How can I tell if an image was generated by AI?

The most reliable way is to read the image's provenance metadata. AI tools embed standardised markers — IPTC DigitalSourceType, C2PA Content Credentials, or a generator name in the file. This tool reads all three and shows the exact evidence. The catch: those markers only survive if the file is untouched, so a missing marker does not prove a human made the image.

What is the IPTC DigitalSourceType "trainedAlgorithmicMedia" tag?

It is the official IPTC marker for media created by a generative model trained on sampled content — the standard value for an AI-generated image. Google, Meta, and Adobe write it. Related values are compositeWithTrainedAlgorithmicMedia (an AI element blended into a real photo) and algorithmicMedia. When this tag is present, the image is AI-generated with high confidence.

What are Content Credentials (C2PA) and how do I read them?

Content Credentials are a tamper-evident record of how an image was made, defined by the C2PA standard and embedded as a signed manifest inside the file. They name the tool that created or edited it (claim_generator) and the source type. This tool detects the manifest and reads the generator string; for the full signed history you can also use the official verify tool at contentcredentials.org/verify.

Does removing metadata hide that an image is AI-generated?

Yes — and that is the honest limitation of any metadata-based check. Taking a screenshot, re-saving in a basic editor, or uploading to most social platforms strips the provenance markers. After that, this tool will report no markers found, even for an image that really was AI-generated. Absence of a marker is never proof of human origin.

Can you detect AI images without looking at the metadata?

There are pixel-forensics classifiers that guess from the image content, but they are probabilistic, frequently wrong on compressed or edited images, and cannot be cited to any standard. This tool deliberately avoids them. It reports only what the standardised provenance markers say, so every verdict is backed by a citable source rather than a black-box guess.

Is my image uploaded to a server?

No. The whole check runs in your browser using JavaScript. The file is read into memory, scanned for metadata, and never sent anywhere. You can confirm this by opening your browser's network tab while you run a check — there is no upload request.

Which file types and generators does it recognise?

It reads JPG, PNG, and WebP up to 25.0 MB. Generator signatures cover Adobe Firefly, OpenAI DALL·E, Midjourney, Stable Diffusion / ComfyUI, Google Imagen and Gemini, Leonardo.Ai, Ideogram, FLUX, and more. For Stable Diffusion PNGs it also reads the embedded generation-parameters block (Steps, Sampler, CFG scale).

Why does a real photo sometimes show Content Credentials?

Newer cameras and editing apps attach C2PA manifests to genuine photos to prove authenticity, not to flag AI. When the manifest's source type is a capture rather than AI, this tool reports “Provenance present — source is not AI”. That is a positive trust signal: the image's origin is documented.

AI · Image provenance

AI Image Detector — Is This Photo AI-Generated?

Drop an image to check it for AI-generation markers — C2PA Content Credentials, IPTC DigitalSourceType, and known generator metadata. It reads provenance, not pixels, and runs entirely in your browser, so the file never leaves your device. A missing marker is not proof an image is human-made.

By Induwara Ashinsana— Executive Director, Ryzera TechnologiesUpdated Jun 7, 2026

Check an image for AI markers

Files stay on your device

Add an image to checkDrop an image here, or click to choose oneJPG, PNG, or WebP · max 25.0 MB

Everything runs in your browser. The image is never uploaded to any server.

What this does

Drop a JPG, PNG, or WebP above and this tool reads its provenance metadata — the standardised markers that tools like Adobe Firefly, DALL·E, and Midjourney embed to label AI-generated images: IPTC DigitalSourceType, C2PA Content Credentials, and generator signatures. You get a clear verdict and the exact evidence, each linked to its standard. It checks metadata, not pixels — so a missing marker is not proof an image is human-made. Nothing is uploaded.

How it works

This tool answers one question — does an image carry a standardised marker that says it was made by AI? — by reading the file's embedded metadata. It runs no neural network and looks at no pixels. Pixel-level "AI or not" classifiers are probabilistic and cannot be cited to any authority; provenance markers, when present, are unambiguous and traceable to a published standard. The deterministic steps are:

Sniff the container. The format is identified by magic bytes — not the filename — so a renamed file still reads correctly: FF D8 FF for JPEG, 89 50 4E 47 for PNG, and a RIFF…WEBP header for WebP.
Extract the fields. EXIF, XMP, and IPTC are parsed with the exifr library. PNG text chunks (tEXt/iTXt/zTXt) are walked by a byte-level scanner, and a JUMBF signature scan detects a C2PA manifest anywhere in the stream.
Match against the provenance tables. Three rules, in order of authority:
- IPTC DigitalSourceType in trainedAlgorithmicMedia, compositeWithTrainedAlgorithmicMedia, or algorithmicMedia → AI, high confidence.
- A generator signature — C2PA claim_generator, XMP CreatorTool, EXIF Software, or a Stable Diffusion PNG parameters block — matching a known tool (Firefly, DALL·E, Midjourney, Stable Diffusion, and others) → AI, high confidence.
- A valid C2PA manifest whose source type is a real capture → provenance present, not AI.
Derive the verdict. The highest-confidence match wins; every piece of evidence is listed as Standard → Field → Valuewith a link to its citing standard. When nothing matches, the verdict is "no markers found" alongside the explicit caveat that this is not proof of human origin.

The mapping tables are a maintenance item, last cross-checked against the IPTC and C2PA specifications on 2026-06-07. The three worked examples below are reconciled by a self-test in the source module so the matching rules cannot silently drift.

Worked examples

Adobe Firefly JPEG

AI provenance detected

Format sniffed by magic bytes: FF D8 FF → JPEG.
C2PA manifest found in an APP11/JUMBF segment; claim_generator = “Adobe Firefly 1.0”.
XMP packet carries IPTC DigitalSourceType = trainedAlgorithmicMedia.
Rule 3 fires twice: IPTC authoritative marker + generator signature.
Highest signal = AI marker → verdict AI (high confidence), 2 evidence rows.

Stable Diffusion PNG

AI provenance detected

Format sniffed: 89 50 4E 47 → PNG. No EXIF, no C2PA.
PNG tEXt chunk with keyword "parameters" found.
Text contains Steps: 30, Sampler: DPM++ 2M, CFG scale: 7, Model: sd_xl_base.
Two characteristic tokens present → recognised as a generation-parameters block.
One generator-signature evidence row → verdict AI (high confidence).

Plain phone photo

No AI provenance markers found

EXIF Make = samsung, Model = SM-S911B, Software = S911BXXU…
No IPTC DigitalSourceType, no C2PA manifest, no generator string.
No rule in the matching step fires → 0 AI markers.
Verdict: none. Camera EXIF shown for context.
Caveat surfaced: metadata may have been stripped, so this is not proof of human origin.

Frequently asked questions

Sources & references

The marker tables on this page were last cross-checked against the IPTC and C2PA specifications on 2026-06-07. They are reviewed whenever IPTC revises the DigitalSourceType vocabulary or a major generator changes the metadata it embeds.

Related tools

LiveAI

Language Detector

Detect the language of any text snippet — Sinhala, Tamil, Hindi, English, Chinese, and 100+ more. Unicode-script analysis plus a server-side tri-gram classifier. No API keys, no signup, sources cited.

Open tool

LiveAI

AI Grammar Checker

Paste any English paragraph and get a corrected version side-by-side with the original — each edit highlighted and labelled. LanguageTool engine server-side, deterministic rule baseline cross-check. No upload, no API key, no signup.

Open tool

LiveAI

AI Similarity Checker

Compare two texts and get a 0–100% semantic similarity score from sentence-transformer embeddings — catches paraphrases that word-overlap checkers miss. Server-side inference, lexical Jaccard cross-check, no signup, sources cited.

Open tool

Rate this tool

Be the first to rate

Comments & feedback

Spotted a bug or want an improvement? Tell us — our team reviews every comment, and good ideas get built. Comments are public and anonymous.

Found a generator we don't recognise yet, or a file that reads wrong?

Email me at [email protected] — most fixes ship within 24 hours.