Protecting creativity in the age of generative AI: productive uncertainty, and visible thinking in scholarship and assessment

I frame this paper from my perspective as an art education scholar, focusing on how creativity thrives within the realms of ambiguity, hesitation, and the productive discomfort of not knowing at least, not yet. A persistent question I have been grappling with is whether seeking immediate answers from AI undermines the essential space of uncertainty that nurtures curiosity, creativity, and critical thinking. These are habits we engage with daily in art education, which are crucial for developing new ideas in scholarship.My reflections on this issue deepened after I commented on a colleague's invitation to discuss AI illiteracy and its implications for scholarly credibility on a professional platform. This discussion illuminated a key problem: the challenge is not solely whether to use AI, but rather how we can do so in ways that uphold the thinking habits essential for trustworthy scholarship and teaching.In this article, I argue that credible scholarship and teaching in the age of AI hinge on AI literacy practices that promote inquiry through the formation of meaningful questions, verification against primary sources, transparent disclosure, and assessment designs that make thinking visible. I begin by defining credibility and outlining the risks associated with AI illiteracy. Next, I examine the impact of AI on research practices and assessment. I conclude by explaining the importance of allowing creativity the time it needs to flourish in uncertainty. I conclude by explaining why creativity requires time in uncertainty, and I offer practical guidance for educators and researchers navigating these challenges. I began reflecting seriously on this issue after commenting on a colleague's invitation to discuss AI illiteracy and its impact on scholarly credibility on a professional platform. The subsequent discussion clarified a key problem: the challenge is not merely whether to use AI but whether we can do so in ways that uphold the thinking habits that make scholarship and teaching trustworthy. While examples are primarily drawn from studio practice, the underlying approach can be generalised. Requirements for disclosure and author responsibility in medical publishing, policy guidance for course-level declarations, reflective activities in higher education, and AI-aware assessment trials in environmental data science all demonstrate that verification, disclosure, and transparent assessment processes are being implemented beyond the realms of art and design 5, 14-16, 19-20, 22-23, 26-27.Two interconnected issues are central to the challenges of AI illiteracy and concerns regarding credibility in research. The first issue is the allure of rapid publication driven by metrics, which can lead to a fragmented approach. In this scenario, authors may simply take what an AI model generates and insert it into their articles, bypassing the rigorous yet rewarding processes of experimenting with ideas, testing theories, and refining arguments through iterative work. This practice can result in polished prose that includes fabricated or distorted citations if the outputs are not properly verified, ultimately undermining the integrity of the research record 1,2.The second issue pertains to teaching and learning contexts. Discussions about assessment are often driven by anxiety, leading to significant efforts focused on bans and detection methods. Comparative studies indicate that AI text detectors frequently misclassify human writing and disproportionately target non-native English speakers, resulting in potential unfair outcomes and diverting attention from meaningful learning design 3,4. This is a concern that many non-English-speaking countries can relate to, particularly in my context in South Africa, where the country boasts 11 official languages. This situation poses a significant risk for many of our students from diverse linguistic backgrounds. The unintended consequence is an assessment process that lacks pedagogical soundness and fails to provide adequate challenge. Rather than accurately gauging whether learning has occurred and the extent to which students have mastered the material, tasks often devolve into mere compliance checks.As an art education scholar, I am particularly concerned with how creativity flourishes in ambiguity, hesitation, and the productive discomfort of not knowing. A persistent question arises: does turning to AI for immediate answers erode that vital space of uncertainty that fosters curiosity, creativity, and critical thinking.Studies show that empirical evaluations reveal significant rates of hallucinated or inaccurate references and confidently declared errors in AI-generated text when authors neglect to verify the outputs 1,2. Incorporating such unverified material undermines essential scholarly practices, including close reading, triangulation, and transparent methodology, thereby weakening the overall scholarly voice. Recent scholarly guidance consistently emphasizes that authors must take full responsibility for accuracy and systematically disclose and verify any AI-assisted steps involved in their work 5,22 .Curiosity serves as a fundamental driver of lifelong learning. Research indicates that it activates midbrain reward centres and enhances hippocampus-dependent memory for both the main information and incidental details encountered during exploration 6,7. Furthermore, intellectual curiosity has been shown to predict academic success beyond mere cognitive ability and conscientiousness 8. If the use of AI consistently reduces the time scholars spend in a state of uncertainty, it may systematically undermine the conditions that promote originality and careful judgment.Meaningful inquiry begins with identifying problems rather than jumping straight to solutions. Research into human interaction with automation reveals that individuals often over-rely on algorithmic suggestions an issue known as automation bias or algorithm appreciation 9,10. However, there is insufficient critical examination of the potential risk that novelty may be entirely lost when problem identification is neglected. When the initial framing of a problem is delegated to a model, scholars might focus more on refining the model's questions instead of engaging in the challenging process of formulating their own questions. This can lead to premature conclusions that overlook essential critical thinking. This consideration led me to reflect on creativity and the conditions necessary for it to flourish in the scholarly journey.Research consistently demonstrates that tolerance for ambiguity is positively correlated with the development of creativity. Design studies indicate that novelty often emerges at the boundaries of coherent solution spaces, where uncertainty is highest 11. Large-scale experimental studies reveal that while AI can enhance average creativity ratings among individuals, it tends to limit the diversity (novelty) of outputs across a group. This lack of diversity can jeopardise credibility within the field if many creators adopt similar patterns and approaches 12,24. When numerous writers follow identical formulas, the field risks losing the variety of styles and ideas that are crucial for signalling credibility and fostering innovation. Currently, the evidence gathered in studio settings is limited. There is a lack of cohort-level evaluations in art and design that explore whether shared prompts or common models lead to stylistic convergence in the work being assessed. Utilising dispersion metrics for outputs and conducting longitudinal studio studies would be beneficial in clarifying the effects of such factors on originality within the field.Credibility is best safeguarded when AI enhances inquiry rather than bypasses it. Practical habits that can help include starting with questions and criteria generated by humans, using Formatted: Font: Bold AI to identify counterarguments and highlight disagreements, and systematically auditing all substantive claims against primary sources. Additionally, maintaining a concise verification log to record prompts, outputs, checks, and corrections is essential. It's also important to disclose the use of AI while taking full responsibility for accuracy. These practices align with contemporary expectations in medical publishing, where journals mandate explicit disclosure of AI assistance and hold authors accountable for the accuracy of their work 5,22. A simple diagnostic test can guide practice: did the tool extend the duration of inquiry, or did it merely accelerate movement past it? Credibility is best maintained when AI enhances inquiry rather than circumventing it. Practical habits that support this approach include: starting with human-generated questions and criteria; leveraging AI to identify counter-arguments and map areas of disagreement, followed by a thorough audit against primary sources; keeping a concise verification log that documents prompts, outputs, checks, and corrections; and transparently disclosing howDetector performance remains highly inconsistent, with documented false positives and notable fairness issues, especially for non-native English speakers 3,4. Over-reliance on detection technologies promotes the creation of tasks that are easy to monitor rather than assessments that genuinely require interpretation and disciplinary judgment. Credibility in teaching increases when institutions move from a policing approach to a learning-focused approach that prioritises educational outcomes.Detector performance continues to exhibit significant inconsistencies, characterised by documented false positives and notable fairness issues, particularly affecting non-native English speakers 3,4. An over-reliance on detection technologies can lead to the development of tasks that are easier to monitor, rather than assessments that genuinely demand interpretation and disciplinary judgment.Credibility in teaching is enhanced when institutions transition from a policing approach to one focused on learning and prioritising educational outcomes. However, existing studies often overlook how detection-centric policies may influence students' trust, their willingness to take intellectual risks, the formulation of meaningful assessment activities that challenge students, and the validity of assessments in multilingual contexts 15,16,[192026].Effective feedback should enhance learning outcomes rather than simply document policy violations. Hattie and Timperley's influential model highlights that effective feedback Formatted: Font: (Default) Arial, Bold addresses three key questions for learners: Where am I going? How am I progressing? Where should I go next? 13. Research indicates that feedback is most effective at the task, process, and self-regulation levels, while it is least effective when it relies solely on general praise.The feedback requirements outlined are aligned with evidence suggesting that effective feedback should focus on task-related aspects, process improvements, and self-regulation, rather than relying solely on general praise. These elements can be seamlessly integrated into AI-aware coursework by incorporating explicit criteria, justifications for rejected suggestions, and verification logs 13,17.In an AI-aware classroom or studio, these feedback levels can be implemented through specific requirements. These include connecting accepted AI suggestions to clear criteria, justifying at least one rejected suggestion, explaining strategies for source integration or image curation, and maintaining a brief verification and reflection log. This log should record checks and future actions, with at least one component completed without AI assistance to ensure cognitive engagement.While some reports offer guidance on implementing these requirements, evidence regarding their effects remains limited. There is a lack of comparative studies across disciplines that assess the impact on learning outcomes, academic integrity, and equity. Future research should include randomized or quasi-experimental comparisons testing portfolios with annotations, short oral examinations, and uncertainty statements against traditional written tasks. Such studies should also evaluate validity, reliability, workload, and student trust in addition to integrity metrics 13-16, 19-20, 25-26] Authentic assessments require students to apply disciplinary criteria to complex problems while making their processes transparent. Documented practices in environmental data science integrate permitted AI tool usage with high-stakes evaluations and explicit disclosures. These practices utilise formats such as process portfolios, annotated drafts, brief oral examinations of decision-making, and statements of uncertainty that detail what was verified and the methods used 14. Policy-oriented research in higher education advocates for course-level declarations of AI usage and structured reflective activities aimed at fostering metacognitive control, rather than relying solely on detection mechanisms 15,16,[192026].I propose that we can gain valuable insights from studio learning, which fundamentally relies on risk-taking, iteration, and critique. To ensure that curiosity and originality remain central when students engage with text-to-image or code-based tools, programs should implement several strategies. First, they can mandate human problem identification prior to any prompting, requiring students to articulate the challenges they wish to address. Second, process annotations should be required to document the tools used, the prompts generated, the choices made in curation, and any post-processing steps undertaken. Finally, incorporating both blind and labelled critiques can illuminate how cues of origin influence judgments of creativity. These approaches align with research indicating that creativity flourishes in contexts of ambiguity and that fostering group-level diversity requires careful balancing, especially when commonly used models dominate the discourse 11,12,24. 5,19,20,27.Prohibition and detector-driven design that misclassifies students and encourages compliance over interpretation and judgement; equity risks for non-native writers 3,4.AI-aware feedback at task, process, and self-regulation levels; authentic assessments requiring process evidence, brief oral examinations, and uncertainty statements; student declarations and justifications of AI use; course-level policy clarity 13-16, 19, 20, 22, 23, 25, 26.Over-reliance on model defaults that flattens stylistic diversity and reduces risk-taking; reduced time in exploratory making and creative uncertainty 11,12,24.Human problem finding before prompting; comprehensive process annotations and disclosure; blind and labelled critiques to examine origin effects; deliberate divergence requirements topreserve cohort-level diversity 11,12,24.Blanket bans and detector-only responses that shift attention from learning design and create inconsistency across courses 15,16,26.Course-level declarations of permitted uses and disclosure norms; alignment of assessment with policy; reflective activities to support metacognition and responsible use 15,16,19,20,26.The shift towards generative AI in education has highlighted two key issues: first, that detector-led approaches can lead to unfair outcomes; and second, that credible assessment in this context relies less on policing and more on designing tasks that make student thinking observable. In my own practice, I started recognising these implications early during the widespread adoption of AI tools.When AI began gaining traction in educational settings, I assigned a research-based project to my second-year students. While marking the submissions, I noticed that some of the work was written in a style that did not align with the typical capabilities of students at this level in my teaching context. The prose was unusually polished and professionally structured. At that time, I relied on Turnitin's AI detection tools to assess whether AI may have been used. This approach led to complications: many students found themselves at a disadvantage due to the limitations of the detector-driven process, especially considering the large enrolments, which made it nearly impossible to thoroughly investigate every submission or conduct individual follow-ups.This experience prompted a shift in assessment design away from increased surveillance. I continued to use written assignments but added an authentication component that emphasizes reasoning rather than mere confessions. Students now submit a short video in which they discuss their work and explain the key ideas, decisions, and sources they utilized. The guidelines explicitly instruct students not to read from their assignments but to articulate their arguments and learning process. This method does not simply aim to "catch" students; rather, it makes learning visible. Students who understand their work can explain and defend it effectively, while those who relied heavily on AI tool often struggle to present a coherent account. Importantly, this approach supports multilingual students more equitably than detector-led judgments, as it assesses whether learning has indeed taken place rather than conformity to a specific assessment layout. What this changes in practice: instead of asking "Did you use AI?", the assessment asks, "Can you demonstrate ownership of your thinking and evidence?"Another experience involved suspected AI misuse that triggered a formal escalation process, with assignments moving from the lecturer to the HOD and into disciplinary structures. In practice, this route often felt tedious and, at times, inhumane, especially when the initial suspicion rested mainly on detector outputs rather than verified evidence. As the limitations of AI detection have including documented false positives and a risk of it is to disciplinary escalation as a more credible and approach is to assessment design as the first of When tasks require process evidence annotations, short oral or video and statements of integrity concerns can be through learning-focused mechanisms before they disciplinary This does not it it into educational In practice, the is that the a and a process, with for unintended consequence of generative AI is that it assessment When assessment mainly on written outputs, it easier for students to thinking. In this AI educators to be more deliberate and to design assessments that and rather than the of This shift beyond an on written or traditional assessment that align with authentic learning, particularly in such as art education, where iteration, and are This shift assessments focus less on simply a and more on the process, and a research I do not as an AI The practical is not that researchers must but that AI literacy must be with curiosity and the ability to work through the problems they In AI can but it the intellectual of uncertainty, and that makes research a scholarly remains one of the most important of academic work. This is particularly when art is used as a when a similar image or originality is in the work through material choices and credible research requires that a remain present in the in how problems are and how evidence is This generative AI tools can text that is often clear and careful the can shift the of what the is to and the of the practical from my own practice is using AI tools to support clarity in writing rather than As a non-native English I often use tools such as and, at times, to and However, I that the tool my style and scholarly as I have that when is not the text polished but and what I to This that AI as a and clarity while the ownership of their and voice. the tool enhances without the originality that scholarly two credibility challenges as reflections of In work on unverified AI outputs can curiosity, the time in productive uncertainty, and the scholarly voice. In education, the of AI can shift assessments detection technologies and tasks that provide insights into what students or can A in both areas in fostering AI which verification, transparent disclosure, and assessment designs that on curiosity highlights the importance of navigating the space of not knowing for learning and Research on automation bias and creativity in the context of AI assistance highlights the of question formation and with ambiguity for credible intellectual work and assessment studies designs that interpretation and at the of from beyond art education demonstrate that similar are being medical disclosure and author for higher education policy development encourages course-level declarations and structured and environmental data science reports assessment formats that process and verification [5, 14-16, These support the that credibility is enhanced when inquiry time, verification, disclosure, and transparent reasoning are integrated into research and several require to assess time in uncertainty both AI and tools to evaluate question and risks in problem studies that diversity in outputs at the studio and levels, evaluations of detection policies in multilingual and studies focused on whether disclosure and verification enhance When AI is to extend inquiry rather than the of maintaining credibility