What compression changes in an AI image check
This research note describes a test we are preparing. It contains no measured SUREvAI detection rate yet. The goal is to measure how common sharing steps affect both real and generated images.
Two mechanisms to examine
Re-encoding may remove ordinary metadata or break access to a signed history. Lossy compression also changes pixel details used by visual classifiers. We will measure these effects separately rather than assuming every compressed file behaves the same way.
Paired evaluation design
Start with labeled camera photos and known generated originals. Produce paired TinyPNG-style, JPEG quality, WebP, resize, crop, and screenshot variants under documented settings. Keep every variant in the same data split as its original.
What the report must show
For each transformation, report AI detection rate, false alarms on real photos, Not Sure rate, and how many valid origin records survive. Include examples of both missed AI and falsely flagged real files, with permission to publish the examples.
What users can do today
Submit the original file when possible and compare it with the shared copy. If only a small or compressed file exists, treat Not Sure as a useful limitation rather than a negative finding.