Best AI image detector picks in 2026
Every detector on this list measures something different. Pick the one whose evidence matches the decision you have to make — not the one with the loudest marketing page.
Why two detectors disagree about the same file
Detection is not one test. Some tools read the file's metadata and provenance records, some look for an invisible watermark written at generation time, some analyse frequency-domain patterns left behind by upsampling, some study noise residuals where a camera sensor would have left grain, and some simply judge whether the scene is physically plausible. A file can fail one of those tests and pass another, which is why two credible tools can hand you opposite answers on the same photo.
The disagreement gets worse after ordinary editing. Resizing, cropping, screenshotting and re-compression all destroy the traces the stricter tests depend on, while metadata-only checks are defeated by any platform that strips EXIF on upload — which is most of them. Treat every result as evidence about one signal, and treat a missing signal as much weaker evidence than a positive one. The table below compares the tools we see used most often in editorial and licensing workflows, with qualitative tiers rather than invented accuracy figures.
| Detector | Type | Best for | Tier | Notes |
|---|---|---|---|---|
| Hive AI Detector | Commercial API | High-volume moderation queues | High | A dedicated AI-generated-media classifier built for pipelines rather than one-off checks. Strongest on generator output it was trained against; weaker on heavily re-compressed files. |
| Google SynthID | Invisible watermark (proprietary) | Confirming Google-generated content | High | Reads a watermark embedded at generation time, so a positive result is meaningful. A negative result proves nothing — most images were never watermarked in the first place. |
| Illuminarty | Browser-based detector | Quick single-image sanity checks | Moderate | Returns a probability plus a suspected generator family. Small images, screenshots and double compression all push the score around, so treat the number as a hint. |
| Sightengine | Commercial API | Teams that need a documented audit trail | Moderate | Combines a synthetic-image score with face and manipulation models. Useful when you must log a decision; the score still depends on resolution and compression history. |
| Platform-native checks | Built into social platforms | End users posting content | Limited | Mostly policy enforcement rather than evidence. Labels and warnings are not reproducible outside the platform, and they are not a substitute for provenance records. |
Tiers as of 2026-09 — Tiers reflect vendor-published benchmarks cross-checked against independent write-ups as of 2026-09. Vendor numbers are measured on their own test sets and typically run higher than what you will see on your own files.
Choosing one for your workflow
If you are a designer receiving files from a client, start with provenance and metadata rather than a score: what does the file itself claim, and can you verify it? If you are a photographer proving your own work is yours, keep your originals and your camera's signed credentials, because a stripped export cannot testify for you. If you are a buyer or an editor clearing assets at volume, you want a documented trail — a tool whose output you can attach to a decision — rather than a single percentage nobody can reproduce.
Whoever you are, run more than one check before you act on a result, and prefer tools that show you the evidence they used. A verdict you cannot inspect is not something you can defend in a review. When a file arrives with no metadata at all, remember that the most common explanation is mundane: social platforms, messaging apps and screenshot tools strip metadata routinely. If you want to see what a specific file actually carries before you judge it, inspect its metadata locally — and if you need to understand the signals themselves, read how AI detection for photos actually works.
What "best" actually means: the criteria that separate useful detectors from misleading ones
A 2023 study by researchers at the University of Waterloo asked 200 participants to label 120 images as real-camera or AI-generated. Human accuracy sat at 61 percent — barely above a coin flip — and that was before subjects had encountered modern diffusion models trained on curated portrait datasets. Accuracy dropped to 54 percent when images were first screenshotted and re-shared once, the kind of chain that happens every time a file moves through a group chat. The finding matters because it sets the baseline: a detector that claims 95 percent accuracy against pristine outputs from a known generator is measuring a different problem from the one you actually face. Re-compression, format conversion, even a single crop applied in a lossy editor destroys the high-frequency patterns that most frequency-domain and noise-residual classifiers depend on. What looks like a confident score on a direct upload becomes noise after two rounds of JPEG at quality 75. The detectors that hold up in re-shared content are the ones that read signals embedded intentionally — watermarks written at generation time, C2PA provenance chains, or blockchain attestations — not the ones that attempt to infer generation from pixel statistics alone.
This creates a counterintuitive rule: the detector that looks most confident on a screenshot is usually the least reliable one to trust. High confidence on degraded input means the model is pattern-matching to compression artifacts or resizing halos rather than genuine generation signatures. When you are choosing a detector, apply this checklist before relying on any result. First, does the tool tell you which signal it used — metadata, watermark, frequency analysis, or scene-plausibility — or does it give you only a percentage? Opaque scores are hard to defend. Second, was the tool tested on re-compressed and re-shared images, not only on pristine generator outputs? Third, does it distinguish between a signal being absent and a signal being negative? A missing watermark is not evidence the image is real; it is evidence that this particular signal was not present. Fourth, can the output be attached to a written decision — is there a report you can save, not just a badge that expires when the tab closes? If you want to start with the metadata layer before reaching for a classifier at all, use the photo metadata viewer to see exactly what the file carries.
Metadata versus pixels: what EXIF, C2PA, and IPTC can prove — and what they cannot
A raw file out of a modern mirrorless camera typically carries four separate timestamp fields, GPS coordinates to within three metres, the lens serial number, the camera body firmware version, and, on Sony and Nikon bodies, a signed certificate linking the capture to a registered device. That is a lot of provenance. The problem is that almost none of it survives a round trip through a social platform. Instagram strips all EXIF before the image is served to visitors. Facebook re-encodes uploads, removing IPTC credit fields and GPS data in the process. WeChat compresses and re-packages image attachments, leaving only a stripped JPEG with no embedded metadata at all. X (formerly Twitter) has stripped EXIF since 2012. What arrives at the other end of a share is, metadata-wise, a blank file — which is why the absence of metadata on a suspicious image tells you almost nothing about whether it is real or AI-generated; both classes of file look the same after one platform hop. C2PA (the Coalition for Content Provenance and Authenticity) standard is designed to survive this: it embeds a cryptographically signed manifest inside the file that is structurally separate from EXIF and is preserved by C2PA-aware tools even across re-encoding. Adobe Firefly, DALL·E 3, and Midjourney v6 began writing C2PA manifests in 2024. However, C2PA is only useful if the chain of custody was never broken by a non-C2PA-aware tool, and most screenshot utilities and all older editing software break the chain silently.
IPTC fields — creator, copyright, description, source — sit in a different part of the file header from EXIF and are preserved by some platforms that strip EXIF. Getty Images and wire agencies rely on IPTC for licensing, so their upload pipelines protect it. But IPTC carries no cryptographic signing; any field can be typed in manually by anyone with an EXIF editor. Claiming IPTC credit for an image you did not create takes about thirty seconds, which means IPTC is useful for tracking provenance in trusted pipelines but is not evidence of authenticity by itself. The honest summary is this: metadata can corroborate a claim when the chain is intact and the source is accountable; it cannot prove a claim when the chain has been broken, and a broken chain is the normal state for images encountered in the wild. For a systematic look at what a specific file actually carries before you reach for a classifier, inspect its fields directly and note which fields are present, which are blank, and which contradict each other — contradictions (a camera date that predates the model's release, a GPS location that conflicts with the claimed scene) are often more informative than the fields themselves.
Step-by-step: what to do when you receive a suspicious image
Start with the file itself before opening any external tool. Check the dimensions: images exported from Midjourney default to 1024×1024 or 1344×768 depending on the aspect ratio preset; images from DALL·E 3 default to 1024×1024 or 1792×1024. A file that is exactly one of these sizes and carries no camera EXIF has already failed one plausibility check. Check the file size relative to dimensions: a real camera JPEG at 1024×768 effective resolution is typically 400–900 KB at quality 85; a diffusion-model output at the same dimensions is often 600 KB–1.2 MB because diffusion outputs have higher entropy in flat areas and compress less efficiently than optical captures. Then open the EXIF: if DateTimeOriginal is present, does it align with the claimed event date? Does the GPS point to the claimed location? Is the camera model a real model with a known release date that predates the image? If DateTimeOriginal is absent on a file claimed to be a direct camera export, that is a yellow flag. Next look at the content itself: AI generators still struggle with text inside images (characters are plausible-looking but not actual words at small sizes), with reflections that obey coherent geometry, and with hands (extra or missing knuckle segments, fingers merging into palms). These are not definitive — edited real photos can show similar artefacts — but they raise the prior before you run any classifier. Screenshot-based AI detection handles the case where you only have a screen capture rather than the original file.
Once you have run the file-level checks, run at least two independent classifiers from different signal families — one that reads provenance or watermark data and one that analyses pixel statistics — and record both results. When they agree, confidence is higher. When they disagree, record the disagreement; it is meaningful information, not a failure of your process. Save the original file, a screenshot of each classifier result, and a timestamped note of what you checked and when: this is your evidence chain. The cost of a false accusation — publicly claiming a photograph is AI-generated when it is real — is significant for the person on the receiving end, and can create legal exposure depending on jurisdiction and context. The evidence chain protects both parties: it documents that you acted on available evidence rather than guessing, and it gives the subject a clear record to rebut if the conclusion turns out to be wrong. Before sharing any conclusion publicly, ask whether you have documented enough to defend it, not just whether you believe it. For the cross-validation step, check the image with a local metadata pass as a final sanity-check before drawing a written conclusion.