Preprint - Bad Computers: How the Wider AI Stack Can Stop the Worst AI Misuse (Veale and Gorwa, forthcoming MIT Press)
Together with Robert Gorwa, I have authored a book on the malicious misuse of AI and how it might be stopped, across the whole stack. It's forthcoming next year with MIT Press. You can read it as a preprint https://homepages.ucl.ac.uk/~ucqnmve/badcomputers/ or https://badcomputers.michae.lv. When it is published, the open access preprint will come down for a period of time.
There are some things that are deeply wrong to do with AI tools, but which are already happening, even at scale. This book is about how these things are or are not stopped, and by whom. When lawmakers wish to prevent the creation of realistic-looking intimate images of ex-partners or of Internet influencers, scamming based on voice or image cloning, or sexual exploitation of children, where do they start? And where should they stop?
In Bad Computers, we argue that in order to understand policies and enforcement in the AI stack, we need to pay much more attention to a much wider array of intermediaries. Rather than focusing on the big AI labs training models, or on evaluations of frontier systems, we examine the past, present and emerging practices of many of the intermediaries which malicious users of AI interact with and through today: cloud providers, operating systems and hardware, network services, model hosting platforms, communications services and more. Code-based “technoregulation,” private governance and law in this area are fast growing and entwining. Trying to animate any single layer of the AI stack to fully achieve a complex policy aim against motivated, malicious users of these computational systems usually brings far too many trade-offs to rights and freedoms, if such an approach even works at all. More promising is what we call full-stack AI governance, where interventions are spread across many actors, coordinating to leave technological “loops” — governance residue like watermarks, logs, hashes and blocking capabilities — that other stack actors can “hook” into. We look back before the generative AI rush to historical analogies that bring significant lessons for today, while deeply probing practices and research we see in the current moment for lessons for the future.