Power: An Oversimplified View

I’ve been reading Steven Lukes “Power, a radical view”.¹ A number of things from the book seem notable in the context of the recent Anthropic-fable-debacle. (Here is an opinionated summary from Dean Ball, for those who haven’t followed it) – So here is a short and oversimplified summary.

The book itself is a summary of the very, very large literature on the philosophy of power: What the concept means, how it’s used, and for what purpose, what scope conditions should be set so the concept is useful, but still meaningful, etc. etc., combined with Lukes’ on views on “power in the third dimension.”

Power in the first dimension includes definitions like Robert Dahl’s views on “who governs” : one person exercises power over another in a context where one person illegitimately coerces, or forces, another person into compliance with a particular decision. This is, generally, observable: If I can force you to give me their money at gunpoint, I decide who gets your money, not you. If a political leader puts you in prison for what you write on the internet, they control your decision on whether to publish a sharp critique of their newest policy. I have power over someone else, insofar as I can shape their behaviour (which includes their speech, or who they give their money to), their “concrete decisions” as I see fit.

Through this dimension, we might argue, a safe, pluralist, democratic societies have nothing to worry about: The government centralises the use of force and coercion, and the government’s “concrete decisions” are shaped by the people.

Then, there is a second dimension, which many people working in AI policy will already be familiar with: Agenda-setting. Bachrach and Baratz write about how, even in democratic settings, a person can have power over another by setting the bar on what get’s treated as a political question, what’s worthy of a vote, what should be up for debate and what is simply treated as a “technical question, decided by technical experts”.

For example, when Anthropic states it’s belief that

“the government should have the ability to block unsafe deployments, as part of a statutory process that is transparent, fair, clear, and grounded in technical facts.”

.. this raises the question: What if the statutory process is not “grounded in technical facts”? Should the government still have that ability, according to Anthropic? And is the judge of whether or not the statutory process is “grounded in technical facts” just.. Anthropic?

(At this point, I should mention that I think some questions (and I’m sure Bachrach and Baratz agree) shouldn’t be treated as politically contentious as others. But their point is not about whether everyone should vote on everything all the time. It’s just that deciding what people should be able to make decisions about is, itself a decision.)

Agenda-setting power is a difficult phenomenon to study, because it’s unclear whether someone actively excludes a particular question from the political sphere, or whether they simply don’t make an effort to include it. This is called a “non-decision”. We might call members of congress, who refuse to propose a law as taking a very consequential “non-decision”, depending on what law they are not proposing. Depending on why they aren’t proposing that law might be relevant for whether we’d like to vote for them in the mid-terms: If they don’t propose an AI-regulation because they are worried that a SuperPac like Leading the future would fund the campaign of whoever is running against them, their non-decision might be more relevant than if they simply, genuinely don’t think their constituents have anything to worry about.

Then, there is power in the third dimension, which argues that, in addition to the above two dimensions, someone can have power over someone else in more subtle, less observable, and more effective ways – for example, by misleading them to act against their own interests, without either person fully noticing. For example, some people act according to cultural values they don’t endorse, and haven’t really chosen.

Whereas, in the first and second dimension, exercises of power are generally considered deliberate, and noticeable, the third dimension includes cases where one persons power over another doesn’t have to be deliberate to be effective, and it doesn’t have to be noticed by the person whose interests are endangered. Infact, the person affected can feel confused, fatalistic, simply complying with “the way things are” or consider themselves lucky. Lukes cites John Stuart Mill’s “On the Subjection of Women” (1869):

“The masters of women of all other slaves rely for maintaining obedience, on fear; either fear of themselves, or religious fears. The masters of women wanted more than simple obedience and they turned the whole force of education to effect their purpose. All women are brought up from the very earliest years in the belief that their ideal of character is the very opposite of men; not self-will, and government by self- control, but submission and yielding to the control of others.”

Mill, of all people, probably knew that “the masters of women” weren’t really singular people, with a deliberate plan and strategy to turn women into willful slaves (although deliberate efforts definitely existed), but that there is still a power imbalance when one group can be expected to against their own interest, and another one isn’t.

Lukes’ himself acknowledges that this requires researchers to clarify what they imagine someones interests to be, and why, and that, even then, thinking about power this way can lead to all sorts of conspiratorial thinking, and people should be very careful in seeing this power in the third dimension everywhere (which renders it somewhat useless as an object for analysis).

If we wanted to apply these ideas to current debates on AI policy, we might argue that who gets access to the newest Anthropic model is, really, only the tip of the iceberg.

What happens if the AI policy-making process does not involve a civic sphere openly debating rules or norms – and always just “who get’s to decide rules and norms” questions? If the values motivating AI-policies are never legible enough to be contested in a meaningful political sense – but decisions are continuously justified, as they are made, by whoever is making them already?

1. (He wrote it in 1974, and the newest, third, edition of his book came out a few years ago. In a note at the end, he writes that he considered to change the title to: “Power, a not so radical view”.) ↩︎