The Metacognitive Demands and Opportunities of Generative AI

In the paper

“In a sample of undergraduate students with programming experience, working with Copilot on realistic programming tasks was perceived to be challenging (although participants still strongly preferred it) [188].”

What the sources say

• Disputed

  The Study Design section says the 30 participants were employees of a US technology company, not undergraduate students. Section 3.3 says they worked with Codex-generated completions, and section 7.2 describes the tasks as timed LeetCode challenges that may not represent day-to-day programming. These passages contradict the claim’s description of the study; they do not establish that participants strongly preferred working with Copilot. (LLM, conf=0.88)

  Checked by a reasoning model — may be wrong.

  cites [188] 188. Priyan Vaithilingam, Tianyi Zhang, and Elena L. Glassman. 2022. Expectation vs. Experience: Evaluating the Usability of Code Generation Tools Powered by Large Language Models. In Extended Abstracts of the 2022 CHI Conference on Human Factors in Computing Systems (CHI EA '22). Association for Computing Machinery, New York, NY, USA, 1-7. https://doi.org/10.1145/3491101.3519665 sourceExperimental — accuracy not yet measured

In the paper

“Their potential stems from a unique combination of model flexibility (in their input/output space), generality (in their applicability to a wide range of tasks), and originality (in their ability to generate novel content) [157].”

What the sources say

• Supported

  Section 2, “A New Kind of Cognitive Tool” (page not provided), explicitly identifies flexibility of inputs and outputs, applicability to a broad range of tasks, and generation of novel content as the three properties that make these models’ interactions unique. This directly supports the attributed claim. (LLM, conf=0.95)

  cites [157] 157. Wout Schellaert, Fernando Martínez-Plumed, Karina Vold, John Burden, Pablo A. M. Casares, Bao Sheng Loe, Roi Reichart, Sean Ó hÉigeartaigh, Anna Korho- nen, and José Hernández-Orallo. 2023. Your Prompt is My Command: On Assess- ing the Human-Centred Generality of Multimodal Models. Journal of Artificial Intelligence Research 77 (June 2023), 377-394. https://doi.org/10.1613/jair.1.14157 source

In the paper

“Studies have shown that improved metacognition helps individuals with management of time, focus, and effort [212], problem-solving [67], academic performance [45,52,106,126,180,214], emotional well-being [200], and overall decision-making [206].”

What the sources say

• ?Not checked — no full text of the sourcecites [212] 212. Barry J. Zimmerman. 2001. Theories of self-regulated learning and academic achievement: An overview and analysis. In Self-regulated learning and academic achievement: Theoretical perspectives, 2nd ed. Lawrence Erlbaum Associates Publishers, Mahwah, NJ, US, 1-37. source
• ?Not checked — no full text of the sourcecites [67] 67. Petros Georghiades. 2004. From the general to the situated: three decades of metacognition. International Journal of Science Education 26, 3 (Feb. 2004), 365-383. https://doi.org/10.1080/0950069032000119401 Publisher: Routledge _eprint: https://doi.org/10.1080/0950069032000119401. source
• Supported

  The citation to this paper is attached to academic performance, not to the sentence’s other outcomes. Section 4.1 reports that interventions with a metacognitive component improved student academic performance at posttest and follow-up. (LLM, conf=0.92)

  cites [45] 45. Hester de Boer, Anouk S. Donker, Danny D. N. M. Kostons, and Greetje P. C. van der Werf. 2018. Long-term effects of metacognitive strategy instruction on student academic performance: A meta-analysis. Educational Research Review 24 (June 2018), 98-115. https://doi.org/10.1016/j.edurev.2018.03.002 source
• Supported

  This source is cited for the academic-performance part of the sentence, not the separately cited outcomes. In §3.2, “Effective strategies” (page unavailable), the meta-analysis reports that interventions including general metacognitive knowledge or planning enhance student performance. (LLM, conf=0.92)

  cites [52] 52. A. S. Donker, H. de Boer, D. Kostons, C. C. Dignath van Ewijk, and M. P. C. van der Werf. 2014. Effectiveness of learning strategy instruction on academic performance: A meta-analysis. Educational Research Review 11 (Jan. 2014), 1-26. https://doi.org/10.1016/j.edurev.2013.11.002 source
• Partly supported

  Section 5 (Domain-Specific Theories Related to Programming Instruction; page unavailable) reports that an intervention promoting metacognitive awareness improved productivity and independence in a programming camp. Section 4.3.1 also describes a relationship between metacognitive strategies and academic outcomes. These passages support parts of the claim in educational settings, but the paper does not establish the broader benefits claimed for emotional well-being or overall decision-making. (LLM, conf=0.86)

  cites [106] 106. Jennifer A. Livingston. 2003. Metacognition: An Overview. Technical Report. https://eric.ed.gov/?id=ED474273 ERIC Number: ED474273. sourceExperimental — accuracy not yet measured
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [126] 126. Daniel Muijs and Christian Bokhove. 2020. Metacognition and Self-Regulation: Evidence Review. Technical Report. Education Endowment Foundation. https:// eric.ed.gov/?id=ED612286 Publication Title: Education Endowment Foundation ERIC Number: ED612286.
• Partly supported

  The discussion under “Classifier C” links participants’ understanding of their own expertise and confidence to effective AI-reliance strategies in a simple image-classification task. That supports a narrow connection between self-assessment and decision-making, but the paper does not show that improving metacognition improves overall decision-making. Citation [206] is attached only to the decision-making part of the citing sentence. (LLM, conf=0.82)

  cites [180] 180. Kimberly D. Tanner. 2012. Promoting Student Metacognition. CBE-Life Sciences Education 11, 2 (June 2012), 113-120. https://doi.org/10.1187/cbe.12-03-0033 Publisher: American Society for Cell Biology (lse). sourceExperimental — accuracy not yet measured
• ?Not checked — no full text of the sourcecites [214] 214. Anat Zohar and Sarit Barzilai. 2013. A review of research on metacognition in science education: current and future directions. Studies in Science Education 49, 2 (Sept. 2013), 121-169. https://doi.org/10.1080/03057267.2013.847261 Publisher: Routledge _eprint: https://doi.org/10.1080/03057267.2013.847261. source

In the paper

“Information that is easily or fluently processed (e.g., in terms of reading) triggers less further cognitive processing relative to information that is more difficult to process [3].”

What the sources say

• Partly supported

  In “Level 3: Momentary Experience” (page unavailable), the paper says easy problems can be solved quickly, while challenging problems may require substantial effort. That supports a narrower relationship between task difficulty and effort, but not the claim that fluent information generally triggers less further processing. In “Box 2” (page unavailable), it distinguishes reading-related perceptual fluency from processing fluency and reports that font readability affected neither judgments nor performance. (LLM, conf=0.82)

  cites [3] 3. Rakefet Ackerman. 2019. Heuristic Cues for Meta-Reasoning Judgments: Review and Methodology. Psihologijske teme 28, 1 (May 2019), 1-20. https://doi.org/10. 31820/pt.28.1.1 Publisher: Filozofski fakultet u Rijeci. sourceExperimental — accuracy not yet measured

In the paper

“We posit that these demands are exacerbated by GenAI's non-determinism and model flexibility (not to be confused with metacognitive flexibility) in terms of (a) the wide range of explicit and implicit parameters that users can adjust, and (b) systems' ability to work with prompts at a wide range of abstraction [156,157].”

What the sources say

• ?Could not verify

  Split into 2 atoms — 0 confirmed, 0 partially supported, 0 disputed, 0 misattributed, 2 unverifiable. Expand the claim to see per-atom verdicts.

  cites [156] 156. Advait Sarkar, Andrew D. Gordon, Carina Negreanu, Christian Poelitz, Sruti Srinivasa Ragavan, and Ben Zorn. 2022. What is it like to program with ar- tificial intelligence? https://doi.org/10.48550/arXiv.2208.06213 arXiv:2208.06213 [cs]. sourceExperimental — accuracy not yet measured
• Partly supported

  Split into 2 atoms — 0 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 1 unverifiable. Expand the claim to see per-atom verdicts.

  cites [157] 157. Wout Schellaert, Fernando Martínez-Plumed, Karina Vold, John Burden, Pablo A. M. Casares, Bao Sheng Loe, Roi Reichart, Sean Ó hÉigeartaigh, Anna Korho- nen, and José Hernández-Orallo. 2023. Your Prompt is My Command: On Assess- ing the Human-Centred Generality of Multimodal Models. Journal of Artificial Intelligence Research 77 (June 2023), 377-394. https://doi.org/10.1613/jair.1.14157 sourceExperimental — accuracy not yet measured

In the paper

“End-user studies suggest that prompting is challenging, with nonexpert users making various errors and adopting ineffective strategies-a reflection of the demand on users' metacognitive monitoring and control [33,41,43,85,103,174,205].”

What the sources say

• ?Could not verify

  Sections 5 and 6 discuss how human intuitions affect the use of machine explanations and agreement with model predictions, not how users write prompts. Searches for prompting errors, ineffective strategies, and metacognitive monitoring found no passage addressing those claims in our copy. (LLM, conf=0.76)

  cites [33] 33. Chacha Chen, Shi Feng, Amit Sharma, and Chenhao Tan. 2023. Machine Expla- nations and Human Understanding. https://doi.org/10.48550/arXiv.2202.04092 arXiv:2202.04092 [cs]. sourceExperimental — accuracy not yet measured
• Partly supported

  Section 8.4 (page unavailable) reports that participants struggled to formulate instructions, and Section 8.3 describes one participant who had little success with instructions and changed approach. These passages support prompting being challenging for some non-expert users, but they do not document various user errors or establish ineffective strategies as a general finding. The paper also does not characterize these difficulties as demands on metacognitive monitoring and control. (LLM, conf=0.84)

  cites [41] 41. Hai Dang, Sven Goller, Florian Lehmann, and Daniel Buschek. 2023. Choice Over Control: How Users Write with Large Language Models using Diegetic and Non-Diegetic Prompting. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (CHI '23). Association for Computing Machinery, New York, NY, USA, 1-17. https://doi.org/10.1145/3544548.3580969 sourceExperimental — accuracy not yet measured
• Partly supported

  Section 5.1 (p. 3) says prompt design is largely trial and error for users, and §5.2 (p. 3) describes hidden rules that can make effective prompting difficult. The paper does not report end-user studies showing non-experts making particular errors or using ineffective strategies, nor does it explain the difficulty in terms of metacognitive monitoring and control. (LLM, conf=0.85)

  cites [43] 43. Hai Dang, Lukas Mecke, Florian Lehmann, Sven Goller, and Daniel Buschek. 2022. How to Prompt? Opportunities and Challenges of Zero-and Few-Shot Learning for Human-AI Interaction in Creative Applications of Generative Models. https://doi.org/10.48550/arXiv.2209.01390 arXiv:2209.01390 [cs]. sourceExperimental — accuracy not yet measured
• ?Not checked — no full text of the sourcecites [85] 85. Ellen Jiang, Edwin Toh, Alejandra Molina, Kristen Olson, Claire Kayacik, Aaron Donsbach, Carrie J Cai, and Michael Terry. 2022. Discovering the Syntax and Strategies of Natural Language Programming with Generative Language Models. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems (CHI '22). Association for Computing Machinery, New York, NY, USA, 1-19. https://doi.org/10.1145/3491102.3501870 source
• ?Could not verify

  The available copy’s abstract (p. 2) discusses human-centered LLM transparency but does not address end-user prompting errors, ineffective strategies, or metacognitive monitoring and control. Searches for those terms returned only the abstract and publication metadata; the linked article text is not present in the indexed copy, so the claim cannot be checked here. (LLM, conf=0.75)

  cites [103] 103. Jenny T. Liang, Chenyang Yang, and Brad A. Myers. 2023. Understanding the Usability of AI Programming Assistants. https://doi.org/10.48550/arXiv.2303. 17125 arXiv:2303.17125 [cs]. sourceExperimental — accuracy not yet measured
• ?Not checked — no full text of the sourcecites [174] 174. Jiao Sun, Q. Vera Liao, Michael Muller, Mayank Agarwal, Stephanie Houde, Kartik Talamadupula, and Justin D. Weisz. 2022. Investigating Explainability of Generative AI for Code through Scenario-based Design. In 27th International Conference on Intelligent User Interfaces (IUI '22). Association for Computing Machinery, New York, NY, USA, 212-228. https://doi.org/10.1145/3490099. 3511119 source
• Partly supported

  Sections 6.2 and 6.3 describe prompting as deceptively tricky and say that expressing one's needs to an LLM requires prompt-programming expertise. The passages do not report nonexpert users making various prompting errors or adopting ineffective strategies, and they do not attribute the difficulty to demands on metacognitive monitoring and control. (LLM, conf=0.86)

  cites [205] 205. Frank F. Xu, Bogdan Vasilescu, and Graham Neubig. 2022. In-IDE Code Generation from Natural Language: Promise and Challenges. ACM Trans- actions on Software Engineering and Methodology 31, 2 (March 2022), 29:1-29:47. https://doi.org/10.1145/3487569 sourceExperimental — accuracy not yet measured

In the paper

“Metacognitive demand contributes to cognitive load, but so do other aspects related to cognitive processing at the object (non-meta) level (i.e., metacognitive demand is sufficient but not necessary for increasing cognitive load).”

What the sources say

• ?Could not verify

  The paper’s “Introducing Metacognition” section describes planning, monitoring, and evaluating learning, but does not address whether metacognitive demand increases cognitive load or whether object-level processing can increase it independently. Searches for cognitive load, metacognitive demand, object-level processing, and mental effort found no passage establishing that relationship. (LLM, conf=0.74)

  cites [178] 178. John Sweller, Jeroen J. G. van Merrienboer, and Fred G. W. C. Paas. 1998. Cogni- tive Architecture and Instructional Design. Educational Psychology Review 10, 3 (Sept. 1998), 251-296. https://doi.org/10.1023/A:1022193728205 sourceExperimental — accuracy not yet measured

In the paper

“It is a hallmark of creative problem-solving [144] and has been deemed essential for organizing and integrating a rapidly changing body of information [120].”

What the sources say

• Supported

  The citation to Preiss is attached to the first clause, about creative problem-solving, rather than the later claim about organizing information, which carries a different citation. In section 3, Preiss says creative achievement rests on cognitive flexibility guided by metacognitive strategies and describes shifting attentional focus as necessary to complete the creative process. (LLM, conf=0.88)

  cites [144] 144. David D. Preiss. 2022. Metacognition, Mind Wandering, and Cognitive Flexibility: Understanding Creativity. Journal of Intelligence 10, 3 (Sept. 2022), 69. https: //doi.org/10.3390/jintelligence10030069 Number: 3 Publisher: Multidisciplinary Digital Publishing Institute. source
• ?Not checked — no full text of the sourcecites [120] 120. Lynn Meltzer. 2014. Teaching Executive Functioning Processes: Promoting Metacognition, Strategy Use, and Effort. In Handbook of Executive Functioning, Sam Goldstein and Jack A. Naglieri (Eds.). Springer, New York, NY, 445-473. https://doi.org/10.1007/978-1-4614-8106-5_25 source

In the paper

“Similarly, manufacturing designers cocreating with GenAI were uncertain about how to interpret outputs and whether users or the system were responsible for addressing errors [70].”

What the sources say

• Supported

  Section 5.2.1 reports that designers struggled to understand generated results and determine whether unusual features were intentional or glitches. The same discussion says they were sometimes uncertain whether the AI or the user was responsible for fixing a problem. Page numbers were not supplied. (LLM, conf=0.94)

  cites [70] 70. Frederic Gmeiner, Humphrey Yang, Lining Yao, Kenneth Holstein, and Nikolas Martelaro. 2023. Exploring Challenges and Opportunities to Support Designers in Learning to Co-create with AI-based Manufacturing Design Tools. In Pro- ceedings of the 2023 CHI Conference on Human Factors in Computing Systems (CHI '23). Association for Computing Machinery, New York, NY, USA, 1-20. https://doi.org/10.1145/3544548.3580999 source

In the paper

“For example, whereas diegetic prompting is easier, it affords less control than the alternative, with preferences differing across users [41].”

What the sources say

• Supported

  Section 9.3.2 says effective non-diegetic prompts can be hard to write. Section 7.5.2 reports that instructions significantly improved users’ ability to influence suggested text, while Section 8.3 records users who preferred diegetic prompting and others who disliked it. Together, these support the stated trade-off and differing preferences. (LLM, conf=0.91)

  cites [41] 41. Hai Dang, Sven Goller, Florian Lehmann, and Daniel Buschek. 2023. Choice Over Control: How Users Write with Large Language Models using Diegetic and Non-Diegetic Prompting. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (CHI '23). Association for Computing Machinery, New York, NY, USA, 1-17. https://doi.org/10.1145/3544548.3580969 source

In the paper

“For example, in AI-assisted decision-making in chess, participants' reliance on the AI was only significantly predicted by their self-confidence, and not by their confidence in the AI [35].”

What the sources say

• Supported

  Section 3.3 reports that, in the chess-puzzle study, self-confidence significantly predicted whether participants accepted AI suggestions (P<0.001), while confidence in the AI did not (P=0.5). (LLM, conf=0.95)

  cites [35] 35. Leah Chong, Guanglu Zhang, Kosa Goucher-Lambert, Kenneth Kotovsky, and Jonathan Cagan. 2022. Human confidence in artificial intelligence and in them- selves: The evolution and impact of confidence on adoption of AI advice. Com- puters in Human Behavior 127 (Feb. 2022), 107018. https://doi.org/10.1016/j.chb. 2021.107018 source

In the paper

“This dovetails with He et al. [77] who find that poor performers in a logical reasoning task tend to be overconfident-the Dunning-Kruger effect-leading to under-reliance on AI”

What the sources say

• Supported

  Section 5.3 reports that 46 of 63 low-performing participants overestimated their performance and identifies this pattern as the Dunning–Kruger effect. The RQ1 discussion says participants with DKE relied less on AI systems; Section 3.1 identifies the task as logical reasoning. (LLM, conf=0.92)

  cites [77] 77. Gaole He, Lucie Kuiper, and Ujwal Gadiraju. 2023. Knowing About Knowing: An Illusion of Human Competence Can Hinder Appropriate Reliance on AI Systems. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (CHI '23). Association for Computing Machinery, New York, NY, USA, 1-18. https://doi.org/10.1145/3544548.3581025 source

In the paper

“Likewise, end-user programmers have been reported to”

What the sources say

• ?Could not verify

  The Results (Measuring Information Search; page unavailable) concern confidence and information-seeking in a perceptual decision task. The Discussion (page unavailable) says it is unknown whether that behavior represents real-world search. Searches for end-user programmers, AI outputs, natural-language queries, and overconfidence in program accuracy found no passage addressing the attributed claim. (LLM, conf=0.76)

  cites [156] 156. Advait Sarkar, Andrew D. Gordon, Carina Negreanu, Christian Poelitz, Sruti Srinivasa Ragavan, and Ben Zorn. 2022. What is it like to program with ar- tificial intelligence? https://doi.org/10.48550/arXiv.2208.06213 arXiv:2208.06213 [cs]. sourceExperimental — accuracy not yet measured
• ?Could not verify

  The Challenge 2 passages discuss how people assess and rely on AI advice, and the Challenge 3 passage discusses natural-language dialogue with AI. Neither addresses end-user programmers ‘eyeballing’ query outputs or suggests that doing so deepens overconfidence in their programs’ accuracy. Searches for those terms found no relevant passage in our copy. (LLM, conf=0.76)

  cites [169] 169. Sruti Srinivasa Ragavan, Zhitao Hou, Yun Wang, Andrew D Gordon, Haidong Zhang, and Dongmei Zhang. 2022. GridBook: Natural Language Formulas for the Spreadsheet Grid. In 27th International Conference on Intelligent User Interfaces (IUI '22). Association for Computing Machinery, New York, NY, USA, 345-368. https://doi.org/10.1145/3490099.3511161 sourceExperimental — accuracy not yet measured

In the paper

“They can also help them encode information in personalized representational schemas [141], which can help users understand the underlying mechanisms of a GenAI system, its capabilities, and failure points.”

What the sources say

• Partly supported

  Section 5.4 (Design Implications; page not available) says users build a mental model of how Copilot works and recommends explanations that help them see into its “black box.” It does not say that the approaches help users encode information in personalized representational schemas, the specific point to which the citation is attached. (LLM, conf=0.79)

  cites [141] 141. Peter Pirolli and Stuart Card. 2005. The Sensemaking Process and Leverage Points for Analyst Technology as Identified Through Cognitive Task Analysis. In Proceedings of international conference on intelligence analysis, Vol. 5. 2-4.Experimental — accuracy not yet measured

In the paper

“For instance, they can introduce subtle, non-intuitive errors that a human would not introduce, further complicating evaluation [156].”

What the sources say

• ?Could not verify

  The Discussion reports that dogmatic participants sought less information under uncertainty and made less accurate judgments; it does not discuss systems introducing subtle errors unlike those a human would make. Searches for the claim’s distinctive wording and for automated-system errors found no passage addressing that comparison or its effect on evaluation. (LLM, conf=0.76)

  cites [156] 156. Advait Sarkar, Andrew D. Gordon, Carina Negreanu, Christian Poelitz, Sruti Srinivasa Ragavan, and Ben Zorn. 2022. What is it like to program with ar- tificial intelligence? https://doi.org/10.48550/arXiv.2208.06213 arXiv:2208.06213 [cs]. sourceExperimental — accuracy not yet measured

In the paper

“Planning can also help users address the 'fuzzy abstraction matching' problem [156], that is, translate their goals and intentions into executable actions-a form of externalization where 'tacit' knowledge is made into explicit prompts [130].”

What the sources say

• ?Could not verify

  The Discussion describes how confidence guides information search in a perceptual decision task, not how planning turns tacit knowledge into explicit prompts. Searches for planning, fuzzy abstraction matching, and translation of goals into executable actions found no passage addressing the claim. The 'fuzzy abstraction matching' phrase is attached to a different citation in the citing sentence. (LLM, conf=0.76)

  cites [156] 156. Advait Sarkar, Andrew D. Gordon, Carina Negreanu, Christian Poelitz, Sruti Srinivasa Ragavan, and Ben Zorn. 2022. What is it like to program with ar- tificial intelligence? https://doi.org/10.48550/arXiv.2208.06213 arXiv:2208.06213 [cs]. sourceExperimental — accuracy not yet measured
• ?Not checked — no full text of the sourcecites [130] 130. Ikujiro Nonaka and Ryoko Toyama. 2003. The knowledge-creating the- ory revisited: knowledge creation as a synthesizing process. Knowl- edge Management Research & Practice 1, 1 (July 2003), 2-10. https://doi. org/10.1057/palgrave.kmrp.8500001 Publisher: Taylor & Francis _eprint: https://doi.org/10.1057/palgrave.kmrp.8500001. source

In the paper

“Chaining also increased users' self-awareness of their goals: the ability to decompose tasks led some participants to create more generalizable outputs better suited to their broader goals [203].”

What the sources say

• ?Could not verify

  The Results section reports that some participants suggested an interactive plugin that could help decompose complex functionality into smaller steps. It does not report chaining increasing users’ self-awareness or producing more generalizable outputs suited to broader goals; searches for those terms found no such finding in our copy. Page numbers were not available in the indexed passages. (LLM, conf=0.78)

  cites [203] 203. Tongshuang Wu, Ellen Jiang, Aaron Donsbach, Jeff Gray, Alejandra Molina, Michael Terry, and Carrie J Cai. 2022. PromptChainer: Chaining Large Lan- guage Model Prompts through Visual Programming. In Extended Abstracts of the 2022 CHI Conference on Human Factors in Computing Systems (CHI EA '22). Association for Computing Machinery, New York, NY, USA, 1-10. https://doi.org/10.1145/3491101.3519729 sourceExperimental — accuracy not yet measured

In the paper

“This is arguably a necessary trade-off within current GenAI systems, as it enables diversity of output [136,156], yet it exacerbates the challenge of confidence adjustment, particularly when working iteratively across prompting and output evaluation (as per §3.1).”

What the sources say

• ?Could not verify

  The citation is attached to the claim that a GenAI trade-off enables diversity of output. The Introduction and §5.1 discuss assessing algorithmic accuracy, not GenAI output diversity; searches for varied outputs and iterative prompting found no passage addressing that point. The later confidence-adjustment clause refers to the citing paper's §3.1 rather than this source. (LLM, conf=0.77)

  cites [136] 136. Shuyin Ouyang, Jie M. Zhang, Mark Harman, and Meng Wang. 2023. LLM is Like a Box of Chocolates: the Non-determinism of ChatGPT in Code Generation. http://arxiv.org/abs/2308.02828 arXiv:2308.02828 [cs]. sourceExperimental — accuracy not yet measured
• ?Could not verify

  The citation follows the claim that a trade-off in current GenAI systems enables diversity of output. The Discussion addresses uncertainty-guided information search and an exploration–exploitation trade-off in human decision-making, not GenAI output diversity. Searches for GenAI output diversity, prompting, and confidence adjustment found no passage addressing that claim. (LLM, conf=0.76)

  cites [156] 156. Advait Sarkar, Andrew D. Gordon, Carina Negreanu, Christian Poelitz, Sruti Srinivasa Ragavan, and Ben Zorn. 2022. What is it like to program with ar- tificial intelligence? https://doi.org/10.48550/arXiv.2208.06213 arXiv:2208.06213 [cs]. sourceExperimental — accuracy not yet measured

In the paper

“Sarkar [154] describes this change as a shift from production to 'critical integration', where”

What the sources say

• Partly supported

  Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  cites [154] 154. Advait Sarkar. 2023. Exploring Perspectives on the Impact of Artificial In- telligence on the Creativity of Knowledge Work: Beyond Mechanised Pla- giarism and Stochastic Parrots. In Proceedings of the 2nd Annual Meeting of the Symposium on Human-Computer Interaction for Work (CHIWORK '23). Association for Computing Machinery, New York, NY, USA, 1-17. https: //doi.org/10.1145/3596671.3597650 sourceExperimental — accuracy not yet measured

In the paper

“Alongside improving the LLM's ability to execute complex tasks, chaining helped participants “think through the task better”, and thereby make more targeted edits to improve their prompting [204].”

What the sources say

• ?Could not verify

  The ‘Shared limitations’ section discusses human decisions chained into sequences serving longer-term goals, not chaining LLM prompts. The ‘Impact of error monitoring on behaviour’ section discusses adjustments after errors, but neither it nor passages retrieved by searching for prompting and targeted edits addresses participants making more targeted prompt edits. (LLM, conf=0.78)

  cites [204] 204. Tongshuang Wu, Michael Terry, and Carrie Jun Cai. 2022. AI Chains: Trans- parent and Controllable Human-AI Interaction by Chaining Large Language Model Prompts. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems (CHI '22). Association for Computing Machinery, New York, NY, USA, 1-22. https://doi.org/10.1145/3491102.3517582 sourceExperimental — accuracy not yet measured

In the paper

“Enabling users to critically evaluate AI outputs can build self-awareness by forcing users to rationalize their decisions to the GenAI system [156], enhance metacognitive flexibility by providing users with different perspectives on their task [15], and adjust confidence by adapting explanations to the users' needs and the models' confidence [110].”

What the sources say

• ?Could not verify

  Split into 3 atoms — 0 confirmed, 0 partially supported, 0 disputed, 0 misattributed, 1 unverifiable, 2 not checked. Expand the claim to see per-atom verdicts.

  cites [156] 156. Advait Sarkar, Andrew D. Gordon, Carina Negreanu, Christian Poelitz, Sruti Srinivasa Ragavan, and Ben Zorn. 2022. What is it like to program with ar- tificial intelligence? https://doi.org/10.48550/arXiv.2208.06213 arXiv:2208.06213 [cs]. sourceExperimental — accuracy not yet measured
• Partly supported

  Split into 3 atoms — 0 confirmed, 3 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  cites [15] 15. Frederic Becker, Maria Wirzberger, Viktoria Pammer-Schindler, Srinidhi Srini- vas, and Falk Lieder. 2023. Systematic metacognitive reflection helps people discover far-sighted decision strategies: A process-tracing experiment. Judgment and Decision Making 18 (Jan. 2023), e15. https://doi.org/10.1017/jdm.2023.16 Publisher: Cambridge University Press. sourceExperimental — accuracy not yet measured
• Partly supported

  Split into 3 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 1 unverifiable. Expand the claim to see per-atom verdicts.

  cites [110] 110. Shuai Ma, Ying Lei, Xinru Wang, Chengbo Zheng, Chuhan Shi, Ming Yin, and Xiaojuan Ma. 2023. Who Should I Trust: AI or Myself? Leveraging Human and AI Correctness Likelihood to Promote Appropriate Trust in AI-Assisted Decision-Making. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (CHI '23). Association for Computing Machinery, New York, NY, USA, 1-19. https://doi.org/10.1145/3544548.3581058 sourceExperimental — accuracy not yet measured

In the paper

“Indeed, in Perry et al. [139], half of the users did not adjust any model parameters, even though many produced insecure code using an AI assistant.”

What the sources say

• Partly supported

  Section 6.2 (Prompt Parameters) reports that 48.5% of participants never adjusted temperature and 51.5% never adjusted response length. It does not establish that half adjusted neither parameter, as the claim implies. Section 1 (Introduction) reports that participants with an AI assistant wrote insecure solutions more often than controls on four of five tasks. (LLM, conf=0.88)

  cites [139] 139. Neil Perry, Megha Srivastava, Deepak Kumar, and Dan Boneh. 2022. Do Users Write More Insecure Code with AI Assistants? https://doi.org/10.48550/arXiv. 2211.03622 arXiv:2211.03622 [cs].Experimental — accuracy not yet measured

In the paper

“Park et al. [138] similarly argue that “slower” interfaces especially enable these benefits, as the waiting time often gets used for reflective thinking about the task at hand (see also [147]).”

What the sources say

• ?Not checked — no full text of the sourcecites [138] 138. Joon Sung Park, Rick Barber, Alex Kirlik, and Karrie Karahalios. 2019. A Slow Al- gorithm Improves Users' Assessments of the Algorithm's Accuracy. Proceedings of the ACM on Human-Computer Interaction 3, CSCW (Nov. 2019), 102:1-102:15. https://doi.org/10.1145/3359204 source
• ?Could not verify

  Section 3, “The Dawn of a New Age of Metacognition Research?” (page unavailable), discusses monitoring and control when people use digital tools. It does not say that slower interfaces enable benefits or that people use waiting time to reflect on a task; searches for those ideas and for delay or latency found no supporting passage. (LLM, conf=0.72)

  cites [147] 147. Leon Reicherts and Yvonne Rogers. 2020. Do Make me Think! How CUIs Can Support Cognitive Processes. In Proceedings of the 2nd Conference on Conver- sational User Interfaces (CUI '20). Association for Computing Machinery, New York, NY, USA, 1-4. https://doi.org/10.1145/3405755.3406157 sourceExperimental — accuracy not yet measured

In the paper

“Explainability for GenAI systems should help users adjust their confidence in their ability to prompt, evaluate outputs, and determine their automation strategy, particularly given the multiple, non-intuitive failure modes common to GenAI systems, and other challenges [104].”

What the sources say

• Partly supported

  Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  cites [104] 104. Q. Vera Liao and Jennifer Wortman Vaughan. 2023. AI Transparency in the Age of LLMs: A Human-Centered Research Roadmap. https://doi.org/10.48550/ arXiv.2306.01941 arXiv:2306.01941 [cs]. sourceExperimental — accuracy not yet measured

In the paper

“For example, coding assistance systems can detect when a user is in a state of flow ('acceleration') or problem-solving ('exploration'), adapting code suggestions accordingly and providing feedback to the user [14].”

What the sources say

• Partly supported

  Section 4 describes acceleration as programming within the user's flow and exploration as using Copilot to help plan code actions. Section 6.2 proposes that a future assistant could become aware of the current mode and adjust its suggestions, but presents this as a conjecture, not an existing detection capability. The cited passages do not establish that such a system provides feedback to the user. (LLM, conf=0.88)

  cites [14] 14. Shraddha Barke, Michael B. James, and Nadia Polikarpova. 2023. Grounded Copilot: How Programmers Interact with Code-Generating Models. Proceedings of the ACM on Programming Languages 7, OOPSLA1 (April 2023), 78:85-78:111. https://doi.org/10.1145/3586030 sourceExperimental — accuracy not yet measured

In the paper

“To this point, [133] found that adaptive and gradual fading out of metacognitive prompts produced the largest performance benefits, as it provided time for students to internalize metacognitive strategies, while ultimately reducing the cognitive load associated with processing now-irrelevant external prompts (see also [134]).”

What the sources say

• ?Not checked — no full text of the sourcecites [133] 133. Matthias Nückles, Sandra Hübner, Sandra Dümer, and Alexander Renkl. 2010. Expertise reversal effects in writing-to-learn. Instructional Science 38, 3 (May 2010), 237-258. https://doi.org/10.1007/s11251-009-9106-9 source
• ?Could not verify

  Split into 2 atoms — 0 confirmed, 0 partially supported, 0 disputed, 0 misattributed, 2 unverifiable. Expand the claim to see per-atom verdicts.

  cites [134] 134. Matthias Nückles, Julian Roelle, Inga Glogger-Frey, Julia Waldeyer, and Alexan- der Renkl. 2020. The Self-Regulation-View in Writing-to-Learn: Using Journal Writing to Optimize Cognitive Load in Self-Regulated Learning. Educational Psychology Review 32, 4 (Dec. 2020), 1089-1126. https://doi.org/10.1007/s10648- 020-09541-1 sourceExperimental — accuracy not yet measured

In the paper

“Designing truly human-centered GenAI systems [34,104] means grappling with these metacognitive demands.”

What the sources say

• Supported

  The citation follows “human-centered GenAI systems,” and the cited paper is explicitly about their design. Section 5.3.1 says designers should consider users’ mental models when employing generative AI; Section 2.2 also discusses uncertain human intentions and the need for appropriate user control. “Metacognitive demands” is the citing authors’ framing, rather than a term used in these passages. (LLM, conf=0.87)

  cites [34] 34. Xiang 'Anthony' Chen, Jeff Burke, Ruofei Du, Matthew K. Hong, Jennifer Jacobs, Philippe Laban, Dingzeyu Li, Nanyun Peng, Karl D. D. Willis, Chien-Sheng Wu, and Bolei Zhou. 2023. Next Steps for Human-Centered Generative AI: A Tech- nical Perspective. https://doi.org/10.48550/arXiv.2306.15774 arXiv:2306.15774 [cs]. source
• Partly supported

  Section 8.6 (Implications for design; page unavailable) recommends helping users develop mental models of how to command the system. Section 6.4 describes the decisions users make after model failures. These findings support attending to users’ cognitive demands in the code-generating interface studied, but do not establish the claim for GenAI systems generally. (LLM, conf=0.82)

  cites [104] 104. Q. Vera Liao and Jennifer Wortman Vaughan. 2023. AI Transparency in the Age of LLMs: A Human-Centered Research Roadmap. https://doi.org/10.48550/ arXiv.2306.01941 arXiv:2306.01941 [cs]. sourceExperimental — accuracy not yet measured

In the paper

“Ultimately, however, as users adapt to working with explainable GenAI systems, we hypothesize that the result should be a net reduction in cognitive load [134].”

What the sources say

• ?Could not verify

  The paper’s Future work section (p. 24) discusses balancing determinism and creativity, while its Implications section (p. 23) discusses the reliability of generated code. Neither addresses whether users adapting to explainable GenAI systems experience a net reduction in cognitive load. Searches for “cognitive load” and related terms found no passage supporting that prediction. (LLM, conf=0.76)

  cites [134] 134. Matthias Nückles, Julian Roelle, Inga Glogger-Frey, Julia Waldeyer, and Alexan- der Renkl. 2020. The Self-Regulation-View in Writing-to-Learn: Using Journal Writing to Optimize Cognitive Load in Self-Regulated Learning. Educational Psychology Review 32, 4 (Dec. 2020), 1089-1126. https://doi.org/10.1007/s10648- 020-09541-1 sourceExperimental — accuracy not yet measured

In the paper

“Apart from facilitating engagement at opportune moments, the sequential paradigm is presumed to encourage independent reflection by the user.”

What the sources say

• Supported

  In Challenge 3, under “When should AI assistance be presented to the human DM?”, the paper says the sequential paradigm gives users AI advice only after they make an independent decision. It explicitly says the paradigm is presumed to encourage independent reflection, while noting that studies differ on whether it improves overall performance. (LLM, conf=0.94)

  cites [171] 171. Mark Steyvers and Aakriti Kumar. 2023. Three Challenges for AI-Assisted Decision-Making. Perspectives on Psychological Science: A Journal of the As- sociation for Psychological Science (July 2023), 17456916231181102. https: //doi.org/10.1177/17456916231181102 source

In the paper

“Another approach is designing the complexity of AI-generated content according to the cognitive load experienced by the user [177].”

What the sources say

• ?Could not verify

  The Introduction (p. 1) describes analysis of Stable Diffusion’s image synthesis, and the Generalized Attribution section (p. 5) describes people rating word–image heat maps. Searches for cognitive load and adapting generated-content complexity found no passage addressing the claimed approach or its role in self-management. (LLM, conf=0.77)

  cites [177] 177. John Sweller. 1988. Cognitive Load During Problem Solv- ing: Effects on Learning. Cognitive Science 12, 2 (1988), 257- 285. https://doi.org/10.1207/s15516709cog1202_4 _eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1207/s15516709cog1202_4. sourceExperimental — accuracy not yet measured

In the paper

“For example, [1] and [164] suggest that in the context of solving logical puzzles, intelligent tutors should offer backwards-oriented workflows (e.g., using prompts to encourage thinking about the negation of the actual solution), rather than focusing on forward-oriented workflows.”

What the sources say

• Partly supported

  The “Two Tutors and Our Interventions” section describes a logic tutor that offers backward chaining, including prompts to switch from its default forward-chaining approach; backward chaining can involve negating the conclusion. But the Abstract says no single strategy is always optimal, and the paper investigates teaching students when to use each strategy. It does not recommend backward-oriented workflows in place of a focus on forward-oriented ones generally. (LLM, conf=0.86)

  cites [1] 1. Mark Abdelshiheed, John Wesley Hostetter, Preya Shabrina, Tiffany Barnes, and Min Chi. 2023. The Power of Nudging: Exploring Three Interventions for Metacognitive Skills Instruction across Intelligent Tutoring Systems. https: //doi.org/10.48550/arXiv.2303.11965 arXiv:2303.11965 [cs]. sourceExperimental — accuracy not yet measured
• Partly supported

  Section 3.3 recommends that AI systems respect a task’s natural sequence and avoid suggestions that interfere with it. The paper discusses this in the context of programming, however; it does not recommend backwards-oriented workflows for logical-puzzle tutors or compare them with forward-oriented workflows. (LLM, conf=0.76)

  cites [164] 164. Preya Shabrina, Behrooz Mostafavi, Mark Abdelshiheed, Min Chi, and Tiffany Barnes. 2023. Investigating the Impact of Backward Strategy Learning in a Logic Tutor: Aiding Subgoal Learning Towards Improved Problem Solving. International Journal of Artificial Intelligence in Education (Aug. 2023). https: //doi.org/10.1007/s40593-023-00338-1 sourceExperimental — accuracy not yet measured

In the paper

“For example, explanations that map each aspect of an output to aspects of the prompt (e.g., using attention visualization [179]), and compare this to examples of effective prompts [25,85], can help users disentangle issues with their prompt from those stemming from model performance, thereby supporting confidence adjustment for prompting ability.”

What the sources say

• ?Not checked — no full text of the sourcecites [179] 179. Raphael Tang, Linqing Liu, Akshat Pandey, Zhiying Jiang, Gefei Yang, Karun Kumar, Pontus Stenetorp, Jimmy Lin, and Ferhan Ture. 2022. What the DAAM: Interpreting Stable Diffusion Using Cross Attention. https://doi.org/10.48550/ arXiv.2210.04885 arXiv:2210.04885 [cs]. source
• Partly supported

  Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  cites [25] 25. Michelle Brachman, Qian Pan, Hyo Jin Do, Casey Dugan, Arunima Chaudhary, James M. Johnson, Priyanshu Rai, Tathagata Chakraborti, Thomas Gschwind, Jim A Laredo, Christoph Miksovic, Paolo Scotton, Kartik Talamadupula, and Gegi Thomas. 2023. Follow the Successful Herd: Towards Explanations for Improved Use and Mental Models of Natural Language Systems. In Proceedings of the 28th International Conference on Intelligent User Interfaces (IUI '23). Association for Computing Machinery, New York, NY, USA, 220-239. https://doi.org/10.1145/ 3581641.3584088 sourceExperimental — accuracy not yet measured
• ?Not checked — no full text of the sourcecites [85] 85. Ellen Jiang, Edwin Toh, Alejandra Molina, Kristen Olson, Claire Kayacik, Aaron Donsbach, Carrie J Cai, and Michael Terry. 2022. Discovering the Syntax and Strategies of Natural Language Programming with Generative Language Models. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems (CHI '22). Association for Computing Machinery, New York, NY, USA, 1-19. https://doi.org/10.1145/3491102.3501870 source

In the paper

“To this point, §3 demonstrated how users struggled to understand GenAI systems and achieve their goals due to GenAI's metacognitive demands.”

In the paper

“We argue that metacognition-the psychological ability to monitor and control one's thoughts and behavior-offers a valuable lens to understand and design for these usability challenges.”

In the paper

“We propose these demands could be addressed by integrating metacognitive support strategies into GenAI systems, and by designing GenAI systems to reduce their metacognitive demand by targeting explainability and customizability.”

In the paper

“We propose these demands could be addressed by integrating metacognitive support strategies into GenAI systems, and by designing GenAI systems to reduce their metacognitive demand by targeting explainability and customizability.”

In the paper

“Metacognition offers a coherent framework for understanding the usability challenges posed by GenAI, and provides novel research and design directions to advance human-AI interaction.”

In the paper

“Firstly, we suggest that current GenAI systems impose multiple metacognitive demands on users; understanding these demands can help interpret and probe the identified and potentially novel usability challenges.”

In the paper

“During prompt formulation, the open-endedness of many current prompting interfaces requires users to have self-awareness of their specific task goals, and be able to decompose their tasks into smaller sub-tasks so as to verbalize these as effective prompts (Figure 2a).”

In the paper

“This demands self-awareness of GenAI's applicability to, and impact on, one's workflow; well-adjusted confidence in manual versus GenAI-supported task completion; and metacognitive flexibility to adapt one's workflows as needed.”

In the paper

“we suggest that a metacognitive lens offers new research and design directions for augmenting GenAI system usability.”

In the paper

“Finally, we discuss the relationship between cognitive load and addressing metacognitive demands, offering ways to manage their balance.”

In the paper

“Analogously, current GenAI systems often require verbalized prompting, demanding self-awareness of task goals, and decomposition of tasks into sub-tasks.”

In the paper

“Not all GenAI systems impose the same type and extent of metacognitive demands due to differences in interface design and interaction modes; where relevant, we point out the implications of this.”

In the paper

“Evidencebased metacognitive support strategies include those that help users in their planning, self-evaluation, and self-management [159].”

What the sources say

• ?Not checked — no full text of the sourcecites [159] 159. Dale H. Schunk and Peggy A. Ertmer. 2000. Self-regulation and academic learning: Self-efficacy enhancing interventions. In Handbook of self-regulation. Academic Press, San Diego, CA, US, 631-649. https://doi.org/10.1016/B978- 012109890-2/50048-2 source

In the paper

“Secondly, we suggest that the perspective of metacognitive demands offers new research and design opportunities for human-AI interaction.”

In the paper

“We suggest that explainability can help offload metacognitive processing from the user to the system, and that existing explainability approaches can be augmented by considering metacognition.”

In the paper

“In manual task completion, many implicit goals and intentions embedded within tasks can remain so without ever being verbalized.”

In the paper

“Likewise, we suggest that a metacognitive perspective can provide insights on approaching the end-user customizability of GenAI systems.”

In the paper

“Metacognitive knowledge and experiences are interrelated [64].”

What the sources say

• ?Not checked — no full text of the sourcecites [64] 64. John H. Flavell. 1979. Metacognition and cognitive monitoring: A new area of cognitive-developmental inquiry. American Psychologist 34, 10 (1979), 906- 911. https://doi.org/10.1037/0003-066X.34.10.906 Place: US Publisher: American Psychological Association. source

In the paper

“Critically, we also highlight how GenAI's model flexibility, generality, and originality can serve as a design solution to these demands.”

In the paper

“System outputs then need to be evaluated, requiring well-adjusted confidence in one's evaluation and prompting abilities, and metacognitive flexibility to iterate on the prompting strategy as necessary.”

In the paper

“Manipulating the activation of these heuristics (e.g., via priming) can be used to improve metacognitive control: increasing subjective processing difficulty (e.g., by using a degraded font) stimulates more metacognitive control and thereby improves participants' performance on reasoning tasks that benefit from a more analytic processing style [7].”

What the sources say

• ?Not checked — no full text of the sourcecites [7] 7. Adam L. Alter, Daniel M. Oppenheimer, Nicholas Epley, and Rebecca N. Eyre. 2007. Overcoming intuition: metacognitive difficulty activates analytic reason- ing. Journal of Experimental Psychology. General 136, 4 (Nov. 2007), 569-576. https://doi.org/10.1037/0096-3445.136.4.569 source

In the paper

“Ackerman [2] found that people's internal confidence threshold for solving reasoning problems decreases as the required effort increases; that is,”

What the sources say

• ?Not checked — no full text of the sourcecites [2] 2. Rakefet Ackerman. 2014. The diminishing criterion model for metacognitive regulation of time investment. Journal of Experimental Psychology: General 143, 3 (2014), 1349-1368. https://doi.org/10.1037/a0035098 Place: US Publisher: American Psychological Association. source

In the paper

“Likewise, evaluating the output requires well-adjusted confidence in one's ability to judge its validity (c).”

In the paper

“The range of possible explanations for a poor output makes confidence adjustment challenging, and the range of possible strategy adaptations demands high metacognitive flexibility from the user.”

In the paper

“However, that participants in [208]”

What the sources say

• ?Not checked — no full text of the sourcecites [208] 208. J.D. Zamfirescu-Pereira, Richmond Y. Wong, Bjoern Hartmann, and Qian Yang. 2023. Why Johnny Can't Prompt: How Non-AI Experts Try (and Fail) to Design LLM Prompts. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. ACM, Hamburg Germany, 1-21. https://doi.org/10.1145/ 3544548.3581388 source

In the paper

“Participants' insistence on a social lens may reflect a lack of self-awareness of their prompting approach, poorly adjusted confidence, and/or inflexibility in their strategies.”

In the paper

“Participants in [208] struggled with this (see also [43,85,205]).They were unable to choose the right prompting instructions, incorrectly expecting human capabilities; in some cases, underestimating the system's capabilities; and insisted on socially appropriate-rather than effective-ways of prompting.”

What the sources say

• ?Not checked — no full text of the sourcecites [208] 208. J.D. Zamfirescu-Pereira, Richmond Y. Wong, Bjoern Hartmann, and Qian Yang. 2023. Why Johnny Can't Prompt: How Non-AI Experts Try (and Fail) to Design LLM Prompts. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. ACM, Hamburg Germany, 1-21. https://doi.org/10.1145/ 3544548.3581388 source

In the paper

“Beyond the local interaction with a GenAI system, there is an overarching demand connected to understanding whether, when, and how to apply GenAI to one's workflows-one's 'automation strategy'.”

In the paper

“At the same time, systems that support users' metacognition can help surface or clarify intentions originally hidden from users' selfawareness, thereby influencing their initial goals or lack thereof (see §4.1.2for details).”

In the paper

“Finally, it also requires metacognitive flexbility to adapt one's workflows as necessary (e).”

In the paper

“This requires self-awareness of how GenAI applies to and affects one's workflows, and well-adjusted confidence in the ability to complete tasks manually and with GenAI (d).”

In the paper

“Depending on the output, iterating on the prompt may be necessary, which requires well-adjusted confidence in one's prompting ability and metacognitive flexiblity to adapt prompting strategies as necessary (b).”

In the paper

“Thus, our framework of metacognitive demands is applicable to many use-cases.”

In the paper

“For example, users 'playing' with a system may be enabled to identify more concrete or diverse forms of play for them to explore.”

In the paper

“For example, users interacting with GenAI systems for non-specific entertainment or exploration may worry less about their prompting strategy.”

In the paper

“From a metacognitive perspective, over-generalization reflects poorly adjusted confidence; rather than maintaining an appropriately low confidence (about their own prompting), and gathering more evidence, participants drew confident conclusions based on limited evidence.”

In the paper

“Thus, users must maintain a well-adjusted level of confidence in their own ability to evaluate this output and not blindly accept generated content.”

In the paper

“Even with use-cases that involve subjectivity, such as creative tasks, users need to adopt an appropriate reliance strategy, which requires welladjusted confidence.”

In the paper

“That is, users must have self-awareness of the applicability and potential impact of using GenAI for their workflow; well-adjusted confidence in the ability to complete a task manually versus with GenAI; and metacognitive flexibility in adapting workflows to GenAI (Figure 2d-e).”

In the paper

“This implies subjectivity in the workflow which, although valid, makes it challenging to adjust one's confidence.”

In the paper

“Most importantly, we hypothesize that improved metacognition should result in a net improvement in output quality.”

In the paper

“Moreover, ready-made prompts still require metacognitive ability to apply appropriately.”

In the paper

“By developing systems that are context-aware, users can be served AI-generated content or prompts at opportune moments during their workflows [74].”

What the sources say

• ?Not checked — no full text of the sourcecites [74] 74. Ken Gu, Madeleine Grunde-McLaughlin, Andrew M. McNutt, Jeffrey Heer, and Tim Althoff. 2023. How Do Data Analysts Respond to AI Assistance? A Wizard- of-Oz Study. https://doi.org/10.48550/arXiv.2309.10108 arXiv:2309.10108 [cs]. source

In the paper

“In a data science context, Gu et al. [74] propose that GenAI tools can offer “a 'think' mode for specific planning suggestions, a 'reflection' mode for connecting decisions and highlighting potential missed steps, and an 'exploration' mode for higher-level planning suggestions”.”

What the sources say

• ?Not checked — no full text of the sourcecites [74] 74. Ken Gu, Madeleine Grunde-McLaughlin, Andrew M. McNutt, Jeffrey Heer, and Tim Althoff. 2023. How Do Data Analysts Respond to AI Assistance? A Wizard- of-Oz Study. https://doi.org/10.48550/arXiv.2309.10108 arXiv:2309.10108 [cs]. source

In the paper

“Note also that self-confidence in output evaluation and confidence in the system's ability are both distinct from users' self-confidence in their prompting ability, discussed above in §3.1.”

In the paper

“A popular workaround to the challenge of prompting is 'prompt libraries' with detailed, task-specific prompts (see[176] for an overview).”

What the sources say

• ?Not checked — no full text of the sourcecites [176] 176. Adam Svendsen and Bruce Garvey. 2023. An Outline for an Interrogative/Prompt Library to help improve output quality from Generative-AI Datasets. https: //doi.org/10.2139/ssrn.4495319 source

In the paper

“Prompting and output evaluation influence each other as users iterate on their task.”

In the paper

“While helpful, ready-made prompts will rarely suit one's context precisely-the devil remains in the details.”

In the paper

“Likewise, we propose that explainability, by partly offloading metacognitive processing from users and onto the system, should reduce the cognitive load associated with metacognitive monitoring and control.”