How to use this check
Get the most original file available
A screenshot or a re-shared copy has usually lost every signal worth reading. Ask the source for the original before doing anything technical.
Validate any C2PA credential
This is the only cryptographic evidence available to you. A verified credential naming a generative source type is close to decisive.
Read the ordinary metadata
Generator names, editing software, and capture data are informative leads, but every field is editable.
Reverse image search
Finding an earlier, higher-resolution, or differently-cropped copy frequently settles origin faster than any watermark check.
Check the context around the image
Who published it, when, alongside what claim, and does any independent source have the same scene from a different angle.
State your uncertainty honestly
If the checks came back empty, say the evidence is inconclusive rather than converting emptiness into a verdict.
Start with the file, not the pixels
The instinct is to zoom in and look for artifacts. That is the least reliable step and it should come last, if at all. The file itself carries far better evidence when it has survived intact.
The single most valuable thing you can do costs no technical skill: obtain the original file rather than a copy. Social platforms re-encode uploads, messaging apps compress, screenshots discard everything. A photograph that arrived through three re-shares has no metadata, no credential, and reduced resolution, which means every subsequent check is running on a degraded sample.
When someone declines to provide an original, that is itself informative. When they provide one, you often get a complete answer in seconds.
What signed provenance can settle
C2PA Content Credentials are the strongest evidence available to a member of the public, because they are cryptographically signed and bound to the file's bytes. If a credential validates and declares a generative digital source type, the image was produced or substantially modified by a generative system, and you can say so with confidence.
Read the recorded actions rather than just the badge. There is a meaningful difference between an asset generated entirely from a prompt and a photograph where a model removed an object, and the actions list distinguishes them where the top-level verdict does not.
The limitation is coverage. Only some generators write credentials, and most platforms strip them. A missing credential is the normal case and carries almost no information.
Why the visual tells stopped working
The famous artifacts — six-fingered hands, garbled text on signs, impossible reflections, melted background faces — were characteristic of a specific generation of image models. Current systems produce these errors far less often, and the ones they do produce are subtler than the checklist circulating online.
Worse, the checklist generates false positives on real photographs. Motion blur produces distorted hands. Compression produces text that looks garbled. Wide-angle lenses produce reflections that look impossible. Unusual lighting produces skin that looks plastic. People have been wrongly accused of AI-generating their own photographs on the basis of exactly these observations.
Some visual signals still carry weight, but they are structural rather than cosmetic: physically impossible geometry that persists under scrutiny, lighting that is internally inconsistent across a scene, or repeated background elements that tile. Even these are hints, not conclusions.
- Unreliable now: finger counts, skin smoothness, 'too perfect' composition, symmetry.
- Still somewhat useful: internally inconsistent lighting, geometry that does not resolve, tiled background detail.
- Never reliable: any single visual tell used on its own.
Why AI-image detectors are not the shortcut
Statistical classifiers that claim to score an image's likelihood of being AI-generated are trained on particular generators and particular image conditions. Their published performance degrades sharply on the images people actually need to check: compressed, cropped, screenshotted, re-shared, or produced by a generator the classifier never saw.
The failure mode is asymmetric and it matters. A false negative means you miss a synthetic image. A false positive means a real photograph — often a person's own work — is labelled fake, with reputational consequences that a probability score does nothing to mitigate.
Confident numbers make this worse rather than better. An output of 'eighty-seven percent AI' reads as a measurement, and people act on it as one, when it is a guess whose calibration does not hold outside the training distribution.
Reverse image search is underrated
It answers a different and often more useful question: has this image existed before, and in what form. An earlier copy at higher resolution usually indicates the version you have is a derivative. A copy in a stock library, an old news story, or a different country's coverage frequently exposes a repurposed image outright.
For claims about current events, this is regularly decisive in a way no watermark check would be. A large share of viral 'AI-generated' images turn out to be real photographs from a different place or a different year, and a large share of viral 'authentic' photographs turn out to be synthetic images that already existed elsewhere.
Run it across more than one engine. Their indexes and crop-matching behaviour differ enough that one will find a match the other misses.
Knowing when to stop
The most common failure in practice is not reaching the wrong answer. It is refusing to accept an inconclusive one and reaching for a tool that will produce a confident-looking number to fill the gap.
If the credential is absent, the metadata is stripped, reverse search finds nothing, and the context is thin, the correct output is: the available evidence does not establish the origin of this image. That is a real finding, and it is far more useful to a reader than a fabricated percentage.
Where the stakes are high — publication, legal process, an accusation against a person — an inconclusive technical result should push you toward non-technical verification rather than toward a more confident tool.
Frequently asked questions
Is there a reliable AI image detector?
Not in the general case. Classifiers work reasonably on images similar to their training data and degrade sharply on compressed, cropped, or unfamiliar-generator images — which is most real-world material.
What is the single best check?
Getting the original file from the source. Every technical signal depends on the file being intact, and re-shared copies have usually lost all of them.
Do AI images always have metadata saying so?
No. Some generators write C2PA credentials or metadata strings, many do not, and platforms strip both routinely. Absence is the normal case and proves nothing.
Can I still trust the hand-and-finger test?
No. Current generators rarely produce that error, and real photographs with motion blur or unusual angles trigger it falsely. It has become a source of wrongful accusations.
What if all the checks come back empty?
Report it as inconclusive. An empty result means the file has been through a pipeline that strips signals, not that the image is authentic or synthetic.
Primary sources
The technical claims on this page follow the published specifications below rather than our own assertions.
- C2PA technical specification 2.3
The normative definition of manifests, claims, assertions, hard bindings, and validation states.
- Content Credentials Verify
The Content Authenticity Initiative's own public inspection tool, useful as a second opinion.
- Google DeepMind SynthID overview
Google's own description of what SynthID marks, which media it covers, and how detection is offered.
- European Commission AI transparency guidance
The EU framework driving machine-readable marking obligations for synthetic content.