A Scan for Research Gaps in Past Notes (2026-10, Fifteen Lenses)

Why “find the research gaps” is not enough

A request that only says “find the research gaps” returns general gaps that would fit any field. In a thread on X (2026-10-09), Hamidullah Khan divided the search for gaps into fifteen specific prompts¹. These prompts include mapping the research, comparing conflicting findings, looking for conclusions that the methods cannot support, examining mismatches in measurement, looking for unexplained mechanisms, and looking for the conditions under which a finding stops holding. The last prompt is a step in which the person who proposed a gap tries to break its novelty.

In this repository, the fifteen prompts were reorganized as lenses L1 to L15 of the research-gap-lenses skill. Each lens was labeled with one of the gap-spotting types distinguished by Sandberg and Alvesson (confusion, in which findings conflict; neglect, in which something has not been studied; and application, in which a theory from another area has not been applied) or with assumption, which questions a premise of the field². The fifteen lenses are tools for gap-spotting, and only L13 questions assumptions. Inverting assumptions and leaps of frame are handled by the methods in Problematizing the wiki corpus: an assumption ledger for four domains and twelve challenges and Where Are the Gaps in AI Research? Seven Voids Found by Reframing. The step in L14 of the original thread that ranks gaps by the duration and budget of a degree program was not used. Paper ideas are evaluated by novelty, and the cost of starting a project is not the first axis.

What was scanned, and how

The scan covered 87 academic notes that have a corpus (source/review/). Notes from the weekly watch, notes on the latest AI developments, notes on industry information, and non-public notes on the author’s own theory were excluded. The 87 notes were divided into five non-overlapping areas.

• A: AI and design practice, design professions, and design research (20 notes)
• B: creativity, AI slop, generative art, and data visualization (16 notes)
• C: learning, education, and end-user development (17 notes)
• D: competence, evaluation, cumulative advantage, social psychology, and gyaru culture (19 notes)
• E: LLM evaluation and sycophancy, field-level trends in AI research, meta-research on novelty, and AI regulation (15 notes)

For each area, one agent tabulated the research rows of the corpus (L1), applied the lenses for which the table had material (L2 to L14), and returned candidates. The agents were instructed to exclude questions already raised in 13 earlier handoffs and in 5 cross-cutting analysis notes. The agents returned 27 candidates, and merging overlaps reduced them to 10. For the 10 remaining candidates, two agents searched externally for the nearest existing studies and assigned verdicts (L15). On that day, OpenAlex and Semantic Scholar had reached their query limits and could not be used, so the search used Crossref, arXiv, Europe PMC, and web search. The bibliographic records of the nearest studies that decided a verdict were checked in Crossref or arXiv, and those whose full text could be obtained were reread in full.

List of verdicts

The axes are the five novelty axes (RW, PD, Mat, Met, Res; each scored 0 to 5, with 3 as the threshold), and the scores are provisional scores from the blog side. The final scoring is left to the novelty-auditor of paper-pipeline. paper-pipeline later discarded all of G1 to G10 (2026-11-06; the reasons for each candidate are in the paper-pipeline ledger).

Gaps and nearest studies for each candidate

G1 Makers’ own evaluation of work co-created with AI

The scans of two areas independently produced this candidate. Telling people that another person’s work was “made by AI” lowers their evaluation of it, and people overestimate things they made themselves. The question is which of these two effects is stronger when the person who had AI make a work and then selected it becomes the evaluator.

In an experiment framed as an AI art competition with prize money (432 participants), Lima, Grgić-Hlača, and Redmiles (2025) reported that people who made images with generative AI rated their own images higher than other people rated the same images. This bias appeared only in the condition with actual prize money. The authors list as a limitation that the effect of egocentricity cannot be separated from the effect of experience with the model. In the abstract of their paper, Mehler et al. (2024) report that the evaluation of generated images rises as human effort increases.

What remains is whether the size of the discrepancy changes with the share of involvement (the share of decisions and edits made by the maker) and against which standard, and by how much, the evaluation deviates. Three standards can be placed side by side: third parties told that the work was “made by AI”, third parties told that it was “made by a human”, and blinded experts. If the share of involvement at which bias against AI and the overestimation of one’s own work cancel each other can be estimated, the direction of the error when makers adopt generated outputs on their own judgment becomes predictable. This question corresponds to the “belief” loop in Do Decision-Makers Without Taste Stop Commissioning Creators? Checking the Problem of Image-Generating AI Through Three Links (Commissioning, Habituation, and the Creator Base) and Relating It to Research on AI Slop. In the check that tries to break the novelty of the candidate (L15), the types that apply are combination (crossing two known effects) and renaming (a rewording of egocentric bias).

G2 Generative AI and impasse, surprise, and reframing

Insight has been held to arise from restructuring the representation after an impasse, and reframing in design has been held to be triggered by surprise (An Academic Map of Methods for Reframing Problems: From Abduction-2 to Problem Structuring). The scan produced the candidate that “because generative AI supplies solution candidates without interruption, impasses and surprises decrease, and so does reframing”.

There are two nearest studies, and the directions of their results conflict. Shin, Polyanskaya, Lucero, and Oulasvirta (2025) assigned 280 people with design experience to four conditions and reported that, in expert ratings, LLMs increased neither the novelty nor the usefulness of reframing. In a protocol analysis of 20 design students (most of them graduate students), Guo et al. (2024) coded solution-to-problem transitions (transitions in which a proposed solution triggers a restructuring of the understanding of the problem) and reported that generative AI fostered the participants’ abductive reasoning. Neither study measured impasses or surprises.

Because the direction of “decrease” in the candidate does not agree with the report by Guo et al., this direction is not posed as a hypothesis. The question is narrowed to whether impasse and surprise events mediate between the presence of generative AI and reframing. The result that quality did not differ and the report that the process fostered abduction can both hold, because they measured different things (the quality of the outcome and the process). Measuring the mediation at the level of the process accounts for this conflict.

G3 Appropriate reliance in design judgments without a correct answer

Research on appropriate reliance on AI has built both its measures and its interventions on judgment tasks with a correct answer. The review by Eckhardt et al. (2025) analyzed 71 papers and pointed out that most used tasks with an objective correct answer and that even studies described as subjective actually asked about correctness. The same review writes that it is unclear what reliance means for generative AI. In writing, attempts have appeared to use the overlap between AI suggestions and a person’s own text as a proxy for reliance (Welzel and Vincent 2026, 47 participants). No study was found that applied appropriate reliance to design judgments or aesthetic judgments, or that defined a criterion separating excessive reliance from appropriate reliance.

The candidate transfers to design judgments the advice-taking procedure, which records a prior judgment of one’s own, the AI’s advice, and a posterior judgment in that order, and it defines appropriate reliance operationally with the distance from expert consensus as the reference point. Treating expert consensus as the correct answer risks a return to tasks with a correct answer. The range in which this reference point can be used therefore needs to be limited by the agreement within the expert panel. Because the review names this topic as part of its agenda, the risk of being preempted is moderate.

G4 Learning of the parts done oneself when part of a task is delegated to AI

Cognitive offloading has two explanations: freed resources improve other processing (reallocation of resources), and with generative AI, metacognitive evaluation itself decreases (overall disengagement) (Does Early Use of Generative AI Inhibit the Formation of Thought? A Literature Map of Cognitive Offloading and Learning, Patterns of Technological Harm to Education, and Their Refutations: A Robustness Assessment Across 25 Works). The two explanations make opposite predictions about the learning of the parts that AI did not take over.

Fan and Chang (2026) randomly assigned 744 undergraduate students to a group in which AI drafted, a group in which AI evaluated, a group in which AI revised, and a group without AI, and they tested the students after removing AI in week 6. The groups in which AI drafted or revised outperformed the group without AI, and the group in which AI evaluated performed at the same level as the group without AI. In the immediate test, overall disengagement was not observed. The authors write that what they measured was performance at the point when AI was removed, not retention over time.

What remains is to measure, for the same person and separately, the skill of the stage that was delegated and the skill of the stage done oneself, using delayed tests. Immediate overall performance cannot distinguish the two explanations.

G5 Four ways of measuring novelty and LLM ideas

Blinded experts rated research ideas from LLMs as more novel (Si et al. 2025, How a Claim of Novelty Is Verified, Defended, and Written: From Pre-Submission Checks to Bundles of Papers). In contrast, Terwiesch et al. (2026) report in their abstract that the product ideas from LLMs were lower both in the novelty of each idea and in the diversity of the set. LLMs that judge novelty by searching the literature reverse their verdicts with small differences in the prompt (Sternlicht et al. 2026, abstract).

There are at least four ways of measuring novelty. They are the degree to which raters feel an idea is new without a search, a rating made after searching the literature, rarity within the sample, and the absence of prior work. No study was found that applied these four side by side to the same set of ideas and measured whether ideas made by LLMs and ideas made by humans differ in how the four measures disagree. If the direction of the conclusion changes with the measure, the question “are LLM ideas novel?” has no determinate answer until the measure is fixed. In the check that tries to break the novelty (L15), the types that apply are renaming (a restatement of Boden’s P-creativity and H-creativity and of the two kinds of measures by Shah et al.) and topic (a neighboring question appeared only in October 2026).

G6 LLM feedback when learners submit work as their own

Sharma et al. (2024) showed that adding “I wrote this” to the prompt makes LLM feedback more positive, for math solutions, arguments, and poems (feedback sycophancy). The core of the candidate has already been answered for text.

What remains is how many of the problems named by experts, and with what weight, an LLM drops when learners submit visual design work and add ownership, attachment, or anxiety. Sharma et al. judged in binary terms whether the feedback was more positive than a baseline and did not measure at the level of individual comments. The verdict is an extension, and its novelty lies in measuring the recall and weight of comments and in placing the study on visual material. This connects directly to settings in which AI is brought into critiques (crits) in design education.

G7 Homogenization by generative AI and homogenization by templates

Goree et al. (2021) measured that web layouts became more similar from 2003 to 2019 (Slop Seen Through Design History: The Profession Began as a Countermeasure to Shoddy Mass Production). On an image-sharing site, a decline in the visual novelty of artists who began to use generative AI has been measured (Zhou and Lee 2024). No study was found that extended the measure of Goree et al. beyond 2020 and separated whether generative AI accelerated an existing trend or added a different component. The slop-design-history note also writes that there is no guarantee that convergence through templates and convergence toward model defaults are the same mechanism. The question is limited to a test for a break relative to the existing trend.

G8 The range of futures in design fiction co-written with LLMs

Kozachek (2026) compared 1,080 future scenarios for Europe written by humans with 3,240 scenarios generated by GPT and reports in the abstract that the GPT scenarios were biased toward technology-centered themes. Lehnert et al. (2026) co-wrote 18 design fictions with GPT-4 and reported that the quality of the stories emerged from the interaction of human framing, prompts, and model behavior, but they did not include a condition without AI. No study was found that compared, for the same participants, how the diversity of the futures they describe and the reproduction of a dominant view of technology change depending on whether they co-write with an LLM. Because one aim of design fiction is to depart from dominant imaginaries, this comparison bears directly on the aim of the method (What Do Science Fiction Prototyping and Design Fiction Actually Do?).

G9 Why AI images are preferred, and the shift of standards through familiarity

When the author is not disclosed, many people prefer AI images (van Hees et al. 2025). Studies have measured that authorship labels shift evaluations (Wang et al. 2026), that placing human works beside AI works raises the evaluation of human creativity (Horton et al. 2023), and what effect online exposure to images labeled as AI-made has (Malecki et al. 2025, 470 participants). No study was found that tested, as mediation with objective measures on the side of the stimuli, whether the unlabeled preference is explained by the typicality or the processing fluency of the stimuli. No study was found either that followed the same people over several weeks of repeated viewing to see in which direction their preferences and their standards for evaluating human works shift. These two questions are preconditions for testing the loop “people become used to low quality” in Do Decision-Makers Without Taste Stop Commissioning Creators? Checking the Problem of Image-Generating AI Through Three Links (Commissioning, Habituation, and the Creator Base) and Relating It to Research on AI Slop.

G10 Appearance and operation in the consistency of design systems

Satzinger and Olfman (1998) manipulated separately the consistency of the screen appearance and of the command language across two applications (114 participants) and reported that inconsistency in the command language increased the number of operations, while inconsistency in appearance increased accuracy. The core manipulation of the candidate was already carried out in 1998. No study was found that tested whether this reversed result reproduces in a modern GUI built from the components of a design system while also measuring exploratory behavior such as the discovery of features (How Much of the Evidence for Adjudicating the Six Design System Conflicts Is in Place?). The verdict is a replication opportunity, and the only provisional axis is Mat. Unless the study is stated explicitly as a replication, its novelty will appear larger than it is.

The 17 dropped candidates

The dropped candidates and the reasons are listed in the table of the corpus. The reasons fall into four groups.

• The candidate is far from the main axis (design theory and pedagogy).
• Earlier handoffs or existing notes already cover it.
• It was likely to be answered already, even before the external search.
• It only copies a limitation written by the authors of a single note.

For example, for a candidate on the licensing and pay of interior designers, the scanning agent itself wrote that the economics of occupational licensing has nearly answered it.

What is thin across the whole corpus

The scans of the five areas named the same questions as having thin research rows. One is the temporal dimension. There are almost no delayed tests, repeated exposures, or longitudinal follow-ups, and changes over time, such as role transitions and deskilling, are often supported by cross-sectional self-reports. The other is research that separates the mechanism of an effect at the level of the process. Although it has been measured that generative AI changes outcomes, few studies separate through what at which stage it changes them. G2, G4, and G9 all came from these two thin areas.

What was corrected during the scan

During the scan of area C, the descriptions of the same paper (Melanou and Beege 2026) in two corpora disagreed. One wrote that structured prompt scaffolding significantly increased knowledge, and the other wrote that there was no significant difference between the strengths of scaffolding. The abstract of the paper states that knowledge increased in every condition but that there was no significant difference between conditions. The erroneous corpus row and the corresponding sentence of the public note Research Currents in AI, Education, and the Learning Sciences: Evidence on Generative AI and Learning (2024–2026) were corrected.

Unverified items

• Kozachek 2026, Lehnert et al. 2026, Goree et al. 2021, Malecki et al. 2025, and Satzinger and Olfman 1998: the full text could not be reached (publisher 403), so the note writes within the scope of the abstracts. [要一次検証: 本文未到達 (full text not reached; abstract only)]
• Mehler et al. 2024: the PDF could not be reached, so the note writes within the scope of the abstract. [要一次検証: 本文未到達 (full text not reached; the AIS eLibrary PDF returns 403)]
• Among the nearest studies named by the auditing agents, the abstract of Tao and Vyas 2025 (CHI) could not be obtained (only the figure descriptions were obtained), so it is not used in this note. It remains in the corpus as the agent’s report.
• Because OpenAlex and Semantic Scholar could not be used, the L15 search was limited to Crossref, arXiv, Europe PMC, and web search. Before paper-pipeline makes its final scoring, the same search queries should be run again in OpenAlex.

Related notes

• Where Do Novel Research Questions Come From? The Four Loci and Their Combinations, Examined Against the Literature: the literature on constructing research questions and gap-spotting
• Problematizing the wiki corpus: an assumption ledger for four domains and twelve challenges: an earlier scan that applied the inversion of assumptions to the same corpus
• Extracting candidates for peer-reviewed papers: a cross-wiki analysis that excludes the author's own theory: earlier work that extracted paper candidates from notes
• Where Are the Gaps in AI Research? Seven Voids Found by Reframing: a method that looks for gaps through leaps of frame
• paper-pipeline への申し送りの行方と落ちた型（2026-11）private: what became of the handoffs that include the candidates of this scan in paper-pipeline, and the types of failure

References

Method

• Sandberg, J., & Alvesson, M. (2011). Ways of constructing research questions: Gap-spotting or problematization? Organization, 18(1), 23–44. https://doi.org/10.1177/1350508410372151

G1

• Lima, G., Grgić-Hlača, N., & Redmiles, E. M. (2025). Public opinions about copyright for AI-generated art: The role of egocentricity, competition, and experience. CHI 2025. https://doi.org/10.1145/3706598.3713338
• Mehler, M., Ellenrieder, S., & Buxmann, P. (2024). The influence of effort on the perceived value of generative AI: A study of the IKEA effect. ECIS 2024 Proceedings. https://aisel.aisnet.org/ecis2024/track09_coghbis/track09_coghbis/6/

G2

• Shin, J., Polyanskaya, A., Lucero, A., & Oulasvirta, A. (2025). No evidence for LLMs being useful in problem reframing. CHI 2025. https://doi.org/10.1145/3706598.3713273
• Guo, J., Yin, Y., Sun, L., & Chen, L. (2024). Empirical study of problem-solution co-evolution in human-GAI collaborative conceptual design. DRS2024. https://doi.org/10.21606/drs.2024.983

G3

• Eckhardt, S., Kühl, N., Dolata, M., & Schwabe, G. (2025). A survey of AI reliance. ACM Computing Surveys. https://doi.org/10.1145/3776528
• Welzel, V. H. A., & Vincent, N. (2026). Overreliance in writing tasks: Exploring similarity-based measures of AI influence on writing and proposing a reflective writing interface intervention. FAccT 2026, 5941–5958 (arXiv:2605.15322). https://arxiv.org/abs/2605.15322

G4

• Fan, M., & Chang, P. (2026). Telling students to evaluate does not make it happen: Task stage, offloading tendency, and error detection in AI-assisted student writing. Behavioral Sciences, 16(9), 1671. https://doi.org/10.3390/bs16091671

G5

• Sternlicht, N., Shahid, S., Jansen, P., Weld, D. S., Siangliulue, P., & Hope, T. (2026). Old ideas, novel problems: The instability of LLM-based novelty evaluation. arXiv:2610.02022. https://arxiv.org/abs/2610.02022
• Terwiesch, C., Meincke, L., Girotra, K., Mollick, E., Nave, G., & Ulrich, K. T. (2026). AI and its impact on creativity and diversity: An empirical study of LLM-generated product ideas. arXiv:2607.27553. https://arxiv.org/abs/2607.27553

G6

• Sharma, M., Tong, M., Korbak, T., Duvenaud, D., Askell, A., Bowman, S. R., et al. (2024). Towards understanding sycophancy in language models. ICLR 2024. https://arxiv.org/abs/2310.13548

G7

• Goree, S., Doosti, B., Crandall, D., & Su, N. M. (2021). Investigating the homogenization of web design: A mixed-methods approach. CHI 2021. https://doi.org/10.1145/3411764.3445156
• Zhou, E., & Lee, D. (2024). Generative artificial intelligence, human creativity, and art. PNAS Nexus, 3(3). https://doi.org/10.1093/pnasnexus/pgae052

G8

• Kozachek, D. (2026). Entering the age of hybrid futures: A comparative study of human and GPT-generated scenarios. Futures, 176. https://doi.org/10.1016/j.futures.2025.103705
• Lehnert, F., Oelscher, L., & Eradze, M. (2026). Co-creating educational futures: How speculative prompting shapes AI-generated design fiction. DIS 2026. https://doi.org/10.1145/3800645.3812953

G9

• van Hees, J., Grootswagers, T., Quek, G. L., & Varlet, M. (2025). Human perception of art in the age of artificial intelligence. Frontiers in Psychology, 15. https://doi.org/10.3389/fpsyg.2024.1497469
• Wang, Y., Cao, Y., Cai, J., & Sun, B. (2026). Where does the bias begin? Cognitive processing differences in the aesthetic evaluation of AI-generated art. Frontiers in Psychology. https://doi.org/10.3389/fpsyg.2026.1827956
• Horton, C. B., Jr., White, M. W., & Iyengar, S. S. (2023). Bias against AI art can enhance perceptions of human creativity. Scientific Reports, 13. https://doi.org/10.1038/s41598-023-45202-3
• Malecki, W. P., Messingschlager, T. V., & Appel, M. (2025). The impact of exposure to generative AI art on aesthetic appreciation, perceptions of AI mind, and evaluations of AI and of art careers. New Media & Society. https://doi.org/10.1177/14614448251344590

G10

• Satzinger, J. W., & Olfman, L. (1998). User interface consistency across end-user applications: The effects on mental models. Journal of Management Information Systems, 14(4), 167–193. https://doi.org/10.1080/07421222.1998.11518190

Corrected during the scan

• Melanou, C., & Beege, M. (2026). Scaffolding generative AI as a tutor: A quasi-experimental study of learning outcomes and motivational, cognitive and metacognitive processes. Education Sciences, 16(4). https://doi.org/10.3390/educsci16040651

Footnotes

1. The thread has 16 posts (an introduction and 15 prompts). The original text and how each prompt was reinterpreted as a lens are in .claude/skills/research-gap-lenses/references/lenses.md. ↩

2. Sandberg and Alvesson analyzed how 52 articles in organization studies constructed their research questions and reported that gap-spotting was the most common way across all positions. The literature on constructing research questions is summarized in Where Do Novel Research Questions Come From? The Four Loci and Their Combinations, Examined Against the Literature. ↩