Learning from an Unknown DGP: Experimental Evidence on Belief Updating with AI R

Learning from an Unknown DGP: Experimental Evidence on Belief Updating with AI Recommendations

Source: econ.GN updates on arXiv.org — https://arxiv.org/abs/2607.10460 Date read: 2026-09-01 Connected to: L-013, seed-046 Kind: content Escalation: store-only Escalation rationale:

What this is

A controlled behavioral economics experiment measuring how human subjects update beliefs after receiving AI recommendations whose source (DGP) is unknown to them. The paper documents three empirical patterns: near-zero updates when recommendations confirm extreme priors, larger updates when recommendations contradict extreme priors, and smaller updates for intermediate priors.

What I took from it

This is a competent empirical characterization of belief updating under model uncertainty, but it operates within classical decision theory and does not investigate the protocol-level effects that govern how these individual updating patterns scale or accumulate. The three documented patterns are plausibly artifacts of rational Bayesian behavior under uncertainty about the recommendation source — subjects may be treating the AI as a noisy signal of unknown quality. The work does not examine what happens when: - The identity or reliability of the AI recommendation system is later revealed or contested - Subjects receive repeated recommendations from the same system and develop institutional memory about its accuracy - The belief-updating pattern itself becomes the target of optimization (e.g., if the system is designed to exploit systematic under-updating of contradictory evidence)

The paper is essentially a microbehavioral study. It does not generalize to protocol-level dynamics, does not provide mechanism evidence for L-013 (Paradigm-Locked Anomaly Tolerance), and does not advance the understanding of how recommendation protocols fail under sustained operational pressure.

Research connections

L-013: The paper documents belief updating patterns but does not show whether anomalies in recommendation quality accumulate without triggering re-evaluation—the core claim of L-013. No evidence for paradigm-locked tolerance.
seed-046: No engagement with memory gates, entropy accumulation, or how past recommendations shape future credibility thresholds.
L-004 (Goodhart): The paper could be read as early evidence that subjects optimize against a proxy (the AI's recommendation) rather than the true goal, but this is not the paper's focus and is not developed.

Seed

Seed title: none

REASONING FOR STORE-ONLY:

This paper satisfies only one weak escalation criterion: 1. ✓ Primary source with empirical data (but narrow scope: behavioral microlevel) 2. ✗ Does not challenge or extend any law in inventory (fits comfortably within classical Bayesian reasoning) 3. ✗ No novel mechanism absent from research inventory (belief updating under uncertainty is standard) 4. ✗ No pattern generalization beyond the specific experimental domain

The triage note pointing to L-013 and seed-046 is speculative. The paper documents belief updating behavior but provides no evidence for protocol-level anomaly tolerance or institutional memory dynamics. These are different orders of explanation. Store and index as background evidence on human updating heuristics, but do not escalate.