Research Note
On Device AI Claim Evaluation Framework
An on-device claim needs a named task, model, version, device, processor, operating system, account, network state, inputs, outputs, telemetry, fallback, and verification
In this article
On Device AI Claim Evaluation Framework
An on-device claim needs a named task, model, version, device, processor, operating system, account, network state, inputs, outputs, telemetry, fallback, and verification date.
Test first use and warm use while connected, disconnected, and partially connected. Record downloads, network requests, feature degradation, latency, power, heat, accuracy, storage, permission, and server fallback.
"Powered on device" can describe one stage while other stages use cloud services. The claim should match the observed data flow rather than the broad product label.
Sources
Follow the evidence.
- blog.google: made by google 2024 collectionblog.google
- c2pa.org: Explainerc2pa.org
- daltonanderson.ghost.io: googles ai magic a pixel 9 gemini deep divedaltonanderson.ghost.io
- iptc.org: photo metadataiptc.org
- open.spotify.com: 4wQhhIrcfJ3wgclijI5256open.spotify.com
- spec.c2pa.org: C2PA Specificationspec.c2pa.org
- store.google.com: phonesstore.google.com
- support.google.com: 6128850support.google.com
- support.google.com: 15532903support.google.com
- support.google.com: find out if your photos have been edited with aisupport.google.com
- support.google.com: pixelphonesupport.google.com
- youtu.be: vmMnqx ZGFEyoutu.be