🤖 AI

Artificial Intelligence Competence of K-12 Students Shapes Their AI Risk Perception: A Co-occurrence Network Analysis

This study of Finnish K-12 students reveals that self-perceived AI competence significantly shapes risk perception, with lower-competence students focusing on personal learning drawbacks while higher-competence students prioritize systemic and institutional concerns, underscoring the need for integrated AI literacy in education.

Original authors: Ville Heilala, Pieta Sikström, Mika Setälä, Tommi Kärkkäinen

Published 2026-06-18

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ville Heilala, Pieta Sikström, Mika Setälä, Tommi Kärkkäinen

Original paper licensed under CC BY 4.0 (http://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

Imagine a classroom where students are looking at a new, powerful tool: Artificial Intelligence (AI). The researchers wanted to know: What are the students afraid of? and Does how "good" they feel at using AI change what they are afraid of?

To find out, they asked 163 high school students in Finland to check off a list of worries they had about using AI for schoolwork. They then used a special map-making technique (called a "co-occurrence network") to see how these worries connected to each other. Think of this map like a spiderweb: if two worries often appear together in a student's mind, they are tied together with a string. The thickness of the string shows how strong that connection is.

Here is what they found, broken down simply:

1. The Two Different Maps

The study discovered that students didn't all see the same "danger zone." Their map of risks changed depending on their self-reported skill level with AI.

The "New Driver" Group (Lower Competence)
Students who felt less confident or skilled with AI were like new drivers who are terrified of the car itself. Their worries were very personal and immediate:

The Fear: "If I use this, will I stop thinking for myself?" "Will I become lazy?" "Will I lose my creativity?"
The Analogy: Imagine a child learning to ride a bike. They are mostly worried about falling off, getting a scraped knee, or the bike breaking. They are focused on personal safety and immediate failure.
The Result: These students had a longer list of worries. They were more likely to connect AI to things like "addiction," "misuse," and "losing my ability to think critically."

The "Expert Driver" Group (Higher Competence)
Students who felt they knew how to use AI well were like experienced drivers. They weren't worried about falling off the bike; they were worried about the traffic laws and the road conditions.

The Fear: "Is the AI lying to me?" "Is the data biased?" "Is it cheating if I use this?" "Are the school rules fair?"
The Analogy: An experienced driver doesn't worry about the bike falling over; they worry about the traffic light being broken, the road being slippery, or someone else breaking the rules. They are focused on systemic issues and fairness.
The Result: These students focused less on personal failure and more on big-picture problems like bias, inaccuracy, and whether the school's rules about AI are consistent.

2. The Most Common Worries (The Big Hubs)

When the researchers looked at the whole group, the biggest "hubs" in the spiderweb (the worries everyone talked about most) were:

Inaccuracy and Bias: "Is the AI telling the truth?"
Cheating: "Is using this cheating?"
Losing Skills: "Will I stop learning or stop being creative?"

Interestingly, worries about copyright laws or strict school policies were the least common. Students cared more about how AI affected them and the truth than they did about legal technicalities.

3. What the "Strings" Tell Us

The network analysis showed how the students' minds connected these ideas:

For the confident students: The strings connected "Cheating" to "Bias" and "School Rules." They saw AI risks as a complex system where fairness and rules matter.
For the less confident students: The strings connected "Creativity" to "Resources" and "Unfair Advantage." They saw AI risks as a threat to their own personal growth and fairness in the classroom.

The Main Takeaway

The paper concludes that how good you feel you are at using AI changes what you think is dangerous.

If you feel unskilled, you worry about you (your brain, your creativity, your habits).
If you feel skilled, you worry about the system (the rules, the truth, the fairness).

The researchers suggest that because students' fears change based on their skills, schools need to teach AI literacy. This isn't just about teaching students how to use the tool, but helping them understand the risks so they don't just fear it blindly, but can navigate it responsibly. As the paper notes, without this education, the gap between those who understand the risks and those who don't might get wider, leading to unequal opportunities.

Technical Summary: Artificial Intelligence Competence of K-12 Students Shapes Their AI Risk Perception

Problem Statement
As Artificial Intelligence (AI), particularly Generative AI (GenAI), becomes integrated into education, understanding how students perceive associated risks is critical for responsible adoption. While AI is recognized as both an opportunity and a risk, the specific relationship between a student's self-perceived AI competence and their perception of AI-related risks remains understudied. Existing literature identifies various risk categories (privacy, bias, accuracy, ethics, etc.), but it is unclear how individual characteristics, specifically competence levels, shape the framing of these risks. This study addresses two primary questions: (1) Which factors do upper secondary K-12 students perceive as AI-related risks? and (2) How does self-reported AI competence shape these risk perceptions?

Methodology
The study employed a quantitative, cross-sectional design utilizing a co-occurrence network analysis approach grounded in complexity science and network psychometrics.

Participants: Data was collected from 163 students (47% female, 51% male, 2% non-binary) in a Finnish general upper secondary school (ISCED level 3) engaged in a two-year GenAI development project. The median birth year was 2008.
Instruments:
- AI Competence: A four-item scale rated on a 5-point Likert scale assessing general AI use, personalized learning support, and information management. Confirmatory Factor Analysis (CFA) confirmed a unidimensional structure with excellent model fit (CFI = 1.00, RMSEA = .041) and reliability ( $\alpha$ = 0.89).
- AI Concerns/Risks: A 14-item binary checklist capturing perceptions of risks across three domains: Systemic (e.g., bias, inaccuracy, resource utilization), Institutional (e.g., cheating, unfair advantage, school policies, copyright), and Personal (e.g., reduced critical thinking, creativity, learning, addiction, privacy).
Analytical Approach:
- Students were stratified into "low" and "high" competence groups using a median split (Mdn = 17).
- A Co-occurrence Network Analysis was performed to model the relational structure of concerns. Binary responses were transformed into a co-occurrence matrix.
- Data underwent $L_2$ -normalization to control for the total number of concerns selected per student.
- Regression-adjusted edge weights were calculated using Ordinary Least Squares (OLS) regression to isolate the effect of competence while controlling for gender and year group.
- Eigenvector Centrality (EC) was computed to identify the relative structural importance of specific concerns within the networks of low- vs. high-competence groups. Edges with absolute values below the 75th percentile were pruned to highlight salient associations.

Key Results
The analysis revealed distinct patterns in how competence levels influence risk perception:

Overall Prevalence: The most frequently cited concerns across all students were AI inaccuracy/bias, AI use for cheating, and risks to learning quality (reduced critical thinking and creativity). Institutional and legal concerns (copyright, inconsistent rules) were least cited.
Competence and Concern Volume: Students in the low-competence group reported a greater number of concerns across nearly all categories compared to high-competence peers. Low-competence students showed a stronger emphasis on personal learning risks and fear-based issues (misuse, addiction).
Network Structure Differences:
- High-Competence Group: The network was characterized by strong positive associations between systemic and institutional risks. Key central nodes included AI_cheating (EC = 0.53) and Biased (EC = 0.52), followed by Teachers_rules and Inaccurate. This indicates that students with higher competence view risks as interconnected systemic issues (e.g., linking cheating to bias and school rules).
- Low-Competence Group: The network centered on personal and learning-related threats. The most central nodes were Less_creative (EC = -0.53) and Less_critical (EC = -0.49). This group strongly connected creativity and critical thinking concerns with fairness and resource issues.
Regression Findings: Positive edge weights (increasing with competence) linked systemic/institutional items (e.g., AI_cheating—Biased). Negative edge weights (decreasing with competence) linked personal/learning items (e.g., Less_creative—Less_critical), suggesting lower-competence students are more likely to conflate cognitive decline with fairness and resource concerns.

Key Contributions

Empirical Evidence of Competence-Risk Link: The study provides empirical evidence that self-reported AI competence is a significant predictor of how students frame AI risks. It moves beyond simple prevalence counts to show how competence alters the structural relationships between different risk perceptions.
Differentiation of Risk Frames: It distinguishes between a "systemic/institutional" risk frame (dominant in high-competence students) and a "personal/cognitive" risk frame (dominant in low-competence students).
Methodological Application: The paper demonstrates the utility of co-occurrence network analysis and Eigenvector Centrality in educational research to visualize the complex web of student beliefs regarding technology.

Significance and Claims
The authors claim that these findings highlight a critical need for educational institutions to incorporate AI literacy into curricula. The study posits that without systematic and equitable teaching of AI literacy, disparities in knowledge may widen, leading to unequal opportunities (the "AI divide").

For Educators: Teachers must move beyond binary "allow/prohibit" policies. For low-competence students, scaffolding is needed to address fears of misuse and cognitive atrophy. For high-competence students, clear guidelines on acceptable use and training in evaluating AI outputs are necessary.
For Policy: The results align with the EU Artificial Intelligence Act's emphasis on AI literacy. The authors argue that failing to institutionalize competence building may leave education systems unprepared for technological shifts.
Modest Scope: The authors explicitly acknowledge limitations, including the single-school context (limiting generalizability), reliance on self-reported competence, and the cross-sectional design which precludes causal claims. They do not propose specific new pedagogical models but suggest that existing research on self-efficacy and risk awareness should inform future interventions.

In conclusion, the paper asserts that AI competence is not merely a skill set but a lens through which students evaluate the risks and opportunities of AI in education. Promoting AI literacy is therefore essential to enable learners to engage critically, responsibly, and productively with AI tools.

1. The Two Different Maps

2. The Most Common Worries (The Big Hubs)

3. What the "Strings" Tell Us

The Main Takeaway

Technical Summary: Artificial Intelligence Competence of K-12 Students Shapes Their AI Risk Perception

More like this