Chapter 2 of Bad Computers by Michael Veale and Robert Gorwa

Cloud: Abusing Distant Computers

Enormous quantities of both legitimate and illegitimate computing happen in large data centers run by cloud giants. Historically, these providers have been able to shrug off responsibility to prevent malicious activity on their services by claiming that they are only neutral, technical middlemen. AI has stressed this logic. As the name suggests, generative AI creates new material itself rather than just processing or analyzing the content of third parties, making claims of just being a passive intermediary much less credible. In this chapter, we look at how cloud firms have proven vulnerable to abuse by malicious actors in practice, and how they have responded with layered, but inherently imperfect, safety measures. We characterize these different layers of safety and content moderation in cloud AI, both in terms of blocking unwanted content and activity and leaving ‘loops’ of governance residue behind that other layers of the stack can hook onto, such as secure metadata, watermarking and logging. Considering the different services that cloud providers sell, we suggest that it will be very difficult to mitigate abuse on the most open service, ‘unmanaged’ compute, where firms rent out hardware in their data centers most directly. Policy responses that might follow from that include limiting the products cloud providers sell to specific, more governable computational services rather than access to raw, difficult-to-monitor hardware, or requiring unmanaged compute to be licensed and overseen like a regulated industry. Neither bodes well for privacy, free expression or computing freedom, nor will either obviously be effective at preventing misuse.

Download chapter 2 as a PDF (pages 19–55, 1.2 MB).

How to cite

Michael Veale and Robert Gorwa, Bad Computers: How the Wider AI Stack Can Stop the Worst AI Misuse (MIT Press forthcoming 2027) ch 2.

Citation data for reference managers: BibTeX or RIS. This is a preprint of a book forthcoming from MIT Press in 2027. Page numbers follow this preprint and will differ from the published edition.