A Dual-Process Perspective on Nudge Susceptibility in LLM-Based GUI Agents
Haya Halimeh, Sascha Kaltenpoth, Kevin B\"osch, Oliver M\"uller
cs.AI · econ.GN · q-fin.EC
Quick verdict
LLM GUI agents follow defaults (OR 8.7) and social cues (OR 37); reasoning redirects influence rather than curbing it; the 'nudge=bias' reading is debatable.
Summary
The paper tests whether LLM-based GUI agents — models that browse and act in human-designed interfaces — are susceptible to digital nudges, and whether their reasoning configuration changes this susceptibility. Adapting a human online-shopping experiment (Ingendahl et al., 2021), the authors built a simulated grocery store accessed via the Browser Use library and ran 3,600 agents (six models from OpenAI, Google and Anthropic, each in flagship and small variants; no-reasoning vs high-reasoning) across control, default-nudge and social-influence-nudge conditions, yielding 21,600 binary product choices analyzed with Bayesian multilevel logistic regressions. Both nudges worked: defaults multiplied the odds of choosing the nudged product by 8.70 and social cues by 37.24, in every model. Reasoning moderated the two nudge types in opposite directions — high reasoning roughly halved default susceptibility (OR 0.51) but amplified social-influence susceptibility (OR 1.71). Exploratory analyses suggest default resistance concentrates in flagship models (Gemini 3.5 Flash falls from 84% to 48%) and social amplification in small variants (GPT-5.4-mini rises from 91% to 99%). A letter-counting manipulation check (80.3%→99.3% accuracy) grounds the System 1/2 reading, though the authors themselves show the toggle is provider-dependent. Main limitations: only two nudge types, ceiling effects in the Claude models, and a contested reading of informative social cues as 'bias'.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
4 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- LLM-based GUI agents are highly susceptible to both default (pooled OR 8.70, CI [7.54, 9.97]) and social-influence (pooled OR 37.24, CI [31.04, 44.20]) nudges, in every one of six models tested.
- High reasoning halves susceptibility to default nudges (OR 0.51) while raising susceptibility to social-influence nudges (OR 1.71): deliberation redirects influence rather than removing it.
- Exploratory analyses suggest the reasoning moderation is structured by model scale — default resistance concentrates in flagship models (Gemini 3.5 Flash: 84%→48%), social-influence amplification in small variants (GPT-5.4-mini: 91%→99%).
- Nudge susceptibility is a behavioural property of agentic AI, making interface design a governance concern for organizations delegating decisions to agents.
Strengths
Large, properly counterbalanced randomized design
3,600 agents × 6 decisions = 21,600 choices across six models, three providers, two reasoning configs and three conditions (Sec. 4.5); the control condition sits at ~50% (CI [47, 53]), empirically validating the A/B counterbalancing against inherent product preferences (Sec. 5.1).
Exemplary manipulation check
Appendix E validates the reasoning toggle on Kahneman's letter-counting task (6,000 API calls): accuracy 80.3%→99.3%, gains concentrated on novel pseudo-words, stratified Cochran–Mantel–Haenszel tests with randomization p-values (pooled OR 266.1, p<0.001), plus dual-labeller trace annotation (κ=0.996) showing decomposition moves into hidden traces (any-channel 59.4%→94.0%).
Honest handling of ceiling/separation
Robust, well-reported inference
Weaknesses
The Type 2 'nudge' carries decision-relevant information
Reasoning manipulation strength is provider-dependent
Scale is confounded with provider
Narrow behavioural coverage
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper asks whether LLM-based GUI agents — models that perceive and act in human-designed interfaces via a perception-plan-act loop — are susceptible to digital nudges, and whether their reasoning configuration moderates this. Three hypotheses: H1, both Type 1 (default) and Type 2 (social influence) nudges shift agent choices; H2a, high reasoning attenuates default susceptibility; H2b, high reasoning amplifies social-influence susceptibility. All three are supported. The headline framing is that deliberation 'redirects rather than removes' choice-architecture influence, with an exploratory claim that the moderation is structured by model scale, and a practical claim that interface design becomes a governance concern for organizations delegating decisions to agents.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 4 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Does giving LLM agents more reasoning make them immune to nudges and dark patterns?
No. In this study of 3,600 agents across six models and 21,600 simulated shopping decisions, high reasoning roughly halved susceptibility to default-option nudges (OR 0.51) but increased susceptibility to social-proof style nudges by about 70% (OR 1.71). Reasoning redirected which nudges work rather than eliminating influence.
How large were the nudge effects on LLM shopping agents?
Pooled across six models from three providers, a preselected default multiplied the odds of choosing the nudged product by 8.70, and a 'most customers prefer this' style label multiplied them by 37.24, compared with a counterbalanced 50% baseline where neither nudge was present.
Is this evidence that AI agents are irrational like humans?
Not straightforwardly. The social-influence nudge carried real information (a stated highest-satisfaction rating), so shifting toward it could reflect rational updating rather than bias. The authors frame the result as a governance issue — interface design steers delegated agent decisions — rather than proof of human-like irrationality, and the dual-process mapping is explicitly functional, not mechanistic.
Can the study be reproduced?
Code, prompts and analyses are public (anonymous.4open.science repository, App. refs in Sec. 4.2), and model configurations are documented. Exact replication requires paid API access to the specific commercial model versions studied, and the authors warn that providers change weights and behaviour under unchanged model names, so absolute susceptibility levels may differ on re-run.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures (Figs. 1–6, B1, C1, D1, E1–E2) could not be analyzed; all quantitative judgments derive from extracted tables and in-text numbers.
- Two extraction gaps exist (part of Section 2.1's mixed-evidence discussion and part of the deployment-safeguards discussion); the surrounding text preserves the argument.
- The public repository's contents could not be verified from the provided text.
- The claim that agents respond more strongly than the human benchmark rests on Ingendahl et al. (2021) numbers not reproduced in the extracted text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.