Complex networks map test anxiety and wellbeing levels in students and ChatGPT

This study utilizes behavioural forma mentis networks to demonstrate that while human students semantically frame exams and grades with negative emotions and test anxiety, STEM experts maintain neutral associations and current AI simulations fail to replicate the specific structural overlap between anxiety and exams found in human populations.

The Gist

Imagine your brain as a giant, bustling city. In this city, every concept you know (like "dog," "pizza," or "math") is a building. The roads connecting these buildings are your memories and how you link ideas together.

This paper is like a city planner's report that maps out how students and an AI chatbot (GPT) navigate this city when they think about exams, grades, and anxiety.

Here is the breakdown of their findings, translated into everyday language:

1. The Map-Making Tool: "Mental Neighborhoods"

The researchers didn't just ask students, "Are you anxious?" (which is like asking a tourist, "Is this city scary?"). Instead, they used a special tool called Behavioral Forma Mentis Networks.

Think of this as a GPS that tracks not just where you go, but how you feel about the places you visit.

• They asked students to say the first three words that came to mind when they heard "Exam."
• Then, they asked how positive or negative those words felt.
• They built a map showing which words are neighbors. If "Exam" is connected to "Failure," "Sweat," and "Panic," that's a negative neighborhood. If it's connected to "Learning" and "Challenge," that's a positive neighborhood.

2. The Main Discovery: The "Exam" Neighborhood is Dark

For almost everyone (high schoolers, college students, and even the AI), the concept of "Exam" and "Grade" lives in a dark, stormy part of the city.

• The Human Experience: When students think of an exam, their mental map immediately leads them to "failure," "pressure," and "anxiety." It's not seen as a chance to learn; it's seen as a high-stakes judgment day.
• The AI Experience: The AI (GPT) actually made this neighborhood look even darker than the humans did. The AI associated exams with very few positive words, suggesting it has absorbed the collective human fear of testing from its training data.

3. The Exception: The "Experts" Live in a Sunny District

There was one group that didn't live in the dark: STEM Experts (professors and researchers).

• For them, "Exam" isn't a scary monster; it's just a task.
• Their mental map of "Exam" is concrete and neutral. They see it as a procedure, a checklist, or a job to be done.
• The Analogy: Imagine a firefighter. To a regular person, a fire truck might look scary and loud. To the firefighter, it's just a vehicle with a specific job. The experts have "professionalized" the exam, stripping away the emotional fear.

4. The "Abstract vs. Concrete" Trap

This is the most fascinating part of the study.

• High-Anxiety Students: When these students think about "Grades," their mental map is foggy and abstract. They see "Grades" as vague concepts like "expectations," "judgment," or "the future." Because it's foggy, they feel like they can't control it. It's like trying to steer a boat in thick fog; you feel helpless.
• Experts: They see "Exams" as concrete. They see specific steps, specific questions, and specific outcomes. Because it's clear, they feel in control.
• The Lesson: Anxiety thrives in the fog. When you can't see the specific steps to solve a problem, your brain fills the gap with fear.

5. Humans vs. AI: The "Clinical" vs. The "Real"

The researchers compared real students to the AI's "digital twin" (a version of the AI pretending to be a student).

• How Humans See Anxiety: Humans link anxiety to real-life situations. "I'm anxious because of that math test," "because of my teacher," or "because I didn't sleep." Their anxiety is rooted in their daily life.
• How AI Sees Anxiety: The AI links anxiety to medical textbooks. It talks about "panic disorders," "psychopathology," and "clinical stress."
• The Metaphor:
  • Humans are like someone saying, "My stomach hurts because I ate too much spicy food." (Specific, lived experience).
  • AI is like a doctor saying, "You have gastritis." (Abstract, clinical label).
  • The AI knows what anxiety is called, but it doesn't understand how it feels to live through it.

6. Wellbeing: Diet vs. Therapy

When asked about "Wellbeing":

• Humans talked about lifestyle: sleeping, eating, walking, and hanging out with friends.
• AI talked about therapy: psychotherapy, mental health diagnoses, and clinical interventions.
• Again, the AI sees wellbeing as something you fix with a doctor, while humans see it as something you live every day.

The Big Takeaway

This paper tells us that test anxiety isn't just a feeling; it's a way of seeing the world.

• For anxious students, the academic world is a foggy, abstract, and threatening place.
• For experts, it's a clear, concrete, and manageable job.
• AI is a great mirror for what society says about stress, but a bad mirror for how humans actually feel it. AI can mimic the vocabulary of anxiety, but it misses the messy, real-life context that makes anxiety so powerful.

What should we do?
To help students, we shouldn't just tell them to "calm down." We need to help them clear the fog. We need to make exams feel more like concrete tasks (checklists, clear steps) and less like abstract monsters, so they can move from the "foggy district" of anxiety to the "sunny district" of control.

Technical Summary

1. Problem Statement

Academic evaluation in STEM fields frequently elicits test anxiety, yet standard assessment methods (grades, test scores) fail to capture the semantic and affective framing students use to process these evaluations. Traditional psychometric questionnaires often reduce anxiety to a single numerical score, missing the complex, associative structure of how students mentally organize concepts like "exam," "grade," and "wellbeing."

Furthermore, while Large Language Models (LLMs) are increasingly used to simulate human behavior, it remains unclear whether they can accurately replicate the cognitive and emotional architecture of human test anxiety. Specifically, do LLMs capture the lived, experience-dependent nature of academic stress, or do they merely reproduce surface-level discursive patterns?

2. Methodology

The study employs Cognitive Network Science using Behavioural Forma Mentis Networks (BFMNs). This approach models the "mindset" of a group as a complex network where nodes are concepts and edges represent memory-based free associations, enriched with affective ratings.

Participants and Data Collection

• Total Sample: 994 participants (994 human + simulated counterparts).
• Human Groups:
  • STEM Experts (N=59).
  • Southern Italian High Schoolers (N=159).
  • Physics Undergraduates (N=10).
  • Northern Italian High Schoolers (N=57).
  • Psychology Undergraduates (N=301), stratified into High-Anxiety and Low-Anxiety groups based on the Math Anxiety Scale–IT (MAS-IT).
• Simulated Groups: GPT-oss (20B) agents were prompted with specific socio-demographic profiles (age, gender, education level, socioeconomic status) to mirror the human subgroups.
• Procedure:
  1. Free Association: Participants were presented with cue words (e.g., "exam," "grade," "anxiety," "wellbeing") and asked to provide up to three spontaneous associations.
  2. Valence Rating: Participants rated the emotional valence (1–5, negative to positive) of both the cue words and their generated associations.
  3. Network Construction: Individual associations were aggregated into group-level undirected graphs. Nodes represent concepts; edges represent associations.

Analytical Framework

• Semantic Frames: Extraction of the immediate neighborhood (directly connected nodes) for target concepts.
• Network Metrics: Calculation of node count ($N_v$), edge count ($N_e$), clustering coefficient ($C_i$), and average shortest path length ($l_G$).
• Valence Aura: A measure of the dominant emotional sentiment (positive/negative/neutral) surrounding a concept in the network.
• Emotional Flowers: Visualization of primary emotions (fear, trust, joy, etc.) using Plutchik's wheel, quantified via Z-scores against a randomized baseline.
• Concreteness Analysis: Comparison of the mean concreteness of semantic frames against a random baseline (using Brysbaert et al., 2024 norms) to determine if concepts are represented as concrete actions or abstract labels.
• Overlap Analysis: Use of the Jaccard Index to measure the semantic overlap between "anxiety" and "exam" frames.

3. Key Contributions

1. BFMN Framework for Anxiety: Demonstrates that BFMNs provide a quantitative, interpretable framework to map the structural organization of test anxiety, moving beyond simple scores to reveal how concepts are clustered and emotionally charged.
2. Human vs. AI Cognitive Divergence: Establishes a critical distinction between human and LLM representations of anxiety. While LLMs capture the negative valence of academic evaluation, they fail to replicate the structural embedding of anxiety within concrete academic contexts.
3. Concreteness as a Mechanism: Identifies "concreteness" as a key cognitive variable. High-anxiety human groups frame evaluation concepts more abstractly (loss of control), whereas experts frame them concretely (procedural tasks), suggesting a mechanism for anxiety reduction through professionalization.
4. Digital Twin Limitations: Provides empirical evidence that current LLMs function as "discursive mirrors" (reproducing what people say about anxiety) rather than "cognitive mirrors" (reproducing how people experience anxiety).

4. Key Results

A. The "Exam" and "Grade" Frames

• Universal Negativity: Across all human student groups, "exam" and "grade" are framed negatively, connected to hubs like "failure," "pressure," and "anxiety."
• Expert Neutrality: STEM experts displayed a neutral and concrete network for "exam" (Z=1.87), associating it with tasks and procedures rather than threat. This suggests experience reduces emotional charge.
• LLM Bias: GPT-simulated students mirrored the negative framing but showed an even stronger negative bias and fewer positive associations (e.g., "effort," "satisfaction") compared to humans.

B. Anxiety and Wellbeing Framing

• Structural Overlap (The Critical Finding):
  • Humans: The semantic overlap (Jaccard Index) between "anxiety" and "exam" was 3–4 times higher in human high-anxiety groups (>0.13) than in GPT groups (<0.04).
  • Interpretation: For humans, anxiety is structurally embedded in the academic context. For GPT, anxiety and exams are conceptually adjacent but structurally distinct.
• Concrete vs. Abstract:
  • Humans: Anchor anxiety in concrete academic triggers (school, grades, time constraints).
  • GPT: Anchors anxiety in abstract, clinical terms (panic, psychopathology, stress).
• Wellbeing:
  • Humans: Linked to concrete lifestyle factors (sleep, diet, exercise).
  • GPT: Linked to abstract clinical interventions (therapy, psychotherapy) and negative psychological constructs.

C. Emotional Patterns

• Fear and Anticipation: Significantly elevated in human student groups regarding exams.
• Trust Bias: A paradoxical significant presence of "trust" was found in "grade" frames for human students, suggesting an affective dissonance where students recognize the legitimacy of grading while feeling threatened by it.

5. Significance and Implications

• Educational Intervention: The finding that high-anxiety groups frame evaluation as abstract (vs. concrete) suggests that interventions should focus on making assessment criteria more transparent, procedural, and actionable to restore perceived controllability.
• AI in Psychology: The study highlights a fundamental limitation of using LLMs as "digital twins" for psychological research. While LLMs can simulate the discourse of anxiety, they lack the situated, experience-dependent cognitive architecture that defines human test anxiety.
• Methodological Advancement: Validates BFMNs as a robust tool for distinguishing between surface-level semantic patterns and deep cognitive structures in educational psychology.

Conclusion: Test anxiety is not merely an emotional reaction but a structurally embedded mindset characterized by abstract, uncontrollable representations of evaluation. While LLMs can mimic the negative sentiment of this mindset, they fail to capture the lived, concrete, and contextually integrated nature of human academic anxiety.