Should AI agents serve one user or balance many interests?

The article argues that "the AI alignment problem has traditionally been cast as the challenge of ensuring that a single AI agent reliably pursues the goals of a single person" — a framing it calls "single-single" alignment — and that this framing is "insufficient" once agents act at scale across many people with "different (and often conflicting) goals." It proposes three replacement governance principles: inclusivity, visibility, and liability. On inclusivity, its central line is that "the goal should not be to build AI agents that exhibit 'single-minded loyalty' but rather agents that aim to promote a more diverse and pluralistic set of interests and values." On visibility, it holds that without "rigorously studying and tracking AI agents," neither consumers nor regulators can trust that agents operate safely, so visibility "should... be a priority for both developers and regulators." The conclusion adds that governance tools must also hold "developers, deployers, and users accountable if and when harm occurs."

The reasoning connects back to agency law's basic question, restated here as a design question: "whose interests should AI agents serve? In other words, who is the principal and to whom (or what) are fiduciary duties owed?" The article's worry about pure single-single loyalty is twofold — it "might embolden malicious actors to use AI agents for nefarious purposes," and even absent misuse, it "can result in artificial agents making decisions and taking actions that come at the expense of broader societal interests and values." Visibility is justified instrumentally, not just normatively: it "can help identify current and anticipated problems," "facilitate interventions," and "assist in evaluating the efficacy of strategies for governing AI agents" — three distinct functions bundled into one principle.

This extends the diagnostic claim in Can traditional governance tools actually control AI agents? into prescription: where that note's claim is that uninterpretable, fast agents break monitoring and enforcement, this note's visibility and liability principles are the article's proposed replacements for those broken tools. It also reframes Does granting agents more autonomy undermine human oversight? as a problem of system design producing the opposite of visibility, and gives that erosion a name — a visibility deficit — plus a stated remedy, rather than leaving it as an unaddressed trend. And it sits in tension with the loyalty-only framing implicit in Does a benign goal actually prevent harmful AI behavior?: that note locates risk in an agent's competence at reasoning about any goal, regardless of whose goal, while this article locates part of the fix in deliberately diversifying whose goals an agent must balance.

The excerpt states principles, not implementations: it does not specify what an inclusivity-respecting objective function looks like, how visibility tooling would work given "the absence of a systematic understanding of the underlying language models," or how liability would be apportioned among developers, deployers, and users in a concrete dispute. The implication it supports is accordingly a call to build infrastructure rather than a solved design — that single-single alignment, even if achieved perfectly, would leave open exactly the harms the article is most concerned about.