You can no longer reliably spot an AI-generated image just by looking at it. The best approach is to stack several independent checks: provenance signals such as Content Credentials and Google's SynthID watermark, a reverse image search, the context around the image, and only then the visual details. This guide shows how to run each check and how much weight to give the result.
Key takeaways
- Classic tells like mangled hands and garbled text still help, but newer generators often avoid them. A clean-looking image proves nothing.
- Provenance checks are the strongest evidence when present: Content Credentials (C2PA) show how a file was made, and SynthID marks content from Google's AI tools. When neither is found, that tells you nothing either way.
- Reverse image search is often the fastest way to expose a fake profile, because many scam accounts still use stolen real photos rather than AI faces.
- AI detector websites give probabilities, not verdicts. Compression, screenshots and cropping make them much less accurate, and they can flag real photos as fake.
- For fake profiles, behavior matters more than pixels: account history, consistency, and requests for money or to move off the platform.
Why AI images are hard to spot now
Image generators improve quickly, and each generation fixes the most obvious mistakes of the last. Early tools gave people six fingers, melted earrings and nonsense signs. Current tools produce convincing hands, legible text and realistic lighting much of the time.
People are not as good at this as they think. In a 2024 online experiment run by Microsoft's AI for Good Lab, about 12,500 participants made roughly 287,000 real-or-AI judgments and were right 62% of the time overall, only modestly better than a coin flip. They did best on portraits of people and worst on landscapes and cityscapes.
The practical lesson: treat your eyes as one input, not the judge. A suspicious image deserves the same layered approach journalists use, which is what the rest of this guide covers.
Visual tells (and why they are fading)
Visual inspection is still worth a minute, especially on lower-effort fakes. Zoom in to 200% or more on the original file, not a thumbnail. The FBI's guidance on AI-enabled fraud lists many of the same signs.
| Where to look | What can give it away |
|---|---|
| Hands and feet | Extra or fused fingers, joints that bend wrongly, fingers melting into objects |
| Eyes and teeth | Mismatched pupils or reflections, teeth that blur into one strip, uneven numbers of teeth |
| Ears and hair | Asymmetric or smeared ears, hair strands that pass through skin or vanish into the background |
| Accessories | Earrings that don't match, glasses with mismatched frames, jewelry that merges with skin |
| Text | Signs, labels or badges with warped letters or near-words |
| Background | Melting architecture, people with distorted faces, repeating patterns, objects with no clear edges |
| Light and physics | Shadows pointing different ways, reflections that don't match the scene, impossibly smooth skin |
Why the tells are fading: newer models are trained specifically to fix them, and people who make fakes can regenerate an image dozens of times and pick the cleanest one, or touch up flaws in an editor. So a lack of tells should lower your suspicion only slightly.
Context checks that still work
Context is harder to fake than pixels. Ask:
- Who posted it first, and when? A dramatic "news" image from an account created last week, with no other original photos, is a warning sign. Our guide to finding the original source of an image walks through this.
- Is there a second angle? Real events usually produce many photos from many people. A single spectacular image that nobody else captured deserves suspicion.
- Do the details match reality? Check weather, signage language, uniforms, license plates and landmarks against what you can verify about the claimed place and date.
- Is there a consistent set? AI-generated personas often have a handful of photos with slightly different faces, or photos that are all head-and-shoulders portraits with no friends, pets or everyday clutter.
- What is the image being used for? Requests for money, investment tips, donations after a disaster or urgent emotional appeals are reasons to slow down, whatever the image shows.
The role of reverse image search
A reverse image search can't tell you whether pixels came from an AI model, but it answers questions that often settle the matter:
- Is it a stolen real photo? Many fake profiles still use real people's pictures. If the same face appears on a model's portfolio or another person's social account under a different name, you have your answer. See our reverse image search guide for how to run one.
- Did it appear on an AI art site first? Images that circulate as "real" sometimes trace back to an AI image gallery, a stock site's AI section, or a post where the creator said it was generated.
- Has it been fact-checked? Viral images frequently already have a debunk from a news organization or fact-checker.
Use more than one engine. Google Lens is broad, TinEye is good at exact copies and dates, and Bing sometimes finds social and regional pages the others miss. Our comparison of reverse image search tools explains the differences. A purely AI-generated face often returns no exact matches at all, which is itself a weak signal when paired with other red flags.
Content Credentials, SynthID and "About this image"
Provenance tools look for information embedded by the software or camera that created the image. When they find something, it is the strongest evidence you can get as an ordinary user.
Content Credentials (C2PA)
Content Credentials are a standard from the Coalition for Content Provenance and Authenticity (C2PA). A signed "manifest" attached to the file can record the app or device used, the edits made, and whether generative AI was involved. Adobe tools, several cameras, OpenAI's image tools and Google's Pixel 10 camera app add them.
- Save the original image file (not a screenshot).
- Open the free Content Credentials Verify site at contentcredentials.org/verify, or Adobe's Content Authenticity site and its Inspect tab.
- Drag the file in. If credentials exist, you will see who or what signed them, the tools used, and any AI-generation steps. Otherwise you will see "No Content Credential".
The big limitation: metadata is easy to lose. Screenshots, re-saving, and many upload pipelines strip it. "No Content Credential" is the normal result for most images online and does not mean the image is real.
SynthID
SynthID is Google DeepMind's invisible watermark, built into the pixels (or audio) rather than the metadata, so it usually survives resizing, recoloring and compression. Google's help page says it can stop being detectable after heavy alteration. To check an image:
- Open the Gemini app or gemini.google.com while signed in.
- Upload the image (one file at a time; crop screenshots tightly and don't upload collages).
- Ask something like "Is this AI-generated?" or "Was this created or edited by Google AI?"
A detected watermark means all or part of the image was made or edited with Google AI. No watermark means Google AI didn't make it, but another tool may have. Google's help page notes a limit of about 10 image checks in a rolling 24-hour period. OpenAI announced in May 2026 that it is adding SynthID to images made with its tools, but Google's help page says Gemini does not yet detect other companies' SynthID watermarks, so check for updates before relying on that.
Google's "About this image" and Lens
In Google Lens and Circle to Search, the About this image tab shows how other sites describe and use an image, related news and fact-checks, and creator-added metadata, including fields that indicate AI generation. In May 2026 Google said you can also ask Lens, AI Mode or Circle to Search whether an image was made with AI, which runs a SynthID check, and that Content Credentials verification is coming to Search and Chrome. Rollout varies by country and device, so you may not see every option yet.
AI detector tools and their limits
Many websites claim to tell you whether an image is AI-generated. They train classifiers on known real and synthetic images and return a score. Used carefully, they are one more data point. Used as a verdict, they mislead.
- Scores are probabilities. "87% likely AI" is the model's confidence, not a measured fact about your image.
- Compression hurts accuracy. When Bellingcat tested one popular detector in 2023, it caught every full-size Midjourney image in its sample but labeled 7 of 10 compressed AI images as real, and it flagged some award-winning real photographs as AI. Images shared on social media and messaging apps are almost always compressed.
- They lag new models. A detector trained before a new generator was released may not recognize its output.
- False positives have real costs. Heavily edited, filtered, upscaled or low-light phone photos can trip detectors. Don't accuse someone of faking a photo on a detector score alone.
- Vendor accuracy claims are self-reported. Headline figures usually come from the vendor's own test sets, which may not resemble the images you are checking.
If you use a detector, upload the largest original you can find, try two or three different tools, and treat agreement between them as a weak signal that still needs support from provenance, search or context. Watermark checkers such as SynthID are different: they look for a specific signal the generator deliberately embedded, so a positive result is far more trustworthy than a classifier's guess.
AI faces in fake profiles
The FBI's Internet Crime Complaint Center warned in December 2024 that criminals use generative AI to create realistic photos for fictitious social media profiles in romance, investment and other confidence scams, to send "private" photos that persuade victims they are talking to a real person, and to produce fake IDs.
AI headshots in profiles often share a look: a centered face, soft blurred background, studio-perfect lighting, and no other people. Older face generators also placed the eyes in almost the same spot in every image. Those patterns are hints, not proof, since plenty of real people use professional headshots.
More reliable checks for a profile:
- Ask for something specific and live. A video call is better than another photo, though real-time face-swap tools exist, so watch for lag, blurring at the edges of the face and refusals to turn their head or move a hand across their face.
- Search every photo, not just the first. Scammers often mix stolen and generated images.
- Check consistency. Does the face change subtly between photos? Do claimed job, location and time zone match how and when they message?
- Watch for the script. Fast intimacy, reasons they can't meet, and any mention of crypto, gift cards or a "great investment" matter more than any pixel.
Our catfish check guide covers profile verification in depth, and romance scam help explains what to do if money has already changed hands.
Step-by-step checklist
- Get the best copy. Save the highest-resolution original you can. Avoid screenshots if possible, since they discard metadata.
- Check Content Credentials. Drop the file into contentcredentials.org/verify. Note any AI or editing steps.
- Check for SynthID. Ask Gemini, or Google Lens where available, whether the image was made with AI.
- Look at metadata. Camera make, model and settings can support a claim; generator names or AI source tags undermine it. Remember metadata is easy to edit. See our EXIF metadata guide.
- Reverse search on two or three engines. Look for earlier copies, other names, AI galleries and fact-checks.
- Check context. Who posted it first, when, and does the scene match the claimed place and date?
- Inspect details at high zoom. Hands, teeth, ears, jewelry, text and background.
- Optionally run two detectors. Treat the result as a weak signal only.
- Decide honestly. "Confirmed AI", "confirmed real source", or "unverified". Unverified is a legitimate answer, and it means don't share it as fact or send money on the strength of it.
Frequently asked questions
Can any tool tell me for certain that an image is AI-generated?
Only in one direction. A detected SynthID watermark or Content Credentials that record AI generation are strong evidence an image was made or edited with AI. No tool can certify that an image is not AI-generated, because watermarks and metadata may be absent or stripped.
Why did a detector say my real photo is AI?
Classifiers make mistakes, especially on compressed, filtered, upscaled or heavily edited photos. A single detector score is not reliable evidence. If you need to prove a photo is yours, keep the original file with its metadata and any earlier versions.
Does Instagram or Facebook label AI images?
Meta said in 2024 that it labels images on Facebook, Instagram and Threads when it detects industry-standard AI indicators such as C2PA and IPTC metadata, or when users disclose AI use. Unlabeled images can still be AI-generated, because those indicators are often missing.
Is a reverse image search with no results a sign the photo is AI?
It is a weak signal. A unique AI-generated face often has no matches, but so do many genuine private photos. Combine it with context: a brand-new profile with no matchable photos and a request for money is very different from a friend's holiday snap.
Sources & further reading
- Google Gemini Help: Verify AI-generated images, videos and audio (SynthID)
- Google: New ways to access About this image (2024)
- Content Authenticity Initiative: Using the Content Credentials Verify tool
- FBI IC3: Criminals Use Generative Artificial Intelligence to Facilitate Financial Fraud (Dec 2024)
- Microsoft AI for Good Lab: How good are humans at detecting AI-generated images? (2025)
- Bellingcat: Testing an AI image detector (2023)
Spotted an error or something out of date? Tell us — we review corrections promptly. Read how we research and update guides in our editorial policy.