# How to check if an image is AI generated: start here

Source: https://creatisimo.net/topic/how-to-check-if-an-image-is-ai-generated-start-here/
Site: Creatisimo, category Is this AI?
Published: 2026-10-23. Updated: 2026-10-23. Replies: 0.

## Question by Oskar Lund (staff), 2026-10-23

No single check proves an image is AI generated or real. Work in layers: find the earliest copy, look for Content Credentials, run a provider watermark check, read the pixels for visual tells, and only then consult a detector. Treat each result as evidence with a limit, and write "unverified" when the layers disagree.

## The method in seven steps

1. **Find the earliest copy.** Run a reverse image search and look at who posted the image first, when, and with what caption. Most debunks are won here, not in the pixels. A picture with a credible, dated origin is a different case from one that appears with no source.
2. **Check for Content Credentials.** Content Credentials are the user-facing name for the [C2PA](https://spec.c2pa.org/specifications/specifications/2.2/explainer/_attachments/Explainer.pdf) standard: a cryptographically signed record of where a file came from and what was done to it. Drop the file into the [Content Credentials Verify tool](https://contentcredentials.org/verify). If a manifest is present and valid, read who signed it and what actions it lists, including any AI step.
3. **Run a watermark check where one exists.** Google describes SynthID as an invisible watermark added when content is created, designed to survive cropping, filters and lossy compression. Its [SynthID page](https://deepmind.google/models/synthid/) says anyone can check whether an image, video or audio file was made with AI from Google or partners it names, including OpenAI, NVIDIA and Kakao.
4. **Look at the file metadata.** EXIF and IPTC fields can show a camera model, an editing program or a generator name. They are also easy to strip or rewrite, so use them as a hint.
5. **Read the picture.** Check hands and fingers, text on signs and labels, reflections, repeated patterns, shadows that disagree about the light source, and background architecture that does not resolve. These tells fade as models improve, so a clean result proves little.
6. **Use a detector last.** A classifier score is a lead worth following. It is not a verdict (see the table below).
7. **Record a conclusion with a confidence level.** "Traced to the original photographer, credentials valid" is a different sentence from "no provenance, one detector says likely synthetic, unverified."

## What each layer can and cannot tell you

| Check | A positive result means | A negative result means |
| --- | --- | --- |
| Content Credentials | A signed history exists and the file has not changed since signing | Very little. Provenance is optional and the metadata can be removed |
| SynthID or another provider watermark | The file was made or edited with that provider's tools | No watermark was found, which does not clear the image |
| Reverse image search | An earlier source or wider context exists | The image may simply not be indexed |
| EXIF and metadata | Hints about camera, software or edits | Often stripped or edited, so absence is uninformative |
| Visual tells | A specific, nameable flaw | Modern models avoid many of the classic flaws |
| Detector score | A reason to look harder | Not proof the image is a photo |

The C2PA explainer is direct about the limits. It says a missing credential should not make an asset universally less trusted, because adding provenance is optional, and that provenance metadata can be removed, which is why the standard also describes soft bindings such as invisible watermarks and fingerprint lookup. It also states that a Content Credential records provenance and does not rule on whether the content is true. The Content Authenticity Initiative says the same in its [FAQ](https://contentauthenticity.org/how-it-works): credentials are a nutrition label, not a lie detector.

## Why detectors are the last step

Researchers who built the AIGIBench benchmark tested eleven detectors and reported that, despite high accuracy in controlled settings, they suffer significant performance drops on real-world data, including images collected from social media. You can read the abstract of [Is Artificial Intelligence Generated Image Detection a Solved Problem?](https://arxiv.org/abs/2505.12335) yourself. The practical reading: a detector tuned on one set of generators can be confidently wrong on a screenshot, a recompressed repost or a new model.

If you do run one, run two or three, note which ones disagree, and never quote a percentage as if it were the probability that the picture is fake.

## A worked example

Say a picture of a flooded market spreads with no caption. You trace it back and find the earliest copy is a screenshot from a post with no byline, so there is no credential to read. SynthID finds nothing, which only tells you it is not a Google-made image. Visual tells are mixed: the signage is readable, but two awnings repeat. One detector leans synthetic, another says photo.

The honest write-up is: origin not established, no provenance data, mixed signals, unverified. That sentence is more useful to an editor or a client than a coin-flip headline.

## Quick rules to keep

- Absence of Content Credentials is not evidence of AI. Presence of a valid one is useful, but check who signed it.
- A watermark hit is strong. A watermark miss is weak.
- Do not accuse a named artist or photographer on the strength of a detector score. Ask for process files or the original capture.
- Save your steps. If the image is later disputed, a record of what you checked is worth more than an opinion.

The categories on this forum go deeper on single topics: detector behaviour, reading Content Credentials and how invisible watermarks hold up. Post your own image checks and what you found.

---
Cite as: Creatisimo, "How to check if an image is AI generated: start here", https://creatisimo.net/topic/how-to-check-if-an-image-is-ai-generated-start-here/, accessed 2026-10-24.