TY  - CHAP
AU  - Veale, Michael
AU  - Gorwa, Robert
TI  - Local: Generation and Regulation On-Device
T2  - Bad Computers: How the Wider AI Stack Can Stop the Worst AI Misuse
SP  - 56
EP  - 92
AB  - Consumer-grade hardware has become increasingly adept at running the type of AI systems capable of being seriously misused. In this chapter, we illustrate this through case studies of scam compounds and their associated underground tool marketplaces, and online communities for the generation of synthetic child sexual abuse material. This poses a problem, as regulating devices is more difficult and in many ways more invasive than governing cloud services, given the private nature of activities on these devices and the more limited control and visibility rule-setters have over them. Despite this, building in limits on the software that can run on individuals’ computers has a long history, stretching from the Nintendo Entertainment System in the mid-1980s to the iPhone today. We identify and characterize two approaches to regulating activities on devices: software gatekeeping, most familiar from the App Store, and operating system surveillance, which has been at the heart of debates on ‘client-side scanning.’ We consider how these might apply to the abuse of AI systems, examining the theoretical possibilities and limitations of approaches including notarization of models, restricting access to on-device machine learning hardware and behavioral monitoring. Given that extending these approaches to every device on Earth is unrealistic, we draw on other areas of technology policy where there have been successful efforts at segmenting ‘safer’ publicly available devices from more specialized open-ended ones. Careful combinations of market design and technical interventions could be deployed to offer a limited — if thorny and potentially dangerous — path forward to disrupting the most harmful forms of locally facilitated AI misuse.
UR  - https://homepages.ucl.ac.uk/~ucqnmve/badcomputers/chapter/3/
L1  - https://homepages.ucl.ac.uk/~ucqnmve/badcomputers/chapter/3/badcomputers-ch3-local.pdf
PY  - 2026
DA  - 2026/10/02
PB  - MIT Press
CY  - Cambridge, MA
LA  - en
N1  - Preprint, 2 October 2026
ER  -