Reverse Image and Video Search

After a wildfire in another state, a photo of a burned playground spread through an invented Idaho group called Canyon Neighbors with the caption "This is what happened at Alder Flats last night." Alder Flats is an invented town. The photo was real, and the fire was real. But the fire was somewhere else and the photo was years old. Someone in the group ran a reverse search, found an earlier copy on a news page, and posted the link. The thread died within the hour.

Reverse image search is the single most useful free trick in this course, and it takes about a minute to learn.

How reverse image search works

An ordinary search takes words and returns pages. A reverse image search takes a picture and returns pages that contain the same picture or similar ones. The service has crawled and indexed many images, and it compares yours to its index. Different services compare in different ways, which is why one can find a match another cannot. Some are tuned to find exact or lightly altered copies. Some use machine learning to match objects, places or overall look. Google describes its tool as returning results for objects in the image, similar images, and websites that host the image or a similar one.

Three consequences follow.

• A match means copies exist, not truth. A hundred copies of a false picture are still a false picture.
• Dates come from pages, not pictures. You learn when a page that holds the picture was posted or crawled. The earliest page you find is a ceiling, not a birth certificate. The picture is at least that old.
• Silence is not a verdict. The index leaves out private pages, new pages, and pages that block crawlers.

Step by step with several engines

A privacy note: when you upload an image to a search service, you are sending it to a company. Do not upload private photos of other people, photos of children, or documents with personal data. For those, describe the scene in words or crop the sensitive part away first.

Crop to ask a better question

An engine matches the whole picture, and a busy picture can confuse it. If the first search finds nothing, try cropping.

• Crop to the object. A distinctive building, bridge, sign or vehicle in the corner may match on its own.
• Crop out the overlay. Watermarks, captions and emoji can throw off a match. Crop to the plain picture beneath.
• Crop a text region. If a sign or a banner is readable, search for the words instead. Business names and street signs can point to a place.
• Flip it. People sometimes mirror an image to dodge matching. Flipping it back and searching again is simple. InVID-WeVerify's magnifier tool offers flipping and sharpening.
• Try a larger version. A tiny thumbnail gives engines less to work with.

Avoid cropping out the evidence. If you are examining a picture to decide whether it shows what a caption says, keep the original uncropped copy for reference.

Video: keyframes, thumbnails and the earliest upload

Search engines index still pictures, so for video you pick the stills. A keyframe is a still from a distinct moment. The free InVID-WeVerify plugin, built within an EU research project, extracts keyframes from videos on platforms including YouTube and Facebook, or from an uploaded file, and sends them to several reverse image search engines. It also reads the metadata of an MP4 video and can link to a map when location data is present, and its thumbnails tab searches YouTube thumbnails. Lesson 4 explains why metadata is a weak clue.

Archive tools

The Internet Archive's Wayback Machine keeps dated copies of web pages. Enter a web address, click Browse History, and a calendar shows captures. In an archived address the number is the capture date and time: year, month, day, hour, minute, second (generally read as UTC, though the Archive's help page does not name the zone). You can use it to see when a page first appeared, whether a caption was changed after criticism, and what a vanished page said.

Limits, according to the Archive's own help page: it does not save password-protected pages, pages the owner has excluded, pages blocked by robots.txt or otherwise inaccessible to its automated systems; captures can take 3 to 10 hours to appear; and Save Page Now saves a single page, not a whole site. If a page matters, save it before it changes. And remember a capture only shows what the page said, not that it was true.

Geolocation basics, conceptually

Sometimes a search finds no earlier copy, but the picture itself carries clues about place and time. Researchers call working these out geolocation and chronolocation. You do not need special software to start. Think of it as matching a scene to a map.

• Fixed landmarks. Mountains, rivers, a water tower, a bridge, a grain elevator, a distinctive church. Compare with satellite and street-level maps of the place the caption names. Skylines and ridge lines are surprisingly specific.
• Signs and language. Road signs, store names, phone number formats, license plate styles, currency, units of measure. A sign that says kilometers is not an Idaho sign.
• Vegetation, terrain and weather. Sagebrush steppe and rainforest look nothing alike. Snow in a picture claimed to be from July is a mismatch.
• Infrastructure. Power pole styles, road markings, guardrails and street light designs differ by country and region.
• Shadows and the sun. Shadows point away from the sun. If you know the place, the direction of a shadow narrows the time of day, and its length relative to the object narrows the sun's height. For a 2-meter post casting a 3-meter shadow, the sun's elevation is the arctangent of 2 / 3, about 33.7 degrees. Whether that matches a claimed date and hour at a given place is a calculation that astronomy tables or sun-position calculators can check, and a mismatch is a good reason to look harder.

Two cautions. First, each clue is a constraint, not a proof, and you want several that agree. Second, do not use geolocation to find where a private person lives. This course is about checking public claims. If a picture shows a private home or a person who has not asked to be public, leave it alone.

Questions and answers

Do I need special software?

No. A web browser and a few minutes are enough. The InVID-WeVerify plugin saves time on video but is optional.

Why do different engines return different results?

They crawl different pages and use different matching methods. That is why you run two or three.

What if I only have a screenshot?

Crop away the interface (buttons, usernames) and search the picture itself. Try to find the original post. The screenshot is easy to fake, as lesson 8 explains.

Can I search a face?

Do not use face search tools on private individuals. Public figures can usually be identified from captions. For this course, the question is where a picture came from, not who an unknown person is.

The picture is real, but is the caption true?

Check the caption separately. Compare the date, place and event to the earliest page and to local official sources.

What if the earliest copy is just another repost?

Keep going back. Look for credits (a photographer, an agency), for dated news stories, and for uploads from the person on the scene. Say "earliest I found."

Is geolocating a photo legal and ethical?

Looking at public pictures of public events to check a claim is ordinary research. Using the techniques to locate private people or homes can endanger them. Do not do it.

How do I handle a video with no obvious frames to search?

Pick frames with distinct features: signs, buildings, vehicles, skylines. Try a thumbnail search too. If that fails, look for the account's earlier posts.

Hard questions

If an image has no earlier copy, does that make it likely authentic?

No. A newly taken photo has no earlier copy. So does a newly made fake. The absence of a match is neutral. You need other evidence: the photographer, the people on the scene, the official record.

Can someone defeat reverse search?

To a degree. Cropping, mirroring, color shifts and re-drawing can lower the odds of a match, and a person with bad intent may try. That is why you use several tools and try crops. It is also why a match is strong evidence and a non-match is weak.

How much weight should I give geolocation?

Treat each clue as a constraint. Landmarks plus signs plus terrain agreeing is strong. A single shadow measurement is weak because the object's true height, the camera angle and lens distortion all add error.

What about search engines changing or shutting features?

They do. Features and limits move often. Learn the idea (compare, crop, trace) and then follow whatever the current tool offers, using the vendor's help page.

When should I stop?

When further checking will not change what you say. If you can truthfully write "this image appears on earlier pages and does not show what the post says," you are done. If you cannot reach that, say what you checked and that it remains unverified.