Shuichiro Ogawa
日本語

Notes · updated 2026-07-22 · Markdown

Recombination of the Known and Nonhuman Reasoning: Can Hallucination Be Treated as a Resource for Creativity?

An essay that reorganizes the argument for treating hallucination as a resource for creativity through two lenses.

Contents (8)
  1. Starting where the neutral map closed
  2. Read against the definition of innovation as recombination of the known, hallucination is variation
  3. Variation supplies only half of value
  4. If the reasoning principles differ, is the added diversity of a new kind?
  5. Orthogonality and compressive degradation are two readings of the same difference in shape
  6. The two faces collide inside the interlock
  7. One experiment to settle the bet
  8. Footnotes

Starting where the neutral map closed

The earlier literature map closed by illuminating the trade-off between performance and diversity1. Detection and mitigation remove hallucinations, and RLHF and SFT remove the diversity of outputs. The map could trace the line of reasoning up to the point that seeds of ideas were mixed into what was removed. However, the original question, that the deviations of older or less accurate models are what help ideation, could not be supported by deduction and remained a leap of frame.

This essay starts from the view that this leap can be rebuilt from a different angle. It moves the question away from “which generation of models is more diverse” in two directions. The first is to read the mechanism that generates hallucinations against the definition of innovation. The second is to read the fact that this mechanism operates on reasoning principles different from those of humans as a source of diversity. Followed far enough, these two readings appear to be two faces of one mechanism. They do appear so, but in the end the two faces collide inside the mechanism itself.

Read against the definition of innovation as recombination of the known, hallucination is variation

The argument begins with the classical definition of innovation. As a classic of this lineage, the references include Schumpeter’s new combinations (neue Kombinationen) (only the bibliographic record was checked this time; [要一次検証: 本文未到達 (full text not reached; bibliographic record only)])2. According to the abstract, Weitzman made the production of new ideas a function of newly reconfigured old ideas and showed with a mathematical model how knowledge builds upon itself in a combinatoric feedback process3. The same abstract states that the ultimate limits to growth lie less in the ability to generate new ideas than in the ability to process an abundance of potentially new ideas into usable form3. The publisher’s description of Arthur’s book states that technologies are put together from pieces that already exist (which are themselves technologies) and that they combine, morph, and combine again to create further technologies4. Placed side by side, these descriptions present innovation as the recombination of known elements.

The mechanism by which generative models produce outputs overlaps structurally with this description. A model samples continuations that were probabilistically possible within its training distribution. The model does not distinguish between continuations that depart from fact and continuations that no one has written yet1. He and colleagues measured hallucination and creativity together at each decoding layer and showed a consistent trade-off between the two across layer depths and model sizes5. If deviation and ideation come from the same mechanism, that mechanism can be called a “probabilistic recombiner of known elements”.

The value of recombination has been shown in quantitative studies. Uzzi and colleagues analyzed the combinations of prior work cited in 17.9 million papers and showed that high-impact papers take a structure in which “a small amount of atypical citation is placed on a base of typical citation”6. Using patent citation data, Fleming showed that new combinations produce more variable inventive outcomes7. If probabilistic deviations at the edge of the training distribution are regarded as “cheap generation of unfamiliar recombinations”, hallucination becomes a device that automatically produces atypical combinations.

Classics of the psychology of creativity support this position from below. According to the abstract of Simonton’s review, Campbell proposed that creative thought should be conceived as a process of blind variation and selective retention (BVSR)89. Simonton represented this process formally with combinatorial models910. Koestler’s bisociation (blending elements drawn from two previously unrelated matrices of thought into a new matrix of meaning) can also be read as two steps: a variation, and a selection in which a new meaning takes shape11. In this scheme, hallucination corresponds to the stage of variation. The strongest form of the first lens is that the machine is a device that performs blind variation in large quantities and at low cost.

Variation supplies only half of value

However, the same two-step scheme immediately limits the reach of this lens. Read within the two-step frame of BVSR, blind variation alone may reach novelty but does not reach creation89. Only when selective retention is added does a variation remain as something worth keeping. Weitzman’s abstract, which places the limits to growth in the ability to process ideas into usable form, is the economic version of the same limit3. In the abstract, Amabile, after considering the definition and assessment of creativity, presented a framework whose components are domain-relevant skills, creativity-relevant skills, and task motivation, and emphasized the previously neglected social factors12. As classics on the side of the definition of creativity, the references also include Runco and Jaeger’s standard definition and Csikszentmihalyi’s systems view. Because no full text or abstract of either could be obtained this time, the essay does not go into their content ([要一次検証: 本文未到達 (full text not reached; bibliographic record only)])1314. In the BVSR scheme and in Weitzman’s abstract, the two sources checked here, value lies outside variation. Hallucination supplies variation but has no selection mechanism that measures whether a variation is effective.

Uzzi’s finding makes this limit more concrete. What appeared in the high-impact papers was not atypical combination itself but a structure in which “a small amount of the atypical is placed on a base of the typical”6. Producing large quantities of random deviations does not reproduce this structure. Fleming’s abstract also states that experimenting with new components and new combinations leads to less useful inventions on average but increases the variability that can result in both failure and breakthrough7. Most deviations are material to be discarded, not breakthroughs.

A deeper limit lies in the range that the mechanism can reach. Boden divided creativity into three types15. They are combinational creativity, which places existing elements in new arrangements, exploratory creativity, which searches within a given conceptual space, and transformational creativity, which restructures the conceptual space itself. Franceschelli and Musolesi mapped this distinction onto machine learning16. In their account, a generative model shows combinational creativity if it can sample new and valuable works that are combinations of real data, and exploratory creativity if the works actually differ from real ones. Transformational creativity appears only when the sampling distribution diverges in some way from the underlying one (through a different training process, by altering the distribution after learning, or by changing the sampling technique)16. They also write that the space of solutions in which pretrained models work is the space derived from real data, and that outputs move outside that space or to its boundaries only when certain techniques are added. For Transformer-based language models as well, they hold that transformational creativity is possible when prompt tuning or reinforcement learning is added, and they note that there is no guarantee that the output is valuable, novel, or surprising16.

If hallucination is a probabilistic deviation that occurs within sampling from the learned distribution, it stays within the support of that distribution (interpolation, in the terms of this essay). An operation that makes the distribution diverge is a procedure added separately from hallucination. The first lens can defend “recombination itself”. However, what it can defend extends only to the supply of combinational variation that leaves the selection of value to humans and stays outside transformational creativity.

If the reasoning principles differ, is the added diversity of a new kind?

The second lens bets not on the position of the mechanism but on its foreignness. Existing diversity is, in the end, diversity among one kind of reasoner: humans. Theories of collective intelligence have formalized why this diversity works. Hong and Page showed that when problem solvers hold different representations of a problem space and different search procedures, their diversity can outweigh high individual ability (diversity trumps ability)17. Page extended this idea in two books to a discussion of cognitive diversity (only the bibliographic records were checked this time; [要一次検証: 本文未到達 (full text not reached; bibliographic record only)])1819. What works here is not the number of heads but differences of position in the problem space.

When LLMs are placed in the same layer as a human group, this “difference of position” splits into two. One is within-source diversity, that is, differences among LLMs. This is small. Wenger and Kenett measured that the creative outputs of more than twenty distinct models resemble each other far more than humans resemble each other20. Kim and colleagues showed across more than 350 models that errors are correlated and that the correlation is stronger for more capable models21. If similar answers come back whichever model is chosen, a group of LLMs does not become “a group of solvers with different representations”.

The other is between-source diversity, that is, the difference between humans and LLMs. This is where the second lens places its bet. The training objective of an LLM is to maximize the probability of the next token over the distribution of training text. Human reasoning is the processing of meaning grounded in purposes, contexts, and bodily experience. The principles differ. The framework in which Rahwan and colleagues called for a research agenda on “machine behaviour”, which studies the behavior of AI systems by adding insights from across the sciences to computer science, provides a basis for dealing with this difference in principle22. Human association is drawn toward the neighborhood in a semantic network. Next-token prediction can also sample from the tails of the probability mass. If so, an LLM can produce juxtapositions that humans do not merely “fail to think of” but have “a practically zero probability of thinking of”.

From the side of ensemble learning, the ambiguity decomposition of Krogh and Vedelsby shows that the generalization error of an ensemble equals the weighted average of the errors of the individual learners minus the weighted average of the disagreement (ambiguity) among their outputs23. When the individual errors are the same, the larger the disagreement, the smaller the ensemble error. In this decomposition, adding the “most different” learner to a human group corresponds to increasing the ambiguity. However, the error of the added learner itself also enters the term for the average error. The strongest form of the second lens is that LLMs add a new axis orthogonal to the human distribution.

Orthogonality and compressive degradation are two readings of the same difference in shape

At this point, the argument turns over from within the second lens. Both camps accept that the distribution of LLMs has a different shape from the human distribution. What they dispute is the sign attached to that difference.

The evidence on the side of the reversal is substantial. Bender and colleagues called LLMs “stochastic parrots” and asked about the risks of ever-larger language models (title and abstract)24. Shumailov and colleagues demonstrated model collapse, in which training recursively on generated data makes the output distribution degenerate and the tails of the distribution disappear25. Peterson showed knowledge collapse, in which knowledge across society narrows because LLM outputs move toward the center of the distribution, using a simple model in which learners or innovators choose between traditional methods and a discounted AI-assisted process26. Taken together, these studies make LLMs look like a biased resample that cuts the tails of the human distribution and shifts toward the majority. The shape certainly differs from the human one. However, this reading holds that the difference points not in “a useful direction that humans do not reach” but in “a direction that has cut away the diversity humans had”.

The two readings come from the same difference in shape, and only the sign of usefulness points the opposite way. The orthogonality of the second lens requires an orthogonality in the direction of adding a new axis. If model collapse and knowledge collapse are correct, the observed orthogonality is an orthogonality in the opposite direction, which cuts diversity. That the shapes differ is compatible with both readings. Which direction holds has not yet been settled empirically.

The same problem of sign appears twice from the side of the theory of collective intelligence. The Hong and Page theorem is not as strong as a naive reading suggests. According to Singer’s abstract, Thompson argued that in the Hong and Page model, the diverse groups win not because of diversity but because of the procedure that creates them at random2728. In their abstract, Grim and colleagues showed that when the model is extended to search landscapes with different degrees of roughness (randomness), the conclusion changes sensitively, and they advised against interpreting such results too broadly29. Singer defended the diversity interpretation and argued that it is necessary to be clearer about how functional diversity is conceived and measured28. However, the more precise the conditions become, the narrower the reach of the generalization that “adding LLMs yields a diversity bonus”. What the theorem requires is relevant diversity that is useful for the problem. Adding solvers whose errors are correlated may produce noise rather than relevant diversity21.

Another mechanism that reverses the sign lies on the side of use. Lorenz and colleagues showed experimentally that even the mild social influence of showing others’ estimates undermines the wisdom of the crowd30. Everyone consulting the same model may work in a way close to this social influence (an analogy drawn here; no study that measures it directly is cited). When Bommasani and colleagues urged caution about the homogenization of foundation models, their concern was that the defects of a foundation model are inherited by all the models adapted from it downstream31. In a 36-participant comparative user study, Anderson and colleagues confirmed that the ideas of different users of ChatGPT were semantically closer to each other than the ideas of different users of an alternative creativity support tool32. As long as LLMs can be kept apart and used as an independent, separate source of reference, there is room for them to act as a new axis. The moment everyone leans on them, the same mechanism turns into something that cuts the existing diversity among humans.

The two faces collide inside the interlock

At this point, each lens has its strongest form and its own limit of scope. What remains is the integrating claim that binds the two into one. The binding line can be drawn as follows. Perhaps the nonhuman reasoning basis described by the second lens is exactly the mechanism that lets the recombination described by the first lens reach combinations that humans do not actually enumerate. If so, the recombination of the known (the first face) and the new, orthogonal diversity (the second face) become two faces of one and the same mechanism. This is the core of the integrating thesis (the interlock).

This way of binding collides with itself from within. The first face was defended as interpolation. Its scope was that it stays within sampling from the learned distribution and, unless an operation that makes the distribution diverge is added, remains outside transformational creativity1516. The second face promises a new axis orthogonal to the human distribution. In the mapping of Franceschelli and Musolesi, sampling within the distribution belongs to combinational or exploratory creativity, and setting up a new axis requires transformational creativity, which makes the distribution diverge16. How can an output that stays inside the support span an axis orthogonal to the distribution that spans that support? If the output is bound to the support, it stays inside the space that the support spans. What stays inside cannot, by definition, span a new axis that points outward.

In other words, accepting the limit of scope of the first face (that it is interpolation) leaves the second face (that it is a new orthogonal axis) without ground. The interlock glues the leap from interpolation to a new axis with the phrase “nonhuman reasoning basis”. The glue word names the leap but does not fill it. What the integrating thesis added was not a new mechanism but a new name.

Stopping at a name casts doubt on this essay itself. Both “variation needs selection” and “homogenization is dangerous” are self-criticisms that earlier research has already written931. If the essay only pasted the names of doctrines onto them again, the lenses would fall into decoration. There are only two points at which the essay steps beyond decoration. The first is that it named the logical contradiction between the limit of scope of the first face and the second face. Earlier research placed the two limits side by side but did not write out the implication that “accepting interpolation makes the new axis false”. The second is that, after accepting that the difference in shape is compatible with both readings, it introduced a separate axis of discrimination: the sign of usefulness. The distinction that the shape differs but the sign falls on the side of collapse does not appear in the binary choice of orthogonal or not.

One experiment to settle the bet

The interlock does not stand as a mechanism. Even so, it is not a question to be discarded. Once the contradiction is removed, the integrating thesis remains as one conditional bet. Does interpolation inside a nonhuman distribution produce outputs that human evaluators judge to be novel and useful, in directions that humans occupy only sparsely? This question cannot be settled on paper. This is also where the second face has room to survive in part. A different generative process places probability mass in regions that humans sample only sparsely. Whether those regions are useful is the only question that remains open.

The form of a decisive test overlaps with the comparative experiment to which the earlier map already pointed1. On the same task, the generation, scale, and temperature of models are varied while the other conditions are kept equal. Three things are measured. They are the diversity of ideas, the rate of useful deviations, and whether the distribution extends in directions that human evaluators judge to be novel and useful. The third measure decides the sign of the orthogonality. The orthogonal contribution to a human group can only be measured as the product of the distance from the human baseline distribution and the sign of usefulness. The ex ante evaluation framework of Azad and Baten provides the closest basis for this measurement33.

There is a gap that should be stated plainly. No peer-reviewed study showing at the level of mechanism that “interpolation inside a nonhuman distribution produces a useful orthogonal axis” was found in either of the two collections. The limits of scope of this essay rest on that gap. Therefore, what remains as the result is not the proposition that “hallucination is innovation”. What remains is the location of the weak point that appears when the two lenses are set up in their strongest forms, and one experiment that would decide that weak point.

The earlier map closed with the gap “are older generations more diverse?”1. This essay replaces that gap with different words. The problem is not the generation but the sign of usefulness attached to the mechanism of recombination in a nonhuman distribution. That sign has not yet been observed. If one question is to be left open, it is this one.

References

The bibliographic records are given in the footnotes with DOIs or URLs. Preprints that have not been peer reviewed and items whose primary pages could not be reached are summarized at the end.

Unverified items

  • Missing mechanism on the affirmative side: No peer-reviewed study showing at the level of mechanism that “interpolation inside a nonhuman distribution produces a useful orthogonal axis in directions that humans occupy only sparsely” was found in either of the two collections. The limits of scope of this essay rest on this gap.
  • Description corrected for Franceschelli & Musolesi 2024: The first draft stated that the two authors argued that “LLMs cannot in principle reach hyperpolation, which transforms the conceptual space”, but the full text (arXiv version) contains no such claim. The text holds that transformational creativity appears only when the sampling distribution diverges from the underlying one, and that it can occur in Transformer-based language models through prompt tuning or reinforcement learning. The section on the internal contradiction of the interlock was corrected to match the text, with the condition “unless an operation that makes the distribution diverge is added”.
  • Works described within the scope of their abstracts because the full text could not be obtained: Weitzman 1998, Uzzi 2013, Fleming 2001, Amabile 1983, Hong & Page 2004, Grim 2019, Singer 2019, Rahwan 2019, Wenger & Kenett 2026 (the number of models from the full text of the arXiv version), Bender 2021, Shumailov 2024, Lorenz 2011, Anderson 2024, and Simonton 2010. Arthur 2009 is described within the scope of the publisher’s description and the table of contents, Koestler 1964 within an encyclopedia entry, and Campbell 1960 and Thompson 2014 within their presentation in the abstracts of other works.
  • Works for which only the bibliographic record was checked: Schumpeter 1934, Runco & Jaeger 2012, Csikszentmihalyi 1988, and Page 2007 and 2017 ([要一次検証: 本文未到達 (full text not reached; bibliographic record only)]). The text does not go into their content and does not use them to support the main claims. The page numbers of Boden 2004 were not confirmed ([要出典確認: 書誌未確定 (bibliographic record incomplete; page numbers)]).
  • Peer-review status of a critique of the Hong-Page theorem: Romaniega (arXiv:2307.04709) is a stronger rebuttal of the theorem, but whether it has been published after peer review was not checked, so it was excluded from the corpus and the text ([要確認: 未試行 (not attempted; peer-review status)]).
  • Dependence of the critique on a canon: The point about the internal contradiction of the interlock relies on Boden, Simon, Schön, Dorst & Cross, and Krippendorff as used by academic-critic (canon: design-theory). It has not been rechecked from canons other than design-theory (within innovation economics, or computational creativity).

Footnotes

  1. The earlier literature map. An Academic Map of Research That Treats Hallucination as a Resource for Creativity (wiki/note/hallucination-as-creativity-resource-literature.md). The full bibliographic records of He 2025, Wenger & Kenett 2026, Kumar 2024, Wright 2025, Mednick 1962, Beaty & Kenett 2023, and others follow that map. ↩ ↩2 ↩3 ↩4

  2. Schumpeter, J. A. (1934). The Theory of Economic Development. Harvard University Press. https://archive.org/details/theoryofeconomic0000schu (German first edition 1912, revised English translation 1934. No book DOI. The Internet Archive page is access-restricted, and only the catalog record was checked. [要一次検証: 本文未到達 (full text not reached; bibliographic record only)]) ↩

  3. Weitzman, M. L. (1998). Recombinant Growth. The Quarterly Journal of Economics, 113(2), 331–360. https://doi.org/10.1162/003355398555595 (DOI resolution checked on 2026-07-22) ↩ ↩2 ↩3

  4. Arthur, W. B. (2009). The Nature of Technology: What It Is and How It Evolves. Free Press. ISBN 9781416544067. https://archive.org/details/natureoftechnolo00arth (reached on 2026-07-22; checked against the table of contents and the publisher’s description on 2026-11-06) ↩

  5. He, Z., Zhang, B., & Cheng, L. (2025). Shakespearean Sparks: The Dance of Hallucination and Creativity in LLMs’ Decoding Layers. arXiv:2503.02851. https://arxiv.org/abs/2503.02851 (the full record follows An Academic Map of Research That Treats Hallucination as a Resource for Creativity. Not peer reviewed) ↩

  6. Uzzi, B., Mukherjee, S., Stringer, M., & Jones, B. (2013). Atypical Combinations and Scientific Impact. Science, 342(6157), 468–472. https://doi.org/10.1126/science.1240474 (bibliographic record checked through the DOI and PubMed on 2026-07-22) ↩ ↩2

  7. Fleming, L. (2001). Recombinant Uncertainty in Technological Search. Management Science, 47(1), 117–132. https://doi.org/10.1287/mnsc.47.1.117.10671 (reached on 2026-07-22) ↩ ↩2

  8. Campbell, D. T. (1960). Blind Variation and Selective Retention in Creative Thought as in Other Knowledge Processes. Psychological Review, 67, 380–400. https://doi.org/10.1037/h0040373 (checked through the DOI and PubMed on 2026-07-22. No abstract is available from the publisher or Europe PMC, and the content is described within the scope of the abstract of Simonton 2010) ↩ ↩2

  9. Simonton, D. K. (2010). Creative Thought as Blind-Variation and Selective-Retention: Combinatorial Models of Exceptional Creativity. Physics of Life Reviews, 7(2), 156–179. https://doi.org/10.1016/j.plrev.2010.02.002 (DOI reached on 2026-07-22; checked against the abstract in Europe PMC on 2026-11-06) ↩ ↩2 ↩3 ↩4

  10. Simonton, D. K. (1988). Scientific Genius: A Psychology of Science. Cambridge University Press. ISBN 9780521352871 (chance-configuration theory. No book DOI) ↩

  11. Koestler, A. (1964). The Act of Creation. Hutchinson (UK) / Macmillan (US). (bisociation. No book DOI; the bibliographic record and the description of bisociation were checked at https://en.wikipedia.org/wiki/The_Act_of_Creation) ↩

  12. Amabile, T. M. (1983). The Social Psychology of Creativity: A Componential Conceptualization. Journal of Personality and Social Psychology, 45(2), 357–376. https://doi.org/10.1037/0022-3514.45.2.357 (on 2026-11-06, Unpaywall and Semantic Scholar confirmed that the title registered for this DOI matches this paper (componential conceptualization). The content is described within the scope of the abstract in OpenAlex) ↩

  13. Runco, M. A., & Jaeger, G. J. (2012). The Standard Definition of Creativity. Creativity Research Journal, 24(1), 92–96. https://doi.org/10.1080/10400419.2012.650092 (DOI reached on 2026-07-22) ↩

  14. Csikszentmihalyi, M. (1988). Society, Culture, and Person: A Systems View of Creativity. In R. J. Sternberg (Ed.), The Nature of Creativity (pp. 325–339). Cambridge University Press. https://link.springer.com/chapter/10.1007/978-94-017-9085-7_4 (the URL is the 2014 reprint. The chapter DOI of the 1988 original was not confirmed. Neither the full text nor an abstract could be obtained from the Springer page, the Wayback Machine, OpenAlex, or Unpaywall. [要一次検証: 本文未到達 (full text not reached; bibliographic record only)]) ↩

  15. Boden, M. A. (2004). The Creative Mind: Myths and Mechanisms (2nd ed.). Routledge. ISBN 9780415314534. https://www.routledge.com/9780415314534 (first edition 1990, revised and expanded 2004. combinational / exploratory / transformational, P/H-creativity. The definitions of the three types were checked through their presentation in the full text of Franceschelli and Musolesi 2024. Page numbers: [要出典確認: 書誌未確定 (bibliographic record incomplete; page numbers)]) ↩ ↩2

  16. Franceschelli, G., & Musolesi, M. (2024). Creativity and Machine Learning: A Survey. ACM Computing Surveys, 56(11), Article 283. https://doi.org/10.1145/3664595 (DOI reached on 2026-07-22. Checked against the full text of the arXiv version, arXiv:2104.02726, on 2026-11-06. The word “hyperpolation” does not appear in the text, and the text does not state that LLMs cannot in principle reach transformational creativity. The description in this essay was corrected to match the text) ↩ ↩2 ↩3 ↩4 ↩5

  17. Hong, L., & Page, S. E. (2004). Groups of diverse problem solvers can outperform groups of high-ability problem solvers. PNAS, 101(46), 16385–16389. https://doi.org/10.1073/pnas.0403723101 (DOI checked on 2026-07-22; the primary page requires authentication) ↩

  18. Page, S. E. (2007). The Difference: How the Power of Diversity Creates Better Groups, Firms, Schools, and Societies. Princeton University Press. ISBN 9780691138541 (diversity-prediction theorem) ↩

  19. Page, S. E. (2017). The Diversity Bonus: How Great Teams Pay Off in the Knowledge Economy. Princeton University Press. ISBN 9780691176888 ↩

  20. Wenger, E., & Kenett, Y. N. (2026). Large language models are homogeneously creative. PNAS Nexus, 5(3), pgag042. https://doi.org/10.1093/pnasnexus/pgag042 (the full record follows An Academic Map of Research That Treats Hallucination as a Resource for Creativity. The number of models, 22, was checked in the full text of the arXiv version, arXiv:2501.19361) ↩

  21. Kim, E., Garg, A., Peng, K., & Garg, N. (2025). Correlated Errors in Large Language Models. ICML 2025. arXiv:2506.07962. https://arxiv.org/abs/2506.07962 (reached on 2026-07-22) ↩ ↩2

  22. Rahwan, I., Cebrian, M., Obradovich, N., Bongard, J., Bonnefon, J.-F., et al. (2019). Machine behaviour. Nature, 568, 477–486. https://doi.org/10.1038/s41586-019-1138-y (DOI checked on 2026-07-22; the primary page requires authentication. Checked against the abstract in Europe PMC on 2026-11-06) ↩

  23. Krogh, A., & Vedelsby, J. (1995). Neural network ensembles, cross validation, and active learning. Advances in Neural Information Processing Systems, 7, 231–238. https://www.semanticscholar.org/paper/910688d01c01856dd20715907af44157de8d3d1d (ambiguity decomposition. Equation (10) was checked in the full text of the PDF in the NeurIPS proceedings on 2026-11-06) ↩

  24. Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? FAccT ‘21, 610–623. https://doi.org/10.1145/3442188.3445922 (DOI reached on 2026-07-22) ↩

  25. Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R., & Gal, Y. (2024). AI models collapse when trained on recursively generated data. Nature, 631, 755–759. https://doi.org/10.1038/s41586-024-07566-y (DOI checked on 2026-07-22; the primary page requires authentication. The content is described within the scope of the abstract in OpenAlex) ↩

  26. Peterson, A. J. (2024/2025). AI and the Problem of Knowledge Collapse. AI & Society (published 2025-01-19). arXiv:2404.03502. https://doi.org/10.1007/s00146-024-02173-x (reached on 2026-07-22; peer reviewed) ↩

  27. Thompson, A. (2014). Does Diversity Trump Ability? An Example of the Misuse of Mathematics in the Social Sciences. Notices of the American Mathematical Society, 61(9), 1024–1030. https://www.ams.org/notices/201409/rnoti-p1024.pdf (direct access to the PDF returned 403, and so did the Wayback Machine. The bibliographic record agrees across several citations. The content is described within the scope of its presentation in the abstract of Singer 2019) ↩

  28. Singer, D. J. (2019). Diversity, Not Randomness, Trumps Ability. Philosophy of Science, 86(1), 178–191. https://doi.org/10.1086/701074 (DOI reached on 2026-07-22) ↩ ↩2

  29. Grim, P., Singer, D. J., Bramson, A., Holman, B., McGeehan, S., & Berger, W. J. (2019). Diversity, Ability, and Expertise in Epistemic Communities. Philosophy of Science, 86(1), 98–123. https://doi.org/10.1086/701070 (DOI reached on 2026-07-22) ↩

  30. Lorenz, J., Rauhut, H., Schweitzer, F., & Helbing, D. (2011). How social influence can undermine the wisdom of crowd effect. PNAS, 108(22), 9020–9025. https://doi.org/10.1073/pnas.1008636108 (DOI checked on 2026-07-22; the primary page returned 403. The content is described within the scope of the abstract in OpenAlex) ↩

  31. Bommasani, R., et al. (2021). On the Opportunities and Risks of Foundation Models. arXiv:2108.07258. https://arxiv.org/abs/2108.07258 (reached on 2026-07-22; a report without peer review) ↩ ↩2

  32. Anderson, B. R., Shah, J. H., & Kreminski, M. (2024). Homogenization Effects of Large Language Models on Human Creative Ideation. ACM Creativity & Cognition ‘24. https://doi.org/10.1145/3635636.3656204 (DOI reached on 2026-07-22) ↩

  33. Azad, N. S., & Baten, R. A. (2026). Ex Ante Evaluation of AI-Induced Idea Diversity Collapse. arXiv:2605.06540. (the full record follows An Academic Map of Research That Treats Hallucination as a Resource for Creativity. Not peer reviewed) ↩

Author: Shuichiro Ogawa (Design Researcher / Design Consultant) About me →