking means and why context matters, a six-step process for checking a claim, techniques for text, images and video, what each major resource can and cannot establish, how OSINT fits in, two worked examples, and the mistakes that most often weaken a conclusion.

Matching Each Claim to the Right Tool

Choosing a tool starts with the input, not the brand. A claim in text, an image, a video frame, a web page and a statistic each call for a different method, and each method produces a different kind of result.

Claim text goes first to a database of published checks, whose result shows whether an organization has already examined it. An unfamiliar source is read laterally, by leaving the page and seeing what others say about it. Images and video frames are searched for earlier copies, which shows where a picture has appeared. Archived pages are compared across capture dates, which shows how a page changed, and a statistic is traced back to the table that produced it. Fact-checking tools differ in speed, but none of them replaces the next step, which is reading the result. The six-step process and the resource comparison below cover both.

What Fact-Checking Means and Why Context Matters

The House of Commons Library defines fact-checking as “the process of verifying the accuracy of reporting or statements”, and its briefing on how to check facts adds that this often means tracing a statement to its original source and considering its original context. The second part matters most online, where material travels without the context it started with.

Verification versus opinion. Statistics, definitions, quotations and dates can usually be verified, while opinion and conjecture cannot. A check applies to the factual proposition inside a post, and it also covers the assumptions that proposition relies on, since accurate figures can still be used to mislead.

Misinformation versus disinformation. In the definitions the United Nations attributes to UNESCO, misinformation is false information shared without an intention to cause harm, while disinformation is false information created deliberately to harm. Intent is rarely visible from a post alone, so a check records what is wrong with the content and avoids assuming a motive.

Context. Genuine media becomes misleading when its date, place or caption changes. SIFT, a method by Mike Caulfield, describes much online content as stripped of context and treats restoring it as the goal: stop, investigate the source, find better coverage, and trace claims, quotes and media back to their original context.

The Fact-Checking Process

The process below provides a practical sequence for checking a claim. Each step has an action, an output for the evidence log and a point at which that step can stop.

Isolating the claim. The action is to rewrite the post as one precise statement, with the date, place and units it implies. The output is that sentence, and the step stops when a stranger could say what evidence would prove or disprove it.

Recording source and date. The action is to save the address, the author or publisher, and the publication and access dates, with an archived copy where appropriate. The output is the first line of the log, and the step stops when the page could vanish without losing the record.

Investigating laterally. The action is to look at who published the claim, which is the “investigate the source” move in SIFT, and at what stronger reporting says about it, which is the “find better coverage” move. PolitiFact’s checklist adds whether the claim has been checked before and what evidence its maker offers. The output is a short note on the publisher and any earlier checks, and the step stops when the claim’s origin is identified or shown to be unclear.

Tracing original evidence. The claim is followed to its first source: the dataset, the document, the full interview or the earliest capture of the image. The original source is inspected wherever possible, then assessed for reliability and for whether it supports the claim. The output is a link to the original with the relevant excerpt, and the step stops when the nearest source is no longer a copy of another.

Corroborating independently. Evidence that does not depend on the first source is sought, which means a different observer or record and not another repost. The Commons Library recommends using at least two sources where possible. Each line of evidence is assessed for reliability, relevance and consistency, and material contradictions are investigated. The output is that assessment, and the step stops when the evidence supports a clearly bounded conclusion, or when what remains unresolved has been stated.

Documenting the conclusion. The conclusion states what the evidence supports, contradicts or leaves unresolved, and it separates fact from inference. The output is that statement plus the next evidence needed, and the step stops when the claim is classified as corroborated, contradicted or unresolved, with the reason.

An evidence log holds the record across all six steps. Each entry has five fields: the claim; the source and capture details, including the address and access date; the action taken and what it found; the assessment; and the next step, which is either an independent source still needed or a clear reason for stopping. Teams that check many claims at once can see how OSINT teams triage and scale verification. These steps apply to any medium, and the techniques below adapt them to text, images and video.

Fact-Checking Techniques for Text, Images and Video

Tracing quotations and statistics. A quotation is searched in quotation marks and followed to the full interview, speech or document. A statistic is followed to the table that produced it, where the unit, the denominator and the period are stated. A match shows that the words or numbers exist at a source, and the surrounding text shows whether they were used accurately. The Commons Library briefing also advises searching named people and organizations to confirm context and attribution.

Reverse-searching images. A reverse image search looks for other places where the same or a similar picture appears, which helps establish whether it has been used before. Different engines index different parts of the web, so one search is a sample and not a census. A match identifies another location where the image appears. Establishing when it appeared requires checking the page, an archive capture or another reliable record, and the earliest copy found is not necessarily the original.

Extracting representative video frames. A video is searched through its parts. Still frames from distinct scenes, including the opening, any change of location and any visible text, are pulled out and searched as images, and the full clip is compared with any longer or earlier version. A short clip may be cut from a longer one, so a missing earlier version is not evidence that none exists.

Comparing location and time clues. Signage, terrain, building styles, shadows and weather can be compared with maps, historical street imagery and weather records to test the claimed place and time. These clues narrow the possibilities, and they are strongest when several of them agree.

Missing or unusual metadata has the same limit as a visual match. Metadata can be edited or stripped, so its absence is a gap to document and not a finding about the image. Taken together, these digital fact-checking methods each produce a lead that still needs corroboration.

Fact-Checking Websites and Verification Resources

Dedicated sites collect the checks that professional organizations have published, and they save time when a claim has already been examined. Their coverage is finite, however. The Duke Reporters’ Lab 2026 census counted 437 active fact-checking projects worldwide as of 1 June 2026, down about 4 percent since the previous June. Not every claim has a published check, and the other resources take over from there.

The techniques above rely on a small set of resources. Each was checked against its official documentation in October 2026, and access terms are included only where the documentation states them. Where it is silent on fees or sign-in, the article says nothing about them, and they should be tested before a tool is relied on. A result is a lead with a defined meaning, so each entry below gives the tool’s best use, its limit and the next check.

Google Fact Check Explorer. Best for finding out whether a claim has already been examined. The input is a keyword or an uploaded image, and the result is a list of published fact-checks by recency, each with the checking organization and its rating. Explorer displays a publisher’s assessment and does not verify the claim itself, so the next step is to open the underlying check and examine its evidence, scope and date. The Google News Initiative lesson describes both search options.

Google Lens. Best for finding other places where an image appears. The input is an uploaded image, an image address or an image selected in Chrome, and the result can include similar images and websites that contain the image or a similar one. A result identifies another page where the image appears, not a creator or a date, so the next step is to open the earliest results and read their captions and dates against another record. Google’s help page notes that it may store the addresses searched.

TinEye. Best for sorting matches by when TinEye first indexed them. The input is an uploaded image, a pasted image or an image address, and the result is matches from its own index, which can be sorted by best match, oldest or newest. The TinEye guide states that a search with no results means no match in its index, not that none exists elsewhere. As with other fact-checking tools, an empty result is a gap to document and not a finding. The oldest result is the oldest one indexed and not necessarily the original, so it still needs checking against a source that dates the picture.

Wayback Machine. Best for seeing how a page looked on a given date. The input is a web address, and the result is captures made by web crawls on specific dates. A capture preserves what the archive collected, not whether it was true or complete. The Internet Archive’s documentation states that pages may not be archived when blocked by robots exclusions or excluded at a site owner’s request, that pages behind passwords are not collected, and that dynamic pages may not work fully. The next step is to compare captures from before and after the claim appeared and to check that the relevant text or image was preserved.

ExifTool. Best for reading what a file records about itself. The input is an image, video or other file, and the result is embedded metadata such as EXIF, GPS, IPTC and XMP fields. The official site describes a library and command-line application for reading, writing and editing metadata on Windows, macOS and Unix systems. Because the same tool can write these fields, a value shows what the file records and not what happened when it was made, so the next step is to compare it with evidence from outside the file.

OSINT Fact-Checking and Open-Source Evidence

Public evidence versus open-source software. The phrase open source fact-checking is ambiguous because three different things share it. Public access means anyone can see the material, free access means no fee is charged, and open-source software means the code can be inspected and modified. A public record can sit behind a paid service, and a free tool can be closed-source. Here, open-source evidence means material that anyone could in principle obtain, whatever software was used to collect it.

Independent corroboration. Public records, satellite imagery, company registers, court filings and archived pages can each be checked without the claimant’s cooperation, which is what makes open sources useful for corroboration. The test is independence. Ten pages repeating one wire report amount to one line of evidence. Separate observers count separately only if each saw the event or held the record directly, since an observer who repeats another’s account or shares a common source does not corroborate the event independently, although the repetition can show how a claim spread.

Preservation. Online material changes and disappears, so evidence is documented at the time of observation, with the capture date and the address recorded. Missing addresses and capture dates make evidence harder to trace and place in time, so both are recorded, together with the archive address and the access date. The scope here is verifying claims, not profiling the people who make them.

A Scenario: Digital Verification in Practice

This scenario is fictional and built for illustration. No real photo, post or city is involved, and it is not presented as a completed investigation.

A post shows a flooded underpass in the invented city of Marlow, captioned “Flooding in Marlow this morning.” The six steps run as follows.

Isolating the claim. The caption becomes one statement: the photo was taken in Marlow on the morning of the post. An earlier appearance of the same photo would disprove it.

Recording source and date. The post’s address, account name and publication time are saved, along with an archived copy and the highest-quality version of the image.

Investigating laterally. The account has no history of local reporting, and no Marlow outlet or city channel mentions flooding that day. That absence is a lead, since it does not show that nothing happened.

Tracing original evidence. A reverse image search returns a copy in a local news gallery, captioned with a different street name. An archive capture of the gallery from three years earlier contains the same photo, confirmed by matching signage and a parked van. That is the earliest copy located, not necessarily the original.

Corroborating independently. The regional weather service’s rainfall records for the post date show dry conditions in Marlow. That fits the earlier-copy finding without proving it, since a burst water main can flood an underpass without rain. The gallery’s different street name also conflicts with the post, and the gallery’s own accuracy has not been checked.

Documenting the conclusion. The claim that the photo was taken this morning is contradicted, because a dated capture shows it online years earlier. Who took the photo and where stay unresolved. The next step in the log is to ask the gallery’s editor or photographer to confirm the original location.

Common Verification Mistakes and Best Practices

Circular sourcing. A claim that appears on many sites may have a single origin. Syndicated reports, quoted posts and articles that cite each other form a loop, and a loop is not corroboration. Following each citation back until it reaches an original observation breaks it.

Treating silence as an all-clear. Fact-checking websites cover only the claims that someone chose to examine, so the absence of a published check says little about a claim. The next step is the six-step process, not a conclusion.

Overconfidence in AI and metadata. Weekly use of AI chatbots for news rose from 7% to 10% in the Reuters Institute’s 2026 survey, and 42% of those users say they always or often click through to original sources. That figure is self-reported behavior and does not show whether they verify what they find, since opening a citation is only the start: the reader still has to check that the source supports the claim and that the answer represents it accurately.

AI-generated content and deepfakes are also hard to settle by detector score alone, so a score is a lead to corroborate with provenance, and specialist help is appropriate when the stakes are high. Metadata carries the same warning, since it can be edited or removed.

Privacy and corrections. Examples and screenshots are redacted of irrelevant personal details, because the face or address of a private individual is rarely needed to test a claim. Conclusions are updated when the evidence changes, with the correction recorded in the log and stated where the earlier conclusion was published.

The Takeaway

Fact-checking tools return results, and reading those results is the work. The next action is the same for every claim: save the source, choose the method that fits the input and record what the result shows and what remains unresolved.

Fact-checking techniques differ by medium, but the conclusion takes the same shape. A claim that ends as corroborated, contradicted or unresolved, with a log that lets another person repeat the steps, is more useful than a confident answer that cannot be audited.

FAQ

Which fact-checking tools should beginners use?

Beginners can start with three: Fact Check Explorer for published checks, a reverse image search such as Google Lens or TinEye for pictures, and the Wayback Machine for pages that have changed. Each answers a different question, so the choice follows the input, as the first section shows.

How can a claim be checked when no fact-check article exists?

The same six steps apply without a published check. The claim is isolated, its source and date are recorded, its origin is traced and independent evidence is sought. If the independent evidence does not support a clearly bounded conclusion, the result is reported as unresolved, with the next source needed.

How can an image or video shared on social media be verified?

The post’s address and date are saved, the image or representative video frames are searched for earlier copies, and the caption, place and time clues are compared with independent evidence. A match or missing metadata is a lead, not a verdict. The worked examples show the log format.

What is the difference between OSINT and open-source software?

OSINT is a practice: evaluating publicly available information to answer a question. Open-source software is code that can be inspected and modified. OSINT can be done with closed-source tools, and open-source software has many other uses, so the two terms describe different things.

Can AI reliably decide whether information is true?

No. AI tools can surface candidate sources, but detector scores and chatbot answers can be wrong or impossible to verify. A conclusion still needs provenance and independent corroboration, and a chatbot’s citations need to be opened and read like any other source.

Want to see how OSINT can help investigators trace a claim to its origin, compare evidence across sources and document what remains unresolved? Book a personalized demo to explore how SL CrimeWall brings public-source data, relationship analysis and documented findings together in one platform.