AI Image Detector API

We use Sightengine as our main AI image detector API and Hive as the backup. We chose them after sending the same 100 images to four commercial APIs, and the deciding number was how many real photos each one flagged as AI: Sightengine and Hive flagged none of 50, while the other two flagged some in small samples.

This page is for developers and teams who want to add AI image checks to their own product. It covers what we measured, the design that kept our call count down, and how we handle a slow or failing engine. We have a stake here, because our checker runs on two of these engines; the accuracy report has the full method and our disclosure.

On our test set, Sightengine was the most accurate AI image detector API at 99%, with 49 of 50 AI images caught and no real photo flagged. Hive reached 95%, also with no false alarms, at a lower list price. AI or Not and WasItAI ran only on small samples within their free allowances, and both flagged real photos.

• Sightengine: 99% on all 100 images, about $0.01 an image, median 2.9 seconds per call.
• Hive: 95% on all 100 images, about $0.006 an image, median 2 seconds.
• AI or Not: 85% on a sample of 20, 1 of 10 real photos flagged, about $0.02 an image, 1.1 seconds.
• WasItAI: 77.8% on a sample of 9, 2 of 5 real photos flagged, about $0.0399 an image, 2.5 seconds.

Prices are worked out from each vendor's published pricing at the time we tested, some of them from monthly plans, so treat them as rough. Every run used free trial credits.

The cheapest API call is the one you don't make. Before calling a model, our checker reads the image's Content Credentials, a signed record some AI tools and cameras write into the file. In our test, credentials from signers on the C2PA trust lists settled 16 of 100 images with no API call at all: 14 AI images and 2 real camera photos.

The other 84 went to a detection API, and two rules kept that safe. First, only signers on the trust lists count: 10 images had credentials from signers that aren't on them and 4 more had an AI label in their metadata, so those 14 went to the model too, which got all 14 right. Second, anything we store or share is read again on our server from the file's own bytes, so a tampered page can't plant a fake credential in a saved result. We also cache results by the file's hash, so the same viral image doesn't cost a second call. Our content credentials checker shows the reading step on its own.

Every detection API returns a score, and you decide what it means. We use two lines with a gap between them: with Sightengine, 0.8 or above is likely AI, 0.2 or below is likely real, and anything in between is Can't tell. On our test that flow judged 98 of 100 right, with 1 Can't tell and 1 wrong.

A single cut-off looks better on paper: counting a Sightengine score of 0.5 or more as AI got 99 of 100 right. But it forces a yes or no on borderline images, and a borderline real photo is exactly where a false alarm hurts someone. The same two-line flow with Hive, at 0.9 and 0.3, got 94 right, 5 Can't tell and 1 wrong. Hive's documentation recommends 0.9 as its AI threshold, so start from each vendor's guidance and then test on images like the ones you will check.

Any API can be slow, down or out of quota, so we call engines in order: Sightengine first, then Hive if Sightengine errors, times out or says its quota is used up. Each engine gets 9 seconds before we move on. The user waits a little longer but still gets an answer.

Three vendor details shaped that design. Sightengine's pricing page says requests that end in an error don't count toward your usage. Hive's self-serve endpoint has a default limit of 100 requests a day. And every successful call costs money, so we log each one with its engine and price and add them up daily. Plan for retries and limits from day one: an AI image detector API that times out in front of a user is worse than one that answers Can't tell.

Not yet. Our checker runs on these pages only, with no public API or bulk upload. If you need to check images at volume, go to the engines directly; Sightengine and Hive both sell API access.

If you only care whether an image came from one company, some vendors check their own signals. OpenAI, for example, offers a Content Provenance API that looks for OpenAI's C2PA credentials and SynthID watermark, and says it doesn't detect other companies' models. For a picture from an unknown source you still need a general detector. To check a single image, the free AI image checker on our home page runs the flow described here, and our page on damage photos in refund claims shows how a support team can use it.

Several have free tiers for testing. When we checked, Sightengine offered 2,000 operations a month (at most 500 a day), and its documentation shows one image check using 5 operations. Hive gave $50 or more in free credits after you add a payment method, AI or Not included $5 in credits with 20 image checks, and WasItAI offered 10 images a month with API access. Check current terms before you build on them.

Not Sightengine's: its documentation says detection is purely pixel-based and ignores metadata, EXIF, C2PA data and invisible watermarks. If you want credentials, read them yourself before you call the API, as we do.

Hive's response includes scores that attribute likely generators, such as Midjourney or Stable Diffusion. Sightengine's documentation says its model can return per-generator scores too, but the responses to our requests carried only the overall score.

In our test the median call took 2.9 seconds for Sightengine, 2 for Hive, 1.1 for AI or Not and 2.5 for WasItAI. Add upload time, and set a timeout for the slow ones.