On Infrastructure (AI, Law, and Otter Things Issue #13)
Today, I venture outside my area of specialisation and offer some thoughts on the recent Facebook outage. This issue is more of an attempt to think out loud about the incident than a complete opinion on anything, and so counterpoints and new perspectives are particularly welcome!
Quite a bit has happened in the tech world since my latest newsletter:
Amazon announced Astro, a mobile robot that finds a neat intersection between the various risks associated with personal assistants and those related to robotics, with a side-dish of increased anthropomorphism. Though, to be fair, the strongest reasons why I’ll sit this one out are, in no particular order, the considerable financial cost and the limited added value: at least for me, the “benefits” section of the cost-benefit analysis feels relatively thin.
An interview with Timnit Gebru and Margaret Mitchell discussed the limits of AI Ethics in software development practice. While both specialists diverge on the possibility of actually achieving positive outcomes within corporate structures, they provide fascinating and resonating insights into the practical challenges and power structures that reduce the reach of ethical design. Indeed, we cannot hope that ethical thinking will address issues better addressed through other regulatory instruments, such as the law, even though ethical thinking may still be relevant for AI regulation.
As Reuters put it, China has announced its plans to release rules for algorithm governance, which would require algorithms to “uphold core values of socialism”. This approach suggests some comparisons with the European approach to AI enshrined by the EU AI Act, as well as with Article 25 GDPR. Given that I do not speak Chinese and my familiarity with Chinese law is limited at best, this is an analysis I should not pursue on my own, but any suggestions of China experts would be welcome!
Broader coverage of relevant AI news would require a dedicated newsletter (fortunately, there are very good options). However, the most salient item in tech news is not centred on AI. For a few hours on Monday, Facebook faced a global outage in its internal and external services, which was normalised by the time I woke up on Tuesday. While we still don’t have a definitive account of what happened at Facebook, a Cloudflare blog post provides an account of how the outage impacted other internet systems, with an introduction to the relevant technical concepts. Based on that and the limited information we have so far, I would like to present some moderately hot takes on what the Monday incident reminds us about technological infrastructure.
The first issue that comes to my mind about this Facebook downtime is decentralisation. As Nicolas Petit argues, the Internet mostly survived the outage: people were somewhat disrupted in their workflows, but in many cases, it was possible to rely on email and other social networks to conduct business, make jokes, share cute animals or spread disinformation, such as hackers claiming merit for what seems to be an internal failure or apocalyptic claims that the Internet was about to collapse. Even though Facebook’s services — especially WhatsApp, which plays important roles in business and social life almost everywhere outside China and the Anglosphere — were missed by many people, their relatively prolonged outage seems to have been more of a nuisance than an existential threat to the digital economy.
(Disclosure: I am more of a Twitter person than a citizen of the Facebook-verse, so my perspective is somewhat biased on this. I look forward to empirical studies that provide us with a better view of this outage and its global effects.)
Yet, we should not mistake this decentralisation for resilience. The Cloudflare blog post shows how the outage impacted other internet services, which had to handle additional traffic. Not all Facebook substitutes were adequately equipped for the increased demand: for example, Telegram users reported considerable slowdowns, while Twitter’s strategies for scaling had some limitations. And here we’re talking about large-scale competitors to Facebook services, which already have the technological know-how and the computational resources for coping with access surges. A newcomer to the market might have been crushed under the increased demand while lacking the brand recognition that will likely allow Facebook services to bounce back to their pre-outage popularity.
The outage thus reminds us that decentralised systems can still be profoundly interdependent and brittle. As Paul N. Edwards reminds us (‘Platforms Are Infrastructures on Fire’ in Thomas S Mullaney and others (eds), Your Computer Is on Fire (The MIT Press 2021).), modern platforms represent but one layer in a complex networking infrastructure. Each layer involves various technical choices and compromises, which are often made to compensate for shortcomings of the layers below (see TCP/IP). When we say platforms such as Facebook or WhatsApp are now part of our technological infrastructure, this means (at least in part) that people are building upon these platforms, either by creating code that interfaces with the platform APIs or by incorporating it into human-centric workflows. For example, consider a Facebook game and a small restaurant that receives orders and reservations through WhatsApp. In both cases, failures on the platform or any of its support layers can disrupt the user experience through no fault of the system the user interacts with.
Infrastructure brittleness can have practical effects even if one is not aware of the existence of the failing system. These effects, as one might expect, are not distributed uniformly among the affected users. In our case, Twitter had little trouble in dealing with the side-effects of the Facebook outage, while the small restaurant might have lost a sizable fraction of its revenue as WhatsApp faltered. This seems in line with an important warning from the decentralisation literature (see this excellent review article): one must pay attention to the actual structures of a given decentralised network rather than presupposing that decentralisation is in itself enough to mitigate or contain the centralisation of various forms of power.
Speaking of power, the fragility of digital infrastructures should remind us that comparisons between corporations and states are often far-fetched. To the extent that corporate power relies on the exercise of power through digital means, it is subject to the interruptions caused by technical failures and state action that may impose more permanent controls over technological infrastructure or disrupt it altogether. This is not to say that we do not need better models to understand — and contain — the political influence of large technology corporations, both by shaping discourse and exercising direct influence over legislators and policy-makers. But claims that place corporations such as Facebook as sovereign powers turn out to be facile, at least if we only look at regulation by code, which is built upon much more foundations than the ones underpinning state power. Megazord is a puny Leviathan, and any attempts to understand the power of technology corporations must look at the interplay between technology and the law.