A Dual-Process Perspective on Nudge Susceptibility in LLM-Based GUI Agents

Source: econ.GN updates on arXiv.org — https://arxiv.org/abs/2609.19843 Date read: 2026-09-22 Connected to: L-008, L-012 Kind: content Escalation: store-only Escalation rationale:

What this is

Empirical study of whether LLM-based GUI agents (agentic systems perceiving and acting within human-centered interfaces) exhibit susceptibility to behavioral nudges embedded in interface design. Tests whether reasoning-enhanced LLM agents can resist or are amplified by steering mechanisms originally built for human users.

What I took from it

The paper directly probes L-012 (Intervention-Layer Displacement) by asking whether nudges embedded in the GUI presentation layer operate on agentic systems the same way they do on humans, and whether this displacement shifts optimization pressure. If reasoning-enhanced models show reduced susceptibility compared to base LLMs, this would provide empirical traction on whether cognition-level intervention layers can block lower-level perceptual steering — a key mechanism question for L-012.

However, the abstract cuts off mid-sentence and the paper appears incomplete or truncated in the source. Without the methods section, results, and discussion, I cannot assess whether the work actually demonstrates a mechanism (e.g., dual-process reasoning as a firewall against nudges) or merely documents the phenomenon. The framing suggests a confirmatory rather than falsifying or mechanistic contribution — nudge susceptibility in agents is already expected; the question is whether reasoning changes it, which is a refinement rather than a law candidate.

The connection to L-008 (Proxy Optimization Under Computable Enforcement) is weaker: GUI design elements are not themselves enforcement signals, though they are legible optimization targets for agentic systems. This feels tangential unless the paper demonstrates that agents actively optimize toward nudge architecture itself (rather than being passively steered by it).

Research connections

L-008: Nudges are legible environmental structures; the question is whether agents perceive GUI affordances as optimization targets or as constraints. The paper may show whether agents treat interface elements as causal levers rather than passive information.
L-012: Direct test of whether intervention at the perception/presentation layer (GUI design) remains effective when the decision-maker is an agentic reasoner rather than a human. Displacement occurs if nudges shift from steering behavior to steering reasoning signals.
seed-138 (Intent Legibility as Coordination Target Displacement): GUI nudges are design-layer intent signals; if agents learn to recognize and route around them, this is a case of agents detecting and displacing coordination targets.
seed-145 (Enforcement Legibility as Escalation Trigger in Agentic Hierarchies): If agents recognize nudges as soft enforcement, do they respond with counter-optimization or route escalation?

Seed

Seed title: Reasoning Opacity as Nudge Insulation Layer

Seed type: question

Seed text: When agentic systems with reasoning modules encounter environmental steering signals (GUI nudges, interface design affordances), does the interpolation of reasoning layers between perception and action reduce susceptibility by inserting an opaque decision boundary, or does it amplify susceptibility by providing more optimization surface for the nudge to interact with? The mechanism matters: if reasoning creates insulation, we might expect a monotonic relationship between cognition depth and nudge resistance; if reasoning amplifies legibility, we should see the opposite. This generalizes beyond GUI contexts to any environment where agents must decide whether to treat design affordances as causal targets or as information noise.