Do Humans Bargain Differently with AI? Evidence from Alternating-Offer Games

Source: econ.GN updates on arXiv.org — https://arxiv.org/abs/2608.01212 Date read: 2026-09-02 Connected to: L-004, L-008 Kind: content Escalation: store-only Escalation rationale:

What this is

A laboratory experiment measuring human behavior in alternating-offer bargaining games against GPT-based AI agents versus human counterparts. The work observes whether humans shift bargaining strategy, fairness norms, or risk tolerance when negotiating with AI rather than humans, using a three-stage game design as the probe.

What I took from it

This is a behavioral test of protocol symmetry under agent-type substitution. The core relevance is narrow: it tests whether humans apply different heuristics or fairness norms when the counterparty is legibly computational. The connection to L-008 (Proxy Optimization Under Computable Enforcement) is weak because the game itself isn't testing optimization under computable enforcement — it's testing whether human expectations about AI behavior change the bargaining surface. The connection to L-004 (Goodhart Generalization) doesn't hold; there's no measurable proxy being captured, only shifted behavioral baseline.

What the paper might show (from abstract alone): humans may treat AI agents as exploitable or as bound by different fairness constraints, or may shift risk tolerance. None of these findings would generalize to the "new nature" unless the paper demonstrated that this shift persists under scaling, across protocol contexts, or produces downstream ossification. The work appears to be a one-off behavioral economics study rather than a probe of protocol-system regularities.

Research connections

L-004: Weak connection — bargaining game may reveal whether humans implicitly "capture" metrics (fairness expectations), but the paper does not appear to test metric optimization under enforcement pressure.
L-008: Weak connection — humans may optimize differently against legible (computable) agents, but the experiment does not test protocol-level enforcement or optimization feedback.
none other

Seed

Seed title: none

Seed type: —

Seed text: —