Modeling AI-Assisted Plagiarism in Academic Social Environments Using Qualitative Plausibility Assessment Supports of the Simulation by Large Language Models

Abstract

This study investigates how AI-assisted plagiarism changes dishonest academic behavior in a socially interactive learning environment under different educational conditions. To this end, this study develops a scenario-based simulation to examine how AI-assisted plagiarism influences dishonest academic behavior in socially interactive learning environments. The model represents students as autonomous agents embedded in local peer networks who adapt their weekly behavior under academic pressure, institutional intervention, and available cheating options. Two behavioral scenarios are considered: a conventional plagiarism environment, in which agents choose between honest submission and direct copying, and an AI-augmented environment, in which AI-assisted plagiarism is introduced as an additional dishonest strategy. Intervention is modeled through environmental and institutional conditions, specifically detection probability and sanction severity, rather than through direct internal reward manipulation. Q-learning is used as a simplified adaptive mechanism for repeated agent choice. Experimental results show that the possibility of producing and assessing a simulation to see the availability of AI-assisted plagiarism substantially changes the behavioral composition of misconduct by increasing total dishonest behavior and shifting a large share of it toward the AI-assisted category. In the simulation, active intervention reduces dishonest behavior overall but does not eliminate AI-assisted plagiarism as the dominant dishonest strategy in the AI-augmented environment. These observations in the simulation suggest that academic misconduct in the AI era should be understood not only as a problem of deterrence but also as a problem of behavioral adaptation under changing technological and institutional conditions. To support the realism assessment of the simulation design, the study also conducts a structured qualitative plausibility review using multiple large language models under a shared prompt. Across these reviews, the model is judged to be acceptable as a first-stage stylized baseline, while important limitations are identified in agent heterogeneity, social influence depth, and the use of Q-learning as a simplified adaptive heuristic to reproduce the behaviors of actors in there.

1. Introduction

Academic plagiarism remains a persistent challenge in higher education, affecting fairness, learning quality, and institutional trust. Prior studies have shown that plagiarism is associated with academic pressure, weak citation literacy, insufficient policy awareness, and perceived low risk of punishment [1,2,3]. Academic integrity is therefore not only a matter of rule enforcement but also a core condition for maintaining credible educational assessment and supporting quality education more broadly [4,5,6]. A preliminary implementation of the basic design of the simulation (without AI-augmented plagiarism or LLM plausibility assessment) was reported by the authors [7], and related analysis of cloud-based educational platforms using multi-agent learning was reported in the earlier work as well [8].

The rise of generative artificial intelligence has changed this problem in important ways. Students now have access to AI-assisted methods for producing academic work, creating a new dishonest pathway that differs from conventional direct copying in accessibility, detectability, and strategic use [9,10]. This shift is particularly important because institutional responses to plagiarism have traditionally been designed around direct copying, textual similarity, and source overlap, even in these cases we have observed various types of academic misconduct and the existence of engagements to them from the students, depending on the types of misconduct [11,12]. In AI-augmented learning environments, however, misconduct may no longer be limited to copying from peers or documents but may instead include AI-supported production that is more individualized and potentially less transparent. While a similar effect might already be seen in the context of “contract cheating” [13], which outsources the assignments outside by a contract, the costs for doing will be significantly lower in the case of the use of AI. While the use of generative AI is not dominant in the performance of the work (e.g., in the business field, [14,15,16]), shutting out access to AI-based assistance in an academic educational environment is not meaningful since it also links to the necessity for constructing ethical governance in there.

This development raises a new question for computational research on plagiarism. The central issue is no longer only whether anti-plagiarism interventions reduce misconduct but also whether they change the form that misconduct takes. In particular, interventions aimed at reducing direct copying may suppress one dishonest strategy while leaving open the possibility that students adapt by shifting toward AI-assisted plagiarism. Thus, plagiarism in the AI era should be understood not only as a problem of deterrence but also as a problem of behavioral adaptation under changing technological and institutional conditions.

Existing research provides useful but incomplete foundations for studying this issue. Detection-centered studies have made strong progress in identifying copied or transformed content [17,18]. Academic integrity studies emphasize institutional culture, policy clarity, and student awareness [3,4,5]. Agent-based modeling offers tools for analyzing how local interaction rules produce population-level social diffusion [19,20,21,22], while reinforcement learning provides a framework for adaptive sequential decision-making under reward and penalty [23,24,25]. However, these strands are often treated separately, and relatively little work explicitly models AI-assisted plagiarism as a distinct behavioral option within a socially interacting adaptive population [26,27,28].

A further limitation in some earlier computational approaches is the tendency to represent intervention as direct modification of internal reward preferences. While such techniques may be useful in some reinforcement learning contexts, they are less appropriate for a social–behavioral interpretation of academic misconduct. In actual educational settings, institutions do not directly alter student conscience. Rather, they alter the surrounding environment through policy, monitoring, sanctions, and access conditions. This study therefore adopts a revised modeling perspective that focuses on environmental and institutional change instead of morality-like internal manipulation.

Based on this perspective, this paper develops a scenario-based simulation of plagiarism behavior in academic social environments, using Q-learning as a simplified heuristic adaptation mechanism for reproducing repeated student decision-making. The simulation distinguishes between honest submission, direct copying, and AI-assisted plagiarism as separate behavioral choices, while representing intervention through detection probability and sanction severity. This design allows the study to examine whether AI-assisted plagiarism changes the distribution of dishonest behavior, whether intervention reduces misconduct overall or redirects it toward less detectable forms, and how these effects interact with peer influence.

Here, another important issue in this topic is the difficulties of precisely observing and investigating the AI-assisted plagiarism behaviors in an academic environment. This issue will also make it difficult to verify how the simulation-based reproduction matches what the actual situation would be in an actual environment. To overcome this issue, we introduce a Cross-LLM qualitative plausibility assessment to assess the simulation setting and analysis of the results on there.

Accordingly, the main purpose of this study is to examine under what realistic educational conditions dishonest academic behavior, including AI-assisted plagiarism, is suppressed or sustained in a multi-agent learning environment. More specifically, the study asks whether the availability of AI-assisted plagiarism changes the proportion of honest submission, direct copying, and AI-assisted cheating; whether classroom or institutional intervention reduces dishonest behavior overall; whether such intervention reduces AI-assisted plagiarism specifically rather than only direct copying; and whether differences in social interaction conditions influence the spread and persistence of dishonest behavior.

The contribution of this study is threefold. First, it introduces AI-assisted plagiarism as a distinct behavioral option in a plagiarism diffusion simulation. Second, it represents intervention through environmental and institutional conditions rather than direct internal reward manipulation. Third, it provides a first-stage baseline simulation for examining suppression, persistence, and substitution of dishonest behavior in AI-augmented academic environments.

2. Related Work

2.1. Academic Plagiarism and Integrity in Higher Education

According to [11], academic misconduct is defined as “the exchange of test and homework answers among students”. Also in [11], a survey is conducted to assess the engagements to various forms of misconduct, and approximately 85 percent of the respondents of the survey have engaged in some form of them, such as copying answers to a test or copying someone’s homework. Also in [11], the engagement degrees were different depending on the types of misconduct (e.g., a specific form reported approximately twice than another form) as well as other factors such as demographic variables. On the other hand, to the best of the authors’ knowledge, the actual records of punishments for the associated academic misconduct were not openly available, due to privacy protection or other purposes. In a case of a faculty in a university, only a few cases seemed to happen over 30 years, and only one or two may be connected to the use of AI. Because of these significant gaps among the survey-based observations and the actual potential records to be used, it is quite hard to know the probabilities of having academic misconduct on each test or homework. Due to these restrictions and difficulties of knowing the necessary parameters in simulating these activities, we have started our work using qualitative assessment supports from large language models (LLMs).

On academic misconduct, typically it also includes the activity to give answers to some other students, rather than obtaining the answers. In [11], it clearly states that such ‘giving’ activities are also included in academic misconduct. However, in case of our context in academic plagiarism, it also includes the activities copying from external sources such as the Internet or the content generated by a generative AI. In this sense, the research in the context of academic misconduct covers some more about academic plagiarism in this paper1. Academic plagiarism has long been recognized as a persistent issue in higher education, with implications for fairness, assessment credibility, and learning quality. Prior studies have linked plagiarism behavior to academic pressure, low citation literacy, weak policy understanding, and perceived low risk of punishment [1,2,3]. Research on academic integrity also emphasizes that institutional culture, explicit principles, and consistent enforcement shape student compliance behavior [4,5]. In this sense, plagiarism should be understood not only as an individual decision problem but also as a socially and institutionally situated behavior.

2.2. Plagiarism Detection and Institutional Response

A large body of research has focused on plagiarism detection, including lexical, semantic, cross-source, and obfuscation-aware approaches. Systematic reviews show substantial progress in the automated detection of copied and transformed content [17,18]. However, most detection-focused work is retrospective: it aims to identify plagiarism after it has occurred rather than to model how students adapt their future behavior under changing detection conditions. As a result, detection research alone is insufficient for analyzing whether institutional policies suppress misconduct or alter the strategic form it takes. Notice that these plagiarism detection approaches were often tackled by obfuscating activities, e.g., by slightly modifying the words or phrases in the original texts by the students to avoid detection from specific words or phrases. Furthermore, sometimes a plagiarism detector wrongly detected the fair use of some materials which were even clearly noted to be allowed by the teachers. In this paper, we assume that the accuracy of detecting plagiarism will not be 100 percent. Also we assume the threshold of dividing plagiarism and the fair use of materials is clearly given in some ways and shared with the teachers and students, while this matter is not quite as easily done in an actual case. In some cases, we can provide a clear boundary between acceptable AI assistance and AI-assisted plagiarism within the descriptions of an assignment. Consider the case when an assignment to the students clearly states where the AI should be used with the actual prompts to use for an AI, such as “Ask the AI with the following prompt, [Please provide the three major difficulties on the use of Linux OS.], and copy the output from the AI to the head of the report, and then give the extra explanations for the three by yourself without using an AI, the explanations should include what actually happened on your computers”. In this case, when a student follows the process of giving the extra explanations with the help of AI, it is clearly judged as AI-assisted plagiarism. However, even in this case, the clear declaration does not give us a clear way to detect such AI-assisted plagiarism. The use of these assumptions is an important limitation in this study.

2.3. AI-Assisted Plagiarism as an Emerging Educational Challenge

The emergence of generative artificial intelligence has introduced new uncertainty into education and research practice [9,29]. In higher education, this development has intensified concern over academic integrity, assessment validity, and the changing boundary between acceptable support and dishonest assistance [10,27,30,31]. Recent literature also shows that the intersection of artificial intelligence and academic integrity is rapidly becoming a distinct research area in higher education [26,27]. Student-side perspectives further suggest that generative AI is already being interpreted in diverse ways within higher education, ranging from acceptable assistance to a source of integrity risk and uncertainty [32]. Earlier related work by the authors examined educational cloud platforms through multi-agent learning [8] and also the basic design of the simulation implementation [7,27].

AI-assisted plagiarism differs from conventional direct copying because it can be generated more individually, may leave weaker similarity traces, and may be perceived by students as a less direct form of misconduct. This makes AI-assisted plagiarism especially relevant in current educational settings, where dishonest behavior may shift from peer-dependent copying toward AI-supported production. From the perspective of educational quality and integrity, this issue is aligned with the broader goals of SDG 4, which emphasize inclusive and equitable quality education and the credibility of learning systems [6]. Accordingly, plagiarism models that focus only on conventional copying risk overlooking a major transformation in the academic misconduct environment.

2.4. Agent-Based Modeling for Social Diffusion

Agent-based modeling (ABM) is widely used to study emergent collective behavior arising from local interaction rules [19,20,21]. In socio-technical and educational contexts, ABM is particularly useful for representing heterogeneous actors, bounded rationality, and peer-based interaction [22,33,34]. These properties make ABM well suited for studying how dishonest behavior diffuses through local social networks, how norms may persist, and how institutional conditions interact with peer exposure.

2.5. Reinforcement Learning for Adaptive Decision-Making

Reinforcement learning (RL) provides a framework for sequential behavior under uncertainty, where agents learn from repeated consequences [23,24]. In multi-agent and socially interactive settings, RL has been applied to coordination, adaptation, and emergent behavior [25,35,36,37]. For academic misconduct modeling, RL is relevant because it allows agents to adapt over time to the institutional response, social environment, and available strategic options. This makes RL especially useful for studying whether students persist in, reduce, or redirect dishonest behavior when the cheating environment changes. Q-learning provides a standard and interpretable baseline for adaptive decision-making in repeated environments [23,24]. While QS-learning is known to effectively reproduce human behaviors [38,39], the use of QS-learning requires a precise modeling of how a similar situation can be seen as similar in one’s mind. This will give a serious complication in the modeling of simulation and this modeling is largely problem specific. In the present study, to remove this complexity of the modeling in the analysis, Q-learning is used as the simplified adaption mechanism to reproduce humans’ learning behaviors in their decision making.

2.6. Research Gap and Positioning of This Study

The existing literature provides strong foundations in plagiarism detection, academic integrity, ABM, and RL, but these areas are often studied in parallel rather than in an integrated model of adaptive misconduct. Detection studies mainly address identification performance; integrity studies emphasize policy and culture; ABM captures social diffusion; and RL captures adaptive behavior. According to the bibliometric analysis given in [26], relatively little work explicitly models AI-assisted plagiarism as a distinct strategic option in a socially interacting adaptive population. Also another systematic review analyzed the impact of AI in online education to the learner engagement [28], as ethical concerns including the detection and prevention of academic misconduct is important future work.

This study addresses that gap by integrating ABM-based social interaction and reinforcement learning adaptation in a framework that distinguishes between conventional and AI-augmented plagiarism environments. Rather than treating intervention as a direct internal moral modification, the study models intervention as a change in environmental and institutional conditions, including action availability, detection probability, and sanction severity. In doing so, it aims to analyze not only whether intervention reduces dishonest behavior but but also whether it changes the form that dishonesty takes.

A further practical limitation in this area is that the realism of classroom-oriented simulation assumptions is difficult to assess directly, especially in sensitive domains such as plagiarism where student self-reports and direct observational validation may be limited. Recent advances in large language models create a new opportunity for structured qualitative plausibility checking during simulation design, allowing researchers to obtain systematic feedback on whether state definitions, action options, intervention conditions, and social assumptions are reasonable as first-stage educational abstractions. In the present study, this capability is treated as a methodological support tool for model refinement rather than as a substitute for empirical validation.

3. Research Methodology and Experiment Design

This study develops a scenario-based simulation of plagiarism behavior in an academic social environment. Q-learning is used as a baseline adaptive mechanism for repeated agent choice rather than as the main research contribution. The simulation is designed to reflect environmental and institutional conditions that can realistically be changed, rather than directly manipulating morality-like internal reward preferences. In particular, the model focuses on four components: scenario, state, action, and intervention.

The scenario defines the behavioral environment available to student agents. Two scenarios are considered. The first is a conventional plagiarism environment, in which agents may choose either honest submission or direct copying. The second is an AI-augmented environment, in which agents may additionally choose AI-assisted plagiarism. This distinction allows the study to examine whether the presence of AI-assisted cheating changes the diffusion and adaptation of dishonest behavior.

The state represents the current condition of an agent within the academic environment. Each student occupies one of three states at time t: regular, under_pressure, or sanctioned. The regular state represents a normal academic condition. The under_pressure state represents temporary academic or social pressure that may increase the attractiveness of dishonest shortcuts. The sanctioned state represents an institutional consequence after a student has been caught engaging in misconduct. This state design treats misconduct as a behavioral choice rather than as a fixed moral identity.

The action represents the behavior selected by the student each week. Depending on the scenario, agents choose from the following action set:

• Honest submission, representing legitimate completion of the academic task;
• Direct copying, representing conventional plagiarism from peers or existing sources;
• AI-assisted plagiarism, representing the dishonest use of generative AI or related tools.

The intervention represents the institutional policy environment. In the present model, intervention affects the probability that dishonest behavior is detected and the severity of the sanction applied after detection. Two intervention conditions are considered: no_intervention and active_intervention. These conditions do not alter agent morality directly; instead, they modify the external policy setting under which agents adapt their behavior.

Figure 1 presents the simulation cycle used in the classroom misconduct model. In each week, a student agent occupies a current academic state under a given scenario, observes the set of available actions, selects one action using a baseline adaptive policy, undergoes detection depending on the intervention condition, receives a reward, transitions to the next state, and updates its Q-value. The scenario determines which behavioral options are available to the agent, while the intervention condition determines the probability of detection and the severity of institutional sanction. This cycle is repeated across the semester in order to examine how dishonest behavior is suppressed, sustained, or redirected under different educational conditions. Simulation Cycle 1 summarizes the weekly simulation procedure, including scenario-dependent action selection, intervention-based detection, state transition under academic pressure and social influence, and repeated independent runs across random seeds.

Simulation Cycle 1 Simulation cycle for AI-assisted plagiarism under scenario and intervention conditions.

1: for each scenario in {conventional, AI-augmented} do   2:     for each intervention in {no intervention, active intervention} do   3:         for each social influence level $p_{inf}$ do   4:            for each random seed r do   5:                Initialize random seed r   6:                Initialize Q-table $Q_i(s,a)$ for each student i   7:                Pretrain $Q_i(s,a)$ for 100 iterations using scenario-available actions and intervention-specific detection probabilities   8:                Initialize all agent states $s_i^1\leftarrow$ regular   9:                Initialize sanction timers and weekly records 10:                Build social groups and peer connections 11:                for week $t=1$ to T do 12:                    Determine task type and task weight $w_t$ 13:                    for each student i do 14:                        Observe current state $s_i^t$ 15:                        Determine the set of actions available under the current scenario 16:                        Select action $a_i^t$ using $ϵ$-greedy policy 17:                        Check whether $a_i^t$ is detected under the current intervention condition 18:                        Compute reward $r_i^t$ based on state, action, task type, and detection outcome 19:                        if $a_i^t$ is dishonest and caught then 20:                           Set next state $s_i^{t+1}\leftarrow$ sanctioned 21:                           Start or refresh sanction timer 22:                        else 23:                           Observe recent dishonest behavior among connected peers 24:                           Update next state using academic pressure and social influence 25:                        end if 26:                        Update $Q_i(s_i^t,a_i^t)$ using temporal-difference learning 27:                    end for 28:                    Decay exploration rate ${ϵ}_t$ 29:                    Record weekly counts of honest submission, direct copying, AI-assisted plagiarism, and detected cases 30:                end for 31:                Store aggregate outcomes for the current {scenario, intervention, $p_{inf}$, seed} run 32:            end for 33:            Aggregate results across random seeds for the current {scenario, intervention, $p_{inf}$} setting 34:         end for 35:     end for 36: end for

3.1. Agent Learning Policy

Each agent maintains a Q-table $Q_i(s,a)$ that stores the expected utility of taking action a in state s. Action selection follows an $ϵ$-greedy rule:

$$a_i^t=\left\{\begin{array}{cc}\mathrm{random}\mathrm{action},\hfill & \mathrm{if}u<{ϵ}_t\hfill \\ \arg \underset{a\in \mathcal{A}}{\max }{Q}_i(s_i^t,a),\hfill & \mathrm{otherwise}\hfill \end{array}\right.$$

(1)

where $u\sim \mathcal{U}(0,1)$, ${ϵ}_t$ is the exploration rate at week t, and $\mathcal{A}$ is the set of actions available in the given scenario.

The exploration rate decays over time according to

$${ϵ}_{t+1}=\max ({ϵ}_{\min },{ϵ}_t\lambda )$$

(2)

where ${ϵ}_{\min }$ is the minimum exploration rate and $\lambda$ is the decay factor.

After receiving reward $r_i^t$ and transitioning to the next state $s_i^{t+1}$, the Q-value is updated using the temporal-difference rule:

$$Q_i(s_i^t,a_i^t)\leftarrow Q_i(s_i^t,a_i^t)+\alpha \left[r_i^t+\gamma \underset{a^{\prime }}{\max }{Q}_i(s_i^{t+1},a^{\prime })-Q_i(s_i^t,a_i^t)\right]$$

(3)

where $\alpha$ is the learning rate and $\gamma$ is the discount factor.

3.2. Detection and Intervention Logic

The model distinguishes direct copying from AI-assisted plagiarism through action-specific detection probabilities. Honest submission is assumed not to trigger detection. Direct copying is modeled as more detectable, while AI-assisted plagiarism is modeled as less detectable or more uncertain to detect.

For a dishonest action $a_i^t$, the probability of detection depends on both the action type and the intervention condition. Let $c\in \{\mathit{no}\_\mathit{intervention},\mathit{active}\_\mathit{intervention}\}$ denote the current intervention condition. Then detection is defined as

$$P({\mathrm{caught}}_i^t=1\mid a_i^t=\mathrm{direct}\mathrm{copying},c)=p_{\mathrm{direct}}^{\left(c\right)}$$

(4)

$$P({\mathrm{caught}}_i^t=1\mid a_i^t=\mathrm{AI}-\mathrm{assisted}\mathrm{plagiarism},c)=p_{\mathrm{AI}}^{\left(c\right)}$$

(5)

Accordingly, the model distinguishes between lower-detection and higher-detection policy environments by assigning different values to $p_{\mathrm{direct}}^{\left(c\right)}$ and $p_{\mathrm{AI}}^{\left(c\right)}$ under no_intervention and active_intervention. In the same way, sanction severity is also conditioned on the intervention setting. This design allows the simulation to examine whether institutional intervention reduces dishonest behavior overall or instead shifts it toward a less detectable cheating strategy. The exact parameter values used under each intervention condition are summarized in

Table 1. Main parameters used in the simulation.

Symbol/Name	Meaning	Value	Notes
N	Number of student agents	30	Total number of agents in each simulation run.
T	Number of weeks	16	Represents one academic semester.
Seeds/runs	Number of independent runs per condition	100	For the main experiment, seeds were initialized from 42 to 141.
$\alpha$	Learning rate	0.1	Q-learning update step size.
$\gamma$	Discount factor	0.9	Weight assigned to future reward.
${ϵ}_0$	Initial exploration rate	0.2	Initial probability of exploratory action choice.
${ϵ}_{\min }$	Minimum exploration rate	0.01	Lower bound for exploration during decay.
$\lambda$	Exploration decay factor	0.98	Controls reduction of exploration over time.
$p_{inf}$	Social influence level	$\{0.0,0.2,0.4\}$	Low, medium, and relatively stronger peer influence settings.
$p_{\mathrm{trigger}}$	Background pressure trigger probability	0.15	Additional probability of entering a pressured state outside high-stakes weeks.
$p_{\mathrm{direct}}^{\left(\mathrm{no}\right)}$	Detection probability for direct copying under no intervention	0.55	Lower-detection institutional setting.
$p_{\mathrm{direct}}^{\left(\mathrm{active}\right)}$	Detection probability for direct copying under active intervention	0.80	Higher-detection institutional setting.
$p_{\mathrm{AI}}^{\left(\mathrm{no}\right)}$	Detection probability for AI-assisted plagiarism under no intervention	0.25	Lower-detection institutional setting.
$p_{\mathrm{AI}}^{\left(\mathrm{active}\right)}$	Detection probability for AI-assisted plagiarism under active intervention	0.45	Higher-detection institutional setting.
Penalty(no)	Sanction penalty under no intervention	10.0	Applied when dishonest behavior is detected.
Penalty(active)	Sanction penalty under active intervention	18.0	Stronger institutional response after detection.
$D_s$	Sanction duration	2 weeks	Duration for which the sanctioned state is maintained.
$G_{\max }$	Maximum peer-group size	6	Random group sizes vary between 2 and 6 students.
Network structure	Peer network generation	Random local cliques	Students are randomly assigned to groups; each group forms an undirected clique.
Week schedule	High-stakes weeks	Quiz: 4 and 12; Midterm: 8; Final: 16	These weeks affect task type and pressure dynamics.
Task weights	Assessment reward multipliers	Homework: total 0.20; Quiz: 0.05 each, total 0.10; Midterm: 0.30; Final: 0.40	Homework weight is distributed evenly across non-exam weeks. The two quiz weeks jointly contribute 0.10.
Pretraining rounds	Initial Q-table warm-up	100 iterations	Used to initialize Q-values before each run.

.

3.3. Reward Structure

The reward function reflects the relative attractiveness and risk of the available behavioral options under different academic conditions. Honest submission yields a positive reward, reflecting legitimate task completion. Direct copying and AI-assisted plagiarism yield higher short-term benefits when not detected because both represent effort-reducing dishonest shortcuts. However, when detected, these actions incur punishment-based penalties.

Let $w_t$ denote the weight of the academic task at week t. Conceptually, the reward design follows three principles:

• Honest submission yields a positive reward that varies by state, with lower value under academic pressure and moderate value under sanction.
• Direct copying yields a higher short-term reward when not caught but incurs a penalty when detected.
• AI-assisted plagiarism also yields a positive short-term reward when not caught but differs from direct copying in its detection probability and strategic role.

This formulation reflects practical academic trade-offs such as effort reduction, submission benefit, and institutional consequence, rather than moral judgment embedded directly into the reward signal. The implemented reward function is specified as follows.

Let c be the intervention condition and $S^{\left(c\right)}$ be the sanction penalty under intervention condition c. The reward is defined as a task-weighted function of the current state, selected action, and detection outcome:

$$r_i^t=w_{t}R(s_i^t,a_i^t,z_i^t,c),$$

(6)

where $z_i^t\in \{0,1\}$ indicates whether the agent is detected. The unweighted reward component is

$$R(s,a,z,c)=\left\{\begin{array}{cc}10.0,\hfill & a=\mathit{honest}-\mathit{submission},s=\mathit{regular},\hfill \\ 7.0,\hfill & a=\mathit{honest}-\mathit{submission},s=\mathit{under}-\mathit{pressure},\hfill \\ 8.0,\hfill & a=\mathit{honest}-\mathit{submission},s=\mathit{sanctioned},\hfill \\ -S^{\left(c\right)},\hfill & a\in \{\mathit{direct}-\mathit{copying},\mathit{ai}-\mathit{assisted}-\mathit{plagiarism}\},z=1,\hfill \\ 14.0,\hfill & a=\mathit{direct}-\mathit{copying},z=0,\hfill \\ 12.0,\hfill & a=\mathit{ai}-\mathit{assisted}-\mathit{plagiarism},z=0.\hfill \end{array}\right.$$

(7)

Thus, honest submission receives a state-dependent positive reward, dishonest actions receive a larger short-term reward when undetected, and detected dishonest actions receive a negative reward proportional to sanction severity. The values are intentionally stylized rather than estimated from individual-level student data, and they are used to create a transparent first-stage behavioral simulation.

3.4. Social Interaction Structure

Students are connected through a simplified peer network constructed from random social groups. Each group forms a local clique, meaning that students within the same group are mutually connected. This produces an undirected and unweighted local interaction network rather than a fully connected classroom. A simplified procedure of peer-network construction algorithm is shown in Procedure Simulation Cycle 1. We use this simplified procedure since we use a small number for N (i.e., 30 as shown in Table 1) in our simulation. When we extend the size of a class, further well-studied network construction algorithms should be applied [34].

Social influence is modeled through recent dishonest exposure. Let ${\mathcal{N}}_i$ denote the peer set of agent i. If a higher share of the agent’s peers engaged in dishonest behavior in the previous step, the agent is more likely to experience social pressure in the current step. This design allows dishonest behavior to diffuse through observed peer behavior rather than through direct internal manipulation.

At the start of each simulation run, students are randomly partitioned into local peer groups with group size between 2 and 6. Each group forms an undirected clique, and the resulting peer network is fixed throughout that run. The peer set of agent i is denoted by ${\mathcal{N}}_i$. The network is regenerated for each independent random seed but not regenerated weekly. This design represents stable local classroom peer circles rather than a fully connected classroom or a dynamically changing social network.

Social influence is based on exposure to dishonest peer behavior in the previous week. Let

$$h_i^{t-1}={\displaystyle \frac{1}{|{\mathcal{N}}_i|}}\sum _{j\in {\mathcal{N}}_i}\mathbb{I}(a_j^{t-1}\ne \mathit{honest}-\mathit{submission})$$

(8)

be the share of agent i’s peers who selected a dishonest action in the previous week. If ${\mathcal{N}}_i$ is empty, $h_i^{t-1}=0$. The social-pressure event is then modeled as

$$g_i^t\sim \mathrm{Bernoulli}\left(p_{\inf }{h}_i^{t-1}\right),$$

(9)

where $p_{\inf }$ is the social influence level. Thus, social exposure affects the probability of entering the under-pressure state, but it does not directly force the agent to choose a dishonest action.

3.5. State Transition Logic

State transition depends on detection, academic pressure, sanction duration, and social influence. If an agent is caught engaging in dishonest behavior, it transitions to the sanctioned state for a fixed sanction duration. Once the sanction period ends, the agent returns to the regular state.

If the agent is not caught, the next state is determined by a combination of academic pressure and social exposure. During high-pressure weeks, such as quiz, midterm, or final assessment periods, or when dishonest peer exposure is sufficiently high, an agent may transition to the under_pressure state. Otherwise, the agent remains in or returns to the regular state.

This transition design treats institutional status and academic context as the main drivers of state change. As a result, the model represents misconduct as adaptive behavior under changing environmental conditions rather than as a permanent agent identity.

3.6. Experimental Setup

The experiment is designed to examine whether the introduction of AI-assisted plagiarism changes behavioral patterns in a multi-agent learning environment and whether institutional intervention suppresses dishonest behavior or redirects it toward less detectable strategies. To keep the model minimal and interpretable at the current stage, the experiment is organized around a small set of scenario, intervention, and social influence conditions.

Two behavioral scenarios are considered. The first is a conventional environment, in which agents may choose between honest submission and direct copying. The second is an AI-augmented environment, in which agents may additionally choose AI-assisted plagiarism. This scenario comparison allows the study to test whether the availability of AI-assisted cheating alters the balance between honest and dishonest behavior.

Two institutional conditions are also considered. Under the no_intervention condition, dishonest actions are detected with lower probability and result in a weaker sanction. Under the active_intervention condition, both direct copying and AI-assisted plagiarism are detected with higher probability and punished more strongly. This comparison is intended to test whether intervention reduces dishonest behavior overall or instead shifts it from conventional plagiarism toward AI-assisted misconduct.

The simulation is conducted over a 16-week academic semester. The weekly schedule includes homework, two quizzes, a midterm examination, and a final examination. Task weights are incorporated into the reward structure so that high-stakes academic periods carry greater consequence than ordinary homework weeks. Homework jointly contributes 0.20 of the task-weight scale, the two quizzes jointly contribute 0.10 with 0.05 assigned to each quiz, the midterm contributes 0.30, and the final examination contributes 0.40. These periods also serve as natural points at which academic pressure may increase.

Social influence is varied across three levels, represented by $p_{inf}\in \{0.0,0.2,0.4\}$. These values represent low, medium, and relatively stronger peer influence on behavioral adaptation. Students are grouped into local peer networks rather than a fully connected classroom, allowing dishonest behavior to spread through neighborhood-level exposure. Here, we assume that the peer-network in the class is stable in each 16-week academic semester.

For each combination of scenario, intervention, and social influence level, the simulation is repeated over multiple random seeds to reduce sensitivity to stochastic variation and improve reproducibility. Note that we keep the peer-network structure in a simulation with each 16-week academic semester. The primary purpose of this first-stage experiment is not to exhaustively optimize all parameters but to establish a healthy and interpretable baseline model for the revised research direction.

3.7. Model Parameters

Table 1 summarizes the main parameters used in the simulation, including learning settings, classroom structure, intervention conditions, and temporal configuration. The key parameters have been decided based on the authors’ experience on their educational activities, especially on the detection probabilities and reward values, which should not produce inconsistent behaviors to our intuition in the simulation. These values were kept intentionally simple in order to support a first-stage interpretable baseline rather than exhaustive parameter optimization. To give the right initial setup to the Q-tables used in the Q-learning, we also applied a pretraining step before starting the actual simulation. This is necessary to avoid producing completely random behaviors in the initial steps of learning, which would be far from the actual human’s behaviors. Regarding the pretraining step shown in Simulation Cycle 1, we performed 100 iterations of pretraining. This pretraining iteration parameter was set to warm up the Q-tables to work, while it was expected to be lower than their complete convergence to give the agent further learning in the actual simulation cycles including the case of receiving penalties in there. We decided this hyper-parameter heuristically to well mimic the students’ behaviors while keeping the learning working as well. As mentioned in the related work section, we could not find any good reference to decide this value from the literature. Further investigations regarding these hyper-parameter settings without heuristics is among our future work. We also prepared two sensitivity analysis setups, which fixes one of $p_{\mathrm{AI}}^{\left(\mathrm{no}\right)}$ or $p_{\mathrm{AI}}^{\left(\mathrm{active}\right)}$ and changes another from 0 to 1 in 0.001 steps, to see the sensitivity of these parameter setups.

The source code for all simulations and the associated data are available in the supplementary file.

3.8. Outcome Measures

The experiment records both weekly and aggregate outcomes. The main behavioral measures are:

• Honest submission: the number of agents choosing legitimate submission;
• Direct copying: the number of agents choosing conventional plagiarism;
• AI-assisted plagiarism: the number of agents choosing AI-supported dishonest behavior;
• Total dishonest behavior: the combined number of direct copying and AI-assisted plagiarism actions.

In addition, the experiment records:

• Caught direct plagiarism: the number of direct copying actions detected by the institution;
• Caught AI-assisted plagiarism: the number of AI-assisted dishonest actions detected by the institution;
• Average cumulative reward: the mean reward accumulated across agents.

Among these measures, the most important are the proportions of honest behavior, direct copying, and AI-assisted plagiarism because they directly indicate whether the availability of AI-assisted cheating changes the behavioral composition of the system. Total dishonest behavior is used to assess whether intervention suppresses misconduct overall, while the separate direct and AI-assisted counts are used to assess whether misconduct is redirected from one dishonest strategy to another.

3.9. Qualitative Plausibility Review

To support the realism assessment of the simulation design, this study conducted a structured qualitative plausibility assessment using multiple large language models (LLMs). The purpose of this procedure was not to validate the simulation empirically but to examine whether the model assumptions were judged plausible as a first-stage behavioral and educational simulation of classroom misconduct.

Five LLM families were used: ChatGPT 5.4, Gemini 3 Flash, Claude Sonnet 4.6, DeepSeek R1 7B, and Qwen 3.5 4B. For each model, the same prompt was submitted in ten independent runs, resulting in 50 review outputs in total. Each run used the same simulation description, research question, state and action definitions, intervention settings, social interaction assumptions, and requested evaluation structure. No follow-up prompts were used during the assessment.

The prompt explicitly instructed the models to evaluate the setup as a behavioral and educational simulation rather than as a machine learning benchmark or optimization problem. The requested evaluation categories were: overall realism judgment, realistic elements, unrealistic or weak elements, the three most important realism gaps, policy realism, recommended minimum revision, and writing advice. The responses were analyzed in two stages. First, the ten responses from each model family were summarized to identify within-model stability. Second, the five model-level summaries were compared to identify cross-model consensus and model-specific differences.

This procedure is treated as a qualitative plausibility and repeatability check of the LLM assessment for the simulation description. Therefore, the LLM assessment is used to clarify the defensible scope and limitations of the model, rather than to claim that the model accurately predicts real student behavior.

4. Experiment Results and Discussions

4.1. Results

This section examines whether the introduction of AI-assisted plagiarism changes the behavioral composition of academic misconduct and whether institutional intervention reduces dishonest behavior overall or shifts it between dishonest strategies. The analysis focuses on the proportions of honest submission, direct copying, and AI-assisted plagiarism, together with total dishonest behavior and detected cases.

The results show that our simulation can produce the scenario that the availability of AI-assisted plagiarism substantially changes agent behavior. In the conventional scenario under no intervention, the mean weekly number of honest submissions was $26.68\pm 0.54$, while direct copying accounted for $3.32\pm 0.54$ actions per week. In contrast, under the AI-augmented scenario with no intervention, honest submissions decreased to $19.20\pm 1.37$, while total dishonest behavior increased to $10.80\pm 1.37$ actions per week. Most of this increase was attributable to AI-assisted plagiarism, which reached $8.42\pm 1.27$ actions per week, compared with $2.37\pm 0.52$ for direct copying. These results indicate that, under the assumed detection and reward settings, the introduction of AI-assisted plagiarism does not merely add a new option to the system but changes the behavioral composition of misconduct. This behavioral shift is summarized in

Table 2. Summary of average weekly behavioral outcomes by scenario and intervention condition, reported as mean ± SD across random seeds and social influence settings.

Scenario	Intervention	Honest	Direct Copying	AI- Assisted	Dishonest Total
Conventional	No intervention	$26.68\pm 0.54$	$3.32\pm 0.54$	$0.00\pm 0.00$	$3.32\pm 0.54$
Conventional	Active intervention	$27.34\pm 0.40$	$2.66\pm 0.40$	$0.00\pm 0.00$	$2.66\pm 0.40$
AI-augmented	No intervention	$19.20\pm 1.37$	$2.37\pm 0.52$	$8.42\pm 1.27$	$10.80\pm 1.37$
AI-augmented	Active intervention	$25.78\pm 0.60$	$1.85\pm 0.33$	$2.37\pm 0.54$	$4.22\pm 0.60$

and illustrated in Figure 2.

Institutional intervention reduced dishonest behavior in both scenarios under the baseline parameterization. In the conventional scenario, active intervention reduced direct copying from $3.32\pm 0.54$ to $2.66\pm 0.40$ actions per week, while honest submissions increased from $26.68\pm 0.54$ to $27.34\pm 0.40$. In the AI-augmented scenario, total dishonest behavior decreased from $10.80\pm 1.37$ under no intervention to $4.22\pm 0.60$ under active intervention, while honest submission increased from $19.20\pm 1.37$ to $25.78\pm 0.60$. Thus, within the tested intervention setting, higher detection probability and stronger sanction severity reduced misconduct overall. The overall intervention effect across social influence conditions is shown in Figure 3. Also we have performed the necessary statistical tests for the above and summarized in

Table 3. Limited paired t-tests for planned main contrasts.

Planned Contrast	Mean Difference	95% CI	t(99)	p
AI availability under no intervention	$7.48$	$[7.21,7.75]$	$55.74$	<0.001
Active intervention in conventional scenario	$-0.66$	$[-0.77,-0.54]$	$-11.56$	<0.001
Active intervention in AI-augmented scenario	$-6.58$	$[-6.83,-6.32]$	$-51.32$	<0.001
AI-assisted vs direct copying under no intervention	$6.05$	$[5.80,6.29]$	$48.92$	<0.001
AI-assisted vs direct copying under active intervention	$0.53$	$[0.40,0.65]$	$8.12$	<0.001

and

Table 4. Paired t-tests for the effect of social influence probability on total dishonest behavior.

Comparison	Mean Difference	95% CI	t(99)	p
Conventional, no intervention: $p_{inf}$ 0.0 vs. 0.2	$-0.03$	$[-0.08,0.02]$	$-1.19$	$0.238$
Conventional, no intervention: $p_{inf}$ 0.0 vs. 0.4	$-0.04$	$[-0.11,0.02]$	$-1.28$	$0.205$
Conventional, active intervention: $p_{inf}$ 0.0 vs. 0.2	$0.00$	$[0.00,0.00]$	$0.00$	$1.000$
Conventional, active intervention: $p_{inf}$ 0.0 vs. 0.4	$0.00$	$[0.00,0.00]$	$0.00$	$1.000$
AI-augmented, no intervention: $p_{inf}$ 0.0 vs. 0.2	$0.13$	$[-0.03,0.29]$	$1.58$	$0.117$
AI-augmented, no intervention: $p_{inf}$ 0.0 vs. 0.4	$0.21$	$[0.04,0.39]$	$2.43$	$0.017$
AI-augmented, active intervention: $p_{inf}$ 0.0 vs. 0.2	$-0.01$	$[-0.05,0.03]$	$-0.31$	$0.760$
AI-augmented, active intervention: $p_{inf}$ 0.0 vs. 0.4	$0.02$	$[-0.03,0.08]$	$0.88$	$0.382$

.

Because intervention acts through detection probability and sanction severity, detected-case outcomes were also examined.

Table 5. Summary of average weekly detected cases by scenario and intervention condition, reported as mean ± SD across random seeds and social influence settings.

Scenario	Intervention	Caught Direct Copying	Caught AI- Assisted
Conventional	No intervention	$1.81\pm 0.34$	$0.00\pm 0.00$
Conventional	Active intervention	$2.14\pm 0.38$	$0.00\pm 0.00$
AI-augmented	No intervention	$1.31\pm 0.33$	$2.12\pm 0.41$
AI-augmented	Active intervention	$1.48\pm 0.29$	$1.06\pm 0.32$

summarizes the average weekly detected cases for direct copying and AI-assisted plagiarism. In the conventional scenario, active intervention increased detected direct-copying cases from $1.81\pm 0.34$ to $2.14\pm 0.38$ actions per week, even though total direct copying decreased. In the AI-augmented scenario, detected AI-assisted plagiarism decreased from $2.12\pm 0.41$ under no intervention to 1.06 ± 0.32 under active intervention because the total number of AI-assisted plagiarism actions was reduced. These results indicate that intervention changes both the frequency of dishonest behavior and the number of detected cases within the tested parameter setting.

However, under the baseline parameterization, active intervention did not fully remove AI-assisted plagiarism as the larger component of dishonest behavior in the AI-augmented environment. Under active intervention, direct copying averaged $1.85\pm 0.33$ actions per week, whereas AI-assisted plagiarism averaged $2.37\pm 0.54$ actions per week. This indicates that, when AI-assisted plagiarism is modeled with lower detection probability than direct copying, it may remain behaviorally attractive within the model even under stricter detection and stronger sanction settings.

Figure 4 and Figure 5 show the results of the two sensitivity analysis setup, which fixes one of $p_{\mathrm{AI}}^{\left(\mathrm{no}\right)}$ or $p_{\mathrm{AI}}^{\left(\mathrm{active}\right)}$ and change another from 0 to 1 in 0.001 step, to see the sensitivity of these parameter setups. With the decrease in each value, the effect of the associated parameter dropped smoothly.

The substitution analysis further supports this conditional interpretation. Under no intervention, the share of AI-assisted plagiarism within total dishonest behavior ranged from approximately $0.776$ to $0.785$ across the tested social influence levels. Under active intervention, this share declined but remained approximately $0.561$ to $0.564$. Therefore, intervention reduced total dishonest behavior, but under the baseline detection setting the AI-assisted option continued to account for more than half of all dishonest actions. The relative balance between direct copying and AI-assisted plagiarism is shown in Figure 6.

The results also show that the current social influence variation produced only modest changes in aggregate behavior. Across the tested values of social influence probability, the mean weekly levels of honest and dishonest behavior remained relatively stable within each scenario–intervention setting. For example, in the AI-augmented scenario under active intervention, total dishonest behavior remained between approximately $4.21$ and $4.24$ actions per week across the tested influence levels. Similar stability was observed under no intervention, where total dishonest behavior remained between approximately $10.68$ and $10.89$ actions per week. This suggests that, in the current minimal implementation, the effect of AI availability and intervention is more pronounced than the effect of the present social influence mechanism.

The weekly trend analysis is consistent with this interpretation. AI-assisted plagiarism remains visible throughout the semester in the AI-augmented environment, although at lower levels under active intervention than under no intervention. The temporal pattern of this behavior is illustrated in Figure 7, which shows that intervention suppresses but does not remove the AI-assisted cheating pathway.

Taken together, these findings support the view that AI-assisted plagiarism changes the structure of behavioral choice in the academic misconduct environment. The model shows that once AI-assisted plagiarism becomes available, dishonest behavior increases and becomes concentrated in the AI-assisted category. Although intervention can reduce misconduct overall, it does not fully remove the strategic attractiveness of AI-assisted plagiarism. At the current stage, the strongest result of the experiment is therefore the behavioral shift introduced by AI-assisted cheating and the partial, rather than complete, corrective effect of institutional intervention.

4.2. Discussion

The results suggest that our simulation can produce the scenario that the introduction of AI-assisted plagiarism changes academic misconduct not only quantitatively but also structurally. In the conventional scenario, dishonest behavior is limited to direct copying, and intervention reduces that behavior in a relatively straightforward manner. In the AI-augmented scenario, however, the action space becomes richer, and the system adapts by allocating a substantial share of dishonest behavior to AI-assisted plagiarism. This indicates that the presence of AI-assisted cheating is not simply an additional option but a qualitatively important change in the behavioral environment.

A central implication of the findings is that anti-plagiarism intervention should not be evaluated solely in terms of whether total misconduct decreases. In the current model, active intervention reduces dishonest behavior overall, especially in the AI-augmented scenario. However, even after intervention, AI-assisted plagiarism remains more prevalent than direct copying. This suggests that institutional measures may suppress misconduct without fully eliminating the strategic attractiveness of less detectable cheating modes. In practical terms, a reduction in one form of misconduct does not necessarily imply a reduction in all forms.

This interpretation is particularly relevant in contemporary educational settings. Traditional plagiarism control mechanisms are often designed around direct copying, source similarity, and textual overlap. By contrast, AI-assisted plagiarism may be more individualized, more flexible, and less dependent on peer access. The model reflects this difference through distinct detection probabilities, and the results suggest that such differences are sufficient to change the behavioral balance of misconduct. Accordingly, academic integrity policy should be interpreted not only as a deterrence mechanism but also as a factor that may unintentionally reshape the strategic landscape of cheating.

The experiment also highlights the importance of distinguishing between external environmental factors and internal moral assumptions. In the previous reward-shaping direction, part of the intervention logic risked appearing as though the simulations were directly altering student conscience. The revised model instead treats the environment, available cheating options, detection conditions, and institutional sanctions as the primary intervention channels. This makes the model more consistent with a social–behavioral interpretation of academic misconduct, where behavior changes in response to opportunity structure and policy environment rather than through direct internal moral programming.

At the same time, the results show that the present implementation of social influence has only a modest effect on aggregate outcomes. This does not necessarily mean that peer effects are unimportant in real educational settings but rather that the current minimal influence mechanism is weaker than the effects of scenario design and institutional intervention. In other words, the model at this stage answers the AI-availability and intervention questions more clearly than the social diffusion question. This is acceptable for an early-stage experiment, but it also suggests an important direction for improvement. Future versions of the model may require stronger or more expressive peer mechanisms, such as weighted networks, directed influence, or explicit observation of peer sanction events.

The repeated multi-LLM plausibility assessment produced a stable cross-model judgment. Across five LLM families and 50 total outputs, the simulation was consistently judged to be acceptable as a stylized first-stage baseline but not as a strongly realistic model of student psychology or classroom behavior. The strongest positive consensus was that separating AI-assisted plagiarism from direct copying is a realistic and valuable design choice. The strongest repeated limitations concerned the use of Q-learning as a simplified adaptive heuristic, limited agent heterogeneity, and the need for richer treatment of AI-related task utility and social dynamics.

Table 6. Cross-LLM qualitative plausibility assessment summary across ten repeated runs per model.

Model	Stable Judgment Across Runs	Distinctive Emphasis
ChatGPT 5.4	Acceptable as a stylized first-stage baseline but not as a strongly realistic classroom model.	Balanced realism critique, including Q-learning limitations, missing heterogeneity, and the need to distinguish dishonest AI use from acceptable AI support.
Gemini 3 Flash	Stable baseline judgment with repeated caution about behavioral realism.	Missing effort or time cost of honest work, and stronger support for AI-resistant assessment design than simple AI restriction.
Claude Sonnet 4.6	Highly stable reviewer-like judgment that the model is acceptable only as an exploratory baseline.	Theoretical defensibility, careful validity framing, and avoiding overclaims about realism or prediction.
DeepSeek R1 7B	Broadly stable judgment that the basic state-action setup is acceptable, with less detailed theoretical critique.	Longitudinal behavior, integrity history, behavioral inertia, and evolving peer influence.
Qwen 3.5 4B	Stable judgment that the study is more defensible as a mechanism-oriented simulation than as a behavioral realism model.	AI-related task utility, effort structure, and why AI-assisted misconduct may be behaviorally attractive.

summarizes the stable judgments and distinctive emphases observed across ten repeated runs for each of the five LLM families2.

The model-specific emphases also provide useful guidance for interpreting and extending the simulation. ChatGPT 5.4 provided a balanced critique across multiple realism dimensions. Gemini 3 Flash emphasized the missing effort cost of honest work and the practical value of AI-resistant assessment design. Claude Sonnet 4.6 emphasized theoretical defensibility and the risk of overclaiming. DeepSeek R1 7B repeatedly emphasized longitudinal student history and dynamic peer influence. Qwen 3.5 4B emphasized the behavioral attractiveness of AI-assisted misconduct through task utility and effort reduction. Taken together, these findings support the use of the present simulation as an exploratory, hypothesis-generating baseline while clarifying that stronger behavioral realism would require richer modeling of student heterogeneity, task design, and social learning. These convergent judgments support the decision to present the current model as a first-stage exploratory simulation while also clarifying where future refinement is most needed. There should be further discussions regarding the use of LLMs for a plausibility assessment on behalf of humans.

Actually, some works started trying to use LLMs on behalf of humans for evaluation in the situations where they would be very hard to be assessed by humans (e.g., [40] in a healthcare field and [41] in a smart electricity control). However, these were mainly focused on comparing and benchmarking the performance of the implemented systems or methods, and therefore they deeply depend on the use of a specific LLM for easier comparisons under equal conditions. Therefore, directly applying these results into our approach is difficult. In this paper, we have tried to utilize different families of LLMs in a different model size. Here, we would note one thing about the limitations regarding the obtained consistent results from these LLMs. We believe having consistent results in this context would be good. However, in [42], it has been reported that having consistent results from multiple AIs would lead to more acceptance of these AIs’ recommendations for decision-making by humans. Therefore, even in the case of having consistent results from different LLMs for a given matter, the use of these results in a practical decision-making context should be carefully reviewed. Our primary aim of using LLMs in this context is just to provide some help for assessing the design of a simulation but not to replace the empirical evidence to support them. This is one of the limitations of our work.

Overall, the present experiment should be understood as a first-stage plausibility assessment under the revised research direction. Its main contribution is to demonstrate that introducing AI-assisted plagiarism materially changes the adaptive landscape of misconduct and that institutional intervention reduces dishonest behavior only partially when AI-assisted strategies remain available. The findings therefore support the idea that plagiarism in the AI era is best understood as a problem of behavioral adaptation under changing technological and institutional conditions, rather than as a static problem of direct copying alone.

5. Conclusions

This study presented an approach to produce and assess a simulation to examine how AI-assisted plagiarism changes dishonest academic behavior in a socially interacting learning environment under different educational conditions. Using a scenario-based simulation with Q-learning as a baseline adaptive mechanism, the study compared a conventional plagiarism environment with an AI-augmented environment and evaluated how institutional intervention influences the resulting behavioral composition in a simulation.

The results show that the implemented simulation is able to produce the scenario where the introduction of AI-assisted plagiarism materially changes the structure of dishonest behavior. When AI-assisted plagiarism is available, total dishonest behavior increases and a large share of misconduct shifts toward the AI-assisted category rather than remaining concentrated in conventional direct copying. This can be used to make the suggestion that AI-assisted plagiarism should not be treated merely as an additional cheating option but as a qualitatively important change in the academic misconduct environment.

The results also show that intervention reduces dishonest behavior overall but does not fully eliminate AI-assisted plagiarism as a persistent strategic pathway in the simulation. In the AI-augmented environment, stricter detection and stronger sanctions will substantially reduce misconduct, yet AI-assisted plagiarism will remain behaviorally attractive and continues to account for a large share of dishonest actions as seen in the simulation. The simulation results can also be used to start arguing that institutional responses should be evaluated not only in terms of overall deterrence but also in terms of whether they unintentionally reshape misconduct toward less detectable forms.

At the same time, the present model should be interpreted as a first-stage exploratory simulation rather than a validated predictor of real classroom behavior. The qualitative plausibility review using multiple large language models supported the acceptability of the model as a stylized baseline, while also highlighting important limitations, especially the points that the LLM-based plausibility assessment itself does not replace any empirical validations, the use of Q-learning as a simplified adaptive heuristic, limited agent heterogeneity, and a relatively shallow social influence mechanism. Accordingly, the findings should be understood as directional and hypothesis-generating rather than as direct empirical claims about student psychology.

Future work may strengthen the realism of the model by expanding agent heterogeneity, refining the social interaction mechanism, and incorporating more explicit classroom-side policy conditions such as AI-resistant task design or discussion policy.