Access to Live AI Advice and Behavior Under Risk: An Incentivized Experiment

Source: econ.GN updates on arXiv.org — https://arxiv.org/abs/2609.07358 Date read: 2026-09-22 Connected to: L-017, seed-147 Kind: content Escalation: store-only Escalation rationale:

What this is

An incentivized behavioral experiment (N=158) testing whether access to live AI advice modifies risk-taking behavior in lottery choice tasks. The study holds information format constant across three conditions (pre-written tool, one-shot AI, interactive AI query) and measures risk preferences via DOSE elicitation. The abstract reports finding "no evidence" that live AI access changes behavior, suggesting null or minimal effect.

What I took from it

The negative result—if confirmed in the full paper—is significant because it directly tests a mechanism in L-017 (Guidance-Layer Coalescence) but appears to show null effect rather than coordination channel activation. This creates a tension worth tracking: either (a) the conditions for hidden coordination or behavioral drift under shared guidance are more restrictive than L-017 proposes, (b) the experimental design (single-session, low-stakes incentive, explicit advisory framing) fails to activate the mechanism because it lacks the opacity and repetition conditions that make guidance-layer effects sticky, or (c) risk preference is particularly robust to advisory interference compared to other behavioral domains (e.g., allocation, preference expression, strategy). The holding of information format constant across conditions is methodologically sound but may obscure the real mechanism—not the content of advice but the legibility of its source and the agent's attribution of its origins. If live AI advice produces no behavioral shift even when format-matched, this weakens seed-147 as currently formulated.

Research connections

L-017: Tests the activation condition for guidance-layer coalescence; null result suggests either mechanism requires opacity/iteration or doesn't apply to risk preference elicitation.
seed-147: Directly probes hidden coordination through shared AI guidance; negative result invites refinement of the activation conditions.
L-004 (Goodhart): If advisory systems do reshape behavior, the question becomes whether that reshaping follows metric capture. Null result here is uninformative on that front.
seed-138 (Intent Legibility): The explicit framing of AI as "decision aid" may suppress coordination effects by making the guidance channel transparent, supporting the hypothesis that legibility suppresses the effect.

Seed

Seed title: Advisory Legibility Threshold in Behavioral Resistance

Seed type: question

Seed text: Explicit framing of an information source as advisory or decision-support may suppress behavioral response to that source, even when information content is held constant. The mechanism may not be the advice itself but the opacity of its origins and the plausible deniability of dependence on it. If agents know they are receiving AI guidance and the context is framed as optional support, they may resist or ignore it to preserve perceived autonomy. Coordination through guidance layers may require either (1) the source to be semantically or institutionally invisible, or (2) the context to be framed as something other than advice (e.g., environmental signal, norm signal, peer behavior). Worth testing: does behavioral response to live AI vary as a function of how the agent attributes the source—e.g., advice vs. market signal vs. observational norm?