How to Tell If an Image Was Made by AI

Image generators have improved enormously, and the old giveaways like six-fingered hands are less reliable than they were. Still, a careful look often turns up clues. No single sign proves anything; look for several together.

• Hands, teeth, and ears: extra or merged fingers, uneven teeth, or earrings that don't match.
• Text and signs: lettering that's garbled, almost-words, or logos that look slightly wrong.
• Background logic: railings that melt into walls, windows that don't line up, or objects that blend into each other.
• Lighting and shadows: shadows pointing in different directions, or reflections that don't match the scene.
• Too-perfect skin and surfaces: a glossy, airbrushed look with no pores, dust, or wear.
• Repeated patterns: crowds where faces look similar, or textures that tile unnaturally.
• Context: the location, weather, or event doesn't match what you know, or the image only appears in reposts with no original source.

A reverse image search through Google Lens or TinEye can help too. If the earliest copy comes from an AI art account or prompt-sharing site, that's a strong clue.

Every image file can carry hidden information called metadata. Real photos from phones and cameras usually include EXIF data such as the camera model, lens, exposure settings, and capture time. (Our guide to what EXIF data is explains these fields in detail.)

AI-generated images often look different on the inside:

• No camera information at all, where a genuine phone photo would normally have it.
• Software fields naming a generator or an AI-assisted editing tool.
• Text parameters such as prompts, seeds, or sampler settings, which some open-source generators write into PNG files.
• XMP or IPTC fields describing the image as AI-generated, which several major generators and stock libraries now add.
• C2PA Content Credentials describing how the image was created or edited.

C2PA stands for the Coalition for Content Provenance and Authenticity, a group that includes Adobe, Microsoft, Google, camera makers, and news organizations. They created a standard called Content Credentials.

Think of Content Credentials as a tamper-evident nutrition label for an image. When a camera, editing app, or AI generator supports the standard, it attaches a signed record to the file saying what made it and what changes were applied — for example, "created with an AI image generator" or "captured on this camera, then cropped in this editor".

The record is cryptographically signed, so if someone alters the image or the label after the fact, the signature no longer checks out. Some news sites and social platforms now show a small "CR" icon you can click to view the credentials.

What C2PA can and can't tell you

• It can show where an image says it came from, signed by the tool that made it.
• It can't prove an image without credentials is fake — most genuine photos online don't have them yet.
• It disappears if someone takes a screenshot or uses software that strips metadata.

You can inspect a file in seconds with our AI Image Metadata Detector. It runs in your browser, so the image you're checking — which might be private — never leaves your device.

1. Save the original image file. Downloading from the source gives better results than a screenshot or a copy pulled from a messaging app.
2. Open the AI Image Metadata Detector and drop in the JPG, PNG, or WebP file.
3. Review the findings: camera fields, software names, generator signatures, XMP or IPTC AI labels, and any C2PA markers.
4. Weigh the evidence together with visual tells and context. A generator name or prompt text is a strong signal; missing camera data alone is not.

If you're about to share your own photos, it's worth doing the opposite check. PrivacyClean strips GPS and device details so your images don't reveal more than you intend.

It's important to be honest about what detection can and can't do.

• Metadata can be stripped. Social networks, messaging apps, screenshots, and simple re-saving often remove it — for both real and AI images. An empty result means "unknown", not "real".
• Metadata can be faked. Anyone with the right software can add camera fields to an AI image.
• Real photos are edited with AI. Phone cameras and editing apps now use AI features to sharpen, relight, or remove objects, blurring the line between "real" and "generated".
• Visual detectors make mistakes. Online tools that claim to recognize AI images from the pixels alone produce false positives and negatives, and can be wrong with confidence.

The safest approach combines several signals: the source, the context, visual clues, metadata, and — where it exists — signed Content Credentials. If the stakes are high, such as a news story or a legal matter, ask for the original file and its source.

When an image is surprising, emotional, or too good to be true, run through these steps before you like, share, or act on it:

1. Pause. Images designed to provoke outrage or excitement spread fastest, and that's exactly when people skip checking.
2. Find the source. Who first posted it? A named photographer, news agency, or official account is far more trustworthy than an anonymous repost.
3. Reverse search it. Google Lens, Bing Visual Search, and TinEye can show earlier copies, other versions, or the AI art community where it first appeared.
4. Look closely. Zoom in on hands, text, edges, reflections, and background details.
5. Check the file. If you can get the original, inspect its metadata and any Content Credentials.
6. Look for corroboration. A real event is usually photographed by many people from different angles. A single dramatic image with no other coverage is a warning sign.
7. Check fact-checkers. Organizations such as AFP Fact Check, Full Fact in the UK, AAP FactCheck in Australia, and US outlets like PolitiFact often debunk viral images quickly.

Why this matters

AI images are used in scams (fake products, fake charity appeals, fake celebrity endorsements), in political misinformation, and in fraudulent insurance or refund claims. They're also used harmlessly in art and marketing. The goal isn't to distrust every picture, but to apply a little more care when an image is being used to persuade you of something.

Can you always tell if an image is AI-generated?

No. The best AI images can fool experts, and metadata is often missing. You can gather strong evidence, but certainty is rare.

Do AI generators label their images?

Many major generators now add metadata labels or C2PA credentials, and some add invisible watermarks. Open-source tools vary, and labels are easily removed.

Does a missing EXIF block mean an image is fake?

No. Most platforms strip EXIF when you upload, so many genuine photos have none.

Is it safe to upload images to an AI detector website?

Only if you're comfortable with a third party keeping a copy. Our detector reads the file locally and never uploads it.

What's the difference between a watermark and Content Credentials?

A watermark is a mark hidden in the pixels themselves; Content Credentials are a signed record attached to the file. Both can help, and both can be lost or removed.