Risk-aware users can make LLMs look worse: when privacy, safety, or ethical concerns prompt limited disclosure and shallow interaction, model outputs decline and trust falls, creating a self-reinforcing cycle of guarded use. At scale this loop can narrow visible applications and strengthen negative narratives, with implications for organizational design, governance, and AI system interfaces.

The guarded engagement loop: risk salience and interaction-driven underperformance in generative AI adoption

Connie Syharat, Arash Zaghi, Sarira Motaref · June 19, 2026 · Frontiers in Research Metrics and Analytics

openalex theoretical n/a evidence 7/10 relevance Full text usable extracted full text DOI Source PDF

The paper introduces the 'guarded engagement loop'—a conceptual model where elevated risk salience leads users to disclose less context and iterate less with LLMs, degrading output quality and trust and thereby reinforcing cautious adoption at both individual and societal levels.

Generative AI adoption is often framed primarily as a question of learning technical skills. It is thought that if users learn better prompting and evaluation practices, useful outputs will follow, leading to greater reliance on the technology. This perspective overlooks a defining feature of large language models (LLMs): their output quality depends heavily on how users engage with them. Because LLM performance varies substantially with depth of disclosure, contextual richness, and iterative refinement, user interaction strategies directly shape perceived usefulness and observed performance. This paper develops a conceptual framework that proposes how risk salience may shape these interaction dynamics. Drawing on research in trust in automation, privacy calculus, algorithm aversion, and the social amplification of risk, we propose the guarded engagement loop, a multilevel feedback mechanism in which risk perceptions may shape interaction strategies that influence observed performance and, in turn, recalibrate trust in generative AI systems. At the micro level, elevated risk salience related to privacy, safety, or ethical concerns may lead users to adopt guarded interaction strategies characterized by reduced contextual disclosure and limited iteration. These constrained interactions can lower output quality and increase the likelihood of visible errors, which may further erode trust and reinforce cautious engagement. At the macro level, values-driven withdrawal from AI use has the potential to narrow the diversity of visible applications, amplifying risk-focused narratives, reinforcing perceptions of harm in public discourse. The guarded engagement loop framework conceptualizes generative AI adoption as a feedback process in which risk perceptions may shape interaction conditions that, in turn, can influence observed performance and subsequent trust calibration. We articulate testable propositions and discuss implications for organizational governance, AI system design, and institutional conditions that enable bounded openness and calibrated reliance.

Summary

Main Finding

The paper proposes the "guarded engagement loop," a multilevel conceptual framework showing that salience of AI-related risks (privacy, safety, ethics, accountability) causes users to adopt guarded interaction strategies with generative AI (LLMs). These constrained interactions—reduced contextual disclosure, limited iteration, and avoidance of exploration—lower interaction richness, produce poorer or more error-prone outputs (or outputs perceived as poor), and thereby reduce trust. Lower trust in turn reinforces guarded engagement, producing a self‑confirming feedback loop. At the macro level, values-driven withdrawal can narrow the visible set of applications and amplify risk-focused public narratives, further sustaining underperformance and limited adoption.

Key Points

Core constructs
- Risk salience: cognitive prominence of potential negative consequences of AI use.
- Guarded engagement: a multidimensional behavioral orientation (reduced disclosure, limited iteration, avoidance of exploration).
- Interaction richness: informational and iterative depth of the user–AI exchange (contextual detail + iterative refinement).
- Objective vs. perceived performance: actual output quality vs. user evaluation of usefulness/reliability.
- Trust calibration: alignment between user trust and system capabilities; drives future engagement.
Micro-level mechanism
- Higher risk salience → guarded engagement → reduced interaction richness → lower objective/perceived performance → trust erosion → more guarded engagement.
Macro-level mechanism
- Collective withdrawal or concentrated use by risk-tolerant populations narrows observable applications, feeding media and institutional narratives that emphasize harms and risks (social amplification of risk), which increases risk salience for others.
Consequence: underperformance in deployed settings can be interaction-driven (not solely capability-driven). Training in prompt skills alone may be insufficient if users are deterred from the interaction conditions needed for high-quality outputs.
Related literatures synthesized: trust in automation, privacy calculus/information disclosure, algorithm aversion, media/social amplification of risk, media richness theory, prior findings that prompt/context sensitivity drives LLM output quality.
The paper articulates testable propositions (micro and macro) and discusses organizational, system-design, and institutional interventions (e.g., bounded openness, privacy-preserving features, psychological safety, governance) to break the loop.

Data & Methods

This is a hypothesis-and-theory / conceptual paper (no novel quantitative dataset).
Methods: integrative literature synthesis and conceptual modeling.
- Draws on empirical and theoretical findings from multiple literatures (e.g., Dietvorst et al. on algorithm aversion; Featherman & Pavlou on privacy calculus; Lee & See on trust; Kasperson et al. on social amplification of risk; prompt/LLM sensitivity literature).
- Uses conceptual argumentation to specify constructs, causal links, and feedback dynamics (includes a table of constructs and explicit propositions).
- References qualitative supporting examples (e.g., student interactions with an AI Virtual Mentor) and prior empirical work to motivate and ground the propositions.
The product is a theoretically grounded framework and a set of testable propositions (suitable for experimental, field, or observational follow-up).

Implications for AI Economics

Measurement and evaluation of AI-driven productivity
- Observed underperformance in field deployments may understate LLM potential if users constrain inputs/iteration because of risk concerns. Productivity estimates that ignore interaction conditions will be biased downward.
- Evaluations should record interaction richness (context provided, number of iterations, openness of tasks) alongside output quality to separate capability from usage effects.
Technology adoption and diffusion models
- Standard adoption models (e.g., TAM, UTAUT) assume stable system performance after learning; generative AI challenges this by making performance interaction-dependent. Diffusion dynamics may exhibit self-reinforcing stagnation if risk salience is high.
- Aggregate adoption curves and welfare gains from AI may be flattened by guarded engagement; heterogeneous risk perceptions across subpopulations can produce differential uptake and widening gaps in productivity gains.
Labor markets and returns to skills
- If risk-averse workers or organizations underuse AI, returns to AI-complementary skills will vary by risk environment and institutional context. This can exacerbate heterogeneity in productivity and wages across firms/sectors.
Market structure and competition
- Firms or groups willing to tolerate higher risk or that can credibly reduce risk salience (via governance, contracts, or technical safeguards) may capture disproportionate gains, producing winner-take-most dynamics.
- Narrowing of visible applications through withdrawal may slow discovery of high-value uses and reduce market size for complementary tools and services.
Policy and organizational design
- Interventions should go beyond training in prompts: reduce risk salience or mitigate its effects by providing privacy guarantees, accountability frameworks, bounded-open interfaces (mechanisms that allow safe contextual disclosure and iterative refinement), and psychological safety in organizations.
- Regulatory/standards work should account for interaction effects—e.g., certification or evaluation protocols should measure performance under realistic, rich-interaction conditions rather than single-shot prompts.
Empirical research agenda for AI economics
- Experiments and field studies: randomize risk salience (e.g., framing, privacy policies), required disclosure levels, or iteration incentives, and measure objective and perceived performance, trust, and subsequent use.
- Measurement development: produce scalable metrics of interaction richness (context tokens, number of refinement rounds, content sensitivity) to include in productivity studies.
- Macro-level analysis: study media/institutional shocks (scandals, regulation) as natural experiments to identify effects on adoption, visible application diversity, and aggregate productivity.
- Distributional studies: examine how risk salience varies across socio-economic groups, industries, and countries and how that shapes differential adoption and inequality in AI benefits.

Short takeaway: economists and policymakers should treat generative-AI performance as endogenous to user interaction choices shaped by perceived risk. Accounting for the guarded engagement loop is necessary to avoid underestimating the technology’s potential, to design effective interventions, and to predict distributional impacts of AI adoption.

Assessment

Paper Typetheoretical Evidence Strengthn/a — This is a conceptual/theoretical paper that synthesizes existing literatures and proposes a framework and testable propositions but presents no new empirical data or causal identification; therefore empirical evidence strength is not applicable. Methods Rigorn/a — The paper offers a structured theoretical integration across multiple literatures (trust in automation, privacy calculus, algorithm aversion, social amplification of risk) and articulates clear propositions, but it does not apply empirical methods, precluding standard methodological rigor assessment. SampleNo empirical sample; the paper develops a conceptual framework (the 'guarded engagement loop') by synthesizing prior theoretical and empirical findings from research on trust in automation, privacy calculus, algorithm aversion, and risk communication. Themeshuman_ai_collab adoption governance org_design skills_training GeneralizabilityNo empirical validation — applicability depends on future behavioral and field evidence., Mechanisms may vary across tasks, domains, and LLM capabilities (e.g., coding vs. creative writing)., Cultural, organizational, and regulatory contexts may alter risk salience and engagement strategies., Heterogeneity across user skill levels and incentives (novice vs. expert; individual vs. organizational use) is not empirically characterized., Temporal dynamics (novelty effects, learning over time) are theorized but untested.

Claims (8)

Claim	Direction	Outcome	Confidence & Evidence	Details
Generative AI adoption is often framed primarily as a question of learning technical skills, and this perspective overlooks a defining feature of large language models (LLMs): their output quality depends heavily on how users engage with them. Adoption Rate	negative	adoption_rate	Reading fidelity high Study strength speculative	not reported 0.02
LLM performance varies substantially with depth of disclosure, contextual richness, and iterative refinement — i.e., LLM output quality depends heavily on how users engage with them. Output Quality	positive	output_quality	Reading fidelity high Study strength speculative	not reported 0.02
Risk salience may shape interaction dynamics with LLMs via a multilevel feedback mechanism called the 'guarded engagement loop', in which risk perceptions shape interaction strategies that influence observed performance and, in turn, recalibrate trust in generative AI systems. Task Allocation	mixed	task_allocation	Reading fidelity high Study strength speculative	not reported 0.02
At the micro level, elevated risk salience related to privacy, safety, or ethical concerns may lead users to adopt guarded interaction strategies characterized by reduced contextual disclosure and limited iteration. Automation Exposure	negative	automation_exposure	Reading fidelity high Study strength speculative	not reported 0.02
These constrained (guarded) interactions can lower output quality and increase the likelihood of visible errors, which may further erode trust and reinforce cautious engagement. Output Quality	negative	output_quality	Reading fidelity high Study strength speculative	not reported 0.02
At the macro level, values-driven withdrawal from AI use has the potential to narrow the diversity of visible applications, amplifying risk-focused narratives and reinforcing perceptions of harm in public discourse. Market Structure	negative	market_structure	Reading fidelity high Study strength speculative	not reported 0.02
The guarded engagement loop framework conceptualizes generative AI adoption as a feedback process in which risk perceptions may shape interaction conditions that, in turn, can influence observed performance and subsequent trust calibration. Adoption Rate	mixed	adoption_rate	Reading fidelity high Study strength speculative	not reported 0.02
The paper articulates testable propositions and discusses implications for organizational governance, AI system design, and institutional conditions that enable bounded openness and calibrated reliance. Governance And Regulation	positive	governance_and_regulation	Reading fidelity high Study strength speculative	not reported 0.02