Uravation Inc., based in Bunkyo, Tokyo, released the results of its "AI Image Checker judgment log survey" on November 21, 2026. The survey covers 31,734 checks completed over 47 days, from September 16 to November 1, with the company's free AI Image Checker. Only 5.8% of the images carried C2PA, a standard that records an image's provenance. Even among images with generative-AI records, visual analysis alone rated just 39.9% as "likely AI-generated."
PNG Files Carried C2PA 14.6% of the Time, JPEGs Only 1.9%
According to the company's announcement, C2PA (a technical standard that records how content was created and edited) was found in 1,845 images. PNG files stood out, while JPEG and WebP files rarely carried it.
| Format | Checks | With C2PA | Share |
|---|---|---|---|
| PNG | 9,949 | 1,452 | 14.6% |
| JPEG | 21,035 | 391 | 1.9% |
| WebP | 750 | 2 | 0.3% |
Of the images with C2PA, 96.1% (1,773) also contained generative-AI records. Fifty-four images had both camera metadata and C2PA, or 2.9% of the C2PA images. The company notes that it counted only whether C2PA data was present and did not verify signatures.
Nearly 90% of Images Had No Records, Leaving Visual Analysis as the Only Option
Images with no generative-AI, camera or editing records and no C2PA numbered 27,689, or 87.3% of the total. By record type, 6.7% (2,140) had generative-AI records, 3.4% (1,088) had camera metadata, and 2.5% (799) had editing-software records.
The share without records varied by image type: 94.2% for screenshots, 90.7% for illustrations and 83.8% for photos. The company explains that records can be lost when images are posted to social media, captured as screenshots or exported.
Even With Generative-AI Records, Only About 40% Rated "Likely"
The AI Image Checker first checks metadata such as C2PA and EXIF (information embedded in the image file) as evidence. A vision AI then analyzes the image from five angles, including texture and light and shadow, and shows a reference value. The survey treats an on-screen value of 70 or higher as the threshold for "likely AI-generated."
Of the 2,139 images that had generative-AI records and a visual-analysis value, 39.9% scored 70 or higher on visual analysis alone, while 45.7% scored below 40. By type, 57.7% of illustrations scored 70 or higher, compared with 29.8% of photos. The company says this group also includes photos partially edited or composited with AI.
Across all checks, 33.6% scored 70 or higher. By image type, 3DCG ranked highest at 54.4%, followed by photos at 35.6%, illustrations at 34.8%, graphics at 25.1% and screenshots at 11.4%. Among the 1,088 images with camera metadata, 8.0% scored 70 or higher.
Weekly Checks Climbed to 9,562
The number of checks rose every week, counting weeks from Monday. The week of September 21 saw 1,918 checks, and the week of October 26 (through November 1) reached 9,562, which the company describes as 5.0 times the first week. October averaged 879 checks a day, and on 8 of the days between October 1 and 16 the daily cap was reached, leaving some hours when checks could not be run.
The tool is free, requires no sign-up, and accepts JPEG, PNG and WebP images up to 8MB. However, the company states that it cannot detect invisible watermarks such as Google's SynthID.
The Company Recommends Checking Original Files and Sources
Suguru Sato, the company's CEO, said C2PA is a strong clue when present, but only 5.8% of images had it, and nearly 90% can be judged only by visual analysis. He added that companies should set internal rules to verify original files and sources rather than decide based on the numbers alone.
For high-stakes uses such as contracts, journalism and academic work, the company recommends combining three steps: obtaining the original file, checking C2PA and confirming how the image was obtained. It also states clearly that the result is only a reference value and does not establish that an image was made with generative AI.
Our Take
The survey is valuable because it puts numbers on how much C2PA actually survives in practice, based on more than 30,000 images that real users submitted. However, the sample likely consists mainly of images that users already suspected were AI-generated, so it may differ from the trend across images online as a whole. Image types were also classified by AI, and the visual-analysis values themselves depend on how the company's tool works.
Even so, the finding that nearly half of the images with generative-AI records scored below 40 on visual analysis is significant. It shows how hard it is to judge authenticity by appearance alone. Systems that keep provenance records such as C2PA intact from capture or generation through editing and sharing are likely to become even more important.
Summary
Uravation's survey found that only 5.8% of the images submitted for checking carried C2PA, and 87.3% were judged by visual analysis alone because they had no records. The findings matter to companies that use generative-AI images in their work, and to newsrooms and schools that need to verify whether images are authentic. The company positions its results as reference values and urges users to combine them with checks of original files and sources.