Disclaimer: This article is for education and for lawful, authorised professional research. Use these methods only where you have a legitimate purpose and a lawful basis, and follow the laws and platform terms that apply to you, including data protection law such as the UK GDPR and EU GDPR. See our Terms of Service.
TL;DR
- To verify AI-generated images with OSINT methods, accept that no single test tells you whether an image is AI-generated. Detectors miss a lot on real-world content, and visual "tells" fade with every new model.
- Work in a fixed order: provenance (C2PA data and watermarks), then reverse image search across several engines, then metadata, then visual checks, then context such as weather, shadows and landmarks.
- Treat AI detector scores as one weak signal. A 2025 benchmark found open-source image detectors lost about 45% of their performance on images circulating in 2024.
- The UserSearch Picture Search type combines reverse image search, geolocation and synthetic-image checks in one place, alongside the rest of your case.
- Your conclusion should be a graded judgement with the evidence behind it, not a yes or no.
When "Looks Real" Stops Being a Useful Test
To verify AI-generated images, you gather independent evidence in a fixed order (provenance, reverse image search, metadata, visual checks and real-world context) and end with a graded judgement rather than a yes or no. Two years ago, many AI images gave themselves away: six fingers, melting text, jewellery fused into skin. Those days are mostly gone. Current image models produce clean hands, legible signs and believable lighting, and the images that reach a newsroom or a claims desk have often been cropped, compressed and re-shared several times, which destroys the fine detail older detection tricks relied on.
So the question has changed. It is no longer "does this look fake?" but "what can I prove about where this image came from, and does its story hold up?" That is a research question, and it has a research answer: a repeatable workflow that gathers independent evidence and ends in a graded judgement.
This guide sets out the workflow we recommend for two groups who face the problem daily: journalists verifying images before publication, and fraud teams checking photos submitted as evidence of loss or damage. We have ordered the steps so the cheapest, most decisive checks come first.
What Is an AI-Generated Image, and What Does Provenance Mean?
An AI-generated image is one created wholly or partly by a generative model from a text prompt or another image. The category runs from fully synthetic scenes to real photos with AI edits, such as added damage, removed objects or a changed number plate. For verification purposes, "partly AI" is often the harder and more important case.
Provenance is the documented history of a piece of media: what created it, and what has been done to it since. The Coalition for Content Provenance and Authenticity (C2PA) publishes an open standard for this. A C2PA-enabled camera, phone or editing app attaches a cryptographically signed manifest to the file, recording the capture device or generator and later edits. Many AI image services now attach these manifests to their output, and some also embed imperceptible watermarks, such as Google's SynthID, that survive light editing.
Provenance data is strong evidence when present. Its absence proves very little, because most platforms strip metadata on upload and most cameras in circulation still do not sign images.
Why Do Synthetic Images Now Reach Every Desk?
For journalists, conflict has become the proving ground. During the 2026 Iran war, fact-checkers found AI-generated war imagery spreading alongside recycled footage. PolitiFact's March 2026 guide to fake war imagery describes AI images of a supposedly sunk US aircraft carrier with vessel details that did not match real photographs, and notes that some images still carried visible Gemini logos. Its advice is the workflow in this article: reverse image search first, look for watermarks, then check the visual details against reality.
For fraud teams, the issue is claims evidence. In February 2026 Insurance Business reported on SAS research showing that generative AI can produce doctored claims images in seconds: adding damage to real photos, altering number plates and removing bystanders. The signs the researchers pointed to were shadows that fall the wrong way, damage inconsistent with the impact, and backgrounds that look unusually clean.
The tooling is moving too. In November 2025 Google brought SynthID checks to the Gemini app, so a user can upload an image and ask whether it was made with Google AI. Useful, but by Google's own description it only covers images made or edited with Google's tools. That is the pattern everywhere: each check covers part of the problem. Teams in newsrooms and claims and fraud prevention need a workflow that combines them.
How Do You Verify AI-Generated Images? A Five-Step Manual Workflow
Always start from the best copy you can get. Ask the source for the original file, sent as a document rather than through a messaging app that recompresses it. Hash it on arrival (sha256sum image.jpg) so you can show later that you analysed exactly what you received.
Step 1: Check Provenance First
Provenance checks are quick and, when they return something, decisive. Upload the file to the Content Authenticity Initiative's Verify tool, which reads C2PA manifests and shows the signer, the tool that produced the image and any recorded edits. For repeatable, offline checks, use the open-source C2PA command-line tool:
brew install c2patool
c2patool image.jpg
c2patool image.jpg --info
c2patool image.jpg -dThe first command installs it on macOS (prebuilt binaries exist for Windows and Linux). Run with just a file path, c2patool prints the manifest store as JSON; --info gives a short summary of whether C2PA data is present and valid; -d prints the detailed manifest, including assertions such as the generator used. If the image may come from Google's tools, also ask the Gemini app about SynthID.
How to read the result: a valid manifest naming an AI generator settles the question of origin. A valid manifest from a camera is strong evidence of a real capture, but check that the scene matches the claim. "No data found" is the most common result and means nothing on its own.
Step 2: Run Reverse Image Search Across Several Engines
Many "AI images" are actually old real photos, and many viral AI images can be followed back to a first poster who labelled them as AI art. Search engines index differently, so use at least three:
- Google Lens is strongest on objects, products and similar scenes.
- TinEye finds exact and near-exact copies. Sort by "Oldest" to find the earliest known appearance.
- Yandex Images is often best for buildings and scenery in Eastern Europe and Central Asia.
Crop the image and search again. Searching only the part that matters (a sign, a building, a vehicle) often finds matches that the full frame misses. Also search for the claimed event with a date filter:
"retail park" flooding before:2026-09-25 after:2026-09-20
site:reddit.com "AI art" OR "midjourney" OR "made with AI" "harbour fire"before: and after: limit Google results to a date window, quoted phrases require exact matches, and site: restricts the search to one platform. The second query looks for earlier posts that labelled a similar image as AI-made.
Step 3: Read the Metadata, and Know Its Limits
Metadata can support a story or contradict it. It cannot prove authenticity, because it is easy to edit. Use ExifTool, available from the ExifTool GitHub repository or your package manager:
exiftool -a -G1 -s image.jpg
exiftool -Software -Make -Model -CreateDate -GPS:all image.jpg
exiftool -XMP:all image.jpg-a -G1 -s prints every tag with its group and short name. The second line pulls the fields most useful here: editing software, camera make and model, creation date and any GPS data. The third shows XMP data, where some AI tools and editors record their name or a "digital source type" value. Things to note: a phone model that did not exist on the claimed date, editing software on a supposedly untouched photo, or no camera data at all on a file the sender says came straight from their phone. Screenshots, messaging apps and social platforms strip most of this, so absence is normal for shared images.
Step 4: Examine the Pixels, Carefully
Visual checks still catch weaker fakes and partial edits. We treat them as a way to decide where to look harder, not as proof. Look at the places models find hard: text on signs and number plates, reflections in windows and puddles, repeating patterns such as brickwork and fences, hands, and the joins between objects. For edited photos, error level analysis and noise analysis in Forensically can show regions that were compressed differently from the rest.
Be disciplined about what these checks can prove. Compression, sharpening and resizing all create artefacts in genuine photos, and new models produce clean images that pass visual inspection. A visual anomaly is a question to answer, not a finding.
Step 5: Test the Context Against the World
This is the step that most often decides a case. An image makes claims about a time and a place, and the world keeps records of both.
- Shadows and sun. SunCalc shows the sun's direction and elevation for any place and time. If shadows point north-east in a photo claimed to be taken at 3pm in London, something is wrong.
- Weather. Check the claimed date against historical records from Meteostat, which publishes weather station data. Dry pavements during recorded heavy rain are a strong signal.
- Landmarks and layout. Compare buildings, road markings, signage and street furniture with street-level imagery from around the claimed date. AI scenes often get the general style right and the specific layout wrong.
- Source history. Who posted it first, what else have they posted, and were they plausibly there?
Where AI Detectors Fit, and Where They Fail
It is tempting to upload an image to a detector and accept the percentage. We see teams do it every week, and we understand why. Resist it. The Deepfake-Eval-2024 benchmark, built from real deepfakes circulating on social media in 2024, found that open-source image detectors lost about 45% of their performance (measured as AUC) compared with older academic benchmarks. Commercial detectors did better, but none of those tested reached 90% accuracy.
There is a second problem: detectors are trained on the generators that existed when they were built. When a new model is released, detection lags until the detector is retrained, and you rarely know which generator produced the image in front of you. In practice we apply three rules. Use detectors only as one signal among several. Never report a detector score as a finding. And be most suspicious of confident scores on heavily compressed images, where detectors are least reliable.
Running the Workflow Inside UserSearch
Done by hand, the workflow above means five or six browser tabs, two command-line tools and a spreadsheet to record what each one said. That is manageable for one image. It gets messy when a claims team has forty photos across twelve files, or when a newsroom needs the same check done identically by whoever is on shift.
UserSearch 2.0 includes a Picture Search type that brings reverse image search, geolocation and synthetic-image checks together. Because it sits inside the same platform as our other 17 Search types, the account that posted an image can be researched in the same session: its username footprint across platforms with Username Intelligence, a linked domain with Domain Intelligence, or a Telegram channel with Chat Messaging.
- Cases in Forensic Mode store the searches and bookmarks behind each judgement, so an editor or claims manager can see how you reached it.
- OneScan runs one input across several data sources you select and merges the results with source attribution.
- SargeBot can draft a PDF report from the case. You verify its content before it goes anywhere.
- Forensic Capture, our free browser extension, preserves the original post with a SHA-256 fingerprint and two independent timestamps before it is deleted.
Our reverse image lookup knowledge base covers the Picture Search options in more detail. As with any source, results are leads to corroborate, and coverage depends on the underlying providers.
Advanced Techniques for Hard Cases
Handle Partial Edits Differently From Full Synthesis
For claims photos, the likeliest manipulation is an edit to a real photo, not a fully synthetic image. Reverse image search is your best tool: find the original listing, dealer photo or earlier social post, and compare. The difference between the two images is the edit.
Ask for a Second Image
A person with a real camera and a real scene can usually produce another angle, a wider shot, or a short video with a reference object in frame. Someone who generated one image often cannot produce a consistent second. This is a simple, fair request for claims teams and for newsroom sources.
Check Series for Consistency
When you have several images of one event, compare them with each other. Generated series drift: a window changes shape, a logo moves, a vehicle gains a door. Real series stay consistent in the details nobody thinks to fake.
Research the Poster, Not Just the Picture
Account age, posting history and links to other accounts often decide a case faster than any pixel. Our Telegram channel research guide shows how to review a channel's history, a common first source for conflict imagery.
Two Verification Scenarios From Start to Finish
Scenario: A Harbour Fire Image in a Newsroom
Context. A fictional regional newsroom receives a dramatic image of a fire at a harbour, sent by @example_handle an hour after reports of smoke.
Actions. The Verify tool finds no C2PA data. Reverse image search across three engines returns no earlier copies. SunCalc shows that the long shadows in the image match early evening, but the reports came in at midday. Meteostat records overcast skies, yet the image shows strong sunlight. The warehouse roofline does not match street-level imagery of the harbour, and a crane in the image has no equivalent on the real quay. The sender's account was created that morning.
Outcome. The newsroom does not publish the image. It runs the story with its own photographer's pictures, and notes in its internal log that a likely synthetic image was circulating, with the evidence saved.
Scenario: Storm Damage Photos in a Home Insurance Claim
Context. A fictional insurer receives a claim for storm damage to a conservatory, supported by three photos.
Actions. Metadata on two photos shows a phone model and dates matching the storm. The third has no camera data and an editing app in its XMP fields. A cropped reverse image search on the third photo finds the undamaged conservatory in a property listing from two years earlier. Forensically shows a differently compressed region exactly where the damage appears.
Outcome. The handler asks the customer for further photos and a video call. The customer confirms the first two photos and withdraws the third, explaining that a relative "made the damage clearer." The claim continues on the genuine evidence, and the file records every step.
Ethical Limits on Image Verification
Verification is about the image and its story. Keep the focus there.
- Verify content, not people. Use context, provenance and public posting history. Do not try to establish who is shown in an image from their appearance.
- Handle submitted images as personal data. Claims photos often show homes, cars and other people. Store them securely, share only what a decision needs, and apply your retention policy.
- Be careful with third-party uploads. Some online detectors keep uploaded images. Check a service's terms before uploading sensitive or embargoed material.
- Word findings carefully. "Consistent with AI generation, based on the following evidence" is honest. "Proven fake" rarely is.
- Know the rules on disclosure. Article 50 of the EU AI Act requires deployers of AI systems that generate deep fakes to disclose that the content is artificial, which shapes how publishers label synthetic material they use themselves.
Frequently Asked Questions About Verifying AI-Generated Images
How do you verify AI-generated images?
Verifying AI-generated images works best as a fixed sequence: check provenance such as C2PA data and watermarks, run reverse image search across several engines, read the metadata, examine the pixels for weak spots, then test the context against shadows, weather and landmarks. The result should be a graded judgement with the evidence behind it, not a yes or no.
Can AI detectors reliably identify AI-generated images?
AI image detectors are one weak signal, not proof. The Deepfake-Eval-2024 benchmark found open-source image detectors lost about 45% of their performance on real deepfakes circulating in 2024, and none of the commercial detectors tested reached 90% accuracy. Detectors also lag behind new image models and are least reliable on heavily compressed images.
What is C2PA and how does it help verify images?
C2PA is an open provenance standard from the Coalition for Content Provenance and Authenticity. A C2PA-enabled camera, phone, editing app or AI service attaches a signed manifest recording the device or generator and later edits. The Content Authenticity Initiative's Verify tool and c2patool read these manifests. A missing manifest proves little, because most platforms strip metadata on upload.
Does image metadata prove a photo is real?
Image metadata can support or contradict a photo's story, but it cannot prove authenticity because it is easy to edit. ExifTool shows camera make and model, creation date, editing software and XMP fields where some AI tools record their name. Screenshots, messaging apps and social platforms strip most metadata, so its absence is normal for shared images.
What is SynthID?
SynthID is Google's imperceptible watermark for images made or edited with Google's AI tools, designed to survive light editing. In November 2025 Google brought SynthID checks to the Gemini app, so a user can upload an image and ask whether it was made with Google AI. SynthID does not cover images from other generators, so it answers only part of the question.
What tools do journalists and fraud teams use to verify AI-generated images?
Common tools for verifying AI-generated images are the Content Authenticity Initiative's Verify tool and c2patool for provenance, Google Lens, TinEye and Yandex Images for reverse image search, ExifTool for metadata, Forensically for error level and noise analysis, SunCalc for shadows and Meteostat for historical weather. No single tool settles a case on its own.
Turn Image Checks Into a Repeatable Process
Verifying AI-generated images is less about spotting fakes and more about building a case: provenance, earliest appearance, metadata, pixels and context, each recorded, weighed and combined into a graded judgement. The tools will keep changing. The order of work should not.
UserSearch puts the pieces in one place: the Picture Search type for reverse image search, geolocation and synthetic-image checks; 17 other Search types for the accounts, domains and channels behind an image; Cases that keep your audit trail; and SargeBot to draft the report you then check. One account gives you 100+ data sources, so your team does not need a separate account with each provider.
Stop guessing. Start researching with UserSearch. Try UserSearch on the next image that lands on your desk.