Beyond Acceptance and Resistance: Digital Media Students’ Judgements of Generative AI Assistance and Perceived Creative Replacement

Abstract

This study investigated how Digital Media students judged acceptable and problematic forms of generative artificial intelligence (GenAI) involvement in creative learning. An embedded mixed-methods analysis examined 48 completed and consented baseline survey responses from an accessible population of 120 students at an Australian university. Forty-seven responses were valid for most Likert items, with qualitative content analysis of 37 benefit responses and 41 concern responses. Participants identified potential value in troubleshooting, coding support, grammar checking, information organisation, and error identification. They reported concern about AI-system training practices (93.6%), reduced employment opportunities (89.4%), the ethical acceptability of AI-supported coursework (83.0%), environmental effects (80.9%), and possible reductions in creative-skill development (76.6%). These percentages represent perceived risks, not demonstrated outcomes. Integration indicated that respondents differentiated bounded technical, corrective, and organisational assistance from GenAI involvement perceived to displace ideation, research, planning, understanding, or capability-building practice. Conditional legitimacy is proposed as a provisional framework linking delegated function, retained capability-building activity, learner agency and responsibility, and ethical and professional acceptability. The contribution is a student-informed account of the criteria underlying the assistance–perceived replacement distinction in Digital Media education.

1. Introduction

Generative artificial intelligence (GenAI) can perform explanation, coding, visual ideation, drafting and revision (Kasneci et al., 2023; Miao & Holmes, 2023). In educational contexts, these activities may also serve as evidence of student thinking and production. The educational issue is therefore not only access to assistance. A technically accomplished artefact may provide limited evidence of what a student understood, decided or produced when GenAI has performed part of the intellectual or creative activity through which capability is developed or demonstrated.

In this exploratory study, a group of Digital Media students at an Australian university were surveyed to examine their judgements of different forms of GenAI involvement in education and future professional practice. Quantitative and qualitative analysis identified patterns in how students distinguished GenAI uses and functions. The relatively small, self-selected sample positions the resulting framework as proposition-generating rather than population-representative.

Adoption and academic-integrity perspectives illuminate different parts of this problem but do not resolve it. Technology-adoption research commonly examines perceived usefulness and behavioural intention (Davis, 1989), while integrity policy addresses permission, disclosure and fair representation. Neither alone determines whether delegated activity remains aligned with the learning outcome. Analysis must therefore consider the function performed by GenAI, the capability-building activity retained by the learner and the evidentiary meaning of the resulting work.

The distinction matters because adoption, permission and legitimacy operate at different analytical levels. Adoption describes willingness, intention, or reported behaviour. Permission concerns whether a practice complies with institutional rules. Legitimacy concerns whether the practice can be justified in relation to the purposes of learning, the claims made by an assessment, and the interests of people affected by the technology. A student may therefore regard a function as convenient without considering it educationally appropriate, or may know that a function is permitted while objecting to its effects on authorship, employment or environmental responsibility. Treating these positions as a single continuum obscures the reasoning through which students evaluate particular forms of GenAI involvement. This study consequently treats student judgement as more than an attitude towards a technology. It is examined as an appraisal of what is delegated, what the learner continues to do, and whether the resulting arrangement remains defensible within a specific educational task.

This framing also addresses a limitation in broad acceptance measures. Aggregate acceptance scores can conceal internally coherent distinctions between functions, with prior studies showing that acceptability varies by use and assessment context (Johnston et al., 2024; Kizilcec et al., 2024). A student might welcome assistance with grammar, debugging, or the organisation of material while rejecting generated concepts, images, or finished code. Such a response is not necessarily inconsistency or technological resistance. It may reflect a judgement that the functions contribute differently to the learning purpose. Research reporting only general enthusiasm or concern may therefore miss the object of the judgement. The relevant object is often not GenAI as a whole, but a particular delegated function situated within a task, assessment, or professional practice. Function-specific analysis is especially important when the same system can support low-level correction, offer feedback, propose alternatives, and generate substantial components of a submission. Each function changes the distribution of cognitive and creative work in a different way.

These questions are consequential in the Digital Media context examined here, where participating units combine technical production with research, ideation, aesthetic judgement, problem-solving, iteration, communication and reflection. The significance of assistance changes with the task, so tool-level labels such as “AI allowed” and “AI prohibited” are too coarse to define educationally meaningful boundaries.

Research across higher education reports benefits in explanation, feedback, writing and coding alongside concern about accuracy, overreliance, integrity, privacy and independent thinking (Chan & Hu, 2023; Cotton et al., 2024; Farrokhnia et al., 2024; Gruenhagen et al., 2024; Johnston et al., 2024; Kim et al., 2025; McGrath et al., 2025). Acceptability also varies by function and assessment type (Johnston et al., 2024; Kizilcec et al., 2024). Positive and negative judgements may therefore coexist without representing indecision: students can recognise practical value while rejecting particular forms of delegation.

Creative and design research sharpens this issue. Graphic design and architectural education studies report pragmatic acceptance alongside concern about authorship, professional change, and displacement of creative processes (Fleischmann, 2024, 2025; Huh et al., 2025). Experimental findings are similarly conditional. ChatGPT-supported divergent-thinking activities have improved some measured creativity dimensions, whereas exposure to generated images has increased design fixation and reduced idea variety under other conditions (Habib et al., 2024; Wadinambiarachchi et al., 2024). The relevant boundary is not a fixed division between human and machine production, but the alignment among learning outcome, delegated function, and retained judgement.

Creative and design education places value on both the resulting artefact and the processes through which ideas and decisions are developed (Fleischmann, 2024, 2025; Habib et al., 2025). A polished artefact is not the sole educational aim. Students are expected to learn how to interpret a brief, investigate a context, generate and discard possibilities, justify aesthetic and technical decisions, respond to critique and revise work through informed iteration. These activities develop a capacity to act when a future problem differs from the one previously encountered. If a generated artefact conceals how decisions were reached, the product may provide weak evidence of the learner’s capability even when it appears professionally convincing. Conversely, GenAI involvement may deepen learning when students must diagnose defects, compare alternatives, interrogate assumptions or explain why a generated response should be rejected. The educational issue is therefore not reducible to the amount of automation. It concerns whether the arrangement preserves meaningful opportunities for the learner to exercise the capability that the activity is intended to develop.

The literature also suggests that creative consequences are not uniformly positive or negative. Generated examples may support exploration and expose students to alternatives (Chen et al., 2025; Habib et al., 2024), yet they may also anchor early ideas and narrow divergence (Wadinambiarachchi et al., 2024). These effects depend on task design, learner expertise, and the way generated material is handled. A requirement to compare, modify and justify alternatives is educationally different from accepting the first plausible output. This variability supports a relational approach in which the value of GenAI is considered through the interaction between function, learner activity, task purpose, and consequence. It also cautions against describing all involvement as either enhancement or replacement before examining what students actually retain responsibility for doing.

Learner agency, evaluative judgement and longer-term learning provide a theoretical basis for examining that boundary (Boud & Falchikov, 2006; Molenaar, 2022). Agency concerns intentional action, consequential decision-making, and responsibility rather than manual completion of every step. Evaluative judgement concerns the capacity to judge the quality of one’s own and others’ work (Tai et al., 2018). Cognitive offloading is not inherently harmful (Risko & Gilbert, 2016). Applied to learning, the educational risk may be greater when the offloaded activity is itself the target capability. Current Australian guidance similarly emphasises evaluative judgement, critical thinking and ethical reasoning in GenAI-integrated higher education (Lodge et al., 2026).

Productive GenAI involvement often depends on the ability to recognise quality rather than merely to produce text, code, or images. Applied to GenAI, evaluative judgement includes identifying factual or technical errors, recognising superficial coherence, testing whether an output meets a brief, comparing alternatives and deciding what requires revision. These activities presuppose standards, disciplinary knowledge and a willingness to remain accountable for the final work. Novices may therefore face a paradox: the students most attracted to rapid assistance may have the least-developed basis for verifying it. The present study does not measure evaluative judgement, but it examines whether students recognise the importance of retained judgement and whether they distinguish assistance that invites evaluation from generation that appears to bypass it.

Sustainability is used here only as an interpretive lens. The study did not measure sustainable outcomes. It examined student concerns relevant to durable learning, ethical practice, professional responsibility and environmental effects. These considerations should not be collapsed into one score because an educationally helpful practice may remain ethically contested or professionally common while creating privacy risks. This bounded framing aligns with human-centred guidance that prioritises meaningful and responsible integration rather than adoption for its own sake (Miao & Holmes, 2023).

Existing research has established mixed student evaluations of GenAI and variation in acceptability across functions and assessment contexts (Chan & Hu, 2023; Johnston et al., 2024; Kizilcec et al., 2024). The narrower gap concerns how Digital Media students connect particular functions with capability development, retained responsibility, and wider ethical or professional acceptability across technical and creative contexts. This study addresses that gap through the provisional concept of conditional legitimacy.

Three unresolved issues follow from this gap. First, it remains unclear which criteria students apply when they move beyond broad approval or disapproval. Second, benefits and concerns are frequently reported as parallel themes (Chan & Hu, 2023; Johnston et al., 2024; McGrath et al., 2025), leaving scope to examine how students reconcile them within particular educational tasks. Third, ethical and environmental concerns are often reported alongside educational attitudes, but their role in judgements of educational legitimacy remains less developed. Task-sensitive assessment approaches already emphasise aligning permitted GenAI use with learning outcomes (Perkins et al., 2024; Xia et al., 2024). Building on this, a policy that lists permitted tools cannot by itself explain why one delegation supports learning while another compromises the capability claim attached to an assessment. A more useful account must connect empirical judgements to the distribution of work, responsibility and consequence within the learning activity.

The present study addresses these issues through the concept of conditional legitimacy. It proposes that students’ judgements become intelligible when four questions are considered together: what function is delegated, what capability-building activity remains with the learner, where judgement and responsibility reside, and whether broader ethical or professional conditions are acceptable. The framework was developed through theory-informed interpretation of convergence and tension across the quantitative and qualitative findings.

The analysis also distinguishes value-based opposition from unfamiliarity or low confidence. The term is reserved for responses that state an educational, moral, environmental, or professional objection.

Aim and Research Questions

This study aimed to explain how Digital Media students evaluated different GenAI functions in creative learning and how their judgements connected educational purpose with agency, responsibility, and ethical acceptability. The study examines baseline judgements before planned teaching activities; the post-unit phase will address whether those judgements change with experience.

• RQ1. How do Digital Media students evaluate the educational role of different GenAI functions in creative learning and practice?
• RQ2. What educational, ethical, environmental, and professional considerations are evident in students’ judgements of GenAI legitimacy?
• RQ3. How do students distinguish between GenAI assistance that supports capability development and GenAI involvement perceived to replace creative or intellectual work?

2. Materials and Methods

2.1. Research Design and Context

This article reports an embedded mixed-methods analysis of the pre-unit phase of a broader pre–post research project. The approved project invited students to complete an anonymous survey during the opening weeks of the teaching term and a later survey after unit grades were released. The present article uses only the baseline survey and should therefore be understood as cross-sectional. It does not examine individual change, intervention effects or post-unit outcomes.

The survey combined Likert-scale items with two optional open-ended questions. Quantitative data provided the main descriptive structure, while qualitative responses were used to explain, qualify, and complicate the numerical patterns. This design was selected because the research problem concerned both the distribution of attitudes and the reasoning attached to those attitudes. The qualitative component was embedded rather than treated as an equal standalone strand because the open responses were brief and collected within the same instrument.

This design does not imply that qualitative comments validate the percentages or that numerical prevalence validates an interpretation. The two forms of evidence answer different questions. Quantitative distributions show how respondents selected among pre-specified positions. Qualitative comments show how some respondents articulated the meaning of those positions. Convergence strengthens confidence that both forms of evidence refer to a similar issue, complementarity adds explanatory detail, and divergence identifies where a pre-specified item may not capture the full reasoning. The mixed-methods contribution lies in these relationships rather than in the simple presence of numbers and quotations (Fetters et al., 2013).

The design also preserves asymmetry between breadth and depth. Every retained respondent could contribute to the item-level distributions, whereas only respondents who chose to answer an open question contributed qualitative material. Qualitative categories therefore cannot be read as population frequencies. Their role is explanatory and conceptual. This distinction is important because a concise concern expressed by a minority may still expose a limitation in an aggregate interpretation, while a common category may remain too briefly described to support a strong theoretical claim. The analysis consequently considered prevalence, specificity, and conceptual relevance separately.

The study was situated within Digital Media programs at an Australian university. Participating units included introductory coding, visual design and animation contexts, an intermediate mobile application development context, and an advanced collaborative project context. These units involve different combinations of technical production, creative decision-making and professional responsibility. This range provides disciplinary breadth, but it limits simple developmental comparisons because differences may reflect unit purpose rather than student progression.

The disciplinary setting shaped the interpretation without being treated as homogeneous. Introductory coding emphasises syntax, debugging and problem decomposition. Visual design and animation place greater emphasis on ideation, composition, style and iterative production. Mobile application development combines interface decisions, code, testing and integration. Client-oriented project work adds communication, accountability and the management of external expectations. These contexts differ, but each requires students to connect decisions with an artefact and to explain or revise those decisions. This shared concern with traceable capability made the program a useful setting for examining the perceived boundary between assistance and replacement. The study does not assume that every respondent encountered identical tasks or held the same conception of creativity.

2.2. Participants and Sample

The accessible population comprised 120 active students enrolled across the Bachelor of Digital Media (n = 101), Diploma of Digital Media (n = 13), and Associate Degree of Digital Media (n = 6). All 120 active students were invited to participate. The Qualtrics dashboard recorded 68 submissions. Seven were incomplete submissions that were not retained in the exported file, leaving 61 exported response records. Two survey-preview records and 11 incomplete non-preview records were excluded. Forty-eight completed, consented responses remained, producing a completed-response rate of 40.0%. One completed respondent provided no substantive answers to most study items, leaving 47 valid responses for most Likert analyses, equivalent to 39.2% of the accessible population.

Among the 47 respondents with substantive data, 33 identified the Bachelor of Digital Media, five the Associate Degree of Digital Media, and three the Diploma of Digital Media. Six selected “Other” when reporting their award but belonged to the eligible Digital Media population. Participation was voluntary and self-selected; students with particularly strong positions may have been more likely to participate. The sample is therefore interpreted as exploratory and context-specific rather than statistically representative.

Unit level was used only to describe educational context. It did not verify each participant’s chronological year of enrolment, prior study or sequence through the program. Introductory, intermediate and advanced labels therefore refer to the position of the unit in the program structure rather than to confirmed student year. Because subgroup sizes were small and uneven, unit-level comparisons were descriptive and secondary. Table 1 summarises the response screening and analytical sample.

Table 1. Response screening and analytical sample.

Stage		Number
Submissions recorded on Qualtrics dashboard		68
Response records present in supplied CSV export		61
Incomplete dashboard submissions not retained in export		7
Survey preview records excluded		2
Incomplete non-preview records excluded		11
Completed, consented, non-preview responses		48
Valid responses for most Likert items		47
Characteristic	Category	n
Award	Bachelor of Digital Media	33
Award	Associate Degree of Digital Media	5
Award	Diploma of Digital Media	3
Award	Other	6
Open-ended response	Expected benefits	37
Open-ended response	Concerns or uncertainties	41
Accessible population		120
Completed-response rate (48/120)		40.0%
Valid-response rate for most Likert items (47/120)		39.2%

2.3. Recruitment, Consent and Ethics

Students were invited through the approved program communication process to complete an anonymous online survey during Weeks 1 to 3 of the teaching term. The participant information explained the study purpose, the voluntary nature of participation, the approximate completion time, and the intended use of findings. Consent was recorded electronically before the survey began. Students could skip optional questions or leave the survey before submission.

The research team included educators who taught within the program. The participant information explicitly recognised that this relationship could create perceived pressure. The approved safeguards were anonymity, voluntary participation, and the assurance that teaching staff would not know who had or had not participated. Participation or non-participation had no effect on grades, feedback, enrolment, progression, university services or relationships with teaching staff. No names, student numbers or direct identifiers were collected. Because submitted responses were anonymous, they could not be withdrawn after submission.

The project was approved by the relevant university Human Research Ethics Committee. Institutional details and the approval reference are withheld for anonymous peer review and will be restored following acceptance.

2.4. Survey Instrument

The baseline survey examined: prior GenAI experience; AI-assisted ideation; prompting; anticipated professional application; confidence; refinement of generated outputs; integration with personal work; creative voice; and ethical practice. Further items addressed perceived effects on creative-skill development, misconduct accusations, uncertainty about permitted coursework practices, employment, personal values, AI-system training data and practices, and environmental effects.

Agreement items employed a five-point scale from Strongly disagree to Strongly agree. Two optional open-ended questions invited expected benefits and concerns or uncertainties. These prompts enabled participants to explain function-specific judgements without requiring general endorsement or rejection of GenAI.

The researcher-developed instrument was aligned with the research questions and domains identified in the literature, but it was not psychometrically validated. Many conceptually distinct domains were represented by single items. Results were therefore analysed at item level; no composite measure of acceptance, sustainability or conditional legitimacy was constructed. An internal-consistency coefficient across the heterogeneous item set would not provide meaningful reliability evidence. The concentration of negatively framed concern items is acknowledged as a measurement limitation.

Content alignment was informed by prior higher-education and creative-education research concerning domains such as confidence, ideation, prompting, refinement, creative voice, integrity, and perceived implications of GenAI use (Fleischmann, 2024; Johnston et al., 2024; Kizilcec et al., 2024). This literature alignment does not constitute psychometric or content validation. The instrument was designed for exploratory item-level mapping rather than measurement of a single latent disposition.

The two open-ended prompts broadened this structure by asking respondents to describe expected benefits and concerns in their own words. Their optional format reduced pressure to disclose a view but produced uneven depth. Some responses consisted of a short phrase, while others connected several considerations. The analysis therefore treated frequency cautiously. A frequently coded category indicates that an issue appeared across multiple responses, but it does not measure intensity, priority, or causal importance. Likewise, a detailed quotation illustrates a line of reasoning but is not presented as representative of every participant.

2.5. Quantitative Analysis

The two Qualtrics metadata rows were removed before screening the 61 exported response records. Survey-preview records and incomplete submissions were excluded from the analytical sample; all 48 retained complete records recorded affirmative consent. The seven additional submissions shown on the Qualtrics dashboard were incomplete and were not retained in the export. Frequencies and percentages were calculated using the valid denominator for each item. Table A1 reports the full five-category distributions and item-specific denominators. For concise narrative reporting, Somewhat agree and Strongly agree were combined as agreement, while Somewhat disagree and Strongly disagree were combined as disagreement. Neutral responses were reported where they were substantively important rather than treated as missing or indecisive.

The analysis was descriptive. No inferential tests were used because the total sample was modest and subgroup sizes were small and unequal. This decision also reflected the exploratory nature of the unit-context comparisons and the absence of a probability sample. The purpose of the quantitative analysis was to identify the distribution and relative prominence of reported orientations, not to estimate population parameters or establish causal relationships.

2.6. Qualitative Content Analysis

Qualitative content analysis was conducted through repeated familiarisation, segmentation into meaning units, initial coding, category development and interpretive comparison (Elo & Kyngäs, 2008). Initial codes stayed close to participants’ wording. Related codes were grouped into descriptive categories using a codebook containing definitions, inclusion and exclusion criteria, and representative extracts. Responses could receive more than one code, and positive, negative, mixed, and contrary cases were retained.

Coding decisions were checked through iterative recoding and comparison against the codebook and full response set. No second independent coder was available, so intercoder reliability was not calculated. The codebook and representative extracts provide an audit trail, not evidence of objectively fixed categories. Saturation was not claimed because the responses were brief and could not be probed.

The coding process involved movement between individual meaning units, provisional categories, and the full response context. Categories were revised when an initial label combined conceptually different actions. For example, grammar correction and troubleshooting were initially grouped as general assistance, but were separated during coding because one concerned corrective language support and the other diagnosis of a technical problem. At the interpretive stage, both contributed to the broader category of bounded assistance because respondents described them as supporting rather than assuming the main task. In contrast, concerns about generated ideas, reduced thinking and loss of creative development were connected through their emphasis on displaced capability-building activity. This distinction between descriptive coding and higher-order interpretation reduced the risk of coding every favourable comment as assistance or every concern as replacement.

Reflexivity was important because the primary analyst was also an educator within the program and had prior interests in AI, learning, and assessment. This position provided contextual knowledge but could predispose the analysis towards educational explanations. Several safeguards were used within the constraints of a single-coder analysis: the quantitative and qualitative analyses were undertaken separately before integration; negative and ambiguous cases were retained; descriptive categories were distinguished from the conceptual framework; and claims were repeatedly checked against the complete response set rather than selected quotations. These practices improve transparency but do not replace independent coding or participant validation. Contrary and ambiguous responses were examined rather than absorbed into the prevailing pattern. For example, one respondent reported no concerns, while another opposed unrestricted use but explicitly rejected a complete ban because carefully specified support could remain useful. A benefit response also welcomed feedback and ideas while cautioning against reliance. These cases did not remove the assistance-replacement distinction; they required it to remain conditional and prevented opposition from being interpreted as a uniform rejection of every GenAI function.

Descriptive codes were kept separate from higher-order interpretation. For example, locating a coding error was coded as error identification; its interpretation as bounded assistance occurred during integration. Comments about dependency or reduced thinking were treated as concerns rather than evidence of capability decline. Conditional legitimacy was introduced only after the descriptive categories had been established.

2.7. Mixed-Methods Integration and Reflexivity

Integration occurred during interpretation and reporting. Item-level distributions were compared with qualitative categories concerning related functions or issues, and a joint display records convergence, complementarity, and evidentiary limits. Because the strands used different response formats and denominators, open responses were used to explain and qualify numerical patterns rather than to provide a direct comparative measure of acceptance. They were not treated as representative of all respondents or as validation of the quantitative strand. Researcher teaching roles and prior interest in AI education were addressed through separation of descriptive coding from interpretive claims, retention of contrary cases, and explicit qualification of the proposed framework.

The integration procedure followed three steps. First, each quantitative item and qualitative category was mapped to the research question it could inform. Second, related evidence was compared for convergence, complementarity, or tension. For example, opposition to generated ideation was considered alongside comments about creative development, while confidence in refinement was compared with descriptions of debugging and improvement. Third, the emerging explanations were tested against disconfirming patterns. Acceptance of some forms of assistance prevented the interpretation of concern as general resistance, while great concern about misconduct accusations alongside lower uncertainty about rules prevented policy knowledge from being treated as equivalent to institutional trust. The joint display in Section 3.8 presents the final integrated claims and the limits placed on each claim.

This procedure also established the hierarchy of findings. The assistance-replacement distinction was treated as the principal interpretation because related patterns appeared across quantitative items, benefit categories, and concern categories and directly informed RQ3. Ethical, professional, and environmental considerations were treated as conditions relevant to legitimacy rather than as demonstrated determinants of particular decisions. Prior experience and confidence were retained as secondary contextual findings because the cross-sectional design could not determine whether they influenced the reported judgements.

3. Results

The Results are organised to make their relationship with the research questions explicit. Section 3.2, Section 3.3 and Section 3.5 primarily answer RQ1. Section 3.4 and Section 3.7 answer RQ2. Section 3.6 answers RQ3. Section 3.8 integrates the strands and distinguishes the principal finding of assistance versus replacement from secondary contextual findings concerning prior experience, confidence, and unit context.

3.1. Sample Profile and Prior Experience

Of the 47 respondents with substantive data, 20 (42.6%) reported no previous GenAI experience, 26 (55.3%) reported occasional experience, and one (2.1%) reported engagement most of the time. Among 39 respondents who self-rated their expertise, 20 (51.3%) identified as beginners, 16 (41.0%) as intermediate, and three (7.7%) as advanced. These self-ratings were not performance measures.

Prior experience provides context but does not explain the subsequent judgements. The design cannot determine whether unfamiliarity produced low confidence or whether educational and ethical objections would persist after further experience.

3.2. RQ1 General Orientation Towards GenAI Functions

Opposition was clearest for AI-generated ideation: 33 respondents (70.2%) rejected it as a legitimate part of creative practice, compared with eight (17.0%) who agreed and six (12.8%) who were neutral. This pattern did not extend uniformly to all functions, as the open responses later identified bounded forms of assistance.

Most respondents also resisted equating prompting with established technical skills: 28 (59.6%) disagreed, 15 (31.9%) agreed, and four (8.5%) were neutral. Anticipated professional expectations produced no consensus: 21 (44.7%) disagreed that regular GenAI involvement would be expected, while 13 (27.7%) agreed and 13 were neutral.

3.3. RQ1 Confidence, Integration and Creative Voice

Confidence was limited and uneven. Twenty-two respondents (46.8%) lacked confidence in creative tasks, but 21 (44.7%) believed they could improve generated material. Integration with personal work and maintenance of creative voice produced mixed distributions, providing no sample-level consensus. These self-reports do not demonstrate evaluative competence.

A majority, 25 respondents (53.2%), believed they understood ethical and responsible practice; nine (19.1%) disagreed, and 13 (27.7%) were neutral. This indicates perceived understanding rather than verified ethical judgement.

3.4. RQ2 Educational, Ethical, Environmental and Professional Considerations

Table 2 presents the complete concern profile. The most prominent concerns involved AI-system training data and practices (93.6%), employment opportunities (89.4%), and the ethical acceptability of AI-supported coursework (83.0%). Environmental effects (80.9%) and possible reductions in creative-skill development (76.6%) were also salient. These distributions establish perceived importance, not the occurrence or magnitude of the anticipated harms.

Table 2. Student concerns about generative AI involvement in education and creative work.

Concern	Agree, n	Agree, %
How AI systems are trained	44	93.6
Reduced employment opportunities	42	89.4
AI-supported coursework feels ethically questionable	39	83.0
Environmental effects	38	80.9
Reduced development of creative skills	36	76.6
AI use conflicts with personal values	35	74.5
Misconduct accusation	24	51.1
Uncertainty about acceptable coursework use	11	23.4

Misconduct accusation and policy uncertainty showed different response patterns. Although 51.1% feared being accused of misconduct, only 23.4% were uncertain about what was permitted in coursework. This difference raises, but does not test, whether procedural anxiety involves institutional trust as well as rule comprehension.

3.5. RQ1 Expected Benefits and Accepted Functions

Thirty-seven respondents answered the expected-benefits question. Fourteen stated that GenAI offered no benefit or expressed value-based opposition. Among stated benefits, efficiency or workload reduction appeared in nine responses, exploratory dialogue or ideation support in eight, bounded technical assistance in six, and corrective or organisational assistance in six. Because responses could receive more than one code, these counts overlap and should not be summed. The technical and corrective categories included troubleshooting, error identification, grammar, clarification, and information structuring. Exploratory comments generally positioned GenAI as an aid within an existing decision process rather than as the source of final creative direction:

“The only time I have used ai and thought it was helpful was when finding errors in my coding”.

(P31)

These categories shared a common boundary: assistance was valued when it reduced friction while leaving the main problem, interpretation, or decision with the student. Not all respondents identified benefits, and the responses included restriction, uncertainty, and rejection as well as acceptance:

“I use ai like a secretary, to do grammar check and online research. Sometimes I use it like a consultant to debate my ideas, helps me to find weakness and find dirrection”.

(P26)

3.6. RQ3 Assistance, Replacement and Capability Development

Among the 41 concerned responses, perceived capability displacement appeared in 20, value-based opposition in 14, institutional trust or procedural fairness in 13, ethical or environmental concern in 12, and professional or client responsibility in 12. Multiple coding was permitted, so the category counts overlap. Concerns about replacement focused on thinking, practice, and understanding. Respondents distinguished a successful output from the ability to explain, evaluate or reproduce the work independently. Creative comments similarly treated ideation, experimentation and mistakes as activities through which individual voice and disciplinary judgement develop.

Dependency was framed as a prospective risk: repeated delegation might weaken independent task initiation, persistence, or recognition of poor output. Other responses accepted feedback and problem identification even though these functions also reduced effort. The operative distinction was therefore not effort versus ease, but continued student decision-making versus perceived displacement of capability-building activity:

“I don’t like using it for creative progress because parts of creative progress are thinking and I think thinking is a very important skill…”.

(P31)

3.7. RQ2 Ethical Legitimacy, Professional Responsibility and Institutional Trust

Ethical concerns addressed creators’ work in training datasets, consent, ownership, and imitation of style. Privacy and professional-responsibility comments addressed unpublished client material, commercial ideas, misinformation, and accountability for unverified outputs.

Employment concerns anticipated fewer entry-level opportunities and devaluation of human creative work. Institutional trust comments focused on false detection and on encouragement of technologies that some students considered ethically problematic. These perceptions do not verify legal, environmental, or labour-market effects; they show that legitimacy judgements extended beyond classroom efficiency.

Some respondents expressed value-based opposition grounded in educational, moral, environmental, or professional reasons. This category excludes unexplained dislike and does not imply a permanent identity:

“I will not be using AI in this unit. Frankly, when I see units that use or encourage AI—I trust them less”.

(P23)

“Learning to use AI takes away from learning the actual skills I need for my future career”.

(P29)

3.8. Integrated Findings and Principal Contribution

Across RQ1–RQ3, the principal integrated interpretation was a distinction between bounded assistance and perceived replacement. Quantitative opposition to generated ideation and concern about creative-skill development were considered alongside comments about thinking, experimentation, and voice. Positive comments concentrated on troubleshooting, correction, and organisation. Table 3 displays these cross-strand relationships without treating the different measures as a direct comparison of acceptance. Prior experience, confidence and unit context remain secondary because their influence was not tested.

Table 3. Joint display integrating quantitative patterns, qualitative explanations and cautious interpretations.

Quantitative Pattern	Qualitative Explanation	Cautious Interpretation
70.2% rejected AI-generated ideation as legitimate	Benefits centred mainly on troubleshooting, errors, grammar and organisation	Participants differentiated selected support functions from creative generation
76.6% feared reduced creative skill development	Thinking, experimentation and practice were described as part of learning	Some uses were judged according to whether capability-building activity remained with the student
93.6% expressed concern about training practices	Comments referred to ownership, consent and creators’ rights	Immediate educational usefulness did not settle ethical acceptability
51.1% feared misconduct accusation	Some comments raised false-flag and procedural concerns	Rule clarity and trust in integrity processes may be distinct issues

Two additional integrations qualified this central pattern. First, misconduct anxiety exceeded uncertainty about permitted practice, raising a question about whether procedural trust and policy knowledge are distinct. Second, ethical concerns remained salient alongside reported uncertainty about permitted practice, but the study did not test the relationship between institutional permission and ethical acceptability.

These patterns informed the provisional framework of conditional legitimacy elaborated in Section 4.2. The framework organises the dataset-level interpretation but does not constitute a validated scale or imply that every participant applied an identical decision rule.

4. Discussion

4.1. Assistance Versus Replacement as the Central Finding

The assistance-replacement distinction offers an interpretation of how practical benefit and principled objection coexisted at the sample level. Respondents did not simply occupy positions on a single acceptance continuum; their reported judgements varied according to what GenAI performed and what remained for the learner to understand, decide, and demonstrate. This extends prior evidence of function-specific acceptability (Johnston et al., 2024; Kim et al., 2025) by connecting the distinction with capability development, responsibility, and ethical acceptability in Digital Media education.

This interpretation differs from a binary classification of acceptable and unacceptable tools. Assistance and replacement describe perceived educational roles, not fixed properties of particular applications. A chatbot that supplies a complete concept may be experienced as replacement in an ideation task, while the same system may function as assistance when it generates deliberately imperfect alternatives for students to critique. Similarly, code generation may displace foundational practice when accepted without understanding, yet support advanced learning when students must test, refactor, and justify the code within a larger system. The boundary therefore depends on the relation among tool function, task purpose, and retained learner activity. This relational account is more precise than categorising entire technologies as supportive or harmful.

The finding also clarifies why efficiency cannot serve as the sole criterion for educational value. Respondents acknowledged that GenAI could make work faster, correct errors, or organise material, but several objected when efficiency removed the thinking or practice they believed the course should develop. In an educational context, productive difficulty may be part of the outcome rather than an avoidable cost. The relevant question is not whether GenAI saves effort, but whether the saved effort is peripheral to the intended capability or constitutes the activity through which that capability is formed. This distinction can explain why apparently small forms of correction were viewed more favourably than ideation despite the latter’s potential to accelerate production.

“Ambivalence” would understate this pattern because many positions were differentiated rather than indecisive. The contribution lies not in showing that students distinguish among functions, but in explaining the criteria that made those distinctions intelligible across the dataset.

4.2. Conditional Legitimacy as a Theory-Informed Framework Grounded in the Evidence

Table 4 formalises conditional legitimacy as a provisional, theory-informed framework grounded in the integrated evidence. The framework evaluates a particular function within a particular task; it does not classify a tool as inherently legitimate or illegitimate. Its four dimensions should be read together rather than combined into a single score.

Table 4. Theory-informed dimensions of conditional legitimacy grounded in the integrated evidence.

Dimension	Empirical Basis	Interpretive Boundary
Delegated function	Different responses to troubleshooting, grammar assistance, ideation, and generated creative work	Legitimacy concerns what AI performs, not the tool label alone.
Retained capability-building activity	Concern about reduced thinking, practice, understanding, and creative-skill development	Perceived risk may be greater when delegated activity is central to the intended capability.
Agency and responsibility	Comments about verification, decision-making, creative voice and accountability	Manual production is not required in every case, but consequential judgement must remain attributable to the learner.
Ethical and professional acceptability	Training-data, privacy, environmental, employment and client-responsibility concerns	Institutional permission does not automatically establish ethical legitimacy.

The first dimension, delegated function, identifies the operation transferred to GenAI. This avoids treating technology as a unitary intervention. The second, retained capability-building activity, asks whether students still practise and demonstrate the knowledge, judgement, or creative process that the task is intended to develop. The third, agency and responsibility, examines who frames the problem, evaluates alternatives, makes consequential decisions, and remains answerable for the result. The fourth, ethical and professional acceptability, extends the judgement beyond immediate task performance to concerns such as provenance, privacy, labour, environmental effects, and professional obligations. The framework proposes that conditional legitimacy is strongest when these dimensions align. A function may be efficient but remain educationally weak if it removes the central practice, or educationally promising but ethically contested if its inputs or consequences are unacceptable.

The dimensions are analytically distinct and conceptually connected through patterns in the dataset. Delegating a function may change what capability can be observed and where judgement and responsibility reside. Ethical and professional conditions may then affect whether an arrangement is considered defensible even if learning remains possible. These relationships were not directly measured, so the framework should not be converted into a checklist in which four positive answers automatically establish legitimacy. It is a structure for reasoned evaluation rather than a tested decision model.

The empirical status of the framework also requires care. The four dimensions were developed through theory-informed interpretation of a modest, context-specific dataset through a single-coder analysis. They are grounded in patterned relationships in the evidence, but they have not been validated as a measurement model or tested across disciplines. Future work should examine whether students and educators recognise the same dimensions, whether additional conditions emerge, and whether the framework helps explain judgements across task types. It may also be useful to compare stated legitimacy with actual decisions when students encounter specific assessment scenarios.

The boundary between assistance and replacement is task-dependent. Central GenAI involvement can support learning when generated material becomes an object of comparison, critique, revision and justification. Conversely, a seemingly minor contribution may displace the reasoning a novice task was designed to develop. The framework therefore directs attention to alignment among delegated function, learning outcome, learner expertise, and the judgement still required.

This interpretation accommodates contrasting evidence from creative education: structured GenAI engagement can support achievement or creative exploration, while generated examples can also constrain originality (Chen et al., 2025; Fleischmann, 2024, 2025; Huh et al., 2025; Wadinambiarachchi et al., 2024). The framework does not resolve that empirical question; it identifies the conditions future studies should examine.

Learner expertise may be an important boundary condition because verification presupposes disciplinary knowledge. The present study did not test verification performance, so expertise remains a hypothesis for future work rather than an explanatory finding.

4.3. Agency, Evaluative Judgement and Responsibility

Agency concerns retained problem framing, evaluation, consequential decision-making and accountability, not manual completion of every production step. Evaluative judgement is implicated when students compare outputs, identify weaknesses and justify choices (Tai et al., 2018). The findings suggest that efficiency becomes educationally problematic when it removes those decisions or weakens evidence that the learner can make them.

Agency can be diminished even when a student remains visibly active. Repeated prompting, selecting, or formatting may involve substantial interaction without requiring the learner to establish standards, diagnose weaknesses, or take responsibility for the direction of the work. Conversely, a learner may delegate a technically demanding operation while retaining strong agency through purposeful specification, critical testing, and justified revision. Observable activity is therefore an inadequate proxy for agency. What matters is whether the student exercises consequential judgement and can account for how the final work responds to the task.

This interpretation has implications for authorship and assessment evidence. Authorship in creative education need not mean that every component was produced manually. It does require a defensible relationship between the learner and the decisions represented in the work. When substantial elements are generated, educators need evidence that the student can explain the brief, evaluate the output, identify limitations, and make informed changes. Process records, oral explanations, critique, version histories, and supervised demonstrations may provide such evidence, although the present study did not test these approaches. The point is not to add surveillance to every task, but to align evidence with the capability claim being made.

The survey did not measure evaluative judgement directly. Positive comments about debugging and refinement therefore indicate perceived opportunities for evaluation, while dependency concerns identify a possible risk to independent practice. Both remain propositions for subsequent assessment research.

Responsibility also extends to professional contexts, where errors, confidentiality breaches and unlicensed material can affect clients and teams. Participants’ privacy and accountability concerns connect educational judgement with prospective professional responsibility, although actual behaviour was not observed.

4.4. Policy Permission, Ethical Legitimacy and Value-Based Opposition

Policy permission and ethical legitimacy answer different questions. Permission concerns institutional rules, disclosure, and procedure; ethical legitimacy concerns whether the practice is defensible in relation to creators’ rights, training data, privacy, environmental effects, employment, and responsibility. Clear rules are necessary for fairness but cannot resolve these ethical judgements.

This distinction helps interpret the gap between uncertainty about what was permitted and fear of being accused of misconduct. A student may understand a rule yet remain unsure how their process will be interpreted, particularly when detection practices are opaque or when acceptable assistance is defined broadly. Procedural trust depends on more than the publication of a policy. It requires intelligible expectations, proportionate evidence, opportunities to explain practice and safeguards against treating probabilistic detection as proof. The findings cannot establish whether participants had experienced false accusations, but they show that confidence in procedural fairness should not be inferred from reported knowledge of rules.

Ethical legitimacy also exceeds compliance because participants considered people and consequences outside the immediate assessment, consistent with wider accounts of ethical responsibility in educational AI (Chan, 2023; Holmes et al., 2022). Concerns about training data, creative labour, imitation, confidentiality and environmental cost were reasons to question a practice even when no course rule prohibited it. Institutions do not need to endorse every student claim as empirically settled in order to take the reasoning seriously. They can distinguish verified evidence from uncertainty, explain the basis of institutional decisions, and create space for justified alternatives where learning outcomes permit. This approach treats ethical literacy as part of professional education rather than as an optional addition to integrity policy.

The study does not determine the accuracy of participants’ legal, environmental, or employment expectations. Their analytical significance is that institutional authorisation did not exhaust the criteria through which some students judged GenAI involvement.

4.5. Sustainability-Related Interpretation

Sustainability remains an interpretive lens, not an outcome. The data most directly support a pedagogical question about continued opportunities to develop and demonstrate capability. Ethical and professional considerations organise concerns about training data, privacy, accountability, and preparation for AI-influenced work; they were not measured as sustainability constructs.

Within that lens, sustainability has at least three relevant but non-equivalent meanings. Pedagogical sustainability concerns whether students continue to develop capabilities that remain available when assistance is absent or inappropriate. Professional sustainability concerns whether practices preserve accountability, trust, and responsible participation in creative work. Environmental sustainability concerns the material and energy demands associated with digital systems. The survey provides direct evidence only of student concern about these matters. It does not compare technologies, calculate environmental costs, or demonstrate long-term capability loss. Keeping these meanings separate prevents a broad sustainability vocabulary from converting perceptions into outcomes that were not measured.

4.6. Implications

Theoretical implications. The study positions the function delegated to GenAI, rather than tool adoption alone, as a useful unit of analysis. Conditional legitimacy links function-specific judgement with capability development, agency, responsibility and ethical acceptability. It should remain a provisional framework until tested across disciplines, tasks, and stakeholder groups.

The framework further suggests that acceptance research and assessment research can be connected without collapsing their purposes. Acceptance research explains willingness to engage, while assessment research asks what conclusions can be drawn from student work. Conditional legitimacy proposes that willingness may be related to the perceived integrity of the capability claim. This proposition should be tested directly rather than interpreted as a demonstrated explanation of resistance.

Methodological implications. Mixed-methods studies of GenAI perceptions should avoid collapsing heterogeneous items into a single acceptance score when students may approve one function and reject another. Item-level distributions, function-specific qualitative prompts and joint displays can make these distinctions visible. Future instruments should test the four dimensions directly and examine their content validity and reliability.

Practical and curriculum implications. Educators can specify which functions may be delegated, which decisions must remain attributable to students, what verification is expected, and how AI involvement should be disclosed. Critical AI literacy should extend beyond prompting to include output evaluation, uncertainty, privacy, training-data practices, attribution, and professional accountability (Tzirides et al., 2024). Where GenAI can perform a capability-defining activity, assessment may require proportionate evidence such as decision rationales, drafts, demonstrations, critique responses, or oral explanation (Lodge et al., 2023; Perkins et al., 2024).

Task-specific guidance should therefore identify the educational purpose before listing acceptable functions. For a foundational coding exercise, assistance might be limited to explaining an error message or posing diagnostic questions, while generation of the required solution may undermine the task. In a later integration task, generated code might be permitted if students must test, refactor, and defend it. In visual ideation, educators might permit generated references for critique but require students to document how concepts were transformed and why alternatives were rejected. These examples are not rules derived from the dataset, but they are consistent with task-specific approaches to GenAI-supported assessment (Furze et al., 2024). They illustrate how the framework can translate a capability claim into a reasoned delegation boundary.

Critical AI literacy should include opportunities to examine provenance, privacy, error, bias, labour, environmental claims, and the professional consequences of relying on generated material, as well as practice in verification and disclosure. Whether value-based objections persist as understanding increases remains an empirical question. Education should support informed judgement, including the capacity to decline a function and explain why.

Policy implications. Clear rules remain necessary, but permission statements alone may not address ethical objections or fear of false accusation. Institutions should explain the educational rationale for boundaries, distinguish permitted assistance from evidence of learning, and maintain fair processes for suspected misuse. Student perspectives should inform these decisions without being treated as the sole basis for policy.

These implications are propositions for evaluation rather than demonstrated intervention effects. The present study measured baseline perceptions and cannot establish that any recommended practice improves learning, integrity, or sustainability outcomes.

4.7. Alternative Explanations and Boundary Conditions

Conditional legitimacy is one plausible interpretation of the pattern, but several alternative explanations remain. Most participants had limited or occasional prior experience, which may have contributed to low confidence or scepticism. Voluntary participation may have attracted students with strong concerns. The instrument also contained a substantial number of negatively framed items, potentially increasing the salience of risk. The current design cannot determine how far responses were shaped by prior experience, self-selection, item framing, or experiences and information outside the classroom.

Unit-specific expectations may also matter. Students in coding, design, and client-project contexts encounter different learning outcomes and risks. A respondent’s judgement may therefore reflect the immediate assessment environment rather than a stable view of GenAI. The study did not collect enough contextual detail to separate these influences.

The interpretation is also bounded by the visibility of AI functions. Conditional legitimacy is easier to apply where a student can identify what the tool performed. AI features are increasingly embedded within software, search, and editing systems, making delegation less visible. The present study did not test whether students can recognise and evaluate such embedded contributions, so this remains a direction for future research.

Finally, the study does not establish a developmental sequence from resistance to acceptance or from novice to sophisticated use. Unit-level comparisons were too small and confounded by subject matter. The post-unit phase may show change, but anonymous survey design may limit individual matching. Conditional legitimacy should therefore remain an interpretive proposition open to revision.

4.8. Limitations and Future Research

The study was conducted in one Australian institutional context with a 40.0% completed-response rate and a modest, voluntary, self-selected sample. The characteristics of non-respondents were unavailable, so non-response bias could not be assessed. Limited demographic information also constrained examination of other influences.

The researcher-developed instrument was not psychometrically validated. Several domains were represented by single items, and the balance of concern items may have shaped the pattern of response. The analysis reports perceptions rather than actual behaviour, capability, creativity, employment effects or environmental outcomes. Causal conclusions are therefore not warranted.

The qualitative responses provided useful explanations but were brief and could not be probed. The analysis did not claim saturation, and strongly expressed concerns may have been more visible than moderate positions. Researcher involvement in teaching and AI education created a need for reflexive caution even though participation was anonymous.

The timing of the baseline survey creates a further limitation. Responses captured expectations near the beginning of the teaching period rather than judgements formed after a common learning intervention. Participants may have interpreted GenAI functions through different prior experiences, public debates, or assumptions about unit requirements. The study can therefore describe the starting conditions educators encounter, but it cannot determine how instruction, assessment design, or practical experience changes those conditions. The planned post-unit phase is important because stability and change may differ by dimension. Confidence may increase while ethical objections remain, or students may become more selective rather than simply more accepting.

Future research should analyse the post-unit survey, conduct interviews or focus groups, and compare stated positions with actual AI practices and assessment outcomes. Multi-stakeholder work should examine whether students, educators, assessment specialists, and creative-industry representatives apply similar criteria to particular functions.

5. Conclusions

Digital Media students in this exploratory sample evaluated GenAI according to the function it performed rather than simply accepting or rejecting it. Respondents described potential value in bounded technical, corrective, and organisational assistance, including troubleshooting code, checking grammar, and organising information. In these situations, GenAI was described as reducing operational friction while leaving the main problem, interpretation, or decision with the learner. Substantial opposition was evident when GenAI involvement was perceived to substitute for foundational thinking, ideation, or capability-building practice. Judgements also incorporated ethical, environmental, professional, and institutional considerations, suggesting that institutional permission did not exhaust the criteria through which some students judged educational legitimacy. These findings concern perceptions rather than demonstrated learning loss, environmental effects, or employment change.

To explain this pattern, the study proposes conditional legitimacy as a provisional, theory-informed framework linking four interdependent dimensions: delegated function, retained capability-building activity, human agency and responsibility, and ethical and professional acceptability. Rather than classifying entire applications as permitted or prohibited, the framework directs attention to what a tool performs within a particular task, what capability-building activity remains with the learner, and what the resulting work can credibly demonstrate. It is a structure for reasoned evaluation rather than a validated scale or decision model.

The study’s primary contribution is an empirically grounded, student-informed account of the criteria through which respondents distinguished bounded assistance from GenAI involvement perceived to replace capability-building activity in creative education. As an exploratory baseline study, the findings reflect student perceptions and expectations before targeted instruction in one institutional context. Future research should test and refine the provisional framework using post-unit data, interviews or focus groups, performance evidence, longitudinal designs, and comparisons across disciplines, assessment types, and institutional contexts.