The study employs a convergent parallel mixed-methods design, aimed at investigating the perceptions and competencies of trainee teachers regarding the integration of AI in assessment. Data collection was conducted between April and May 2025 via digital administration.

The convenience sample consists of N = 662 teachers enrolled in qualifying courses (60 ECTS) at the University of Foggia (n = 523; 79.0%) and the University of Bergamo (n = 139; 21.0%).

The demographic composition (Table 1) highlights a prevalence of female participants (n = 424; 64.0%) compared to the male component (n = 234; 35.3%). The distribution by age group reveals a heterogeneous population: although the modal group is represented by the 25–29 age range (n = 176; 26.6%), a significant presence of trainee teachers in the over-45 range (n = 162; 24.5%) is observed, delineating a multigenerational training context.

In terms of academic background, the majority of participants hold a Master’s Degree (70.0%), followed by postgraduate qualifications or research doctorates (7.4% overall). Regarding professional experience, the sample is primarily divided between teachers with less than 5 years of service and those with experience ranging between 6 and 10 years (51.0%).

The data were collected between April and May 2025 using a Computer-Assisted Web Interview (CAWI) methodology. The survey was self-administered via a secure online platform (Microsoft Forms), and the link was distributed to participants through the institutional Learning Management Systems (LMS) of the involved universities. Participation was voluntary and anonymous, with an estimated completion time of approximately 20 min. The instrument consisted of 57 items organized into four logical sections, which were subsequently operationalized into five analytical dimensions (Trust, Competence, Relational Impact, Training, Perceived Utility) for the analysis:

To ensure maximum measurement sensitivity, the response scales were adapted to the nature of the constructs. Specifically, Trust, Technical Comfort, Training Adequacy, and Relational Risks were measured using a 10-point estimation scale to capture subtle variations in sentiment and avoid the flattening effect of shorter scales. Conversely, specific constructs related to Perceived Utility (e.g., for writing correction) and traditional pedagogical background utilized a standard 5-point Likert scale. Table 2 provides a detailed overview of the key variables and the measurement scales used in the present study.

The analysis plan was conducted entirely using the R statistical computing environment (via the RStudio interface) and was structured into distinct procedures for the two types of data. Regarding the quantitative component, descriptive statistics (frequencies, mean, and standard deviation) were initially calculated to characterize the sample profile. Subsequently, to address the research questions, Spearman’s bivariate correlation analysis (ρ) was applied using the psych and tidyverse packages; this methodological choice was dictated by the ordinal nature of the variables (measured on Likert scales) and the non-normality of the data distribution, thereby ensuring greater robustness compared to parametric coefficients. On the qualitative front, open-ended responses were subjected to Inductive Thematic Analysis following the model by Braun and Clarke (2006). To ensure the trustworthiness and reduce the subjectivity of the inductive coding process, a methodological triangulation strategy was adopted. The themes emerged from the manual qualitative analysis were cross-validated against the results of the automated Text Mining analysis performed in R. Specifically, the high-frequency lexical clusters identified by the software (using tidytext and tm packages) consistently matched the semantic categories manually identified by the researchers. This convergence between human interpretation and algorithmic extraction confirms the robustness of the coding frame and mitigates the risk of researcher bias.

Regarding the first research question, aimed at investigating the transformation of assessment in the interaction with AI, the data outline a scenario characterized by a deep ambivalence between technical competence and pedagogical trust. Descriptive analysis highlights that, despite a moderate level of reported comfort in the general use of intelligent tools (M = 6.04; SD = 2.52), trust in delegating assessment support to these systems suffers a significant decline, settling on mean values below the sufficiency threshold (M = 4.83; SD = 2.21). Even more critical is the data regarding the perceived utility of AI for self-correction, which records a minimal mean (M = 2.57 on a 5-point scale), suggesting widespread skepticism toward the current didactic efficacy of such tools. Sample polarization is further confirmed by the distribution of responses on AI’s capacity to support assessment (M = 5.18), where the modal frequency coincides with the scale’s minimum value (1), indicating the presence of certain resistance. However, inferential analysis, conducted via Spearman’s correlation coefficient, clarifies that trust is not independent of competence: indeed, significant positive associations emerge between the level of trust and the perception of the support role (ρ = 0.56), as well as between trust itself and the degree of technological comfort (ρ = 0.54). The qualitative analysis allows for decoding the underlying reasons for these numerical trends, revealing four semantic areas that explain the tension between acceptance and rejection:

Instrumental Efficiency (Acceptance): On the pragmatic axis, a view oriented towards optimization consolidates. Teachers recognize AI as a valuable resource for time management, the automation of repetitive correction tasks, and the enhancement of feedback speed.

Relational Depersonalization (Rejection): Conversely, on the emotional axis, a pervasive fear of “dehumanization” emerges. Participant narratives describe the risk of a drift towards an “arid,” “sterile,” and “distant” assessment, deprived of the empathetic component deemed essential for education.

Epistemic Opacity (Distrust): A third area concerns the lack of transparency. The recurring references to the inability to explain the algorithm’s judgment (“I do not know how it decides”) fuel the distrust recorded in the quantitative data, linking resistance to the “Black Box” phenomenon.

Human Irreducibility (Defense): Finally, the tension resolves in the claim of a “human reserve.” Assessment is described as a hermeneutic act requiring context and sensitivity, qualities that participants define as exclusively human and non-delegable to a machine.

This dialectic resolves into an explicit demand for a principle of pedagogical subsidiarity, according to which AI is legitimized exclusively as an ancillary support tool, without ever being able to replace professional judgment and the educational centrality of the teacher (see Table 3).

Regarding the second research question, aimed at delineating the central training dimensions for a teacher capable of integrating AI without educational compromises, the descriptive data primarily indicate a significant gap in perceived preparation. The adequacy of the training received is evaluated as insufficient, with a mean of 4.29 (SD = 2.63) and a modal frequency situated at the scale’s minimum value (1), pointing to a widespread perception of structural deficiency in current training pathways. Nevertheless, participants express moderate confidence in their capacities for critical discernment (M = 5.23) and ethical management (M = 5.13), as well as in the competence to mediate the use of AI while preserving the educational relationship (M = 5.39). The inferential analysis, conducted on a sample of 662 subjects using Spearman’s correlation, offers a crucial interpretive key for understanding these self-perceptions. A network of statistically significant relationships (p < 0.001) emerges, delineating a complex framework of technological acceptance. A finding of particular relevance concerns the role of training: while the training received correlates positively with the sense of technical comfort in using the tool (ρ = 0.46), it shows a decidedly weaker association with trust in assessment (ρ = 0.36) and a marginal association with the perception of AI as an effective didactic support (ρ = 0.20). Conversely, trust in AI as an assessment tool appears to be driven much more strongly by the perception of concrete pedagogical support (ρ = 0.57) and by the actual use of the tool (ρ = 0.54), as well as by its perceived pragmatic utility (ρ = 0.48) (see Figure 1).

These results suggest the existence of a “training paradox”: current pathways appear effective in fostering technical confidence (“knowing how to use the tool”), yet prove less impactful in constructing an integrated pedagogical vision (“knowing how to assess with the tool”). Consequently, teachers’ trust appears to stem not so much from theoretical instruction as from direct and pragmatic experience. This highlights the necessity to reorient training toward the development of a Critical AI Literacy that anchors technical competence in ethical and decision-making responsibility.

The qualitative inquiry, supported by lexical evidence emerging from text mining, provides a vivid snapshot not only of what teachers think but of how they position themselves professionally vis-à-vis technology. The analysis of the textual corpus allows for the construction of training needs into three interconnected dimensions: operational uncertainty, ethical urgency, and the defense of pedagogical humanism.

First and foremost, the most immediate finding is the pervasiveness of doubt. The recurrence of dubitative linguistic forms, specifically the lemma “Uncertainty,” signals a state of operational paralysis (Figure 2). This term, appearing alongside the explicit demand for “Training” and technical “Knowledge,” suggests that the issue is not merely a lack of tools, but a difficulty of pedagogical imagination: many respondents struggle to concretely envision how AI can be integrated into their routine without disrupting it. This void of “vision” implies that current training may have been too theoretical, leaving teachers devoid of role models or practical examples of sustainable integration.

Once the initial bewilderment is overcome, the training request crystallizes in a precise direction: the demand for meaning prevails over the demand for tools. The prominence of terms such as “Critical Thinking” and “Ethics,” contrasted with the high frequency of “Risk” and “Loss” (Figure 3), indicates that teachers are not asking to become computer technicians, but are claiming a deontological compass. The dominant concern does not regard the mechanics of the algorithm, but the management of its implications: data privacy, the recognition of cognitive biases, and, above all, the ability to discern when the use of technology is an added value and when it represents a loss of educational context.

This evidence supports the hypothesis that Critical AI Literacy is not an academic luxury, but a pragmatic need for those who must manage educational responsibility in the classroom.

Finally, the semantic analysis of instructional strategies reveals a clear defensive line of demarcation. The strong recurrence of the term “Relationship,” flanked by “Human” and “Educational” values, unveils the teachers’ adaptation strategy (Figure 4): AI is welcomed if it acts as a “prosthesis” to enhance attention to the individual (e.g., by creating differentiated materials), but is rejected when it attempts to surrogate the relationship. In the free narratives, assessment is described as a hermeneutic act, an interpretation of student growth that requires empathy and context. Consequently, the “educational centrality” invoked by participants translates into the will to guard the emotional space of learning, delegating procedural execution to the machine while firmly retaining the monopoly on value judgment and educational care.

The evidence collected in this study suggests that the observed resistance is not a barrier to be dismantled, but a meaningful signal to be heeded, translating into specific implications for policy and practice. Consequently, we propose a strategic realignment of teacher training and school governance based on the “Training Paradox” and the demand for ethical control emerged from the data.

Regarding curriculum design, the discrepancy between high technical comfort and low trust indicates that professional development programs must urgently shift their focus from mere operational interaction (“prompt engineering”) to a rigorous “Output Validation Literacy.” Teachers need to be trained not as passive users but as critical evaluators capable of detecting hallucinations, logic errors, and algorithmic biases in AI-generated content. Crucially, this literacy must be grounded in a deep ethical understanding of the tool; to mitigate the distrust fueled by “epistemic opacity,” training modules should demystify the “Black Box” nature of Large Language Models, clarifying their probabilistic rather than logical functioning. Only by understanding the mechanical limits of the machine can teachers exercise true epistemic responsibility without fearing professional abdication.

At the institutional level, these insights call for the standardization of “Human-in-the-loop” (AI²L) workflows. School policies should explicitly regulate that while AI may assist in preliminary screening or feedback generation, the final assessment and its semantic explanation must remain the exclusive prerogative of the human teacher, thereby formalizing the “Principle of Subsidiarity” demanded by participants. In operational terms, we recommend adopting “Triangulated Assessment” practices, where AI is utilized not as a grader but as an adversarial critic or a “second opinion” to refine the teacher’s own judgment. This approach ensures that the integration of AI enhances rather than replaces the relational and hermeneutic value of educational assessment.

Despite the significance of the findings and the sample size, the present investigation entails certain limitations that warrant a cautious interpretation of the results. First, the non-probabilistic nature of the sampling strategy and the geographical concentration of participants in two specific universities (Foggia and Bergamo) preclude the statistical generalizability of the data to the entire Italian teaching population, although they offer a representative cross-section of the current state of initial teacher training. Second, the sample composition, characterized by a prevalence of novice teachers or those in the qualification phase with less than 5 years of experience, may not fully reflect the professional stances, resistances, or consolidated pedagogical competencies of tenured teachers with greater seniority, who might approach technological innovation through different dynamics. A further limitation lies in the data collection methodology based on self-reported measures. While effective for investigating perceptions and beliefs, this approach may be subject to social desirability bias, particularly regarding sensitive topics such as ethics and digital competence. Consequently, it remains to be verified through future observational or experimental studies the extent to which declared attitudes translate into actual instructional practices in the classroom. Finally, the rapid evolution of Generative AI tools necessitates viewing these results as a snapshot situated within a specific historical-technological moment, inevitably destined to evolve as teachers’ familiarity with these new devices progresses. Additionally, it must be noted that the correlational nature of the study allows us to identify significant associations between variables (e.g., Training and Trust) but does not establish direct causal links. Therefore, the interpretative models proposed (such as the Training Paradox) should be understood as probabilistic frameworks to be further verified through experimental designs.

In light of the emerging evidence and the outlined limitations, several promising avenues for future research arise. First, it would be beneficial to extend the investigation through longitudinal studies monitoring the evolution of the perceptions and assessment practices of the same subjects during their professional induction, verifying whether field experience modifies trust or resistance toward AI. Second, it appears crucial to complement the survey of perceptions with direct observational or experimental analysis of classroom practices, investigating how teachers effectively integrate AI into assessment design and formative feedback. Finally, a promising line of inquiry concerns the comparison between STEM and Humanities teachers, to verify whether the epistemological nature of the subject taught influences the propensity to delegate or retain specific assessment functions. Such in-depth investigations could provide the empirical foundations for constructing an AI Assessment Literacy framework that is increasingly refined and grounded in educational reality.