Does LLM moral advice persuade through reasoning or just recommendations?

In three pre-registered studies on everyday, non-sacrificial moral dilemmas — the kind people are actually likely to ask an LLM about, like whether to attend a birthday party or pick up litter in the park, rather than trolley-style sacrificial scenarios — Landes, Francis, and Everett find that LLM-generated moral advice is "equally persuasive" whether participants are told it comes from an AI or from a human philosopher (Study 1), even though participants still "rate human advisors as more trustworthy." In Study 2, manipulating a real LLM's track record (giving it a history of appropriate versus tangential advice) changed how trustworthy participants rated it, but the manipulation "did not have an effect on its persuasiveness." In Study 3, holding the LLM's quality constant and varying only the reasons behind its recommendation, "high-quality reasons do not increase persuasion relative to no reasons" — a bare recommendation was just as persuasive as one backed by good reasons — but "bad reasons may actively undermine it."

The authors read this pattern as deference rather than reasoned uptake. People are not updating their moral judgments because they judge the LLM trustworthy, nor because they evaluate its stated reasons as good reasons — if either were true, quality or reason manipulations would have moved persuasion, and they didn't. Instead, "users defer to the LLM on a response-by-response basis," a pattern driven by "uptake of the LLM's recommendation rather than uptake of its generated reasons." In the paper's own borrowed psychological language, the recommendation functions as "a peripheral cue rather than a central cue": removing reasons entirely doesn't weaken it, but a reason bad enough to be "patently absurd" does. Their stated conclusion is blunt: a heuristic of "this advice seems good enough" is not the way people should approach moral advice, because deference that isn't blind is still deference, with the same risks of moral deskilling and stunted growth that philosophers have long attributed to deferring to any advisor, human or AI.

This complicates Do people prefer AI moral reasoning when they don't know the source?, which found a trolley-problem anti-AI bias: participants who preferred LLM-generated content nonetheless agreed with it less once told the source was AI. Landes et al.'s Study 1 finds no such attribution penalty for everyday dilemmas — AI and human-philosopher labels produced equal persuasion despite the same trust gap the Moral Turing Test paper documents. The two results aren't necessarily contradictory; the dilemma types differ (sacrificial trolley scenarios inviting explicit utilitarian justification versus everyday egoism/altruism tradeoffs), which suggests attribution penalties on moral persuasion are dilemma-dependent rather than a stable anti-AI bias that generalizes across all moral content. The finding also runs alongside Can sycophantic AI advice still push people away from polarized views?: both locate persuasion in something about the advice itself rather than simple deference to source reputation, but this paper's Study 3 goes further, showing that the advice's content need not be reasoned through at all to do the persuading — the bare recommendation carries the effect that reasons were assumed to carry.

The excerpt doesn't establish whether this response-by-response deference holds up over repeated interactions or at real personal stakes rather than in single-shot vignettes — the paper's own cited related work on conspiracy-belief reduction found an effect that persisted at two months, but these three studies measure a single exposure. It also doesn't explain why bad reasons undermine persuasion while good reasons add nothing beyond the bare recommendation — whether that is a floor effect, triggered suspicion, or something else is left open. If the pattern generalizes, it implies that safeguards aimed at source transparency or published AI track records may protect moral judgment less than assumed, since in this design neither source credibility nor reasoning quality is what the persuasion is actually riding on.