The Week in Examples #13 [18 November]
Regulatory misalignment, a theory of AI governance, and alignment assemblies
This week, we have a policy brief on the AI regulatory alignment problem from the folks at Stanford, a paper from the Legal Priorities Project reviewing problems, options, and proposals in AI governance, and initial lessons from the Collective Intelligence Project’s alignment assemblies work. As usual, it’s hp464@cam.ac.uk to send me feedback or ideas for future editions.
Three things
1. Regulatory misalignment
What happened? Researchers at Stanford University’s Human-Centred Artificial Intelligence group released a policy brief organised around the idea that “calls to regulate create their own regulatory alignment problem, where proposals may distract, fail, or backfire.” To make this case, the authors consider the technical and institutional feasibility of four commonly proposed AI regulatory regimes: disclosure, registration, licensing, and auditing to argue that each could suffer from regulatory misalignment.
What’s interesting? They argue, for example, that a regime to prevent the release of personally identifiable information (PII) may instead provide incentives to configure algorithms to “aggressively mine data” without adequately protecting PII. The group also looks to 1) disparities in hiring to suggest bad regulation may see firms discard the feature most predictive of job performance, and 2) the prevention of bioweapon construction aided by large models to suggest that it may lead to the use of proprietary models whose deployment removes visibility over proliferation risks.
What else? The idea that regulation has unintended consequences is an old one, though it doesn’t tend to get much airtime in the world of AI policy. For those interested, Air Street Capital has a good essay defending the UK approach to regulation based on this concept. You’ll have to read the whole thing for more, but I’ll leave you with a quote from the political philosopher Chris Freiman that sums up the problem: “Perfect states beat imperfect markets, but that doesn’t establish the superiority of state solutions any more than finding that omnivorous non-smokers have lower rates of cancer than vegan smokers establishes the superiority of an omnivorous diet. We should compare like-to-like.”
2. The state of AI governance
What happened? Matthijs Maas of the Legal Priorities Project has written a mammoth state of play of the current AI governance landscape. The work aims to identify key features of the governance space (e.g. technical, political), draw into focus governance options within this environment (e.g. actors, actions, and mechanisms for influence) and a review of work aimed at putting research into practice in order to improve the governance of advanced AI.
What’s interesting? I especially like the final section, policy proposals, which gives a great rundown of the (many) types of proposals that have emerged in the last couple of years. These include proposals to regulate AI using existing laws or policies (e.g. strengthen or reorganise existing international institutions, extend or apply existing principles and regimes in international law); or proposals for new policies, laws, or institutions (e.g. pausing AI development, creating licensing regimes, establishing governance over AI inputs like compute and data).
What else? One of the handful of goals, as described by Maas, is to “Improve the field’s accessibility & reduce some of its ‘research debt’ [to] help those new to the field understand the existing literature, in order to facilitate a more cohesive and coordinated research field with lower barriers to entry.” I think the work certainly succeeds on that count, and I’d encourage anyone who wants to really get up to speed with the governance of frontier AI to carve out some time to read this paper.
3. Aligners assemble
What happened? The Collective Intelligence Project released early lessons from its Alignment Assemblies work with partners including the UK Frontier AI Taskforce (soon to be the AI Safety Institute), OpenAI, Anthropic, the US Summit for Democracy, the Taiwan Ministry Of Digital Affairs, and the Creative Commons Foundation. For those of you who aren’t familiar with the group, the Collective Intelligence Project aims to build the mechanisms by which public attitudes, beliefs, preferences, and perspectives can be introduced into the AI development process.
What’s interesting? In the research efforts with OpenAI and Anthropic, CIP engaged 1,000 people proportionally representative of the USA across age, gender, income, and geography. While they generally found it to be a useful approach, they also “recognise[d] that finding and engaging publics will vary from issue to issue and that representative sampling may not be best in every situation. For some, for example, we may want to guarantee particularly affected voices or key stakeholders are represented in the deliberation.” This is one of the central issues that I expect will receive lots of attention as these types of projects continue to get underway, primarily because the extent to which certain voices ought to be uplifted within a system of democratic representation is a long running and important question in political philosophy. Finding the right balance between different voices also poses some interesting methodological challenges, and I look forward to hearing about CIP plans on tackling them.
What else? The basic idea underpinning this work is that AI should be reflective of the people who use it. There’s tonnes of work about the extent to which these moves represent meaningful input (see a great paper from last month that gets under the skin of different approaches) but CIP’s work here looks promising. I also like that they argue that “the opportunity for public input into AI goes beyond policy-making; developers, designers, engineers, users and other people make important and socially impactful decisions throughout the AI lifecycle.” For my part, I see the key questions here as how to demarcate the boundaries between certain publics, what to do when little consensus emerges from public discussion, the extent to which input ought to be used throughout different parts of the AI value chain, and the level at which it is appropriate to localise the values of particular models.
Best of the rest
Friday 17 November
Unlocking the UK’s pensions for science and tech growth (Onward)
AI is coming for our jobs! Could universal basic income be the solution? (The Guardian)
OpenAI announces leadership transition (OpenAI)
AI chief quits over 'exploitative' copyright row (BBC)
Billionaires Niel, Saadé and Schmidt Invest in €300 Million AI Lab (Bloomberg)
Thursday 16 November
Rishi Sunak’s AI plan has no teeth – and once again, big tech is ready to exploit that (The Guardian)
Identifying AI-generated content with SynthID (Google DeepMind)
Regulating AI foundation models is crucial for innovation (Euractiv)
Introducing Emu Video and Emu Edit, our latest generative AI research milestones (Meta AI)
UK will refrain from regulating AI ‘in the short term’ (FT)
Wednesday 15 November
Is Argentina the First A.I. Election? (New York Times)
The Cambridge Dictionary Word of the Year is ‘Hallucinate’ (Cambridge Dictionary)
Interview with Dominc Cummings (Dwarkesh Patel)
Debate Helps Supervise Unreliable Experts (GitHub)
Underage Workers Are Training AI (WIRED)
Tuesday 14 November
Cybersecurity experts on BBC Radio 4 (BBC)
Will AIs fake alignment during training in order to get power? (arXiv)
DHS AI roadmap prioritises cybersecurity and national safety (AI News)
Baidu CEO warns China's rush to develop AI models risks wasting resources (Reuters)
How The Public Sector Can Benefit From Generative AI (Forbes)
Monday 13 November
Despite its big OpenAI push, Microsoft's Bing search market share decreases year-over-year (Windows Central)
Use Specialized Juries in AI Litigation (The Regulatory Review)
SAG-AFTRA Release Factsheet on Deal (SAG-AFTRA)
Giant AI Platform Introduces ‘Bounties’ for Deepfakes of Real People (404 Media)
White faces generated by AI are more convincing than photos, finds survey (The Guardian)
One for the road
