How to tell if a photo was generated or edited with AI. A practical guide for 2026.
AI-made images already fool most people. In this guide we review what to look for, which tools to use and what their limits are, and how Vouch runs every step in a single analysis.
Vouch teamOctober 21, 20268-minute read
In short
- Look at the image for errors in anatomy, physics, text and context.
- Check the metadata and Content Credentials, knowing they may be missing.
- Search for the photo online and in your past cases.
- Use automatic detectors as a lead, not a verdict.
- If the photo decides a payment, combine every signal and record your reasoning.
- Vouch runs all these steps in under 30 seconds per photo and explains the reason behind each result.
Why looking is not enough
accuracy for participants in a study classifying images as real or AI-generated: worse than a coin toss.1
Study published in Communications of the ACM, 2025.accuracy across 287,000 evaluations by more than 12,500 people in a real-or-fake image game.2
Microsoft Research, 2025.of insurance claims professionals feel very confident they can spot a deepfake.3
Survey of 300 claims professionals, Dec 2025–Jan 2026.pieces of content already carry Google’s invisible watermark. It only covers what Google’s own tools create.4
Google, May 2025.Every study reaches the same conclusion: looking closely helps, but it is not enough as the only test.
1. Look closely at the image
Northwestern University researchers grouped the most common errors in AI-generated images into five categories.5 They don’t always show up, but when they do they are the fastest clue.
Anatomy
Hands with extra or missing fingers, teeth, ears or eyes that don’t match, bodies with odd proportions.
Style
Skin or surfaces that are too smooth, a plastic sheen, evenly blurred backgrounds or a sharpness that doesn’t fit a phone camera.
Function
Objects that couldn’t work: unreadable text, number plates or labels with warped characters, handles or cables that merge into other things.
Physics
Shadows pointing in different directions, reflections that don’t match the surroundings, breaks or liquids that ignore gravity.
Context
Details that don’t fit the place or time: sockets, signs, uniforms or packaging from another country.
Partial edits
If only damage was added, look at the edge of the edited area: changes in sharpness, noise or color between that area and the rest of the photo.
These clues help rule out crude fakes. Newer tools make fewer and fewer visible mistakes.
With Vouch: image analysis looks for these signals automatically and highlights the suspicious area, with the reason.
2. Check the file’s metadata
Metadata stores details like the date, the camera model and the software that edited the photo. Some AI tools also label the image as “created using generative AI” with an IPTC standard that Google shows in its results.6
On Windows
Right-click the file, Properties, Details tab.
On Mac
Open the photo in Preview and choose Tools, Show Inspector.
On a phone
Open the photo in your gallery and tap the info icon or swipe up.
What to look for
A date that doesn’t match the event, an editing program in the “Software” field or the AI-generated content label.
Its absence proves nothing: many social networks and messaging apps strip metadata when a photo is uploaded or sent,7 and an AI-generated image may have none at all.
With Vouch: metadata is checked in every analysis, without opening the file, as one more signal in the risk score.
3. Look for Content Credentials and watermarks
Some cameras and AI tools sign their images with the C2PA standard, also known as Content Credentials. That signature records how the photo was created and edited.
Content Credentials
Upload the photo to the Verify tool at contentcredentials.org/verify to see whether it has a C2PA signature and what it says.8
Phones that sign
Some phones already sign photos at capture.9
Google’s watermark
Google marks what its tools generate with an invisible watermark and offers a detector, but it only recognizes content created with Google’s own AI.4
A photo without a signature or watermark is not fake: it simply has no record of origin.8
With Vouch: since most photos carry no signature, the analysis doesn’t depend on one: it looks at the image itself.
4. Search for the photo online
Many fake photos aren’t generated: they’re recycled. A reverse search shows whether the image existed before the event it claims to prove.
Google Lens
Upload the photo or paste its link in Google Images or the Google app.
Bing Visual Search and TinEye
Useful as a second opinion, because each engine indexes different images. TinEye lets you sort results by date.
Crop the photo
Searching only part of it, like an object or the background, finds matches the full image doesn’t.
Your own cases
Search your past claims too: the same photo may show up in another account or another claim.
With Vouch: every photo is searched online and in your previous claims, even if cropped or rotated.
5. Use detectors as a lead, not a verdict
Automatic AI detectors and techniques like error level analysis (ELA) can flag suspicious areas, but they have known limits.
They fail on compressed photos
When an image goes through WhatsApp, social media or a screenshot, it loses the fine traces many detectors rely on, and their accuracy drops.10
They fail on new generators
Detectors work well with the tools they were trained on, but lose accuracy against generators they haven’t seen.11
ELA mistakes texture for editing
Error level analysis is unreliable on images that have been rescaled or compressed several times, and flags edges and highly detailed areas as suspicious.12
A “90% AI” result with no explanation doesn’t work as evidence. What matters is knowing what was detected and where.
With Vouch: the result doesn’t depend on a single detector. It combines three layers, explains the reason and routes doubtful cases to a specialist.
6. Look at the context
A photo never arrives alone. Who sends it, when and why are signals too.
Does it match the story?
The time, weather, place and objects in the photo have to fit what is being claimed.
Who is sending it?
A new account, many claims in a short time or the same photo across several accounts are risk signals.
Ask for another photo
If in doubt, ask for a new photo, taken in the moment, from another angle or with a reference object. It is hard to fake in minutes.
With Vouch: the photo is cross-checked with the customer’s history and claim frequency.
7. Which tool to use
Each free tool solves part of the problem. To review a single photo, combining them by hand is enough. When the photo decides a payment, and hundreds arrive every day, you want one that runs every step together.
| Tool | Detects AI or editing | Searches the photo online | Searches your cases | Uses customer context |
|---|---|---|---|---|
| Reverse image search (Google Lens, TinEye) | – | – | – | |
| Content Credentials (Verify) | Signed photos only | – | – | – |
| Google’s watermark (SynthID) | Google AI only | – | – | – |
| Generic AI detectors | – | – | – | |
| Vouch |
Comparison based on each tool’s main function, as of October 2026.
Vouch combines every step in this guide into a single analysis of under 30 seconds, returns a 0–100 risk score with the reason and leaves only the doubtful cases to people.
8. Checklist: before accepting a photo as evidence
- Look at hands, text, shadows and reflections.
- Check the metadata: date, camera and editing software.
- Verify whether it has Content Credentials.
- Search for the photo online and in your past cases.
- Use detectors as a lead, not as proof.
- Check that the photo matches the story.
- If in doubt, ask for a new photo taken in the moment.
- Record what you checked and why you decided.
How Vouch solves it
Vouch analyzes every photo in under 30 seconds with three layers: detection of images generated or edited with AI, a search for the image online and in previous claims, and the customer’s context. It returns a 0–100 risk score with the reason, and gray-area cases go to a specialist. It works from Vouch Web, with no integration, or via API.
Frequently asked questions
Can you know for sure whether a photo was made with AI?
Not with a single test. People do little better than chance, and detectors fail with compressed photos or new generators. The most reliable approach is to combine image analysis, metadata, reverse search and context.
What is the best tool to detect AI-edited photos?
It depends on the use. To review a single photo, reverse search and metadata may be enough. For photos used as evidence, as in insurance, delivery or e-commerce claims, you want a tool that combines every signal. Vouch analyzes the image, searches for it online and in previous claims, cross-checks the customer’s context and explains the reason behind each result, with human review for doubtful cases.
How do I see a photo’s metadata?
On Windows, right-click, Properties and Details; on Mac, open the photo in Preview and choose Show Inspector; on a phone, use the info icon in the gallery. If the photo went through social media or a messaging app, it may have no metadata.
Do AI image detectors work?
They are useful as a lead, not as proof. They work best with uncompressed images and with the generators they were trained on, and lose accuracy outside those conditions.
What are Content Credentials (C2PA)?
A standard that signs an image with information about how it was created and edited. You can check it at contentcredentials.org/verify. A photo without credentials is not necessarily fake.
How do you verify photos at scale, for example in claims?
With a tool that combines every signal automatically. Vouch, for example, analyzes every photo in under 30 seconds, searches for it online and in previous claims, returns a 0–100 risk score with the reason and routes doubtful cases to a specialist.
Sources
This guide is based on public sources only; every figure can be verified.
- Communications of the ACM. As Good as a Coin Toss: Human Detection of AI-Generated Content (Oct 2025)
- Microsoft Research. How good are humans at detecting AI-generated images? Learnings from an experiment (2025)
- Verisk. AI Editing Tools Are Fueling a New Era of Insurance Fraud, According to New Research from Verisk (Mar 2026)
- Google. SynthID Detector — a new portal to help identify AI-generated content (May 2025)
- Kamali et al., Northwestern University (arXiv). How to Distinguish AI-Generated Images from Authentic Photographs (Jun 2024)
- IPTC. Google AI transparency based on IPTC standards
- IPTC. IPTC study shows some social media networks remove rights information from photos (Mar 2013)
- Content Authenticity Initiative. Using the Verify tool
- Google Security Blog. Pixel and Android: trusted images with C2PA Content Credentials (Sep 2025)
- arXiv. DailyBench: A Unified Benchmark for AI-Generated and Manipulated Images from Modern Generative Models (Jul 2026)
- arXiv. AI-GenBench: A New Ongoing Benchmark for AI-Generated Image Detection (Apr 2025)
- 29a.ch. Image Error Level Analysis with HTML5 (Apr 2012)