Esta página solo está disponible en inglés por ahora. El resto del sitio está en tu idioma.

How to Check If an Image Is AI-Generated

2026-08-11 · 7 min

The short version

Open the file and read what is written inside it. Most AI images carry a declaration of their own origin — a signed manifest, a one-line origin tag, or a block of generation parameters. Reading that declaration is fast, deterministic and verifiable.

What you cannot do is look at the pixels and get a reliable answer. No free tool does that well, including this one. The honest workflow is: read the metadata, then reason about the source.

What is actually checkable

An image file is pixels plus everything written around them. Every dependable check looks at the second part.

Evidence Written by What it tells you
C2PA Content Credentials OpenAI, Adobe, Microsoft, Google, LinkedIn, some camera makers A signed manifest naming the tool and the actions applied to the file
IPTC DigitalSourceType Midjourney, Meta A one-line origin tag in XMP: trainedAlgorithmicMedia means fully synthetic
Generator fingerprints Stable Diffusion, AUTOMATIC1111, ComfyUI, InvokeAI, Fooocus, NovelAI Prompt, negative prompt, seed, sampler, CFG scale, model name, in plain text
Camera EXIF Cameras and phones Make, model, lens, shutter speed, aperture, ISO, focal length, sometimes GPS
Container mismatch Nobody deliberately A file named .png that is really a JPEG — a sign it was re-saved or renamed

None of this involves analysing the image content. That is a real limitation, and it is also why the result is trustworthy when it is present: you are reading a string the generator wrote, not a guess a model produced.

Step 1 — read the file

Drop the image into SynthCheck. The file is parsed by JavaScript in your own browser. Nothing is uploaded, there is no server, no account and no file limit. It handles JPEG, PNG, WebP, AVIF, HEIC and TIFF, and it decompresses PNG zTXt and iTXt chunks so compressed prompt blocks are not missed.

Use the original file if you possibly can. A file that has been screenshotted, sent through a messaging app or downloaded from a social feed has usually lost everything worth reading. More on that in why Instagram strips image metadata.

Step 2 — read the result as evidence, not as a score

There are five possible outcomes, and each one is backed by the specific line from the file that produced it.

Result What it means How strong is it
AI-generated The file declares its own AI origin Strong positive evidence
AI-assisted Markers indicate generative editing over other source material Strong, but narrower — part of the image may be photographic
Camera capture evidence Capture fields present, no AI markers Supporting only. EXIF can be forged or copied between files
Conflicting signals Both AI markers and camera fields are present Read the raw evidence. Often AI content pasted into a real photo
Inconclusive No provenance markers survive Proves nothing either way

That last row is the one people misread. Keep going.

Step 3 — verify the signature, if there is one

A C2PA manifest is cryptographically signed. SynthCheck shows you the manifest, the claim generator string and the recorded actions, but it does not validate the signing certificate chain — that requires trust-list checks that belong in a dedicated verifier. For that step, upload the same file to contentcredentials.org/verify.

The distinction matters. An unsigned or unverified manifest tells you what a file claims. A validated signature tells you the claim came from a key that a known issuer controls. See what C2PA Content Credentials are for how the manifest is put together.

The rule that matters most

Absence of an AI marker never proves an image is authentic.

Metadata is trivially removed. It is not encrypted, it is not required, and almost every intermediate step in an image's life destroys it:

  • Taking a screenshot creates a brand-new file with no history at all.
  • Uploading to a social platform re-encodes the image and drops the metadata block.
  • Cropping in a basic editor rewrites the file without carrying anything over.
  • Converting PNG to JPEG discards PNG text chunks entirely.

So "Inconclusive" is the most common result for images found in the wild, and it means exactly what it says: the file carries no origin information. It is an absence of evidence, not evidence of absence.

Why there is no confidence percentage here

Most sites advertising an "AI image detector" run a classifier over the pixels and print a number: 87% AI. Those numbers are usually uncalibrated. The classifier learned from the output of particular models and degrades badly on models it has not seen, on compressed images, and on screenshots. False positives and false negatives are both common, and the number gives you no way to tell which one you have.

SynthCheck does not run a pixel-forensics classifier and does not output a confidence percentage. It reports what is written in the file, quotes the raw field, and stops.

What about SynthID?

SynthID is Google DeepMind's watermark. It lives in the pixel values, not in the metadata, which is why it survives cropping, resizing and re-compression — the very operations that erase everything else.

SynthCheck cannot detect it. Nobody outside Google can, because there is no public detector; releasing one would hand an optimisation target to anyone trying to defeat the watermark. If a file happens to mention SynthID in its metadata, SynthCheck will show you that string, but a text mention is not the watermark and its absence means nothing.

For an actual SynthID check, use Google's own SynthID Detector portal. The full picture is in SynthID explained, and its limits.

When the metadata is gone

If the file is bare, you are no longer doing file forensics — you are doing sourcing. That is a different skill, and it is often more reliable:

  • Reverse image search. Google Images, TinEye, Yandex and Bing index different corners of the web. Run more than one.
  • Find the earliest copy. Sort by date. The earliest posting is often the generator's own portfolio or a prompt-sharing site.
  • Ask for the original. A real photograph usually has a photographer who still holds the RAW file and the burst around it.
  • Check the context. Who posted it, when, and what else has that account posted?

Visual tells — hands, garbled text, over-smooth skin, impossible reflections — were useful in 2023 and are steadily becoming useless. Treat them as a reason to investigate, never as a conclusion.

The baseline is shifting

Under Article 50 of the EU AI Act, which applies from 2 August 2026, providers of generative AI systems must mark synthetic output in a machine-readable way. Penalties reach EUR 15,000,000 or 3% of worldwide annual turnover. Systems already on the market before that date have until 2 December 2026 to comply with the marking obligation, and content generated before 2 August 2026 needs no retroactive labelling.

In practice that means the proportion of AI images that arrive carrying a readable marker should rise sharply. It does not mean unmarked images become rare: locally run open-source models, non-EU services and every screenshot ever taken will keep producing clean files. Details in EU AI Act Article 50 and image labelling.

FAQ

Can you tell if a JPEG from Twitter is AI-generated? Usually not from the file. Twitter/X re-encodes uploads and strips metadata, so the file arrives bare. Try to find the original posting or the highest-resolution copy elsewhere, then check that.

The tool says "Camera capture evidence". Is the photo definitely real? No. EXIF fields are plain text and can be written by anything. A generated image can be given a full set of plausible camera fields in seconds. Camera EXIF is supporting evidence that fits a story; it does not close the case.

Does editing a photo in Photoshop make it show up as AI? Not by itself. An EXIF Software field naming Photoshop tells you the file passed through an editor, nothing more. It is reported separately from AI markers, and generative fill is tagged distinctly when the tool records it.

Which file should I check if I have several copies? The largest and earliest one. Every re-save loses information and none of them add it back. If someone can send you the file straight from the camera or the generator, ask for that.

Comprueba una imagen ahora

Gratis, sin registro, y el archivo nunca sale de tu navegador.

Abrir la herramienta