Sightengine's headline claim of 98.3 percent accuracy for AI-generated image detection is a real, independently benchmarked figure, produced by researchers at the Universities of Kansas and Rochester testing 80,000 images. It is also a best-case number. On a broader test spanning 17 different AI image generators, Sightengine's accuracy averaged 92.7 percent, ranging as low as 75 percent on some individual generators, and deliberate adversarial edits to an image can drop accuracy from 92.6 percent down to 55 percent.
Is Sightengine accurate?
Yes, by the standards of current AI image detection, Sightengine performs well and has been independently ranked ahead of competing commercial solutions in at least one large-scale benchmark. But "accurate" is doing a lot of work in that sentence, since the actual number depends heavily on which images you are testing and under what conditions. Treating 98.3 percent as a universal, always-true figure would be a mistake; treating it as the strong end of a real range is closer to the truth.
What is Sightengine, and who actually uses it?
Sightengine is a content-moderation and media-authenticity API, not a consumer-facing tool built for students checking a single essay. It analyzes images, video, text, and audio for AI-generated content, deepfakes, and policy violations across more than 120 detection classes, and it is built for developers and platforms to integrate into their own moderation pipelines rather than for someone to paste a single file into a web form, though a hosted checking interface does exist. Pricing runs on a per-operation basis: the free tier includes 2,000 operations a month capped at 500 a day, and AI-image and deepfake detection specifically cost 5 operations per check rather than the standard 1.
Why does accuracy vary so much by generator?
Because different AI image generators leave different statistical fingerprints, and a detector trained heavily on patterns from one generator does not automatically generalize perfectly to another. On the broader 17-generator test, Sightengine scored as high as 98 percent on some generators and as low as 75 percent on others, which is a real and meaningful spread for anyone relying on a single accuracy number to make a decision about a specific image from an unknown source.
What happens to accuracy when someone tries to evade detection?
It drops substantially. Testing with deliberate adversarial modifications, specifically inpainting edits designed to disrupt the patterns a detector looks for, dropped Sightengine's accuracy from 92.6 percent on the unedited set down to 55 percent, close to a coin flip. That is not a flaw unique to Sightengine; it reflects a broader limitation of pattern-based AI detection generally, where a determined effort to evade detection can meaningfully degrade almost any detector's reliability.
That same vulnerability, adversarial evasion lowering detection accuracy across tools generally rather than one product specifically, is part of the broader reliability picture covered in are AI detectors accurate.
For a text-focused comparison of an image-specific detector's claims against independent testing, our review of TruthScan's AI image detector covers similar ground from a different vendor.
How this was made: Sightengine's benchmark figures, per-generator accuracy range, and adversarial-attack vulnerability were checked via search 2026-08-24 against Sightengine's own published benchmark writeup and independent third-party testing sources. Drafting is AI-assisted, with a person verifying each figure against its source before publishing.
Frequently asked questions
What is Sightengine's real accuracy rate?
It depends on the test conditions. An independent 80,000-image benchmark found 98.3 percent, a broader 17-generator test found 92.7 percent on average with real variation by generator, and deliberately edited images dropped accuracy to 55 percent.
Is Sightengine built for students or for developers?
Primarily developers and platforms integrating it into their own moderation or verification pipelines via API, though a hosted checking interface exists. It is not designed as a single-essay classroom checking tool the way a student-facing AI text detector is.
Can Sightengine be fooled?
Yes, under deliberate adversarial editing. Testing found accuracy dropping from 92.6 percent to 55 percent after inpainting-style edits designed specifically to disrupt the patterns the detector relies on.
Does Sightengine detect AI-generated text as well as images?
Sightengine offers text classification and pattern-matching features as part of its broader content-moderation suite, but its most independently verified and benchmarked accuracy claims are specific to AI-generated image detection.
