Updated 2026-10-15 after evidence review.
Short answer: 0 auditable head-to-head test records are available for the earlier ranking. Choose an image detector by the kind of media and manipulation you need to review, then check what evidence the service returns. The earlier numerical rankings have been withdrawn pending a consistent, documented test. An automated label should prompt further investigation of the file's source and history rather than settle a consequential decision.
Evidence checks before choosing an image detector
| Check | Why it matters |
|---|---|
| File types | Still images, video and audio call for different tests. |
| Alterations | A face swap is a different task from a wholly generated image. |
| Provenance | Original files and content credentials can provide evidence outside a classifier score. |
| False alarms | Authentic files can trigger automated flags. |
| Review process | A consequential result needs human corroboration and a route to challenge it. |
Choose an image detector by task
Image detection and deepfake detection overlap, but they are not identical tasks. Some services focus on wholly generated still images. Others analyse face swaps, video or audio. Start with the file type and the question you need to answer. A service that accepts a file is not automatically reliable for every generator or alteration.
DeepfakeDetector.AI covers multiple media types. Hive Moderation offers an API-led moderation workflow. Illuminarty focuses on still-image checks with visual output. These are coverage descriptions, not a performance ranking. Check current product documentation for supported files, access requirements and limits.
What the published figures cannot establish
The earlier guide mixed vendor-reported claims with claims of independent tests. We could not establish a consistent dataset, test date or raw result record for a head-to-head ranking. Those numbers have been withdrawn. A comparison should use the same files, versions and thresholds for every tool and separately report false alarms on authentic images.
Keep the original file where possible. Compression, resizing and reposting can remove clues that a service uses. For consequential decisions, inspect provenance and editing history, seek the original source, and treat automated output as one piece of evidence. The C2PA specifications describe a separate approach to recording media provenance. NIST guidance surveys provenance tracking and synthetic-content detection as complementary techniques.
How to choose an image-checking workflow
Start with the media in hand. A still image, altered photograph, face-swapped video and cloned voice can require different analysis. Check that a service accepts the original file type and explains the scope of its output. A tool that produces a confidence label for a still image should not be assumed to verify a video or its provenance.
Next, preserve the file and its history. An original upload, camera file, source URL or known editing chain can be more useful than a detector label. If only a reposted copy exists, note its compression and cropping. Those changes may affect what a model can observe. Compare the automated result with independent evidence rather than running many similar tools and treating their agreement as proof.
For a new performance comparison, use the same documented file set across services and keep each raw response. Include authentic media so false alarms are visible. Report results by alteration type and explain unavailable responses. Until that run exists, this guide describes product coverage and review steps without an accuracy winner.
Frequently Asked Questions
What is the best AI image detector in 2026?
Can AI image detectors detect Midjourney images?
What is the difference between AI image detection and deepfake detection?
AI image detection identifies fully synthetic images generated entirely by models like Midjourney, DALL-E, or Stable Diffusion with no real photographic source. Deepfake detection identifies manipulated media: typically a real photo or video where a face has been replaced or distorted. Both are forms of synthetic media but require different technical approaches. DeepfakeDetector.AI and Hive Moderation handle both. Illuminarty covers AI-generated still images only and should be paired with a deepfake-capable tool for complete coverage.
Are there free AI image detectors?
Yes. Illuminarty offers 5 free scans per day with no account and heatmap output, the most useful free option. AI or Not and Is It AI are both free with no signup. Sightengine provides 2,000 free API checks per month for developers. Free tiers suit individual checks and evaluation; high-volume production use requires a paid plan across all these tools.
How accurate are AI image detectors in 2026?
Which AI image detectors have an API?
Numerical results on this page are withdrawn pending review of the original inputs, detector responses and calculations. Tool descriptions are editorial information, not a verified accuracy ranking.
Questions to ask before trusting a result
Check whether the provider describes the media and alterations its tool was built to analyse. A service may recognise some generated images but have little information about heavily edited photographs. If the file came from a social platform, try to obtain an earlier copy: cropping, compression and format conversion can change the evidence available to the tool.
Look for an explanation of the output. A single label gives less to review than a report that identifies the suspected regions or manipulation type, but an explanation is still only a model output. Document the source, file history and any independent corroboration before using a result in a consequential decision.
The historical text-detector benchmark covers a different medium. Its figures cannot establish the performance of an image detector. Our methodology explains the records needed before publishing a numerical comparison.