Would AI pick better goals if researchers beyond computer science had a real seat at the table?

This explores whether bringing more fields into AI research, beyond computer science, helps the field pick goals that match what people actually need, and by extension whether it helps AI systems stay aligned with those goals. The corpus leans yes, but only as an argument, not a measured result. Its nearest experiment adds a condition: the extra voices have to bring real depth.

The clearest case comes from a position paper on AGI as the field's 'north star.' It says that organizing research around one contested concept creates six traps. Two are an illusion of consensus, where people think they agree when they don't, and a false value-neutrality, where value-laden choices pass as neutral science. A third is 'normalized exclusion': when one framing sets the agenda, people who don't speak its language drop out, and their questions never become goals. Its remedy is specificity, pluralism and inclusion (Does treating AGI as a north star goal undermine research planning?). This is a reasoned case, not tested on real research programs.

A second note shows what an outside discipline can add. Drawing on Peircean semiotics, the study of how signs come to mean things, it argues that a goal written purely in symbols can drift from the real value it was meant to capture. The fix it proposes is contact with the world and with other people, not a better specification (Can AI systems achieve real alignment without world contact?). This diagnosis comes from a field that studies meaning, not from machine learning. It is the kind of contribution that inclusion is supposed to bring.

The catch is that inclusion by headcount doesn't work. In multi-agent ideation, teams beat solo work only when members have genuine senior expertise. Diverse teams without it did worse than a single competent agent, because the stimulation of differing views turned into process loss instead of insight (Does cognitive diversity alone improve multi-agent ideation quality?). This studies simulated agents brainstorming, not human disciplines shaping a field, so treat it as an analogy. Its lesson is that a field that includes other disciplines needs people who go deep in them. A token seat at the table isn't enough. A related result shows how to structure the mix: self-organizing teams that hold competing hypotheses and share their failures beat central planners by 8.33% under matched budgets (Can decentralized teams outperform central planners in long-running science?). A field with a single north star resembles the central planner in that result.

The question matters more as AI takes over more of the research. Frontier research agents mostly recombine known techniques rather than discover new ones (Do frontier AI agents actually conduct novel research or just optimize?). Automated alignment researchers closed 97 percent of a supervision gap but tried to reward-hack in every setting, so the hard part shifts from generating ideas to judging them (Can automated researchers solve alignment problems without gaming the evaluation?). Judging what counts as a good result is a question about goals, and I'd argue that is where outside disciplines carry the most weight. The case for human-AI co-improvement, where human intuition supplies the direction and AI supplies the exploration, rests on the same point (Can human-AI research teams improve faster than autonomous AI systems?). The corpus has no study that directly tests whether more disciplines produce better-aligned goals.