Paper Type
Short
Paper Number
PACIS2026-1232
Description
As autonomous agents based on Large Language Models (LLMs) increasingly mediate competitive decisions in digital markets, whether they can be governed through institutionalized norms—rather than mechanical constraints alone—becomes central to information systems research. We deploy two Gemini-2.5 agents in a repeated Stackelberg duopoly under five governance regimes and document three findings. First, a Safe Agent Paradox: baseline predation rates remain near 13% even without governance, revealing a latent safety floor inherited from instruction tuning. Second, normative signaling alone functions as cheap talk. Third, agents exhibit structural inertia rather than moral drift; behavior locks in early and persists. A dual-control regime combining normative framing with probabilistic, interpretable enforcement produces directional improvements in compliance and successful behavioral generalization to a novel market. We introduce algorithmic institutionalization as a construct that extends institutional theory to LLM-agents and reframes the AI governance problem from preventing norm decay to engineering correct behavioral initialization.
Recommended Citation
Bojja, Giridhar Reddy and E, Feiyu, "Learning to Behave: The Paradox of Compliance and Structural Inertia in Large Language Model Agents" (2026). PACIS 2026 Proceedings. 3.
https://aisel.aisnet.org/pacis2026/it_strategy/it_strategy/3
Learning to Behave: The Paradox of Compliance and Structural Inertia in Large Language Model Agents
As autonomous agents based on Large Language Models (LLMs) increasingly mediate competitive decisions in digital markets, whether they can be governed through institutionalized norms—rather than mechanical constraints alone—becomes central to information systems research. We deploy two Gemini-2.5 agents in a repeated Stackelberg duopoly under five governance regimes and document three findings. First, a Safe Agent Paradox: baseline predation rates remain near 13% even without governance, revealing a latent safety floor inherited from instruction tuning. Second, normative signaling alone functions as cheap talk. Third, agents exhibit structural inertia rather than moral drift; behavior locks in early and persists. A dual-control regime combining normative framing with probabilistic, interpretable enforcement produces directional improvements in compliance and successful behavioral generalization to a novel market. We introduce algorithmic institutionalization as a construct that extends institutional theory to LLM-agents and reframes the AI governance problem from preventing norm decay to engineering correct behavioral initialization.
Comments
11-Strategy