How Journalists Can Verify AI-Generated Images
Source tracing beats pixel analysis. When an image may be AI-generated, find the earliest version you can, check the uploader history, and look for independent confirmation before you analyze anything. A detector can support the story, but it should never be the finding. Report the evidence you can stand behind, and say plainly when the image is unverified.
The workflow below fits a deadline. It runs strongest check first and tells you when to stop rather than push a weak conclusion out the door.
1. Trace the earliest source
Reverse image search with Google Lens or TinEye, then dig into the first accounts that posted the image. Check when they were created, what else they post, and whether real people or outlets follow them. A “breaking news photo” that appears only on week-old accounts is a warning, not a story. Source tracing is usually the strongest evidence you will have, and it survives a screenshot.
2. Check Content Credentials
Ask for the original file, not a screenshot. Inspect it for Content Credentials, the C2PA standard backed by Adobe, Microsoft, OpenAI, and Google, at contentcredentials.org/verify. Images made with ChatGPT include C2PA manifests, and Google’s Nano Banana images carry an invisible SynthID watermark. A valid record is strong evidence about how the file was made. Because screenshots and social uploads strip metadata, a missing record proves little.
3. Run an AI image check
Score the original file and treat the result as a probability. Our free AI image checker runs in any browser with no account: upload a JPEG, PNG, WebP, or GIF and get a 0-100% score with a color-coded verdict. On mobile, AI Detector returns the same score in under three seconds and never stores your images. Both help you move fast on a deadline without uploading sensitive material to a tool that keeps it.
The confidence score gives your image review a clear, recordable result. Learn how it works in Do AI Image Detectors Actually Work? so you can describe a result accurately in copy.
4. Corroborate or hold
Contact the source and ask how the image was made and where it came from. Look for independent confirmation, on-the-ground reporting, or a matching original. If the evidence does not hold, say the image is unverified rather than implying a conclusion. That restraint is the difference between a correction and a scoop.
How to describe a result in copy
Report the result clearly: “AI Detector returned an 87% AI confidence score.” Link to the image source and record the tool used so readers can follow your findings.
For the visual checks that complement this workflow, see the seven-step image check and Is This Image AI?
Keep a verification record with the story
For each image you may publish, keep the original file alongside a short record that another editor can review:
- Claim: the event, location, date, and what the image is being used to establish.
- Source trail: the earliest upload URL, who supplied the file, and any photographer or agency credit.
- File checked: filename, dimensions, whether it is an original or screenshot, and a SHA-256 hash when your workflow supports it.
- Findings: the detector and check date, provenance records, reverse-search matches, and independent corroboration.
- Editorial decision: what you can substantiate, the remaining questions, and the reviewing editor.
Our published file experiment shows why recording the file matters: one Midjourney image returned 99%, 28%, and 44% with the same detector across three versions. For photos sent by readers, use the local-news intake questions before building the source trail.
Common questions
Can a detector be the sole basis for a story about a fake image?
No. Report the evidence you can stand behind: the original source, provenance data, and expert analysis. A detector score is supporting context, not the finding itself.
What is the first thing to do with a suspicious image?
Find the earliest version you can. Compare timestamps, check the uploader history, and run a reverse image search before you analyze the pixels. Source tracing is usually the strongest evidence.
How reliable are Content Credentials?
A valid C2PA record is strong evidence about how a file was made or edited. But screenshots and social re-uploads strip it, so a missing record proves little. Only a present one is informative.
What should I tell readers about a detector result?
Say what the tool found, that it is a probability, and what else you verified. Avoid presenting a single score as proof, and link to the underlying evidence.
A fast check that does not store your files
AI Detector scores an image in seconds and deletes it immediately after processing.
