Can You Trust an AI Image Detector? How to Assess a Suspicious Picture

Can You Trust an AI Image Detector? How to Assess a Suspicious Picture

Use an AI image detector as a reason to investigate, not as your final verdict. Our recommended approach is to combine the result with the original source, available image history, and evidence for the accompanying claim. If those checks leave a gap, record it as unverified.

This guide is for editors, creators, and readers deciding whether to publish or share a picture. It offers a review process, not a ranking of detectors or a claim that we tested their accuracy.

What does an AI image detector actually tell you?

An AI image detector attempts to classify content as synthetic or non-synthetic, often returning a score. NIST explains that performance depends on the detector, the evaluation images, and factors such as compression and image style. A detector trained on one generator may perform less well on another; resizing and compression can further complicate detection. Its report also distinguishes false positives from false negatives: incorrectly flagging non-synthetic content and missing synthetic content. These limitations are discussed in NIST’s report on synthetic-content transparency, section 3.2 and Appendix E.

Before interpreting a score, ask the provider what it measures. Does the tool explain its threshold, supported image types, and evaluation method? Does it distinguish a fully generated image from a photograph with an edited region? If the documentation leaves those questions unanswered, record the result without translating it into a confident accusation.

Start with the claim, not the pixels

Write down the exact statement you need to check. “This image shows today’s storm at the harbour” presents several separate questions: which harbour, what date, who captured it, and whether the image was generated or altered. Resolve each question separately.

A working checklist for deciding what to investigate

Question	Next check	Record
Where did this copy come from?	Open the post and follow its attribution.	Post URL, account, caption, access date.
Is there an earlier version?	Search by image and inspect matching pages.	Candidate sources and publication context.
What history is available?	Inspect the file’s Content Credentials, if present.	Validation status and specific recorded actions.
What did a detector report?	Read its documentation alongside the result.	Tool, date, file tested, score and limitations.
Is the caption supported?	Seek evidence for the named place, date and event.	Corroboration, contradictions and unresolved gaps.

1. Preserve the copy and trace its source

Save the post URL and its full caption before investigating. Keep the received file unchanged, and use a separate working copy for any crop or annotation. Ask the uploader for the original file and where they obtained it. Treat their answer as something to check, rather than the end of the inquiry.

For reverse image search, Google supports uploading an image, pasting an image URL, or searching a web image with Lens. Results can include similar images and websites containing the image or a similar one. Follow Google’s instructions for searching with an image.

Open promising results instead of relying on thumbnails. Compare the framing, the page’s caption, and any attribution. Label a match “earliest version found in this search” unless you have evidence that it is the original. If the search produces no useful match, leave the source question open.

2. Read Content Credentials as a history record

Content Credentials use signed, tamper-evident structures to record provenance. They can describe creation tools and subsequent changes. The C2PA FAQ also explains that credentials can become separated from a file and that some mechanisms can help rediscover them. Validating one asset does not necessarily reveal the complete history of every ingredient used to make it.

Use a Content Credentials-aware viewer and record exactly what it reports: the validation result, signing implementation, available creation or editing actions, and any missing history. Distinguish an explicit generative-AI action from a generic edit entry. Do not replace the viewer’s specific findings with a broad “authentic” label.

The C2PA explainer makes two essential distinctions: credential validation concerns the integrity and association of provenance information, not whether the depicted story is true; and participation is optional, so missing credentials alone should not determine trustworthiness.

For example, imagine a properly documented photograph being shared with the wrong date. Your review still needs to establish when the event happened. Conversely, if a file has no credentials, write “no credentials found in the file checked.” That is a useful observation without an unsupported conclusion about how the picture was made.

3. Use visual inspection to generate questions

NIST lists inconsistent reflections, shadows, and other perceptible features among possible detection cues in its overview of automated detection methods. Use such observations to direct a closer review, rather than treating an odd detail as a verdict.

Write “the reflection needs explanation,” then compare it with the best available file. Ask what evidence would resolve the question. Avoid a checklist that awards points for how strange an image feels.

For a practice exercise, use a photograph you own and keep a labelled edit log. The image editing prompts collection can help you plan a controlled edit. For an intentionally stylized comparison, try the creative double exposure photo prompts. These are practice materials, not detection tools or evidence about someone else’s image.

4. Record the detector result without overstating it

If you use a detector, choose one with documentation you can inspect. Before uploading private material, check its upload and retention terms. Record the exact file tested, the tool name, the date, its output, and its explanation of that output.

If two tools disagree, preserve both results and return to the source investigation. Do not average their scores unless their documentation establishes a valid method for doing so. Likewise, do not keep changing the input until you obtain your preferred answer.

We recommend a simple stopping rule: after the initial source search, credential check, and documented detector check, list the remaining evidence you need. Seek that evidence or retain an unresolved finding.

A worked example: a dramatic storm photograph

Hypothetical example: a post claims that a striking lighthouse photograph shows a storm this morning. A detector flags the downloaded copy, but a reverse image search leads to an older page with a matching image.

First, verify that the match is the same image and investigate the earlier page’s date and attribution. If the earlier publication is established, report that the image predates the claimed event. Keep the creation-method question separate: the date discrepancy does not establish whether the picture was AI-generated.

If the earlier page cannot be verified and the uploader provides no original, an appropriate working note is: “The claimed date and original source remain unverified. A detector flagged the copy tested; that result does not resolve its origin.” No score needs to be converted into a certainty claim.

Keep a short evidence note

Use this template for a repeatable review. For a broader inquiry, the research prompts collection can help organize follow-up questions.

Image or post URL:
Claim being checked:
File received and date saved:
Original source contacted:
Earlier versions found and checked:
Content Credentials findings:
Detector, date, input file and exact output:
Evidence supporting the caption:
Contradictions and missing evidence:
Current conclusion and next action:

Keep the final conclusion no broader than the evidence: an established source, a documented AI editing action, an incorrect caption, or an unresolved origin. A useful review explains what you checked and what remains unknown.