How accurate is CYPHR's detection?
+
CYPHR achieves 97.4% accuracy on our standard benchmark set of 10,000 images. This covers output from Midjourney, DALL-E 3, Stable Diffusion, Firefly, and Imagen. Real-world accuracy depends on image quality and post-processing — heavily compressed or cropped images may score with lower confidence.
Does CYPHR upload my images to a server?
+
No. All analysis runs on-device using a locally-cached model. The image bytes never leave your browser. CYPHR does not have a backend that receives, logs, or stores image data.
Can CYPHR detect images that have been edited after generation?
+
Yes, with reduced confidence. Post-processing like cropping, filtering, and JPEG re-compression can obscure some signals, but diffusion artifacts in texture and frequency domains are often persistent. CYPHR reports its confidence level — treat low-confidence scans as inconclusive, not as a verdict.
Which browsers does CYPHR support?
+
Chrome and Chromium-based browsers including Edge, Brave, and Arc. Firefox and Safari support is planned for a future release.
What does the confidence score mean?
+
The score represents CYPHR's certainty in its verdict — not the percentage likelihood that an image is AI-generated. A 97% authentic verdict means CYPHR is 97% confident the image is real. A 72% AI verdict means the signals are present but not fully conclusive.
Can I use CYPHR for professional or journalistic work?
+
Yes, and we encourage it. CYPHR Pro was built with journalists and researchers in mind. The shareable report format is designed to be citable. That said, CYPHR is a screening tool — it should support editorial judgment, not replace it.