Is this AI image detector free?

Yes. Guests get 3 checks a day with no account; signing in with Google or an email link raises that to 20 a day, still free with no card. Credits on this site are only spent by the generation tools.

How accurate is the AI photo detector?

On an internal set of photoreal AI images from Seedream, GPT Image and Nano Banana against pre-2019 real photographs, the visual layer scored 92–96%. Heavily retouched studio portraits are the weak spot. The metadata layer, when present, is exact. No detector is proof, which is why every result shows its evidence.

What does the AI-likelihood score mean?

It is the model's calibrated estimate, from 0 to 100, that the image was generated or substantially synthesized by AI. Under 35 reads as likely real, over 65 as likely AI, and the band between is marked inconclusive. A declared generator tag pushes the score to 99.

Can the AI image checker tell which generator made it?

Only when the file says so. Content Credentials and generator tags name the tool — OpenAI, Adobe Firefly, Midjourney, Stable Diffusion. From pixels alone the detector reports the type of evidence, not a model name, because pixel-level attribution is not reliable enough to show.

Is my image stored or used for training?

No. Metadata is read inside your browser, a downscaled copy goes to the analysis model, the verdict comes back, and nothing is written to disk or a database. Nothing is used for training.

Why does it say 'inconclusive'?

The model's confidence was below 65%: signals pointed both ways or there was little to go on — tiny images, heavy compression, blurry crops, minimalist scenes. Try the original file at full size and read the signal list; it usually explains the hesitation.

Does it detect AI videos, screenshots or memes?

It checks single images. Screenshots and reposted memes work but carry no metadata, so only the forensic layer applies. For a video, pause on a clear frame with faces, hands or on-screen text, screenshot it, and check that frame.

My own AI image got flagged. Can I make it pass?

That is what the AI Image Humanizer is for: it re-develops a generated image with real skin texture, uneven natural light and subtle film grain while keeping the subject. Run the result back through the detector to see the difference.

REAL OR AI? · FREE · NOTHING STORED

AI Image Detector

Is this image AI? Upload a photo and the AI image detector returns a verdict, an AI-likelihood score and the exact signals behind it — free, in about three seconds, and your image is never stored.

2-LAYER CHECKMETADATA · C2PASKIN · HANDS · TEXTLIGHT · LENS · GEOMETRY~3 S

Drop an image here or click to choose

JPG, PNG, WebP · up to 25 MB · or paste from clipboard

TRY ONE — ALL THREE ARE AI

Analyzed in memory and discarded. We never save your image, and metadata is read in your browser.

WHAT GETS CHECKED

Two layers, one verdict

First the file's own provenance record (Content Credentials, IPTC digital source type, generator tags, EXIF). Then a forensic pass over the pixels themselves. You get a likelihood score, the signals that drove it, and what to do next.

SkinHandsTextLightingLensMetadata

3 free checks a day as a guest

IN SHORT

What is an AI image detector?

An AI image detector checks whether a picture was generated by AI — Midjourney, DALL·E, Flux, Seedream, Nano Banana — or made by a camera or a person. This one reads the file's provenance metadata first, then runs a forensic vision pass over skin, hands, text, lighting and lens signals, and returns a verdict with the evidence.

KEY FACTS

Method: 2 layers: provenance metadata (C2PA, IPTC, EXIF) + visual forensics by a vision model
Speed: About 3 seconds per image
Benchmark: 92–96% on photoreal AI images vs. real photos (internal 24-image set, Oct 2026)
Price: Free — 3 checks a day as a guest, 20 a day signed in
Storage: Nothing stored; metadata is read in your browser
Formats: JPG, PNG, WebP up to 25 MB; paste from clipboard

FORENSIC PASS

What the AI photo detector looks at

Zoom crops from AI images made on this site. Each is one signal the detector weighs — no single one is proof.

AI image detector signal: airbrushed, pore-less skin on an AI-generated portrait — Airbrushed skin
Even, pore-less skin and perfectly symmetric features are the most common AI tell. Real skin has texture, blemishes and uneven light.

AI image detector signal: fingers merging with a cup handle in an AI photo — Fingers that merge
Hands remain hard for generators: extra joints, fingers fusing with objects, and knuckles that bend the wrong way.

AI image detector signal: garbled engraving on an AI-generated perfume bottle — Garbled lettering
Labels, signs and engravings come out as letter-shaped noise. Real text is legible even when it is blurry.

AI image detector signal: too-perfect rim light and shadow on an AI-generated animal photo — Too-perfect light
Rim light with no source, shadows that disagree with each other, and highlights placed for drama rather than physics.

AI image detector signal: generic synthetic bokeh with no sensor grain — Synthetic bokeh
Generic blur with no sensor grain, no chromatic aberration, and depth of field that does not fall off the way glass does.

$ read-provenance image.png

c2pa = manifest found · signed

digitalSourceType = trainedAlgorithmicMedia

Software = Midjourney 7.0

tEXt:parameters = steps 28 · cfg 4.5 · seed 8192

EXIF Make = — (stripped)

→ verdict: AI (declared)

METADATA

What the file admits

Content Credentials, an IPTC AI source type, generator tags or an embedded prompt settle the question before pixels are inspected.

CONTACT SHEET · REAL RESULTS

Sample AI image checker results

Three unedited checks from this detector: two AI images and one human-made graphic, with the score and the signals it reported.

AI image detector result for an AI-generated muscle car photo: likely AI, 72%

AI PHOTO · SEEDREAM

VERDICT · PHOTO

Likely AI-generated

AI LIKELIHOOD72%

GeometryCar body and grille blend oddly; hood scoop and headlight housings look smearedAI
BackgroundGas station sign and pumps are blurred, with unreadable glyphsAI
LensShallow depth of field and motion blur suit a real long-exposure captureReal

Stylized, over-smooth textures, garbled sign details and smeared car geometry typical of diffusion output, though lighting and motion blur are plausible.

HUMANIZED AI PORTRAIT

VERDICT · PHOTO

Likely AI-generated

AI LIKELIHOOD80%

SkinVery smooth, evenly lit skin with little visible pore detailAI
HandsFingers merge ambiguously where they meet the cup handleAI
OtherSteam wisp and cup rim look plausible; no sensor grain visibleReal

Polished, too-clean studio-style rendering with airbrushed skin, fingers merging with the cup handle and generic synthetic bokeh; several weak signals agree.

HUMAN-DESIGNED BANNER

VERDICT · ILLUSTRATION

Likely real

AI LIKELIHOOD20%

TextAll glyphs legible and correctly spelled, consistent letterspacingReal
StyleConventional promotional-banner layout typical of human graphic designReal
BackgroundFlat near-black background, no melted edges or artifactsReal

A designed logo and text banner with clean, legible type and a vector icon; no generation artifacts, so it reads as human-made graphic design.

RESULTS FROM CLAUDE HAIKU 5.5 · 2026-10-08 · SCORES VARY A FEW POINTS BETWEEN RUNS

HOW IT WORKS

Two layers behind every verdict

Metadata is evidence; the visual pass is an expert opinion. The AI image detector shows both so you can weigh them yourself.

LAYER 01 · IN YOUR BROWSER

Provenance metadata

Read from the original file before anything is uploaded.

▸C2PA Content Credentials — the signed manifest OpenAI, Adobe and Google attach
▸IPTC digitalSourceType = trainedAlgorithmicMedia, the field that literally says 'made by AI'
▸Generator tags and embedded prompts (Midjourney, Stable Diffusion, Firefly), plus camera EXIF

RESULT → a definite yes when the file declares itself AI; neutral when metadata is missing, which is normal for screenshots and social downloads.

LAYER 02 · VISION MODEL

Visual forensics

A forensic vision model inspects a downscaled copy and names what it saw.

▸Skin and hair micro-texture, eyes and teeth, hands and fingers
▸Text and logos, background joins, repeated patterns and geometry
▸Light and shadow logic, sensor grain, chromatic aberration, depth-of-field falloff

RESULT → an AI-likelihood score plus the signals behind it; strongest on photoreal portraits and product shots, weakest on heavily retouched studio photos.

WHEN TO CHECK

When to run an AI image check

The questions people bring to an AI photo detector most often.

Dating and profile photos

Waxy skin, perfect teeth and melted earrings are the classic tells. Check a suspicious profile picture before you reply.

Inspecting a marketplace product photo for AI generation with a loupe — Marketplace listings
Rendered products with garbled label text and impossible reflections are common on resale sites. See what the listing photo really is.

Doubting a viral news image on a phone at night — News and viral posts
Social downloads strip metadata, so the forensic pass does the work — hands, signage and shadow logic are where fakes slip.

A creator checking their own AI renders before publishing — Your own AI renders
Creators run images through the detector before publishing. If it still reads as AI, the humanizer re-develops it with real texture and grain.

DEVELOP PROCESS

How to check if an image is AI-generated

01
Upload or paste the image
Drop a JPG, PNG or WebP, or paste from your clipboard. Use the original file when you have it — metadata lives there and is lost in screenshots.
VERDICT · PHOTO
Likely AI-generated
AI LIKELIHOOD72%
- GeometryCar body and grille blend oddly; hood scoop and headlight housings look smearedAI
- BackgroundGas station sign and pumps are blurred, with unreadable glyphsAI
02
Read the verdict and its evidence
You get Likely AI, Likely real or Inconclusive, an AI-likelihood score, and a signal list that names the area, the finding and which way it points.
FLAGGEDHUMANIZED
03
Act on it
A declared metadata source is as certain as it gets. If it is your own image and you want it to pass as a photo, send it to the AI Image Humanizer and check again.
Open the AI Image Humanizer

FAQ