💻 computer science

Explainable artificial intelligence in interface design influences user outcomes and cybersecurity literacy

This study demonstrates that integrating explainable artificial intelligence (XAI) into interface design tools significantly enhances user outcomes, including creative confidence, perceived learning, trust, autonomy, and usability, during UI design tasks, suggesting that XAI fosters critical engagement with AI recommendations while highlighting the need for future research to directly measure its impact on cybersecurity and privacy literacy behaviors.

Original authors: Ahmed Al-sa'di, Mahmoud Al-Sarayreh, Esraa Ahmad

Published 2026-09-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ahmed Al-sa'di, Mahmoud Al-Sarayreh, Esraa Ahmad

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the digital world, the screens we touch and the buttons we click are not neutral. They are designed with a specific intent, sometimes to guide us gently and other times to steer us toward choices we might not make if we thought more carefully. These manipulative designs, often called dark patterns, exploit the way our brains work. They use visual tricks, confusing layouts, and automatic defaults to make us share personal data, agree to intrusive permissions, or sign up for recurring payments without fully realizing the cost. The problem is not just that these designs exist, but that they rely on our speed and inattention. When a decision feels easy and requires little effort, we often click without questioning. To protect ourselves, we need more than just knowledge; we need interfaces that encourage us to pause, look closer, and understand what is being asked of us.

This is where the idea of explainable artificial intelligence comes in. Imagine a computer program that helps you design a website or an app. A standard version of such a tool might simply offer a suggestion: "Place this button here." It gives the answer but not the reasoning. An explainable version, however, does more. It offers the same suggestion but also provides a clear explanation of why it was made, what trade-offs were considered, and how confident the system is in its choice. The goal is to turn a passive user into an active partner who can evaluate the advice rather than just accepting it. Researchers have long wondered if this extra layer of transparency actually helps people learn, trust the system more, and feel more in control of their own decisions.

A team of researchers set out to test this idea in a controlled experiment involving 128 participants. They asked these individuals to perform a specific task: designing a dashboard for a health research team. The dashboard needed to help researchers track recruitment numbers, deadlines, and reporting for several ongoing studies. The participants had thirty minutes to create interface components for this scenario using an AI-assisted tool. The researchers split the group into two teams. One team used a standard AI tool that offered design suggestions without any extra context. The other team used an explainable AI tool that presented the same suggestions but included a panel explaining the design principles, the context, and the potential trade-offs for each recommendation.

The results showed a clear difference in how the two groups experienced the task. Those using the explainable tool reported a significant boost in their creative confidence. Before the task, their belief in their ability to create ideas was lower, but afterward, it rose much higher than it did for the group using the standard tool. The explainable group also felt they had learned more from the experience, trusted the AI suggestions more, and felt a stronger sense of autonomy over their final decisions. They rated the system as much more usable and engaging. In fact, the group with the explanations interacted with the interface far more often, opening the explanation panels to read the details, whereas the other group clicked on fewer informational elements.

However, the researchers were careful about what these results actually meant. While the participants felt more confident and learned more during the design task, the study did not test whether this experience made them better at spotting deceptive designs in the real world. The experiment did not include any tests of privacy literacy, nor did it measure if the participants could resist dark patterns when faced with them later. The study explicitly ruled out the idea that they had proven that explainable AI creates a shield against manipulation. Instead, the findings suggest that providing reasons alongside AI suggestions improves the immediate experience of using the tool and helps users feel more capable and informed during the design process.

The study also highlighted that the benefits were not limited to those who were already experts in technology. Even participants with less familiarity with artificial intelligence showed gains in confidence and learning when they had access to the explanations. This suggests that the value of transparency is not just for the tech-savvy but can support a wider range of users. Yet, the authors noted that the experiment was a single session lasting about forty-five minutes. They could not determine if these feelings of confidence and trust would last over time or if they would transfer to different situations, such as making decisions about personal data or security.

Ultimately, the research points to a promising direction for how we interact with intelligent systems. By adding explanations to AI suggestions, designers can create tools that make users feel more in control and more willing to engage critically with the technology. This approach aligns with growing regulatory efforts, such as those in the European Union, which demand that high-risk AI systems be transparent enough for users to understand and use them properly. While this study did not solve the broader problem of cybersecurity or privacy manipulation, it demonstrated that when an AI tool takes the time to explain its thinking, people feel more confident, learn more, and trust the process more. The next step for science is to see if this increased engagement can translate into real-world protection against the deceptive designs that threaten our digital lives.

Technical Summary: Explainable Artificial Intelligence in Interface Design Influences User Outcomes and Cybersecurity Literacy

1. Problem Statement

The paper addresses the intersection of interface design, artificial intelligence (AI), and user security literacy. It posits that "dark patterns"—interface designs that manipulate users into choices conflicting with their interests (e.g., data disclosure, intrusive permissions)—exploit attentional constraints and rapid decision-making. While regulations like the EU Digital Services Act and the EU AI Act mandate transparency, mere transparency does not guarantee user literacy or protective behavior.

The core problem investigated is whether Explainable Artificial Intelligence (XAI), when integrated into interface design tools, can foster competencies relevant to privacy and cybersecurity. Specifically, the authors question if providing rationales and trade-offs alongside AI-generated recommendations encourages users to pause, critically evaluate, and engage more deeply with system outputs, thereby potentially building the "critical engagement" necessary to resist deceptive interfaces. The study notes a gap in evidence regarding whether explanation-supported interaction actually develops these competencies or merely increases trust without improving decision-making calibration.

2. Methodology

The study employed a controlled between-groups experiment with 128 participants recruited via Prolific. Participants were randomly assigned to one of two conditions:

XAI Condition ( $n=64$ ): Participants used an AI-assisted UI design tool where every design suggestion was accompanied by an explanation panel. This panel included the design rationale, contextual considerations, potential trade-offs, and a confidence indicator.
Standard AI Condition ( $n=64$ ): Participants used the same tool and received identical design suggestions but without the dedicated explanation panel.

Task:
Participants completed a 30-minute UI design task focused on improving a dashboard for health-research administration (monitoring recruitment and deadlines). The task was explicitly non-clinical to avoid requiring medical knowledge.

Measures:

Creative Confidence: Measured via a 3-item Likert scale pre- and post-task to calculate "gain."
Post-Task Perceptions: Reported trust, perceived autonomy, and perceived learning (3-item Likert scales each).
Usability: Measured using the System Usability Scale (SUS).
Task Quality: Scored by two blinded raters on a 0–100 scale.
Logged Interactions: Automatically recorded counts of user interactions (opening explanation panels in XAI vs. clicking tooltips/metadata in Standard).
Covariates: Prior AI familiarity and UI design experience were stratified during recruitment.

Analysis:
Data were analyzed using independent-samples t-tests for between-group differences, paired t-tests for within-group changes, and Pearson correlations. Hierarchical regression models were used to evaluate predictors of creative confidence gain. Effect sizes were reported using Cohen's $d$ .

3. Key Results

The study found statistically significant differences favoring the XAI condition across multiple self-reported and behavioral metrics:

Creative Confidence: The XAI group showed a significantly larger gain in creative confidence ( $d=0.68$ ) compared to the Standard group. Both groups improved, but the XAI group improved more.
Perceived Learning: The XAI condition yielded a large effect size for perceived learning ( $d=1.17$ ), indicating participants felt they learned more from the task.
Reported Trust: Trust in the AI tool was significantly higher in the XAI condition ( $d=0.97$ ).
Perceived Autonomy: Participants in the XAI condition reported higher perceived autonomy ( $d=0.42$ ).
Usability: The SUS score was significantly higher for the XAI condition ( $M=71.52$ ) compared to the Standard condition ( $M=58.36$ ), with an effect size of $d=1.78$ . The XAI group's mean score crossed the "above average" threshold (68), while the Standard group did not.
Task Quality: The XAI group produced higher quality designs ( $M=82.80$ ) than the Standard group ( $M=75.54$ ), with a large effect size ( $d=0.97$ ).
Logged Interactions: Interaction counts were significantly higher in the XAI condition ( $d=2.66$ ). However, the authors note these counts represent different actions (opening panels vs. clicking tooltips) and should not be treated as equivalent measures of engagement.
Correlations: Across the full sample, logged interaction counts correlated positively with perceived learning, usability, reported trust, and task quality. However, in the Standard group, interaction counts did not correlate with any outcomes.

4. Key Contributions

The paper contributes empirical evidence that XAI-enhanced interface design positively influences user outcomes during AI-assisted tasks. Specifically, it demonstrates that providing rationales and trade-offs alongside AI suggestions:

Increases creative confidence and perceived learning.
Enhances reported trust and perceived autonomy.
Improves system usability ratings and task quality.
Encourages higher levels of interaction with the interface.

The study supports the hypothesis that explanations can serve as an interface feature that promotes critical engagement with AI recommendations, moving beyond simple transparency to active user involvement.

5. Significance and Claims

The authors maintain a modest and precise scope regarding the significance of their findings:

Relevance to Literacy, Not Proof of Resilience: The paper explicitly states that the results do not demonstrate that XAI improves privacy literacy, cybersecurity literacy, or resistance to dark patterns. The study did not measure dark-pattern recognition, avoidance behavior, or trust calibration (the alignment of trust with system accuracy).
Potential vs. Demonstrated Transfer: The findings suggest a potential relevance to privacy and cybersecurity literacy. The authors argue that by fostering critical engagement and perceived autonomy during a design task, XAI may inform future interventions for privacy decisions. However, direct transfer to security contexts (e.g., resisting phishing or deceptive consent flows) remains an open question requiring separate testing.
Trust Calibration: The study found higher reported trust in the XAI group but clarifies that this does not equate to calibrated trust. Higher trust without accuracy measures could potentially lead to over-reliance if the AI is incorrect.
Limitations: The authors acknowledge limitations including the lack of longitudinal data (persistence of effects), the specific nature of the task (non-clinical, non-security), and the recruitment source (Prolific users), which may limit generalizability to populations with lower digital literacy.

In conclusion, the paper establishes that explanation-supported interfaces improve user experience and self-efficacy in AI-assisted design tasks. It frames these outcomes as a necessary foundation for future research into whether such interfaces can effectively build behavioral resilience against deceptive design patterns, rather than claiming to have solved the problem of cybersecurity literacy directly.