Sightengine AI image detector

The web demo takes a dropped or pasted image of up to 12 MB and runs free checks until a "Free limit reached" panel asks for a free account (sightengine.com/detect-ai-generated-images, read 9 November 2026). The API limits are tabulated on the pricing page: images up to 12 MB on Free, Starter and Pro, video up to 50 MB on Starter and 500 MB on Pro, audio 12 MB on Free and 20 MB on paid plans, and a Free tier of 2,000 operations a month with a daily maximum of 500 (sightengine.com/pricing, read 9 November 2026). An operation is the billing unit: "running Nudity Detection on an image counts as 1 operation", and each response reports the operations consumed. Figures below are as displayed in USD on 9 November 2026; a EUR toggle exists.

Retention is stated as deletion after processing. The privacy policy reads: "The images, videos, text and audio you send to our API are only used to return results to you. The only exception is content you choose to send us through the Feedback API, which we use to improve our models", and under "We keep as little as possible": "Your content is deleted once it has been processed, unless you ask us to keep it or the law requires us to. Technical logs are deleted automatically after a limited period" (sightengine.com/policies/privacy, last updated November 6, 2026, read 9 November 2026). The same policy says content customers send through the API falls under the Data Processing Addendum rather than the policy, and that files tested in the demo are kept as logs "deleted automatically after a limited period". The card's retention rank field takes the deletion clause as a stated zero. Ownership is Kozelo SAS, RCS Paris 510 113 509, at 16 bis rue d'Odessa, Paris, with David Lissmyr as director of publication (legal notice, read 9 November 2026).

Sightengine publishes no accuracy percentage. Its accuracy statement points to an outside paper: "In an independent benchmark by researchers at the University of Rochester and the University of Kansas using 80,000 images across multiple generators, our detector achieved the highest accuracy among the tools tested. Accuracy depends on the generator, image quality, and post-processing; no detector is 100% reliable, and results should be combined with human judgment" (sightengine.com/detect-ai-generated-images, read 9 November 2026). The page links arXiv 2404.14581, which is not an index record, so the sentence is recorded as the vendor's statement and no figure is taken from it. The same page claims "Super-human accuracy" on the ground that "models specifically trained to detect AI images typically outperform human performance", and offers an "AI or not" quiz.

The vendor's caveats are specific. The FAQ says AI media detectors "can produce inaccurate results: false positives or false negatives", that the detector "may also struggle with content produced by AI media generators it has never encountered before", and that "there is usually a short window after a model launches when detection accuracy on its outputs is lower than for established generators" (sightengine.com/detect-ai-generated-images, read 9 November 2026). On the score itself: "Even if the API returns a very high or very low score, such as 99% or 1%, it is important to consider the context and use human judgment in the final assessment." On robustness the documentation states that "Detection is robust to re-encoding, compression, resizing and other transforms. Confidence may drop somewhat on heavily degraded images or screenshots with significant UI chrome" (docs, read 9 November 2026). No false positive rate is published.