Guide

How this tool actually works

Unlike the text-based tools on this site, which rely on statistical heuristics like sentence-length variation, this image tool doesn't guess based on pixel patterns at all. Instead, it reads the file's embedded metadata — structured information that cameras, editing software, and AI image generators frequently write directly into the image file itself, separate from the visible pixel data. For PNG files, that means parsing the file's text chunks; for JPEG files, that means parsing the EXIF data block. Both formats store this information in a well-documented, publicly specified structure, which is what makes reading it directly in a browser (with no server round-trip) possible at all.

Why PNG and JPEG are handled differently

PNG and JPEG store metadata in completely different structures, so this tool uses a different parser for each. PNG files store optional text information in labeled chunks called tEXt and iTXt — and notably, several popular AI image generation tools and interfaces (including local Stable Diffusion setups and various generation UIs) write their generation parameters, including the exact prompt used, directly into one of these text chunks by default, under keywords like "parameters" or "Comment." When present, this is about as direct a signal as metadata analysis can provide.

JPEG files instead use the EXIF standard, a structured binary format originally designed for cameras to record shooting details — make, model, exposure settings, timestamp, and (relevant here) the name of any software that last processed the file. This tool reads the EXIF block looking specifically for Make, Model, Software, DateTime, and UserComment fields.

What embedded metadata can and can\'t prove

When metadata does contain a recognizable AI generator signature — a tool name like Midjourney or Stable Diffusion appearing in a text chunk — that's meaningfully stronger evidence than any statistical guess, because it's the software itself declaring what it is, not an inference drawn from analyzing the image. Camera metadata (a Make and Model field, like "Apple" and "iPhone 15 Pro") is similarly direct evidence, though it establishes that a camera was involved at some point in the file's history, not that the entire image is unedited or untouched by AI-based editing tools afterward.

What metadata can't do is offer any signal at all once it's missing — and as covered below, that happens constantly, for images of every origin.

Why metadata so often comes up empty

An empty or inconclusive result from this tool is extremely common, and it's important not to read that as evidence of anything. Metadata routinely gets stripped or never written in the first place for reasons that have nothing to do with an image's true origin:

• Messaging apps like WhatsApp, Telegram, and iMessage strip most or all metadata from images sent through them, by design, for user privacy.
• Social platforms (Instagram, X, Facebook) similarly strip metadata on upload as a standard privacy and file-size practice.
• Screenshots generally carry no camera or generation metadata at all, since they're a rendering of a screen, not an original capture or generation.
• Deliberate stripping tools exist and are trivially easy to use, for anyone specifically trying to remove this kind of evidence.

Because of how common all four of these are, a "no metadata found" result should be read as strictly inconclusive, never as evidence the image is human-photographed.

Detection signals beyond metadata

Metadata analysis is one legitimate signal among several used in image provenance and forensics, not a complete solution on its own. Other approaches — outside the scope of this specific tool — include compression-artifact analysis (AI-generated and heavily-edited images can show statistically unusual compression patterns compared to a straight-from-camera JPEG), physical-consistency checks (unnatural lighting, reflections, or anatomical details, though modern generators have gotten considerably better at avoiding these tells), and cryptographic content-provenance standards like C2PA, which some camera manufacturers and AI tool providers have begun adopting to cryptographically sign an image's creation and editing history. None of these, including metadata analysis, are individually conclusive — provenance is best established by combining multiple signals, not relying on any single check.

Using this responsibly

This tool is well suited to a quick first-pass check — particularly useful when metadata does contain a clear generator signature, since that's a strong, direct finding. It's poorly suited to serving as the sole basis for a confident public claim about an image's origin when the result comes back inconclusive, which, as covered above, will be a large share of real-world cases. Treat an inconclusive result as exactly that: inconclusive, not a green light to assume the image is authentic.

Frequently asked questions