The 30-second checklist

Run these checks in order. The first four look for records of where and how the image was made; the last two can raise or lower suspicion but cannot settle the question. Each step is explained below.

  • Look for a label or disclosure: an AI label on Instagram, Facebook, TikTok or YouTube, or a caption or credit that says how the image was made.
  • Search for earlier copies with Google's About this image, Google Lens or TinEye to see where the picture first appeared.
  • Check the file for Content Credentials and AI metadata with the checker on this page. The file stays in your browser.
  • Upload it to Google's SynthID Detector at synthid.com to look for an invisible watermark from Google or a company that has adopted SynthID.
  • Zoom in on hands, text, reflections, shadows and backgrounds, and note anything that does not add up.
  • If the answer matters, ask whoever shared the image for the original file and how it was made.

Step 1: Begin with the claim

Write down what you are trying to establish before examining pixels. “Was any AI used?” differs from “Did this event happen?” or “Is this an unedited camera capture?” A generated illustration can be openly disclosed, and a camera image can carry a misleading caption. Clarifying the question prevents a score from answering something it cannot measure. Avoid framing a review as an accusation about a named person.

Step 2: Get the best available source

Prefer the original export to a screenshot, preview or copy pasted into a document. Record the source link and any production explanation. If the image came from a post on Instagram, Facebook, TikTok or YouTube, also note any AI label the platform shows, because it records what the platform detected or was told. Ask the creator whether the file was generated, photographed or edited, and which changes were made. Respect permission and privacy when requesting supporting material; investigating a file's origin does not require identifying the people in it.

Step 3: Search for earlier copies

A reverse image search shows where else a picture, or a visually similar one, has been published. If it appeared years before the event it supposedly shows, the caption is wrong whatever made the pixels. If the earliest copies sit in an AI art community or a generator's public gallery, that is a strong lead. Photos of real news events usually appear in several outlets, often alongside other shots of the same moment, so a dramatic image with no earlier trail deserves extra care. Finding nothing is common for new, private or heavily edited images and proves nothing either way. Three free tools cover most searches:

  • Google's About this image, opened from an image result in Google Search or the Google app, shows when Google may have first seen the image or similar ones, pages that used it much earlier than others and, when the file carries it, metadata saying whether it was captured with a camera or generated with AI. Google says the feature is only available in some regions and that this metadata can be modified.
  • Google Lens finds visually similar images and the pages that use them, which helps when a picture has been cropped, flipped or re-edited.
  • TinEye finds copies of the same image and can sort them by the oldest crawl date. That date is when TinEye found the copy, not when it was published, but it still helps you work back toward the first upload.

Step 4: Check Content Credentials

Open the original file in a C2PA validator, such as the Content Authenticity Initiative's Verify tool, or check it in Gemini, which can read Content Credentials. A validated manifest can record whether a file came from a camera, an AI generator or an editing app. A manifest is meaningful only once its signature and signer are validated, and most shared images carry none, so absence is common and inconclusive.

Step 5: Check for an invisible watermark

Some generators embed invisible watermarks. Google's free SynthID Detector at synthid.com checks an uploaded file for SynthID watermarks from Google and from companies that have adopted it, including OpenAI, NVIDIA and Kakao. OpenAI also runs its own verifier at openai.com/verify, and the Gemini app checks for Google's watermark when you ask “Is this made with AI?”. Each tool only detects the signals it supports, so a negative result rules out nothing beyond that coverage.

Step 6: Read the metadata

A metadata viewer such as ExifTool shows camera fields, software tags and IPTC fields. The IPTC Digital Source Type code trainedAlgorithmicMedia is the standard value for media created by a generative AI model, and compositeWithTrainedAlgorithmicMedia marks AI-assisted edits such as inpainting. Some local generators also save prompts in PNG text. Metadata is easy to edit or strip, so treat it as a clue; our metadata guide covers the details.

Step 7: Zoom in on visual clues

Open the largest version you can find and zoom in on the places generators still get wrong. A guide by Northwestern University researchers, built from nearly 200 real and generated images, groups these artifacts into five kinds, and the first five checks below follow that grouping. Any of them can also come from lenses, compression, editing or an artistic choice, and newer generators make fewer of these mistakes, so treat what you find as a reason to verify, not as a verdict. If the image claims to show a news event or a product, compare it with independent records of that specific subject. The examples after the checks are AI images published on Wikimedia Commons, with the detail to look at outlined.

  • Anatomy: count fingers and check how hands hold things; look for mismatched eyes, ears or teeth, and limbs that merge into clothing, furniture or other people.
  • Style: skin that looks waxy or airbrushed, a glossy, overly cinematic finish, or parts of the picture that are far sharper or smoother than the rest.
  • Function: lettering on signs, labels and clothing that turns into near-letters, and glasses, jewelry, zips or straps that merge or could not work.
  • Physics: shadows that fall in different directions, reflections in mirrors, windows, water or eyes that do not match the scene, and perspective lines that do not meet.
  • Context: details that do not fit the claimed time and place, such as the wrong uniform, signage, vehicle or season.
  • Backgrounds: faces in a crowd that blur together, and repeating patterns such as tiles, fences, railings or leaves that warp or change spacing.
AI-generated reef scene in which a blue fish's tail turns into a second, black and yellow striped fish; the fused part is outlined.
Anatomy: check animals as well as hands. In this Stable Diffusion XL reef scene, the blue fish's tail turns into a second, striped fish. Image: MrAlanKoh with Stable Diffusion XL (Fooocus), via Wikimedia Commons (opens in a new tab). Public domain or CC BY 4.0 (opens in a new tab). Cropped and annotated by Image Evidence.
AI-generated fiery logo that was asked to spell Adobe Firefly but reads ADBE FIRLE; the misspelled lettering is outlined.
Function: lettering is a common giveaway. Asked for a logo with the exact text “Adobe Firefly”, Firefly Image 3 wrote “ADBE FIRLE” (June 2024). Image: Jiří Vedral with Adobe Firefly Image 3, via Wikimedia Commons (opens in a new tab). Public domain. Cropped and annotated by Image Evidence.
AI image titled as an astronaut on Mercury, where the outlined Sun fills half the frame above a wide view of the cratered ground.
Context: this DALL-E 3 image is titled as a walk on Mercury. NASA says the Sun would look more than three times as large there as from Earth, which is still a small disc in a view this wide, not one that fills half the frame. Image: Neriex89 with DALL-E 3, via Wikimedia Commons (opens in a new tab). Public domain. Cropped and annotated by Image Evidence.
AI image of a glass dome over a city on Mars whose frame is a regular triangle grid lower down but turns into uneven, cracked cells in the outlined area near the top.
Backgrounds: follow a repeating pattern across the picture. The frame of this DALL-E 3 dome is a regular triangle grid lower down but breaks into uneven, cracked cells near the top. Image: Raresvent with DALL-E 3, via Wikimedia Commons (opens in a new tab). Public domain. Cropped and annotated by Image Evidence.
Watercolor-style AI image of a rhinoceros in a suit beside a monitor that correctly reads STOCK PRICE CRASH!, with the cleanly lettered text outlined.
No obvious flaw: Nano Banana (December 2025) lettered this screen correctly. Newer models often pass these checks, so a clean look proves nothing. Image: MrAlanKoh with Nano Banana, via Wikimedia Commons (opens in a new tab). Public domain. Cropped and annotated by Image Evidence.

Step 8: Weigh any detector score

If you use an AI image detector, check what it was tested on: generator coverage, transformed inputs, false-positive rate, detection rate and how often it declines to answer. Detectors trained on older generators can miss newer ones, and compression can push scores either way. If independent evidence stays weak, keep the conclusion uncertain and note what would help. Never penalize a person solely because a tool assigns a high number to their image.

Is there a reliable way to tell if an image is AI-generated?

There is no single reliable test. The strongest evidence is a validated provenance record or watermark from the tool that made the file, together with the original source. Visual inspection and detector scores help prioritize questions but are error-prone on their own, especially for edited photos and newer generators.

What are common signs of an AI-generated image?

These are the signs people most often notice. Older generators made them frequently; current models make far fewer, and real photos can show similar oddities from lenses, motion or compression, so use them as reasons to look for the source, not as proof.

  • Hands with extra, missing or merged fingers.
  • Garbled text on signs, labels and clothing.
  • Shadows and reflections that do not match the light in the scene.
  • Waxy, airbrushed skin and an overly polished, glossy look.
  • Earrings, glasses and other accessories that do not match or melt into skin and hair.
  • Backgrounds where people, buildings or patterns blend into each other.

How can I check if a photo is AI on my phone?

Open this page in your phone's browser and choose the photo in the checker above: it reads Content Credentials and AI metadata on the phone without uploading the file. Google Lens in the Google app can search for earlier copies of a photo saved on your phone. Use the original file rather than a screenshot, because a screenshot is a new image without the original's metadata, and expect photos saved from social apps to have lost most of it.

Can AI image detectors be wrong?

Yes. Detectors produce false positives on real photos and miss AI images, particularly images from generators missing from their training data or files that have been compressed, resized or screenshotted. Our guide to detector accuracy explains how to read accuracy claims and why base rates matter.