Examining Users' Behavioural Intention to Use OpenClaw Through the Cognition--Affect--Conation Framework

📝 Paper Summary

Human-AI Interaction AI Adoption Trust in AI

This study establishes that user adoption of the autonomous agent OpenClaw is primarily driven by positive attitudes derived from perceived intelligence and personalization, while being significantly inhibited by distrust stemming from privacy and opacity concerns.

Core Problem

While conversational AI adoption is well-studied, there is limited empirical evidence regarding how users evaluate and accept *autonomous* AI agents that execute real-world actions.

Why it matters:

Autonomous agents (like OpenClaw) introduce new risks (execution errors, loss of control) compared to passive chatbots, potentially altering adoption dynamics.
Understanding the dual role of enabling factors (e.g., intelligence) and inhibiting factors (e.g., privacy risk) is crucial for designing trustworthy agentic systems.

Concrete Example: A user might want OpenClaw to automate a workflow but hesitates because they perceive the decision-making process as opaque (algorithmic opacity) or fear data misuse (privacy concern), leading to distrust despite the tool's efficiency.

Key Novelty

CAC Framework for Agentic AI

Applies the Cognition–Affect–Conation (CAC) framework to autonomous agents, mapping user psychology into three sequential stages: beliefs (Cognition), emotional responses (Affect), and behavioral intent (Conation).
Separates antecedents into two distinct paths: an 'enabling' path where positive perceptions build Attitude, and an 'inhibiting' path where negative perceptions generate Distrust.

Architecture

The Conceptual Model (CAC Framework) applied to OpenClaw adoption.

Evaluation Highlights

Attitude is the strongest predictor of behavioural intention (β = 0.49), indicating that positive affect is the primary driver of adoption.
Relative Advantage (β = 0.32) is the most influential cognitive factor for building a positive attitude, surpassing perceived intelligence and personalization.
Perceived Risk (β = 0.28) is the strongest contributor to Distrust, which subsequently reduces usage intention (β = -0.22).

Breakthrough Assessment

4/10

Solid empirical validation of user psychology regarding agents. It confirms expected relationships (risk -> distrust) rather than proposing a new technical method, but provides necessary ground truth for HCI.

⚙️ Technical Details

Problem Definition

Setting: Structural Equation Modelling (SEM) of user survey data to predict Behavioural Intention (BI)

Inputs: User responses to 5-point Likert scale items regarding OpenClaw usage

Outputs: Path coefficients (β) indicating the strength of relationships between cognitive, affective, and conative constructs

Pipeline Flow

Cognitive Evaluation (Beliefs)
Affective Reaction (Emotions)
Conative Intention (Behaviors)

System Modules

Cognitive Factors (Enablers) (Input Processing (Psychological))

Positive beliefs that drive favorable evaluation

Model or implementation: Latent Constructs

Cognitive Factors (Inhibitors) (Input Processing (Psychological))

Negative beliefs that drive skepticism

Model or implementation: Latent Constructs

Affective Response

Emotional evaluation of the system

Model or implementation: Latent Constructs

Conation

Final decision to act/adopt

Model or implementation: Latent Construct

Novel Architectural Elements

Dual-path CAC model explicitly separating 'Enabling' (Cognition -> Attitude) and 'Inhibiting' (Cognition -> Distrust) pathways for autonomous agents

Comparison to Prior Work

vs. TAM: Incorporates 'Distrust' and 'Algorithmic Opacity' as specific inhibitors relevant to autonomous agents, rather than just utility [not cited in paper]
vs. Standard CAC studies: Applies the framework specifically to *agentic* AI (OpenClaw) involving action execution, highlighting 'Perceived Risk' as a major inhibitor unlike in passive conversational AI studies

Limitations

Cross-sectional design captures a snapshot of perceptions, failing to track how trust evolves with long-term use.
Sample restricted to users with prior OpenClaw experience, potentially introducing selection bias compared to the general population.
Self-reported data may be subject to response biases compared to behavioral logs.

Reproducibility

Survey items are fully provided in Table 2. Dataset (N=436) is not publicly available. OpenClaw software is mentioned as open-source but no specific version or repo link for the tool itself is provided in the text (only citations to Shan et al., 2026).

📊 Experiments & Results

Evaluation Setup

Online survey of 436 OpenClaw users recruited via Credamo and online communities.

Benchmarks:

Structural Equation Model Fit (Statistical Goodness-of-Fit) [New]

Metrics:

Path Coefficients (β)
R-squared (variance explained)
CFI (Comparative Fit Index)
RMSEA (Root Mean Square Error of Approximation)
Statistical methodology: Structural Equation Modelling (SEM) using Confirmatory Factor Analysis (CFA) and Path Analysis. Significance tested at p < .05, .01, .001.

Key Results

Benchmark	Metric	Baseline	This Paper	Δ
The structural model results demonstrate strong relationships between cognitive evaluations and affective responses, and ultimately behavioural intention.
SEM	Standardized Path Coefficient (β)	0	0.49	0.49
SEM	Standardized Path Coefficient (β)	0	-0.22	-0.22
SEM	Standardized Path Coefficient (β)	0	0.32	0.32
SEM	Standardized Path Coefficient (β)	0	0.28	0.28
SEM	Standardized Path Coefficient (β)	0	0.24	0.24

Main Takeaways

The 'enabling' path (positive cognition -> attitude -> intention) is stronger than the 'inhibiting' path (negative cognition -> distrust -> intention), suggesting users prioritize benefits over risks.
Perceived Personalisation (β=0.21) significantly boosts attitude, confirming that users value agents that adapt to their specific workflows.
Algorithmic Opacity (β=0.19) is a significant but weaker predictor of distrust compared to Perceived Risk (β=0.28), implying users care more about *outcomes* (risk) than *process* (opacity).
Distrust does not merely reduce attitude; it functions as a distinct negative affective state that directly reduces behavioral intention.

📚 Prerequisite Knowledge

Prerequisites

Structural Equation Modelling (SEM)
Technology Acceptance Model (TAM) concepts
Basics of autonomous AI agents

Key Terms

OpenClaw: An open-source autonomous AI agent framework capable of executing real tasks (file manipulation, API calls) rather than just generating text.

CAC Framework: Cognition–Affect–Conation Framework—a psychological model proposing that behavior evolves from cognitive beliefs to emotional responses (affect) to behavioral intentions (conation).

SEM: Structural Equation Modelling—a statistical technique used to test relationships between observed and latent variables.

Algorithmic Opacity: The perceived lack of transparency in how an AI system processes information and makes decisions.

Perceived Personalisation: The extent to which a user believes the system adapts its content and services to their individual needs.

CFI: Comparative Fit Index—a goodness-of-fit statistic for SEM; values > 0.90 indicate good fit.

RMSEA: Root Mean Square Error of Approximation—a measure of model fit; values < 0.08 indicate good fit.

AVE: Average Variance Extracted—a measure of convergent validity ensuring a construct explains more variance than the error term.