The creation story
TrueFrame is this AI?
Free foreverDrop an image - TrueFrame reads the provenance trail and the pixel fingerprint AI generators leave behind, then answers with an honest, evidence-backed confidence: AI-made, human-made, or uncertain. Analyzed on your device, never uploaded.
A TIMxAI tool · Honesty is the product · Trust but verify · Private by design · No account, no ads, no upload
Drop an image, click to upload, or paste
Original files carry the strongest evidence, but screenshots get a real pixel read too.
Images (JPEG · PNG · WebP · HEIC) · text/HTML (phrasing patterns) · video (MP4 container) - all analyzed locally in your browser.
Two layers, one honest verdict. The provenance layer reads Content Credentials, EXIF and generator tags - near-certain when present. The pixel layer reads the frequency-domain artifacts and missing sensor noise that generators leave in the pixels themselves - so social re-uploads with stripped metadata still get a real read, honestly capped as probabilistic evidence.
Lossy capture path - the metadata layer is unavailable, so this verdict leans on pixel evidence and its confidence is capped. For the best read, analyze the original file.
Middle-of-the-road result - that usually means this is NOT the original file. Screenshots, social-media downloads, and re-saves strip the strongest evidence before we ever see it. If you can, get the original: long-press / right-click and save the actual file (not a screenshot of it), or ask the sender for the file itself, then analyze that. Social feeds re-encode everything they serve - a URL from a feed rarely carries the original's evidence.
The evidence
How honest is this? TrueFrame reads two independent layers. The provenance layer (C2PA Content Credentials, EXIF/XMP software tags, embedded generation parameters, camera capture data) is near-certain when present - and we still never say 100% without a cryptographic tell. The pixel layer reads the fingerprint left in the pixels themselves: periodic frequency-domain artifacts from generator upsampling grids, and the sensor-noise floor real cameras always leave (and renders don't). Pixel evidence is probabilistic, so it is capped below signature-backed verdicts; flat vector-style graphics are declared out of scope instead of guessed at; heavy re-compression is detected and said out loud. When evidence is thin we say uncertain plainly. A false "definitely AI" on a real person's work is the exact dishonesty this tool exists to fight.
Shipping next (build order, spec-public): a learned classifier that re-calibrates on labeled data so published confidence matches measured hit-rate · screen-snip capture · video, text, and URL-page modes.