Research note 02, June 2026

On-Device Models and the Right to Private Assistive Computing

Assistive AI is most useful when it can observe sensitive context. That usefulness creates a privacy obligation that cannot be satisfied by policy language alone.

An assistive model may help a user read personal mail, interpret a medical form, navigate a transit station, summarize a workplace conversation, or operate a device during a moment of vulnerability. These are precisely the contexts in which constant transmission to remote services can become unacceptable. Privacy is not a premium feature in assistive computing. It is part of access itself.

On-device inference changes the architecture of trust. It can reduce exposure of images, speech, location, biometrics, and disability-related routines. It can also make assistance available where network connectivity is unreliable or where cloud latency would make an interaction unusable. The tradeoff is that local models must operate under strict limits on memory, power, heat, and compute.

Those limits are research constraints, not excuses for weak systems. Compression, distillation, quantization, retrieval, and adaptive computation need to be evaluated against assistive tasks rather than only against generic benchmark suites. A local captioning model that misses medication labels or warning signs cannot be defended by average-case performance.

Private assistive computing also requires clear control. Users need to know what is processed locally, what leaves the device, what is retained, and how to disable collection without losing basic functionality. The interface must make these boundaries legible to users who may rely on the same system to read the interface.

The right to private assistive computing is therefore both technical and institutional. It asks researchers to design models that run close to the user, and it asks institutions to avoid architectures that require people to trade privacy for participation in ordinary digital life.