Best AI Image Detectors
By Christopher Elley, Founder, FactHeck · Published 9 June 2026
Written with AI assistance and reviewed for accuracy by the author.
The best AI image detectors each work differently, so the right choice depends on your situation. For a free, no-signup check, Illuminarty and AI or Not accept an image upload and return a probability score. For higher reliability in newsroom or research settings, Hive Moderation offers a free web demo backed by an enterprise model. Sightengine suits bulk API use. FactHeck integrates AI-image detection into a broader fact-check when you have a post URL rather than a standalone image. Run at least two tools and compare — agreement between independent detectors is more meaningful than any single score.
Ranked: the best AI image detectors
This ranking is based on free-tier availability, detection breadth, and documented use in media-verification contexts. No detector is universally best; re-compressed or low-resolution images reduce accuracy across all tools.
- Hive Moderation — Enterprise-grade model covering AI-generated images, video, and audio. The free web demo accepts uploads and returns a probability score with no account required. Widely cited in academic and newsroom contexts. Limitation: the full API requires an enterprise agreement.
- Illuminarty — Free-tier web tool that returns a probability score for AI generation and attempts to identify the source model (Midjourney, Stable Diffusion, DALL-E, and others). Useful for journalists and researchers who want more than a binary result. Limitation: model identification is indicative, not definitive; newer or obscure generators may be misclassified.
- AI or Not — Simple upload interface returning a binary AI/human result with a confidence score. Free tier available. Also covers short video clips, which distinguishes it from image-only tools. Limitation: less granular output than enterprise tools; the free tier has monthly limits.
- Sightengine — API-first platform designed for developers who need to run checks at volume. A free web demo is available. Strong for bulk processing pipelines. Limitation: the web UI is a demo only; production use requires an API key and costs credits beyond the free quota.
- FactHeck — AI-detection module built into a broader fact-checking pipeline. Submit a TikTok, Instagram, or YouTube URL and FactHeck analyses the media for AI generation alongside extracting and verifying the factual claims in the post. Best when you have a post URL and want both an AI check and a claim check in one step. Limitation: requires a public post URL; does not accept standalone image uploads outside the pipeline.
Side-by-side comparison
| Tool | Best for | Free tier | Video support | Key limitation |
|---|---|---|---|---|
| Hive Moderation | Media organisations and researchers needing reliable scores | Web demo | Yes | Full API requires enterprise account |
| Illuminarty | Quick checks + model identification | Yes | No (images only) | Model guesses are indicative, not definitive |
| AI or Not | Simple binary result on an uploaded image or short clip | Yes (limited) | Short clips | Less granular output than enterprise tools |
| Sightengine | Bulk API processing | Small free quota | Yes | Designed for developers; web UI is a demo only |
| FactHeck | Checking AI-generated images submitted as part of a post URL | 5 checks/day | Yes (part of broader fact-check) | Requires a public post URL; no standalone image upload |
How AI image detectors work
Most AI image detectors analyse statistical patterns in an image's frequency domain — the way pixels vary at a fine scale — rather than looking at what the image shows. AI generators introduce characteristic artefacts at this level that human photographers and traditional editing software do not. Detectors trained on large sets of AI-generated and real images learn to distinguish these patterns and return a probability score.
The weakness of this approach is that re-compression degrades these artefacts. Each time an image is saved as JPEG, uploaded to a social platform, or forwarded via a messaging app, the fine-grained signal is partially destroyed. A detector test on an image that has been screenshotted twice and shared on WhatsApp is significantly less reliable than a test on an original file. For the highest accuracy, work from the original image whenever possible.
Independent evaluation of AI image detectors is limited. The Content Authenticity Initiative and the Coalition for Content Provenance and Authenticity (C2PA) are working on provenance standards (cryptographic signing of media at capture) that address the detection problem from the source rather than the artefact — worth watching as a longer-term alternative to probabilistic detectors.
Using multiple detectors
Because each detector is trained differently, their results are not perfectly correlated. When two independent tools agree — both flag an image as AI-generated, or both return a low probability — that agreement is more informative than either result alone. When they disagree, treat the image as uncertain and look for other evidence: Is the claim the image supports plausible? Does the image metadata (if available) tell you anything? Has the image appeared elsewhere with a different caption?
Tools to check EXIF metadata — such as ExifTool or Pic2Map — complement detector outputs by confirming or contradicting claimed capture conditions.
When detectors are not enough
A negative result from a detector does not mean an image is authentic. It means the detector did not find patterns it was trained to recognise. Newer generators and post-processing techniques can reduce detectable artefacts significantly. For high-stakes decisions, combine detector output with source verification: where was the image first published, by whom, and in what context? A reverse image search via Google Lens or TinEye often reveals misattributed images faster than any AI-detection tool.
Frequently asked questions
What is the best free AI image detector?
No single tool is definitively the best because detection accuracy varies by generator, image subject, and re-compression level. Hive Moderation and Illuminarty both offer free tiers and are well-regarded in media-verification circles. Running two tools and comparing results gives you more confidence than relying on one alone. If results disagree, treat the image as uncertain until you gather other evidence.
Do AI image detectors work on images that have been screenshotted or resaved?
Re-saving or screenshotting an AI-generated image degrades the fine-grained frequency artefacts that detectors rely on. The more times an image has been re-compressed — for example by being saved from Instagram or forwarded on WhatsApp — the lower the detection accuracy. A clean result on a heavily re-compressed image is weaker evidence of authenticity than a clean result on an original.
Can AI image detectors identify the specific generator used?
Some tools attempt to identify the source model (Midjourney, Stable Diffusion, DALL-E, and so on) in addition to a probability score. Accuracy on model identification varies widely; treat model guesses as indicative rather than definitive. Detectors trained on a narrow set of generators often miss images from newer or less common models.
Should I trust a result that says an image is not AI-generated?
A 'not AI-generated' result means the detector did not find the patterns it was trained to look for — it is not proof the image is authentic. Detectors have false-negative rates that vary by model and image type. Combine the detector result with visual inspection and context: does the image match the claim being made about it?
Checking a post that includes an image? Paste the URL into FactHeck for an automated AI-detection scan alongside a claim-by-claim fact-check.