Psycholinguistic markers of a comfortable educational environment for students in the context of digital transformation

Full Text

Introduction

The contemporary educational landscape is characterized by an increasing diversification of learning formats, ranging from traditional face-to-face instruction to fully online and blended environments [1; 2; 3; 4]. This transformation, accelerated by technological advancement and global events, has prompted a shift in the understanding of educational outcomes beyond conventional academic metrics toward a more holistic conceptualization of student success. Central to this expanded perspective is educational well-being, a multidimensional construct encompassing students’ emotional experiences, psychological functioning, and subjective satisfaction within learning environments [5; 6].

Literature revie

Within the existing literature, educational well-being is conceptualized in two primary ways: as a facilitator of academic learning, and as an outcome of the educational process itself. These perspectives are not mutually exclusive. For instance, McCallum and Price [7] argue for an “inextricable link between wellbeing and academic achievement,” positioning well-being as both “foundational and integral to learning” [7, p. 2]. Similarly, Seligman et al. [8] contend that enhanced well-being creates a synergistic effect with learning, suggesting that “increases in well-being are likely to produce increases in learning”. This dual role positions well-being not merely as a desirable affective state, but as a core component of effective education. These days, however, students’ subjective well-being is increasingly at risk due to a confluence of academic pressures, including high-stakes testing and workload intensification, alongside financial concerns and a marked reduction in meaningful social interaction [9; 10].

While considerable research has examined social and psychological determinants of well-being in traditional classrooms [11; 12; 13], the proliferation of alternative learning formats necessitates comparative a more differentiated analysis of how these factors are experienced and interpreted by students across contexts.

Self-determination theory provides a useful framework for understanding educational well-being, positing that the satisfaction of three basic psychological needs, i.e. autonomy, competence, and relatedness, is essential for optimal functioning and personal growth across the lifespan [14; 15]. In educational contexts, these needs are realized through concrete interactional conditions: opportunities for participation and choice (autonomy), clarity and feedback supporting mastery (competence), and supportive relationships with teachers and peers (relatedness). These conditions are enacted in practice and are also reflected in how students describe and interpret their learning experiences. Within this framework, teacher – student relationships and assessment practices seem to be key regulators of educational experience. Supportive teacher interactions have been shown to foster engagement and positive affect while reducing academic anxiety [16]. At the same time, the form and function of assessment play a critical role: feedback-oriented assessment can support learning and self-efficacy, whereas unpredictable or punitive evaluation may increase stress and avoidance. Peer relatedness further contributes to students’ sense of belonging and emotional security, though its effects are often intertwined with teacher relationships [17].

The expansion of digital technologies in education, including artificial intelligence, introduces new complexities to these established dynamics. AI systems increasingly act as instructional tools and, in some cases, as interactive partners in the learning process. Questions are being raised about how human – AI interaction shapes students’ psychological experiences [18]. Some scholars conceptualize this evolution as giving rise to entirely new “formats” of educational interaction that transcend traditional distinctions between face-to-face and online learning [19; 20]. In a previous study [21; 22], it was theorized that creative partnerships may emerge, wherein students, teachers, and AI agents engage in collaborative knowledge construction, with each participant contributing unique capacities: human creativity, emotional intelligence, and contextual understanding from teachers and peers, complemented by AI’s computational power, information access, and pattern recognition. The empirical analysis by Otto et al. [23] reveals specific modalities through which such partnerships manifest, including Tutoring, Co-creating, Processing, Coaching, and Simulating. These modalities function primarily to scaffold individual learners’ development of higher-order cognitive competencies, specifically problem-solving and critical thinking. This vision of triadic pedagogy positions the student as an active co-author rather than passive recipient of learning, potentially enhancing autonomy and engagement.

However, while some theoretical models describe novel forms of “triadic” interaction between students, teachers, and AI systems, it remains unclear how such configurations are perceived by students themselves and to what extent they are integrated into their representations of effective academic contexts. Understanding how students perceive these emerging configurations and their implications for well-being represents a critical frontier for educational research. Such inquiry requires methodological approaches capable of capturing students’ subjective experiences, their hopes and anxieties regarding educational technology, and the nuanced ways AI integration may differentially affect diverse learner populations. To capture students’ subjective representations of educational environments, it is important to consider not only thematic content but also the linguistic forms through which experience is expressed. Psycholinguistic markers, such as pronominal structures (e.g., “I” vs. “we”) and modal expressions (e.g., “can,” “must”), can serve as indicators of perceived agency, relational positioning, and normative structure within the learning environment. In this sense, language reflects the “grammar” of interaction, including the degree of compatibility, shared participation, and perceived constraints, providing an additional level of analysis for understanding educational well-being. In this perspective, psycholinguistic markers can be interpreted as indicators of individual experience and as signals of the degree to which interaction is organized as a shared, dialogic space – a necessary condition for the emergence of creative partnerships.

Despite growing scholarly attention to educational well-being, significant gaps persist in the literature. Existing research has largely focused on single learning formats and has mostly relied on quantitative measures, which may not fully record the subjective and interpretive dimensions of students’ experiences. In addition, the linguistic and discursive organization of these experiences is still underexplored. Qualitative approaches, particularly those eliciting students’ narrative accounts of desirable and undesirable learning experiences, offer a valuable means of accessing these phenomenological and psycholinguistic dimensions [24].

This study addresses these gaps by analyzing students’ subjective representations of learning settings through a comparative examination of two text corpora: essays describing an “ideal lesson” and essays describing an “uncomfortable lesson.” The analysis focuses on the thematic structure of these representations; psycholinguistic markers of agency and interaction, including pronominal framing and modality; the role of the teacher and assessment as regulators of the educational experience; and differences in discursive styles across academic contexts (technical vs. humanities). In addition, the study explores students’ perceptions of AI as a potential element of academic settings. This approach allows for the identification of both stable and context-sensitie features of educational well-being, as well as the linguistic patterns through which they are constructed.

The Aim and Hypotheses of the Study

The study is aimed at reconstructing students’ subjective perceptions of educational environments by comparing two text corpora (their essays on the topic of “My Ideal Lesson” and contrasting essays about an uncomfortable lesson). The following parameters are specified: the thematic structure of representations; psycholinguistic markers of the subject position (the pronominal frame and the modality of permission/prohibition) as indicators of the grammar of compatibility and its reduction; the role of the teacher and assessment as regulators of the educational experience; discourse differences between universities with technical and humanities profiles; the place of the digital/AI component in the “fourth format” model of creative partnerships that implies reflected subjectness and dialogic transcription (M. V. Borodenko).

Hypotheses:

H1. The image of an “ideal lesson,” evaluated as realistic and highly comfortable, will elicit a ceiling effect on the realism and comfort scales, limiting the variability of quantitative differences.

H2. Texts about “ideal lessons” will be more detailed than those about uncomfortable lessons, reflecting greater cognitive accessibility of the positive scenarios and/or a defensive economy when describing negative experiences.

H3. The core of the thematic profile of an “ideal lesson” will be based on psychological safety, cognitive accessibility and predictability, dialogicity, and developmental assessment; the profile of an uncomfortable lesson will be formed around the threat of assessment, unpredictable control, stress, pressure, and a disrupted group atmosphere.

H4. The “ideal lesson” will be marked by an increase in permissive modality (“may/can”) and elements of the grammar of compatibility (“we”-frame), while the uncomfortable lesson will be marked by an increased prohibitive modality (“must not/cannot”) and the atomization of discourse (the reduction of “we” to “I” as controlled by the assessor).

H5. The humanities subgroup is expected to exhibit an emotional-communicative discourse style (texts will be more extensive; references to the lesson atmosphere, stress, and supportive relationships will be more frequent), while the subgroup of students of technical and engineering professions is expected to produce an operational-structural style (pace, classroom assignments, lesson organization), with comparable scale ratings of the ideal lessons.

H6. Attitudes toward AI (trust; instrumental use; skepticism) will be conceptually linked to the need for predictability and autonomy: trust will be associated with the experience of an AI agent as a factor in reducing uncertainty and supporting dialogic transcription; skepticism will be associated with the experience of a threat to autonomy and an increased prohibitive modality when discussing the digital mediator.

Materials and methods

Participants

Eighty-five senior students from two Russian universities participated in the pilot study: 60 students from Moscow Institute of Physics and Technology and 25 students from Moscow State Linguistic University. The participants were aged 19 to 26 years (M = 20.52, Median = 20), with the following distribution: 8 students aged 19, 40 aged 20, 30 aged 21, 3 aged 22, 2 aged 23, 1 aged 24, and 1 aged 26. The sample included 52 men and 33 women. Most participants were enrolled in their 3rd (n = 39) or 4th (n = 45) year; one participant was in the 5th year. The data were collected in standard educational settings.

Materials and Procedure

Participants were asked to produce two short essays, one describing their “ideal lesson” and the other defining an “uncomfortable lesson.” The goal was to capture the students’ perceptions of psychological comfort in a learning setting, their emotional states, the role of AI technologies, and the forms of interaction among students, teachers, and digital agents.

The instructions encouraged the participants to address four thematic areas:

• psychological comfort in the educational environment (defined as support from teachers and peers, autonomy, predictability of lesson structure);
• subjective experiences during learning (interest, engagement, stress, resilience when handling complex tasks);
• the role of AI technologies (expectations, concerns, perceived function as a tool or a partner);
• forms of joint activity (dialogic interaction and co-creation between student, teacher, and AI).

The recommended text length was 200–300 words; however, the essays varied substantially. The “ideal lesson” essays averaged 60.3 words (Median = 54, range 1–209), with five essays under 10 words. The “uncomfortable lesson” essays averaged 37.2 words (Median = 28, range 1–154), with 17 essays under 10 words. These findings may indicate that participants more readily and elaborately describe preferred learning experiences than undesirable ones.

Participants also rated, on two 5-point scales, the perceived extent of how possible (attainable) their ideal lessons could be and how comfortable they would feel. For the possibility scale, the mean was 4.58 (Median = 5), and for the comfort scale, the mean was 4.86 (Median = 5). The “ceiling effects” were pronounced in both measures, particularly for comfort, with 90.6% of participants assigning the maximum score. These results presume that students perceive the ideal lesson as both attainable and highly desirable.

All essays were subjected to content analysis. The analysis methodology was specifically designed to identify students’ perceptions of the characteristics of a favorable educational environment, as well as the specifics of human interaction with digital intelligent systems in the educational process.

The content analysis combined deductive and inductive approaches. The deductive component was based on a pre-developed categorical system, formed on the basis of modern research on psychological well-being in education, as well as theoretical concepts of human interaction with artificial intelligence (CAI, Communicative AI [25]) and V. A. Petrovsky’s paradigm of creative partnerships [26; 27; 28]. The inductive component of the analysis allowed for the identification of additional semantic elements arising directly in the students’ texts and not anticipated by the original categorical system. Such elements were recorded during the analysis and, if necessary, formalized as additional subcategories. The unit of analysis was a complete semantic fragment of text, for example, a sentence, a paragraph, or a group of sentences united by a common theme and expressing a single complete thought or a single element of subjective experience.

The categorical apparatus of the methodology included four thematic sections.

1. Components of psychological comfort in the educational environment. This section included the following categories:
   
   1. support: mentions of assistance, attentive attitudes of the teacher or fellow students, availability of feedback;
   2. autonomy: statements about the ability to choose work formats, individual pace, and independent research;
   3. predictability: references to a clear lesson structure, transparency of rules and assessment criteria.
2. Subjective indicators of educational well-being. This section included the following categories:
   
   1. emotional richness: loss of interest, inspiration, and passion;
   2. stressfulness: statements about anxiety, pressure, or fear of making mistakes, as well as indications of their absence;
   3. resilience (resourcefulness): descriptions of confidence, inner balance, the ability to cope with complex tasks.
3. Perceptions of interaction with artificial intelligence. This section included the following categories: expectations from AI – mentions of possible assistance, structuring of material; fears and anxieties – concerns about AI errors, dependence, or substitution of the teacher; trust in AI – statements about the reliability and acceptance of AI as an element of educational practice; functions of AI as a tool – utilitarian use of AI (information search, checking papers); functions of AI as a partner – dialogic interaction, joint analysis of ideas, support for reflection.
4. Characteristics of creative partnership (the “fourth format” of education). This section included the following categories: collaboration – perceptions of cooperation between the student, teacher, and AI; dialogicity – indications of an open exchange of meanings and discussion; co-creation – descriptions of the joint creation of new ideas and solutions, in which the student is an active co-author of the learning process.

The pilot testing of the methodology aimed to verify the conceptual consistency of the categorical system and its sensitivity to key indicators of educational well-being, and to assess the suitability of the methodology for further large-scale research.

In the “ideal lesson” essays, the most frequently coded themes were teacher (81.2%), students/peers (67.1%), assessment (50.6%), atmosphere/comfort (47.1%), interest/motivation (45.9%), active methods (35.3%), and support/respect (31.8%). Ideal lessons were described as safe, supportive, and structured environments, with clear explanations, voluntary participation, and assessment functioning as feedback rather than punishment. Collaborative pronouns (“we”) appeared in 15 essays; first-person singular pronouns (“I”) appeared in 30.

The essays about uncomfortable learning experiences emphasized the teacher as a source of threat (68.2%), stress/tension (29.4%), unpredictable assessment (29.4%), negative group atmosphere (30.6%), and aggression or humiliation (17.6%). Collaborative pronouns were nearly absent (“we” in 1 essay, “I” in 23), underscoring the disruption of joint engagement. Modal expressions contrasted permissive language in ideal lessons (“may,” “can”) with obligation or prohibition in uncomfortable lessons (“must,” “cannot”).

Research results

The results are presented for the main sample after data cleaning (N=63) and for the expanded primary analysis group of subjects (N=85) used in the psycholinguistic block.

The sample characteristics and data cleaning (N=63): the final analysis group included 63 respondents after field normalization and correction of obvious typos. The sample represented two universities: 49 respondents from Moscow Institute of Physics and Technology (MIPT) (77.8%) and 14 respondents from Moscow State Linguistic University (MSLU) (22.2%). Gender: 45 men (71.4%), 18 women (28.6%). Age: 19 years old – 5 (7.9%), 20 years old – 31 (49.2%), 21 years old – 23 (36.5%), 22 years old – 3 (4.8%), 24 years old – 1 (1.6%); M≈20.6, Me=20. After converting the field to numerical form, the study year data was comprised of: 3rd year – 28 (44.4%), 4th year – 33 (52.4%), 5th year – 1 (1.6%), and one value was left blank as it could not be interpreted unambiguously.

Formal characteristics of the texts (N=63): when using a free, expanded format (the guideline was 200–300 words), the actual essay length was significantly shorter and inconsistent between the positive and negative scenarios. The “ideal lesson” essays: average length 55 words, Me=50, min=1, max=211, P25=28, P75=68. The “uncomfortable lesson” essays: average length 37 words, Me=30, min=1, max=156, P25=13, P75=50. The proportion of extremely brief responses (<10 words) varied across corpora: four texts (6.3%) in the “ideal lesson” versus 12 texts (19.0%) in the “uncomfortable lesson,” including forms such as “a day-off”/”no such thing”/“the opposite of ideal,” which provide a “condensed” representation of negative experiences and the reduced observability of thematic markers.

The interuniversity comparison demonstrated differences in the productivity of utterances with comparable scale ratings: at MSLU, the average length of the “ideal lesson” essays was 76.9 words (Me=69), while that of the “uncomfortable lesson” essays was 69.9 (Me=66.5); at MIPT, the figures were 48.3 (Me=41) and 27.9 (Me=21), respectively.

Scale ratings for the “ideal lesson” (N=63): realism (1–5): 3 points – 2 people (3.2%), 4 points – 19 (30.2%), 5 points – 42 (66.7%); M=4.63, Me=5, the proportion of assessments of 4–5 points was 61/63=96.8%. Comfort (1–5): 2 points – 1 (1.6%), 3 points – 3 (4.8%), 4 points – 5 (7.9%), 5 points – 54 (85.7%); M=4.78, Me=5, the proportion of assessments of 4–5 was 59/63=93.7%. The configuration of the distributions corresponds to the ceiling effect (dominance of 4–5), limiting the variance and reducing the sensitivity of the scales to differences within the high-rating subsample. The interscale relationship is weak (r=0.14); Correlations of age with ratings were negative and small (realism r=–0.13; comfort r=–0.25). Gender means indicated a tendency for women to have higher comfort ratings (5.00 vs. 4.69) with similar realism (4.72 vs. 4.60), interpreted as a descriptive difference without the status of a stable effect.

The thematic structure of representations (N=63): in the “ideal lesson,” the core of the image was determined by the teacher-centered organization of the learning environment: the category “teacher/lecturer” appeared in 47 essays (74.6%), “students/groupmates/us” in 37 (58.7%), “assessment and scores” in 27 (42.9%), and “atmosphere/comfort” in 25 (39.7%). The motivational-dynamic contour is represented by “interest/motivation” (20; 31.7%) and “active methods” (19; 30.2%), including discussions, project-based learning, and group work. In the logic of the instrument, these fragments act as operational vehicles for dialogic transcription and potential creative partnerships. “Support/understanding/respect” was identified in 12 essays (19.0%), “stress/fear/tension” in 7 (11.1%), primarily through negation (“no stress”), “homework” in 7 (11.1%), and “online/remote format” in 2 (3.2%).

In the “uncomfortable lesson” essays, the teacher-centeredness was maintained (39 essays; 61.9%) with a redistribution of markers towards the threat: “stress/fear/tension” – 18 (28.6%), “atmosphere/comfort” in the negative modality - 16 (25.4%), “active methods” – 15 (23.8%) with a tendency to reflect on the imposed activity “at gunpoint”, “interest/motivation” in the form of boredom/lack of interest – 14 (22.2%). “Assessment and scores” were mentioned in 11 essays (17.5%), “pressure/ compulsion” – in 8 (12.7%), “boredom/monotony” – in 7 (11.1%), “aggression/insults/shouting” – in 4 (6.3%), “homework” – in 3 (4.8%). The resulting configuration describes not a “lesson deterioration,” but a shift in mode – from developmental interaction to social and evaluative vulnerability.

Psycholinguistic markers and sample differentiation (N=63 and N=85): in the final sample (N=63), the frequency of the thematic category “teacher” was 74.6% in the “ideal lesson” essays and 61.9% in the “uncomfortable lesson” essays, establishing the teacher-centered core of both representations.

In the expanded corpus of the primary analysis (N=85; MIPT – 60, MSLU – 25), the teacher more frequently acted as the central agent of the utterance: 81.2% in the “ideal lesson” corpus and 68.2% in the “uncomfortable lesson” corpus, establishing the causal pattern “teacher – emotional background/pace/feedback – opportunity to learn.” In the ideal scenarios, the characteristics of cognitive accessibility and dialogicity were consistently highlighted: “clarity of explanation” – 51.8%, “dialogue/questions” – 57.6%, “evaluation in the form of feedback” – 62.4%, “psychological safety/respect” – 35.3%; in the uncomfortable scenarios, the markers of social threat and the threat of assessment were: “stress/fear/tension” – 30.6%, “negative atmosphere in the study group” – 30.6%, “unpredictable control/threat of assessment” – 29.4%, “obscurity/haste/overload” – 25.9%, “pressure/humiliation/aggression” – 24.7%.

The pronominal frame in the extended corpus demonstrated an asymmetry in the grammar of compatibility: in the “ideal lesson” essays, the “I”-frame (“I/me/my”) was found in 30 essays, with 15 essays featuring the “we”-frame (“we/us/together”). In the “uncomfortable lesson” essays, the “we”-frame was reduced to a single instance, while the “I”-frame (23 essays) and the teacher-frame (58 essays) were preserved, forming an atomization of discourse (the teacher as a source of control; “I” as a vulnerable participant). Modal markers indicated a norm shift: in the “ideal lesson” texts, “may/can” was found in 30 texts, with “must not” in one text, while in the “uncomfortable lesson” essays, “may/can” was reduced to 5 texts, with “must not/cannot” increasing to 4; the obligation markers (“must/need to/have to”) were more often addressed to the teacher in the positive scenario (24 texts), this parameter weakening in the negative one (7 texts).

Comparative analysis of the subgroups of MIPT and MSLU (N=63): while the average scale ratings of the ideal lesson in the subgroups were similar (realism 4.57–4.65; comfort 4.78–4.79), the discursive styles differed, as shown in the extensiveness of the text and the thematic “optics.” In the “ideal lesson” texts, MSLU students more often coded parameters of emotions and attitudes: “atmosphere/comfort” – 85.7% (12/14) versus 26.5% (13/49) in MIPT students; “stress” (usually as “no stress”) – 35.7% (5/14) versus 4.1% (2/49); “support/understanding/respect” – 35.7% (5/14) versus 14.3% (7/49); “teacher” – 100% (14/14) versus 67.3% (33/49). In the “uncomfortable lesson” texts, the emotionally threatening aspects were similarly intensified in MSLU students: “stress/fear/tension” – 57.1% (8/14) versus 20.4% (10/49), negative “atmosphere” – 42.9% (6/14) versus 20.4% (10/49), “teacher” – 85.7% (12/14) versus 55.1% (27/49). The resulting configuration is consistent with the interpretation that distinguishes between the operational-structural style (MIPT: reduction to the parameters of the organization and the learning process) and the emotional-relational style (MSLU: explication of experience, boundaries and supportive relationships) while maintaining a single teacher-centered core.

Discussion and conclusions

The pilot data presents the “ideal lesson” as an environment of reduced social-evaluative threat that is constructed through permissive normativity: the teacher is predictable, questions are permitted, mistakes are legitimate, and feedback does not undermine self-esteem. This logic manifests itself simultaneously at the thematic and linguistic levels, linking high frequencies of “teacher” and “assessment and scores” (74.6% and 42.9% in N=63) with the predominance of “may/can” and the inclusion (albeit limited) of the “we”-framework in the extended corpus. In terms of psychological safety, such an environment describes a space where interpersonal risks (questions, mistakes, demonstrations of lack of knowledge) do not entail punishment, supporting learning behavior as an acceptable form of agency [29].

The opposition of “developmental and monitoring assessment” acts as the central mode switch. In the expanded corpus, “evaluation in the form of feedback” (62.4%) is part of the stable structure of an ideal lesson, while the uncomfortable scenario produces “unpredictable control/threat of assessment” (29.4%) formed along with stress and a negative atmosphere. At the level of pedagogical logic, this is consistent with the idea of formative assessment as a mechanism that increases learning gains with frequent and meaningful feedback, as found in the study by Black and Wiliam, and with that of punitive assessment as a factor supporting error avoidance and a self-defense strategy [30]. At the level of micromechanisms of feedback, there is a relevant concept of a “gap” between the current and target levels, which is closed by targeted information and is destroyed by shifting attention from solving the task to saving face [31].

The conceptual connection between the threat of assessment and motivational modes can be explained through the model of goal orientations and implicit theories: threatening assessment that makes a mistake a social loss, supports a pattern of helplessness/avoidance, whereas developmental assessment, which normalizes an error as information, supports skill-oriented activity. Carol S. Dweck and Ellen L. Leggett described these differences as consequences of varying goal-oriented behavioral organization and cognitive-affective response chains [32]. In the pilot data, this logic is “read” in the modal frame: the dominance of “may/can” in the “ideal lesson” essays (30 texts with 1 “must not”) versus the increase in prohibitive modality in the “uncomfortable lesson” essays (4 “must not/cannot” with 5 “may/can”), defining the narrative of the impermissibility of error and the impermissibility of “not knowing.”

The neuropedagogical interpretation proposed in the methodology translates this opposition into a resource model: an “uncomfortable lesson,” described through fear, haste, humiliation, and unexpected tests, triggers competition between learning and self-defense, redistributing attention to threat monitoring and response suppression. Physiologically, social-evaluative threat in laboratory models is associated with more pronounced cortisol reactivity when combined with appraisal and low controllability of the situation [33]. Neurocognitive reviews relate how acute, uncontrollable stress can weaken prefrontal cortex function and reduce the quality of voluntary regulation and working memory, that is, the mechanisms that underpin extensive learning activities and dialogic transcription [34]. This transition helps explain why the negative scenario is more often described as “dangerous” (30.6% stress; 29.4% threat of assessment; 24.7% pressure/humiliation) and is accompanied by an increased proportion of extremely short answers (19.0% in N=63): the discursive collapse becomes not just a stylistic feature, but a possible behavioral correlate of the threat mode.

The teacher-centered nature of both representations creates a methodologically significant tension with the stated framework of the “fourth format,” which presupposes creative partnership and reflected subjectness in the “student-teacher-AI” triad, supported by dialogicity and collaborative decision-making. Even with the direct inclusion of sections on AI and creative partnerships in the instructions, students construct their “ideal lessons” through the teacher’s figure (74.6% in N=63; 81.2% as the central agent in N=85), that is, through a locus of power that sets the boundaries of permissibility for error and asking questions. This discrepancy may be interpreted as the boundary for a fourth format’s implementation at the level of routine perceptions: reflected subjectness remains a “design norm,” while experienced safety remains a “launch condition,” still assigned to a human agent. This same logic explains why the digital component in the corpus gets sidelined: the technological characteristic “online/remote learning format” appears in only 3.2% of the “ideal” essays (N=63), giving way to the socio-evaluative architecture.

The interuniversity differences are revealed as differences in discursive norms, not differences in the “concept of the ideal.” With comparable scale scores, MSLU students demonstrate a more frequent and more detailed explication of the emotional-attitude layer (atmosphere 85.7% versus 26.5%; stress 35.7% versus 4.1%), as well as greater text yield (76.9 words versus 48.3 in the “ideal” essays; 69.9 versus 27.9 in the “uncomfortable” essays). This configuration is consistent with the hypothesis regarding the influence of educational culture and professional “linguistic tools”: the humanities subsample has more accessible means of verbalizing relationships and states, while the technical subsample more often “folds” experience into the parameters of structure, pace, and the learning process, without eliminating the demand for safety, but encoding it in other forms.

Attitudes toward AI, embedded in the categorical apparatus (trust; tool; partner; fears/skepticism), happen to be functionally close to the same regulators that govern the image of an ideal lesson: predictability, non-punitive feedback, and the preservation of voice and autonomy. Within this framework, trust in AI is interpreted as a desire for a “safe” assistant that supports dialog and reflection (“helps without judgment,” “can always clarify my questions”), and fears are interpreted as a projection of the experiences of the threat of assessment onto the digital agent, perceived as a source of uncertainty and potential control. This logic aligns with classical models of trust in automation, which describe trust as a regulator of reliability attribution and the choice of the degree of reliance on a system under uncertainty, as well as with contemporary research on trust in AI [35; 36]. In a study [37], the design task was formulated as one of providing “appropriate reliance.” In learning contexts, this means that creative partnerships with AI (that pertain to the fourth educational format) require both the tool’s functionality and conditions for calibrating trust (transparent boundaries, error control, and the preservation of autonomy in decision-making). Otherwise, the AI will be “inscribed” into a threat mode already familiar to the student [36].

In terms of motivation, the interpretation aligns with self-determination theory: the maintained autonomy, competence, and relatedness depend on a social context, which either supports or blocks internal regulation. Edward L. Deci and Richard M. Ryan described social-contextual conditions that enhance intrinsic motivation and well-being, which is methodologically consistent with the permissive normativity revealed in the data (“it’s okay to ask,” “one can make mistakes”) as a linguistic indicator of experienced autonomy within a structured environment [38]. Further support for the interpretation of the digital block is provided by empirical studies of students’ perceptions of generative AI, indicating a combination of instrumental usefulness and caution, requiring regulatory rules and pedagogical “safe use” design [39].

Limitations and Prospects

The limitations of the pilot study are determined by the ceiling effect of the scales (a concentration of ratings of 4–5 with Me=5), the asymmetry of the subgroups (49 at MIPT versus 14 at MSLU), and the high proportion of extremely brief responses in the negative scenario (19.0% with <10 words). This reduces the observability of thematic and psycholinguistic markers and requires a separate strategy for accounting for “minimal elaboration/refusal” as an independent indicator. The difference in the analysis bases (N=63 as the final sample and N=85 as the extended corpus) requires an explicit separation of the inference levels: thematic frequencies and scale ratings are interpreted for N=63, while psycholinguistic markers of agency, pronouns, and modality are interpreted for N=85, specifying that these are indicators from the primary corpus.

Research prospects include further refinement of the instrument, which involves: increasing the discriminativeness of the scales (expanding gradations or including points that distinguish between the “ideal” and the “acceptable norm”), clarifying the requirements for the minimum text length, balancing subsamples and preserving the conceptual core of the method – the grammar of compatibility and the modality framework – as a compact psycholinguistic interface for measuring psychological safety, threat of assessment and readiness for creative partnerships in the fourth format of education.

About the authors

Olga A. Ulyanina

Moscow State University of Psychology & Education; Moscow Institute of Physics and Technology (National Research University)

Email: ulyaninaoa@mgppu.ru
ORCID iD: 0000-0001-9300-4825

Doc. Psych. Sci., Associate Professor, Head of the Federal Coordination Center for the Development of Psychological and Pedagogical Assistance in the Education System of the Russian Federation, Chief Research Fellow of the Center for Applied Linguistic Research and Testing “ISTOK”

Russian Federation, 127051, Moscow, Sretenka Str., 29; 141701, Moscow region, Dolgoprudny, Institutsky Lane, 9, Bld 3

Ekaterina N. Vikhrova

Moscow Institute of Physics and Technology (National Research University)

Email: vikhrova.en@mipt.ru
ORCID iD: 0009-0006-9233-8894

Cand. Philol. Sci., Associate Professor at the Department of Foreign Languages

Russian Federation, 141701, Moscow region, Dolgoprudny, Institutsky Lane, 9, Bld 3

Oksana V. Marunevich

Moscow Institute of Physics and Technology (National Research University)

Author for correspondence.
Email: marunevich.ov@mipt.ru
ORCID iD: 0000-0002-4480-6642

Cand. Philol. Sci., Associate Professor at the Department of Foreign Languages

Russian Federation, 141701, Moscow region, Dolgoprudny, Institutsky Lane, 9, Bld 3

References

Supplementary files