Do People Follow AI Advice? Evidence from a Pension Portfolio Choice Experiment

Source: econ.GN updates on arXiv.org — https://arxiv.org/abs/2608.11371 Date read: 2026-09-02 Connected to: L-004, L-012 Kind: content Escalation: store-only Escalation rationale:

What this is

A behavioral economics experiment (N=400, South Korean pension participants) measuring portfolio choice revision following AI recommendation presentation in a 2×2 design (recommendation content × presence of rationale). The work quantifies advice-following rates and compliance heterogeneity across recommendation types and demographic subgroups.

What I took from it

The experiment sits at the intersection of L-012 (intervention-layer displacement) and L-004 (Goodhart generalization), but delivers neither sustained theoretical argument nor mechanism evidence. The 37% aggregate follow rate is descriptively useful but does not isolate whether compliance is driven by rational delegation, legibility-substitution for missing advisor trust, or optimization-pressure displacement. The paper does not track why subjects revise or whether the AI recommendation functions as a legitimacy proxy rather than information source—the key distinction for understanding whether we're observing delegation or formalization-capture.

The rationale condition (presence vs. absence of explanation) would be the experimental lever for testing seed-072 (explanation-marker decoupling) or seed-069 (transparency-legibility as trust proxy substitution), but the abstract provides no detail on this result, and the design does not appear to separate legibility effects from epistemic authority effects. This reads as a competent behavioral study of compliance elasticity, not a mechanism investigation.

Research connections

L-012: The experiment measures behavioral response to formalized AI input, but does not distinguish between subjects treating recommendation as information vs. subjects treating it as coordination signal or legitimacy marker. No evidence for displacement of optimization pressure.
L-004: The paper does not address whether subjects optimize toward matching the AI recommendation (target capture) or toward the underlying portfolio objective that the AI purports to optimize for. No measurement of proxy divergence.
seed-069: Presence/absence of rationale could test whether transparency substitutes for trust, but results are not reported in abstract and design does not control for advisor credibility or institutional trust separately.
seed-077: No measurement of preference shift post-recommendation (ratcheting effect).

Seed

Seed title: none