How to spot AI-generated images and deepfakes

No single trick reliably exposes an AI-generated image or deepfake: today’s generators rarely leave obvious glitches, and detector tools can be wrong in both directions. What works is a layered check, starting with the context, then the image’s history, then Content Credentials and watermarks, and only then a close look at the pixels. Here’s how to do each step on your phone or computer, and what each result can and can’t tell you.

Why spotting AI images is harder than it used to be

A few years ago, AI images gave themselves away with six-fingered hands and gibberish signs. Current image models make far fewer of those mistakes, so the old tells are no longer dependable.

The U.S. National Institute of Standards and Technology (NIST) describes automated detection as “a constant cat-and-mouse game” in its report on synthetic content. Detectors often work best on images from the generators they were trained on, and as models improve, people have a harder time spotting fakes too: NIST cites a 2024 study in which people’s accuracy at judging images, video and audio was close to chance.

Two more complications:

• Many images are only partly AI. A real photo can have one object removed or a background extended with AI. Most of it is genuine.
• Real photos can look fake. Phone cameras process every shot heavily, and filters, compression and beauty modes can create the smooth skin and odd edges people associate with AI.

So the goal isn’t a yes-or-no verdict from one glance. It’s to gather enough evidence to decide whether to trust, share or ignore an image.

Step 1: Check the context before the pixels

The fastest checks don’t involve the image at all.

• Who posted it? A brand-new account, an anonymous page or an account that posts nothing but viral shocks deserves suspicion.
• Is anyone credible reporting it? If a photo shows a major event, reputable news outlets should have it too. If they don’t, wait.
• Does it push you to act fast? Urgency, outrage and “share before it’s deleted” are classic manipulation signs.
• Is someone famous endorsing something? The FBI warns that criminals use AI-generated videos of public figures to make fraud schemes more believable. A celebrity pitching a crypto deal or a giveaway is a red flag on its own.

Many deepfake scams work like phishing: the fake is just bait. Our guide to spotting phishing covers the same pressure tactics.

Step 2: Trace the image with a reverse image search

A reverse image search looks for the same or similar pictures elsewhere online. It can reveal an older original, a different caption, the real photographer, or a fact-check that has already debunked it.

In the Google app or on Google Search, search for an image, open it in the image viewer, select About, then Learn more about an image. Google says this can show when it may have first seen similar versions, other pages that use the image, and information about whether it was captured with a camera or generated with AI. Google also cautions that this metadata can be modified, so it may not be accurate, and that the feature is only available in some regions. Google Lens, Bing Visual Search and TinEye are useful alternatives.

What to look for:

• An earlier version with a different date or location
• The same scene from other angles, which is hard to fake consistently
• A version with a visible watermark or caption that someone cropped out

Step 3: Look for Content Credentials and watermarks

Some AI tools and cameras now label their output in ways software can read.

Content Credentials are based on the open standard from the Coalition for Content Provenance and Authenticity (C2PA). They’re a signed, tamper-evident record of how a file was created and edited. Google’s Pixel 10 phones add them to every photo taken with the Pixel Camera app, according to Google’s announcement, and Google Photos shows them in the image details. OpenAI says supported images made with ChatGPT, Codex and the OpenAI API carry both C2PA metadata and an invisible watermark.

Invisible watermarks such as Google DeepMind’s SynthID are woven into the image itself. Google says they’re designed to survive common changes such as cropping, filters and compression.

Here are the free checkers worth knowing, as of 2026:

Tool	What it looks for	If it finds something	If it finds nothing
Content Credentials Verify	C2PA Content Credentials in the file	Shows the recorded origin and edit history, including AI tools if logged	Tells you little: credentials may have been stripped or never added
Gemini app	Google’s SynthID watermark and Content Credentials	All or part was made or edited with Google AI	It could still come from another AI tool
OpenAI Verify	OpenAI’s C2PA metadata and SynthID watermarks	It was generated or exported with OpenAI tools	It could still be AI, including from OpenAI if the signal was lost

To use Gemini, upload the image and ask something like “Was this created or edited by Google AI?” Google’s help page lists limits, including files of 100 MB or less and roughly 10 image checks per 24 hours.

What a missing label really means

Almost nothing. OpenAI notes that metadata can be stripped when a file is uploaded, edited, converted or shared, and that watermarks can degrade with compression, cropping and other changes. A screenshot creates a brand-new file without the original metadata, too. And many cameras and AI tools add no labels at all. The C2PA explainer also warns that provenance information alone can’t tell you whether content is true or accurate: a genuine photo can still be staged or miscaptioned.

What about AI detector websites?

Third-party detectors analyze the pixels and give a probability score. They can help, but they can also miss fakes and flag real photos, especially with images from newer generators. Never treat one score as proof, and never use one to accuse someone.

Step 4: Study the image itself

If the image survives the first three steps and still feels wrong, zoom in on the highest-resolution copy you can find. The FBI’s list of warning signs is a good checklist:

• Hands and feet: extra, missing or fused fingers, or joints that bend strangely
• Faces: unrealistic teeth or eyes, or features that look smeared or irregular
• Accessories: glasses or jewelry that look wrong, such as mismatched earrings
• Shadows: shadows that fall in different directions or don’t match the light
• Watermarks: faint logos or corner marks from an AI tool that someone tried to crop or blur

A few more places where AI often slips:

• Reflections: glasses, mirrors and windows that don’t reflect the scene
• Text: warped lettering on signs, labels and license plates
• Backgrounds: bent lines, objects melting into each other, and blurry faces in crowds
• Texture: skin, hair or fabric that looks airbrushed, plastic or too perfect

Remember that these are clues, not proof. A flawless image isn’t automatically real, and one odd detail doesn’t make an image fake.

Deepfake video and voice: extra checks

Video and audio fakes are especially useful to scammers. The FBI says criminals have used AI-generated voice clips to impersonate people and get into bank accounts. It also lists lag, voices that don’t quite match, and unrealistic movement as signs of a fake video. Watch for lips that drift out of sync, odd blinking, and edges around the hairline or jaw that flicker when the head turns.

Voice cloning is even harder to catch by ear. The FTC warns that a scammer needs only a short clip of a family member’s voice, which may be posted online, to fake a call. So don’t rely on your ears:

• Hang up and call back on a number you know belongs to that person, or check with another relative.
• Set up a family code word. The FBI recommends a secret word or phrase to verify identity in an emergency call.
• Watch the payment method. Requests to wire money, send cryptocurrency or buy gift cards are scam signals, according to the FTC.
• Report it at ReportFraud.ftc.gov, or to the FBI’s Internet Crime Complaint Center at ic3.gov.

The FBI also suggests limiting how much of your image and voice is public, for example by making social accounts private. Our cybersecurity coverage has more ways to lock down your accounts.

The bottom line

You can’t reliably spot AI-generated images by eye alone, and no detector is foolproof. Work from the outside in: check who posted it and why, trace its history with a reverse image search, look for Content Credentials or a watermark, and only then study the details. If money or personal information is involved, skip the analysis and verify the person through a channel you already trust.

If you use AI tools yourself, our guide to writing better AI prompts is a good next step. For more plain-English help, browse our AI coverage or the latest guides on RedAndWhiteMagz.com.