AI or Not is one of the better-known AI image detectors, and its free plan is fair: 20 image checks a month. Run out, or want to check something without uploading it, and you need alternatives. The useful thing to know is that "AI detectors" do three quite different jobs. Some guess from the pixels. Some read a signed label inside the file. Some look for an invisible watermark. Each catches different fakes and misses others. This guide compares seven options, shows which one fits which situation, and is upfront about where our own AI Image Detector fits. It reads metadata on your device and makes no pixel-level guess.
From aiornot.com/pricing in October 2026: the free plan includes 20 AI image checks a month and text detection. Pro is $5 a month with 500 image checks and adds video, voice, music and deepfake detection. Business is $100 a month for 6,600 image checks. AI or Not's privacy policy doesn't say whether uploaded images are kept or used to train its models, so don't upload anything you wouldn't post publicly.
| Tool | How it checks | Free tier | Paid from |
|---|---|---|---|
| ToolAspect AI Image Detector | Reads camera data, software tags and timestamps, in your browser | Unlimited, no account | Free |
| Hive AI Detector (Chrome extension) | Pixel classifier; also guesses which generator made it | Free extension; web demo limits not published | API $6 per 1,000 images |
| Illuminarty | Pixel classifier | Free checks, no published cap | $10 a month adds region highlighting and generator ID |
| Sightengine | Pixel classifier via API | 2,000 operations a month, max 500 a day | $29 a month for 10,000 |
| Content Credentials Verify | Reads a signed C2PA label inside the file | Free | n/a |
| Google SynthID Detector | Finds Google's invisible watermark | Free; reportedly about 10 checks a day with a sign-in | n/a |
| Reverse image search | Finds where else the image appears online | Free | n/a |
Prices and limits are from each vendor's pricing page in October 2026. The SynthID daily quota and sign-in come from press coverage of the launch, not from Google's own page.
This is where most people get misled. A pixel classifier gives you a probability, not a fact. A label or watermark check gives you a fact, but only when the label is there. Here's how they play out on common cases.
| Situation | Metadata check | Content Credentials | SynthID | Pixel classifier |
|---|---|---|---|---|
| Original photo straight off a phone | Camera make, model and capture time: strong sign it's real | Only if the phone signs its photos | Nothing to find | Usually reads as real; false alarms happen |
| Image downloaded from ChatGPT or Gemini | No camera data | Often present on the original file | Gemini images, yes; Google lists OpenAI as a SynthID partner | Usually flagged |
| Same image screenshotted or reposted on social media | Wiped; looks like any screenshot | Usually stripped | Google says the watermark survives common edits | Less reliable after compression |
| Image from Midjourney or an open-source model | Nothing conclusive | Only if the tool adds them | Nothing to find | Your main option |
The pattern: proof-based checks are excellent when they hit and silent when they miss. Classifiers always give an answer, but it's an estimate.
Less accurate than their marketing suggests, especially on images found in the wild. A 2025 study by Li and colleagues, "Is Artificial Intelligence Generated Image Detection a Solved Problem?", found detectors suffer significant performance drops on real-world data. Another paper, "Fake or JPEG?", showed some detectors had learned to spot JPEG compression and image size patterns from their training data rather than signs of AI. That means a resized or re-saved image can fool them in either direction.
In practice, a high "AI" score on a heavily compressed image from a group chat is weak evidence. Run two different classifiers. If they disagree, you don't know yet.
Start with what you have. If you've got the original file, check its EXIF data or run our detector: a full set of camera details is a good sign, though not proof, since metadata can be edited. Then try Content Credentials Verify and SynthID, because a positive from either is close to definitive. If the image came from social media, skip straight to a reverse image search to find the earliest copy and its source. Use a pixel classifier last, and treat it as a tiebreaker. For video and voice clips, our deepfake detection guide covers what to look and listen for.
For more on the visual tells, like warped text, odd reflections and hands, see our guide on how to tell if an image is AI generated.
There's a free plan with 20 AI image checks a month plus text detection, as listed on aiornot.com/pricing in October 2026. Pro costs $5 a month for 500 image checks and adds video, voice, music and deepfake detection. Business is $100 a month.
No independent test has crowned one, and accuracy drops for every detector on real-world images that have been compressed, resized or screenshotted. A 2025 study found detectors suffer significant performance drops outside their test sets. Use a detector as one signal and look for proof, like Content Credentials or a SynthID result, where you can.
No. SynthID only finds the invisible watermark added by Google and the partners using its technology, which Google listed as OpenAI, NVIDIA and Kakao in October 2026. Images from other generators won't carry it, so a negative result tells you nothing about them.
No. Social media sites and messaging apps commonly strip metadata, including Content Credentials, when images are uploaded. A real photo and an AI image can both come back with no credentials. Only a credential that is present tells you something.
No. It reads the file's metadata in your browser and nothing is sent to a server. That also means it checks camera data, software tags and timestamps, not the pixels themselves, so treat its result as one clue rather than a verdict.