Language Log

« previous post |

In a comment on "Alignment", Sniffnoy wrote:

At least as far as I'm aware, the application of "alignment" to AI comes from Eliezer Yudkowsky or at least someone in his circles. He used to speak of "friendly AI" and "unfriendly AI". However, the meaning of these terms was fairly different from the plain meaning, which confused people. So at some point he switched to talking about "aligned" or "unaligned" AI.

This is certainly true — see e.g. Yudkowsky's 2016 essay "The AI alignment problem: why it is hard, and where to start".

However, an (almost?) exactly parallel usage was established in the sociological literature, half a century earlier, as discussed in Randall Stokes and John Hewitt, "Aligning actions" (1976):

A substantial body of literature has been developed within the symbolic interactionist tradition that focuses upon various tactics, ploys, methods, procedures and techniques found in social interaction under those circumstances where some feature of a situation is problematic. Mills' (1940) concept of motive talk, Scott and Lyman's (1968) discussion of accounts, Hewitt and Hall's (1973) and Hall and Hewitt's (1970) quasi-theorists, and Hewitt and Stokes' (1975) disclaimers are among the contributions to this literature. In addition, some of Goffman's work (1959; 1967; 1971) addresses itself to a similar set of issues, and McHugh's (1968) analysis of the concept of the definition of the situation is pertinent to the question of how people deal with problematic occurrences.

We refer to these phenomena collectively as aligning actions. Largely verbal efforts to restore or assure meaningful interaction in the face of problematic situations of one kind or another, activities such as disclaiming, requesting and giving accounts, constructing quasi-theoretical explanations of problematic situations, offering apologies, formulating the definition of a situation, and talking about motives illustrate a dual process of alignment. First, such activities are crucial to the process in which people create and sustain joint action by aligning individual lines of conduct when obstacles arise in its path. Second, and of particular import for the present analysis, aligning actions can be shown to play a major part in sustaining a relationship between culture and conduct, in maintaining an alignment between the two in the face of actions that depart from cultural expectations or definitions of what is situationally appropriate.

More from later in the paper:

Much, though not all, that is problematic in everyday life can be conceived in terms of a metaphor of alignment, a term that has a double meaning in the present analysis. First, alignment is a central metaphor in the interactionist analysis of conduct formation. Social interaction is con- ceived as a process in which people orient their conduct toward one another and toward a common set of objects. In this mutual orientation of conduct, an effort is made by participants to align their indi- vidual acts, one to another, in the creation of joint or social acts. 

[…]

The second meaning of alignment — and in the present essay the more crucial one  — revolves around the fact that problematic situations often involve misalignment between the actual or intended acts of participants and cultural ideals, expectations, beliefs, knowledge, and the like. "Alignment" in this sense has to do with perceived discrepancies between what is actually taking place in a given situation and what is thought to be typical, normatively expected, probable, desirable or, in other respects, more in accord with what is culturally normal.

That second sense is exactly what is now meant by alignment in the "AI alignment" discussion, or so it seems to me.

Yudkowsky's 2016 essay doesn't cite the sociological usage, and there's no bibliography to check — according to footnote 1 , "This document is a complete transcript of a talk that Eliezer Yudkowsky gave at Stanford University for the 26th Annual Symbolic Systems Distinguished Speaker series". I don't find a reference in a quick scan of his other publications  either, so presumably he perceived the term as just a normal part of the language of intellectual discourse.

Also unclear to me is the connection between the sociologists' alignment and the D&D version.

But anyhow, as the earlier post noted, "alignment, like journey, is an old word that has been finding new meanings and broader uses over the past few decades".

July 17, 2025 @ 4:03 am · Filed by Mark Liberman under Words words words

Permalink