How to detect AI-generated images
This guide explains how to evaluate AI-generated images without relying on a single visual clue or detector score. The strongest workflow combines provenance, forensic signals, context, and a documented review decision.
Coverage
Text, images, audio, video, and code in one product.
Access
A web checker for direct use and an API for integrations.
Interpretation
Probabilistic evidence that should be reviewed with source context.
First-pass checks
Start with indicators that are fast to inspect and low-risk to document.
- Inspect hands, text, reflections, and object boundaries
- Review metadata and provenance
- Look for texture repetition
- Use detector evidence for ambiguous cases
Detector-assisted review
Use a detector to identify patterns that are hard to inspect manually, then validate the output with source context.
- Check whether evidence is localized or global.
- Compare detector confidence with metadata and provenance.
- Keep borderline cases in a manual review queue.
When not to overclaim
Compression, editing, templates, translation, and human post-production can create signals that resemble synthetic artifacts.
- Avoid public accusations from a single automated result.
- Use confidence bands, not absolute language.
- Document what evidence was present and what was missing.
Common use cases
Newsroom verification before publication.
Marketplace or platform review of suspicious media.
Enterprise fraud and impersonation triage.
Methodology and limitations
How to read a result
Detection output is probabilistic evidence. A high score means the observed signals are consistent with synthetic or manipulated content under the current model and sample conditions. It does not prove authorship or intent.
When review is required
Short samples, heavy editing, compression, translation, re-recording, mixed human and AI content, and new generators can reduce confidence. Use human review before high-impact decisions.
Try ZeroTrue
Run a browser check, inspect the public API example, or create an account to keep results and generate an API key.
Frequently asked questions
Can manual inspection replace detection tools?
Manual inspection is useful, but many synthetic signals are subtle or hidden in metadata, frequency patterns, or frame-level inconsistencies.
What should I do with an uncertain result?
Preserve the evidence, request source material when possible, and route the case to human review instead of making a final claim.
Why do detectors disagree?
Detectors use different training data, features, thresholds, and modality coverage, so disagreement is expected on ambiguous samples.