AI-Mediated Creative Agency: A Framework for Creative Self-Efficacy in Human-AI Collaboration

Creative Self-Beliefs

Creative self-efficacy refers to individuals’ beliefs in their ability to engage in creative activities within specific domains or situations (Tierney & Farmer, 2011). Creative confidence denotes a broader, more generalized sense of oneself as a creative person and as someone willing to take risks in creative situations (Karwowski, 2016; Kelley & Kelley, 2012). These beliefs influence how individuals approach creative problems, how persistently they engage with them, and how they interpret success and failure (Wojtycka et al., 2025). Creative self-efficacy and confidence are related constructs but not interchangeable: self-efficacy shapes moment-to-moment judgments about capability in specific domains or tasks, while confidence influences whether individuals choose to engage in creative work at all and how they narrate themselves as creative people across contexts. Over time, task-specific mastery experiences can accumulate into broader creative self-beliefs, so increases in self-efficacy can support the development of creative confidence (Beghetto, 2006; Beghetto et al., 2021).

In human-AI creative contexts, the most immediate psychological effects often appear at the level of self-efficacy. AI can increase or decrease perceived capability depending on how individuals interpret their contributions relative to the system and on how success is attributed (Newman et al., 2024). Recent research on general and AI-specific creative self-beliefs similarly suggests that AI contexts may differentiate the ways people evaluate their creative capacities rather than increase or decrease a single global belief (Faiella et al., 2025). These shifts can subsequently influence creative confidence and creative identity by altering individuals’ sense of themselves as capable creators across contexts.

Bandura’s self-efficacy theory identifies four primary sources of efficacy information: mastery experiences, vicarious learning, social persuasion, and affective states (Bandura, 1977). In many traditional creative practices, the relation between intention, action, and outcome is sufficiently transparent that individuals can treat successes and failures as meaningful evidence of their own capability.

Generative AI and Self-Efficacy

Generative AI introduces new forms of mediation that complicate the development of creative self-beliefs (Habib & Gardiner, 2026; Newman et al., 2024; Worrall & Collins, 2024). Although many technologies involve processes that are not fully transparent, generative AI can make it especially difficult to trace how human intentions, prompts, selections, and revisions shape the final result (Marrone et al., 2026). This can introduce attribution ambiguity: creators may struggle to determine whether successful work reflects their own capabilities, the system’s contribution, or some combination of both.

Empirical evidence illustrates the psychological effects. Students using generative AI as a brainstorming partner on creative tasks reported that AI was helpful, yet simultaneously reported decreased confidence in their own creative abilities; they could produce outputs with AI assistance but felt less capable as independent creators (Habib et al., 2024). Children who collaborated with AI for storytelling reported heightened self-efficacy during AI-supported activities but experienced difficulty when asked to perform without AI support (Newman et al., 2024). Professional creatives described tensions between empowerment and loss of authorship when AI-generated ideas closely resembled professional work (Worrall & Collins, 2024). These findings suggest that AI reshapes the experiential conditions under which self-efficacy is formed, particularly by decoupling performance outcomes from internalized capability.

A Relational Perspective on Creative Agency

To account for these transformations, it is useful to consider sociocultural, ecological, and distributed approaches that conceptualize creativity as emerging from relations rather than residing solely within individuals. Frameworks such as the Five A’s model (Glăveanu, 2013) and distributed agency theories describe creativity as a sociomaterial process in which actors, artifacts, and audiences co-constitute creative outcomes (Enfield & Kockelman, 2017; Sawyer & DeZutter, 2009). These accounts establish that creativity is not located solely in individuals but emerges through the configuration of relations among actors, tools, artifacts, audiences, and environments. Mediated agency builds from this premise by asking a complementary psychological question: how do creators experience themselves as capable agents within such distributed systems?

The Five A’s framework provides an entry point for understanding what AI mediation changes. AI systems alter the Actor by expanding what is technologically possible; they transform the Action by introducing generative participation that both expands possible directions and complicates the traceability of human intention; reshape Artifacts by producing outputs that blur the boundary between human and algorithmic contribution; modify the Audience by shifting evaluative standards toward AI-augmented benchmarks; and restructure Affordances by reconfiguring what creative moves are available, efficient, or legible within a given context. Distributed approaches illuminate these transformations across the relational field. A complementary question concerns how these transformations are experienced subjectively: how they register in the creator’s formation of self-beliefs and sense of possibility.

Distributed agency and mediated agency address complementary analytical questions. Distributed agency examines how creative agency is configured across sociomaterial systems; mediated agency examines how individuals within those systems interpret AI-mediated creative outcomes as evidence of their own capability. The former provides a relational account of creative action, while the latter focuses on the lived formation of creative self-beliefs within that relational field.

Mediation in Sociocultural Theory

The concept of mediation in this framework is grounded in sociocultural theory, particularly Vygotsky’s (1978) argument that human cognition develops through the use of culturally produced tools and signs. For Vygotsky, higher mental functions such as reasoning, memory, planning, and self-regulation do not emerge in isolation within the individual mind. Rather, they are formed through social interaction and are mediated by psychological tools, including language, writing, counting systems, diagrams, symbols, and other culturally available means.

These tools actively shape the structure and possibilities of thinking itself. Language, for example, does more than communicate preexisting ideas; it reorganizes attention, supports abstraction, and enables forms of reflection that would not be possible in the same way without it. Wertsch (1991) extended this insight through mediated action theory, arguing that human agency is always exercised through cultural tools rather than independently of them. From this perspective, tools participate in action by shaping what individuals can perceive, imagine, decide, and accomplish. Mediation, therefore, can be understood as a constitutive process through which human thought and agency are formed, transformed, and enacted.

The term mediated agency has prior uses in adjacent literatures that this framework both draws upon and extends. In sociocultural psychology, Wertsch and colleagues used the concept to describe agency as enacted through mediational means, arguing that agency is not a property of the isolated individual but is constituted through the cultural tools and signs that structure action (Wertsch et al., 1993). This formulation, rooted in Vygotsky’s account of psychological tool use, establishes the foundational principle that mediational means actively reshape the structure and meaning of action. A related body of work in human factors and automation research shows that machine-mediated action can alter users’ subjective sense of control. For example, Berberian et al. (2012) found that the level of automation affects both implicit and explicit measures of agency, suggesting that technological mediation can reshape the experienced relationship among intention, action, and outcome.

The proposed framework here builds on these traditions while advancing a psychologically distinct application: rather than analyzing how tools configure action or how automation disrupts motor control, it examines how algorithmic mediation reorganizes the experiential and interpretive processes through which creative self-beliefs form. Mediated agency, as used here, is therefore not a departure from these prior accounts but an extension of their core insight: agency is constituted through its mediational conditions.

When AI mediates between intention and outcome, it can reshape how intention is formed, how effort is apportioned, and how outcomes are interpreted, with direct consequences for the formation of self-efficacy. A critical departure from many earlier mediational tools, however, is the combination of opaque operation and generative participation. Many tools shape creative action in ways that are not fully transparent, but generative AI can introduce substantive content into the emerging artifact, sometimes in directions that diverge from the user’s intention. This can expand possibility spaces by surfacing unexpected directions, alternatives, and forms of expression that creators might not have generated on their own. At the same time, it can make it harder to trace the relationship among intention, action, and outcome, with direct consequences for how creators interpret their own contributions and capabilities. Mediated agency offers a vocabulary for examining what happens to Bandura’s sources of self-efficacy when the mediational means not only support creative action but also contribute substantive content whose relation to human intention can be difficult to trace.

Negotiation, Power, and Psychological Outcomes

The relationship between humans and AI is not symmetrical. AI systems carry algorithmic authority: users often perceive their outputs and evaluations as objective or neutral, which grants these systems disproportionate influence over creative decisions (Kern et al., 2022; Shirky, 2010). This creates an uneven power dynamic in which AI can shape outcomes and evaluative standards in ways that may diverge from users’ intentions or developmental needs.

Algorithmic authority operates within power structures that configure how mediated agency unfolds. Training data can embed cultural and aesthetic biases that privilege certain styles or ideas (Liu, 2023). In educational and professional contexts, institutional policies ranging from prohibitions to mandated adoption create distinct psychological conditions that shape how creative self-efficacy develops (Fawns, 2022). These institutional choices sit within broader layers of power, from relationships with specific algorithms to dependence on platforms and employers, which shape how creators can negotiate authorship and control (Nguyen & Mateescu, 2024).

Mediated agency should therefore be viewed as an ongoing negotiation in which humans must actively maintain creative authority against both algorithmic influence and institutional pressures. The degree to which individuals succeed in this negotiation directly affects their creative self-efficacy, as it determines whether they experience authentic control over the process, credible authorship of outcomes, and a stable sense of creative identity within human-AI collaboration. The framework, therefore, captures both the collaborative potential and the contested nature of these partnerships, emphasizing that creative confidence grows when collaboration preserves meaningful agency and diminishes when collaboration displaces control or obscures contribution.

Mediated Mastery Experiences

Mastery experiences provide the most influential source of self-efficacy because they offer direct evidence of capability (Bandura, 1977). In AI-mediated contexts, individuals can achieve impressive creative outcomes, potentially accomplishing tasks they could not complete alone, which constitutes a genuine form of mastery: mastery of human-AI collaboration. However, when success cannot be clearly attributed to one’s own effort and decision-making, the ability to build self-efficacy can be weakened.

Attribution ambiguity is central here. When creators ask, “Did I do this, or did the AI?” they may struggle to treat AI-supported success as evidence of personal capability. In AI-mediated creative work, this can produce a form of pseudo-competence, or miscalibrated self-efficacy: a superficial sense of capability that arises when individuals succeed in supportive conditions without developing the underlying skills required for independent performance. Pseudo-competence is especially likely when impressive outputs are achieved primarily through algorithmic contribution rather than the creator’s own conceptual development, decision-making, or technical mastery. As a result, these successes may not provide reliable evidence of transferable capability.

AI can also weaken mastery experiences when it removes too much of the challenge through which capability is normally tested. Mastery experiences are most potent when creators encounter difficulty, make decisions, revise, and eventually succeed (Csikszentmihalyi, 1988; Young et al., 2024). When AI collapses this process into rapid output generation, creators may produce impressive artifacts without developing equally robust beliefs in their own transferable capability.

However, AI can support genuine mastery when it scaffolds rather than replaces creative processes. Systems that prompt reflection on creative decisions, require iterative refinement, or provide partial solutions that still demand substantial human judgment can facilitate enduring skill development. When creators understand how their decisions guide or constrain AI outputs and can trace the connection between their intentions and the results, mediation can support rather than undermine the accumulation of robust mastery experiences. The key distinction lies in whether AI handles mechanical execution while preserving conceptual challenge, or collapses both dimensions, making creators selectors rather than makers.

Mediated Vicarious Learning

Vicarious learning occurs when individuals observe others similar to themselves succeed and infer that they too possess the capability to achieve similar success (Bandura, 1997). The effectiveness of vicarious learning depends on the perceived similarity between the observer and the model (J. V. Wood, 1996). In AI-mediated contexts, creators observe both AI systems and peers producing impressive AI-augmented outputs, which complicates the identification of appropriate comparison targets.

A first challenge is that AI is not a “similar other.” AI-generated outputs may function as examples or benchmarks, but they do not model the human process of struggling, revising, and developing skill in the way a peer or mentor can. Observing an algorithm generate fluent text or polished designs may expand a sense of what is possible, but can also trigger feelings of inadequacy when human efforts seem slow or clumsy by comparison (Habib et al., 2024). A second challenge is the “double comparison” problem: creators might compare themselves both to AI outputs and to peers whose work is enhanced by AI support. Emerging work on AI and art self-image assessment suggests that exposure to AI-generated or AI-augmented outputs can obscure how individuals assess their own creativity, even when their performance improves (Kellerwessel & Ujhelyi, 2024).

Social comparison theory indicates that comparisons to highly superior standards, particularly in domains central to self-concept, can threaten rather than motivate (Festinger, 1954; Tesser, 1988). In AI-mediated creative work, the evaluative space increasingly encompasses AI outputs and AI-augmented performances, making it difficult to distinguish between developments in human capability and the appearance of capability through technological augmentation.

Vygotsky’s (1978) concept of the zone of proximal development offers additional insight into when AI-generated outputs serve as productive models rather than discouraging comparisons. AI outputs may serve as demonstrations that lie within or beyond an individual’s developmental reach. When outputs fall within the zone of proximal development, representing what learners could achieve with appropriate guidance, they can expand their capabilities through supported exploration and by providing aspirational models. However, when AI-generated work dramatically exceeds a creator’s current capability, falling outside their zone of proximal development, it may instead induce comparisons that feel threatening rather than motivating. This suggests that the developmental appropriateness of AI-generated models can significantly affect self-efficacy, and that creators at different skill levels may require different relationships with AI outputs to maintain confidence as they develop capability.

Mediated agency emphasizes that the psychological impact of these comparisons depends on how AI outputs are framed within a creator’s identity and practice. When AI is framed as a tool that extends human possibility, AI-generated examples can serve as inspirational models for exploration. When AI is framed—or experienced—as a replacement or as an unreachable standard, vicarious exposure can undermine self-efficacy and destabilize creative identity.

Mediated Social Persuasion

Social persuasion involves feedback and encouragement from others that shape self-efficacy by signaling capability and potential (Bandura, 1997). In creative contexts, feedback from teachers, mentors, and peers carries weight when it comes from credible, trusted sources (Hattie & Timperley, 2007). AI participation transforms this source by introducing automated feedback that often appears objective yet lacks contextual grounding.

AI feedback can carry an illusion of objectivity because it emerges from data-driven analysis rather than overtly subjective human judgment. Creators may treat AI-generated evaluations or recommendations as more authoritative, even when the underlying models encode biases or lack contextual sensitivity. At the same time, many AI systems are designed to be supportive, sometimes providing overly positive assessments to maintain engagement, which can erode trust in their feedback as an authentic indicator of capability (Kern et al., 2022; O’Hara, 2021).

Complexities arise when AI feedback is mediated through human instructors or peers. A teacher might rely on AI analysis to identify strengths and weaknesses in student work, then present that feedback in human terms. In such cases, the persuasive impact depends on whether creators perceive the human source as genuinely endorsing the feedback or merely relaying algorithmic judgments. When a mentor’s encouragement is informed by AI analysis, does that encouragement carry the same persuasive weight as feedback based on the mentor’s own judgment? If creators know that praise or critique was algorithmically generated, even if delivered by a human intermediary, how does this knowledge affect the feedback’s impact on self-efficacy?

Effective social persuasion depends on clarity about the source, the authenticity of the relationship, and a genuine belief in the individual’s potential, qualities that algorithmic systems cannot provide in the same relational sense as teachers, mentors, or peers. Feedback practices that clarify the role of AI, foreground human judgment, and situate evaluation within ongoing relationships are more likely to support creative self-efficacy than practices that delegate persuasion to automated systems without clear human interpretation or responsibility.

Mediated Affective States

Physiological and affective states function as sources of self-efficacy information, as individuals interpret emotions such as excitement, frustration, anxiety, or satisfaction as cues about their capability (Bandura, 1997). Creative work traditionally involves a mixture of challenging emotions and satisfying breakthroughs, with periods of productive struggle contributing to a sense of growth and resilience (Csikszentmihalyi, 1997).

AI mediation introduces new affective patterns. Creators may experience excitement when AI generates surprising possibilities and flow-like engagement when collaboration feels responsive; related research on technology-mediated learning and creative confidence suggests that affective and metacognitive experiences matter for self-efficacy and self-beliefs (Smit et al., 2025; Wojtycka et al., 2025). At the same time, they experience frustration when AI misinterprets intentions, anxiety when AI outputs appear superior to their own, and discomfort when the boundary between human and machine contribution becomes unclear (Worrall & Collins, 2024).

A particularly important shift is the emergence of “too easy” creative experiences. When AI dramatically reduces effort and uncertainty, work can feel effortless but also emotionally flat, depriving creators of the satisfaction that comes from overcoming difficulty. Struggle provides evidence of emotional engagement, boundary pushing, and skill development. When AI eliminates this struggle by generating solutions immediately, the emotional texture of creative work changes: impressive results become achievable without the sustained effort that has traditionally been registered as evidence of capability and growth. This creates a paradox: creators can produce compelling work while feeling emotionally empty about the process, a pattern that, if sustained, prevents those affective cues from accumulating into creative identity.

Mediated agency emphasizes that individuals must learn to interpret these novel affective states in ways that support, rather than undermine, self-efficacy. Design and pedagogical practices that intentionally reintroduce manageable challenges, encourage reflection on emotional responses to AI-assisted work, and validate struggle as integral to creative development can help recalibrate affective cues in AI-mediated environments.

Control Over Process

Control refers to the extent to which creators can direct, interrupt, and revise AI contributions in ways that feel responsive to their intentions. Interactive AI interfaces invite a sense of control through prompting and parameter adjustments, yet the underlying generative mechanisms remain inaccessible, creating a structured form of control in which creators have real but bounded influence over outcomes. This differs from the illusion of control (Langer, 1975), which was formulated for chance situations in which individuals exert no actual influence yet believe otherwise. In AI-mediated creative work, influence is genuine but partial: creators direct outputs through prompting, selection, and revision, yet they cannot fully inspect the mechanisms linking their intentions to the resulting artifact. The psychological effect is significant. When the path from intention to outcome is difficult to trace, successful outputs are less readily internalized as evidence of personal capability. Research on algorithmic authority suggests that this interpretive difficulty is both cognitive and social, shaped by the perceived objectivity and authority attributed to AI systems (Kern et al., 2022; O’Hara, 2021).

The negotiation of control involves balancing two risks. Ceding too much control can lead to AI-dominated processes in which human decision-making is reduced to accepting or rejecting suggestions. Exercising excessive control by over-constraining the system can render AI contributions shallow or merely decorative, limiting opportunities for genuine co-creation. Interfaces that surface model uncertainty, provide stepwise explanations, or allow users to iteratively refine outputs with clear traceability can support more meaningful control.

Within mediated agency, control over process is a key dimension through which mediated mastery, vicarious learning, persuasion, and affect are interpreted. When creators experience themselves as actively directing AI contributions, successes, and failures are more readily integrated as evidence of their own capability, supporting creative self-efficacy.

Authorship

Authorship concerns who is perceived as responsible for creative work and which contributions can be legitimately claimed. In human-AI collaboration, authorship becomes ambiguous when ideas, structures, or stylistic elements originate from AI prompts or outputs and are not clearly distinguished from human input.

This ambiguity has direct implications for self-efficacy. If creators cannot identify which aspects of a product reflect their decisions, skills, and efforts, success provides weak evidence of personal capability. Consider a student using AI to help write an essay. The final text reflects both the student’s ideas and the AI’s elaborations. Which sentences express the student’s thinking? Which phrases came from the AI’s training data? If the overall argument structure emerged from AI suggestions, but the student selected which points to develop, who authored the argument?

Professional creatives encounter similar attribution challenges. Synthesizing recent empirical work, Habib and Gardiner (2026) report that designers, writers, and other creative professionals who use generative AI often describe their role shifting from constructing work directly to curating, editing, or refining system-generated options, which can blur the boundary between their own expertise and the system’s contribution. In these cases, highly polished outputs may be achieved, but creators are less certain which aspects of the work genuinely reflect their spatial reasoning, aesthetic judgment, or conceptual development. As a result, even when performance improves, it can be harder for professionals to treat AI-supported successes as clear evidence of their own creative capability.

Mediated agency emphasizes practices that preserve authorship as a source of mastery and identity. These include tracking the provenance of AI-generated elements, encouraging transparent disclosure of AI involvement, and incorporating reflective activities in which creators articulate their contributions, intentions, and rationales. Such practices make it easier for both creators and evaluators to recognize human agency within hybrid outcomes, strengthening the link between performance and self-beliefs.

Creative Identity

Creative identity refers to the centrality of creativity to one’s sense of self and to the narratives through which individuals understand themselves as creative people (Glăveanu & Tanggaard, 2014; Winner, 1982). It develops through sustained creative practice, recognition by others, and the integration of creative experiences into broader life stories (Beghetto et al., 2021; Csikszentmihalyi, 1984). When creative identity is secure, temporary setbacks or transitions do not fundamentally threaten one’s sense of being a creator.

AI mediation introduces new threats and opportunities for creative identity. When technologies perform tasks that previously signaled creative expertise, individuals may question what remains distinctly human in their practice. Encountering AI systems that appear more fluent or technically proficient can provoke feelings of inadequacy, particularly for those whose identities are deeply tied to specific creative skills. Others may adopt new identities as orchestrators, editors, or conductors of human-AI ensembles, redefining what it means to be creative.

From the perspective of mediated agency, sustaining creative identity under AI mediation involves recognizing and foregrounding dimensions of creativity that remain irreducibly human: problem framing, value judgment, contextual adaptation, ethical reflection, and the integration of creative work into meaningful life narratives. Educational and professional practices that acknowledge these dimensions and help individuals narrate themselves as creative agents in relation to AI can support the continuity and renewal of creative identity.

Developmental Pathways

The developmental dynamics of mediated agency involve complex interactions between immediate performance gains and longer-term self-efficacy outcomes. When AI scaffolding enables immediate success, individuals may experience temporary increases in perceived capability. However, when mediation obscures the link between intention, effort, and outcome, these short-term gains may fail to accumulate into durable creative self-efficacy or broader creative confidence—a pattern consistent with research on scaffolding dependence in learning environments (D. Wood et al., 1976).

Sustainable self-efficacy development requires experiences that preserve effort, authorship, and credible attribution. The three possible trajectories introduced earlier, augmentation, dependence, and displacement, describe the range of developmental outcomes these patterns can produce. Augmentation requires that mediation actively preserve the conditions for genuine capability development; dependence emerges when convenience and quality displace those conditions; displacement occurs when the cumulative effect is a loss of creative self-belief and identity.

Expertise likely moderates these developmental trajectories. Novices may rely more heavily on algorithmic scaffolding and experience greater difficulty maintaining attribution clarity, whereas experts with established creative schemas may more effectively integrate mediation into existing creative strategies (Chi et al., 2014). Metacognitive awareness—the capacity to monitor and regulate one’s own cognitive processes—may also buffer against negative effects by helping creators maintain clarity about their contributions even within mediated systems (Schraw & Dennison, 1994).

At the same time, AI expands perceived possibilities in ways that can genuinely enhance creative engagement. Across domains, individuals report increased willingness to attempt complex creative tasks when AI scaffolding is available, suggesting that mediated mastery experiences can broaden creative self-efficacy and creative identity (Beghetto, 2006; Newman et al., 2024). However, evidence also indicates that these expanded possibilities may be fragile when confidence is tied primarily to AI-supported performance rather than transferable creative capability. This tension between expanded short-term possibility and reduced longer-term developmental openness points to a central implication of mediated agency: sustaining creative self-efficacy under AI mediation requires preserving indeterminacy, struggle, and authorship as conditions of possibility, rather than allowing efficiency and polish to crowd them out.