Pith. sign in

REVIEW 2 major objections 5 minor 18 references

When AI design tools show more varied options at once, people pick the middle ones more often, in both preference and typicality judgments.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-13 00:57 UTC pith:P7XGH2LD

load-bearing objection Clean within-subjects result that multi-option design sets with higher variance pull selections toward the center, but the high-variance condition also uniquely adds extreme poles, so ensemble perception is not cleanly isolated from classic extremeness aversion. the 2 major comments →

arxiv 2607.09018 v1 pith:P7XGH2LD submitted 2026-07-10 cs.HC

Central Tendency Bias in Human Selection of AI-Generated Design Variations

Huiyang Chen , Keqing Jiao This is my paper
classification cs.HC
keywords Image-generation AIHuman–AI collaborationEnsemble perceptionDesign selectionCentral tendency biasMulti-option presentationCreative workflows
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper claims that the common grid layout of image-generation AI tools can systematically pull human choices toward the visual center of the set. Drawing on ensemble perception—the automatic extraction of summary statistics from groups of objects viewed together—the authors argue that higher variance among simultaneous design alternatives strengthens that central pull. In a controlled experiment, participants saw high- and low-variance sets of eight poster designs and chose once for personal preference and once for which design best represented the set. Higher variance raised the rate of center-proximal selections in both tasks. The practical stakes are clear: systems built to expand creative options may, through the way those options are shown, shrink the diversity of what users actually select.

Core claim

Higher variance in simultaneously presented sets of AI-style design variations increases selection of center-proximal designs, both when people choose what they like most and when they choose which design best represents the overall style of the set.

What carries the argument

Ensemble perception / central tendency bias: the automatic extraction of a set-level mean representation that anchors evaluation toward options near the perceptual center; the experiment manipulates set variance (high vs. low continua of poster designs) while holding the two center items constant across conditions.

Load-bearing premise

The claim rests on the idea that hand-built poster sets ordered by holistic sparse-to-busy ranking, shown in grids, are a fair stand-in for the variance and presentation structure of real AI image-generation interfaces.

What would settle it

Run the same preference and representativeness tasks with live outputs from commercial image models that truly vary in stylistic spread, or present the same high-variance options sequentially rather than simultaneously; if center-selection rates no longer rise with variance (or fall under sequential presentation), the interface-ensemble account is undermined.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Multi-option grids in generative tools can reduce selection diversity even when generation diversity is high.
  • Raising output variance alone may strengthen rather than weaken the pull toward average-looking options.
  • Self-reported reasons for liking a design may stay stable while actual choices still shift toward the set center.
  • Interface interventions that break simultaneous ensemble extraction (sequential presentation, visual emphasis on extremes) become a design lever for preserving selection diversity.
  • Evaluations of generative AI creative tools should treat selection-interface structure as a first-class variable alongside generation quality.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Default 2×2 or 2×4 grids in tools like Midjourney-style interfaces may be quietly producing more conservative final selections than the raw diversity of the model suggests.
  • The same bias could appear in other multi-option creative UIs (logo variants, layout systems, product configurators) whenever options share a visual continuum.
  • A/B tests that only measure generation diversity or click-through may miss a selection-stage bottleneck that flattens creative outcomes.
  • If sequential or subset presentation reduces the bias, product teams have a low-cost lever that does not require changing the underlying model.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper claims that simultaneous multi-option presentation of AI-generated design variations can induce central tendency bias via ensemble perception: users preferentially select designs near the perceptual center of a set, and this bias strengthens when set variance is higher. In a within-subjects experiment (N=50), participants viewed controlled 8-poster grids under high-variance ({-3..+3}) versus low-variance (clustered near 0) conditions and selected once for aesthetic preference and once for set representativeness. Center-selection rates rose with variance in both tasks (preference 56% vs 34%, t(49)=3.51, p=0.001; representativeness 66% vs 38%, t(49)=4.34, p<0.001), with a successful perceived-diversity manipulation check. The authors conclude that multi-variation interfaces may constrain selection diversity even when generation diversity is high, and discuss interface mitigations.

Significance. If the result holds, it identifies a concrete, interface-level cognitive bottleneck in human–AI co-creation that is currently under-studied relative to generation quality. The work usefully bridges ensemble-perception theory to evaluative selection of complex visual artifacts, reports a clean within-subjects design with counterbalancing, randomized grid positions, and a strong manipulation check, and surfaces a practical tension between output diversity and selection diversity. The dual-task structure (preference plus representativeness) and self-report reason data strengthen the claim that the bias is not merely a conscious typicality strategy. These contributions are relevant to HCI and design-tool research even if the precise mechanism remains partly ambiguous.

major comments (2)
  1. Method (stimulus construction) and Discussion: high-variance sets uniquely include the absolute poles (-3, +3) while low-variance sets do not. Extremeness aversion / compromise-effect accounts therefore predict elevated center selection under high variance without requiring ensemble mean extraction. The paper cites these alternatives but does not hold absolute range fixed while varying only dispersion around a matched centroid, so the reported contrasts (Table I; 56% vs 34%, 66% vs 38%) cannot cleanly attribute the increase to ensemble mechanisms. A load-bearing revision is either an additional condition that equates range or a substantially stronger experimental argument that the present design isolates ensemble coding.
  2. Method and Limitations: stimuli are manually constructed posters ordered by three-rater holistic consensus along a sparse-to-busy continuum, not live model outputs. While control is understandable, this weakens the ecological claim that the results speak to Midjourney/DALL·E-style interfaces. The central applied claim (multi-variation AI interfaces constrain selection diversity) therefore rests on an untested transfer assumption that should be either tested with real generative outputs or more tightly scoped in the abstract and conclusions.
minor comments (5)
  1. Table I and Results: report effect sizes (e.g., Cohen’s d or equivalent) alongside the paired t-tests so readers can judge practical magnitude beyond p-values.
  2. Figure 1 caption and Method: clarify whether the two near-centroid items (0 and 0') were visually distinct enough that participants treated them as separate options rather than near-duplicates; any recognition of duplication could itself bias center rates.
  3. Self-reported reasons (Table II) are pooled across variance conditions; a brief breakdown by high vs low variance would help confirm that reason distributions remain stable under the manipulation.
  4. Related Work / Discussion: a short note on whether positional randomization fully eliminates known grid-position biases (center-of-screen preference) would strengthen the design description.
  5. Minor typography: “USERSTUDY”, “RELATEDWORK”, and similar concatenated headings should be spaced for readability.

Circularity Check

0 steps flagged

No circularity: empirical user study with independent variance manipulation and center-selection DV; no fitted parameters, self-definitional equations, or load-bearing self-citation chains.

full rationale

This is a controlled within-subjects HCI experiment, not a theoretical derivation. The independent variable (high vs. low set variance) is operationalized by selecting different subsets from a pre-ordered perceptual continuum ({-3,-2,-1,0,0',+1,+2,+3} vs. {-1,-1,0,0,0',0,+1,+1}), validated by independent raters and a manipulation check (perceived diversity t(49)=19.57). The dependent variable is the participant-level proportion of trials selecting the two shared center positions (0/0'), compared via paired t-tests. These quantities are definitionally independent: the continuum ordering and set construction do not encode or force the selection rates. Ensemble-perception citations ([1]–[7]) supply background motivation and are external; the paper does not import a uniqueness theorem or ansatz that forces the result. Alternative accounts (extremeness aversion, compromise effect) are openly noted in Discussion as complementary explanations that future work should disentangle; acknowledging them does not make the reported contrast circular. No parameters are fitted to data and then re-presented as predictions, no self-referential equations appear, and the central claim remains an empirical observation about selection frequencies under the two constructed conditions. Score 0 is therefore the correct, proportionate finding.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 0 invented entities

Empirical HCI study resting on standard experimental assumptions and the domain claim that ensemble perception extends to complex design artifacts. No free parameters fitted to produce the main effect; no invented physical entities. Load-bearing premises are the validity of the hand-constructed continuum and the ecological mapping to real AI interfaces.

axioms (3)
  • domain assumption Observers automatically extract summary statistics (including central tendency) from simultaneously viewed sets of complex visual designs, analogous to established ensemble coding for simpler stimuli.
    Invoked throughout Introduction and Related Work to motivate the hypothesis; treated as transferable from face/size/orientation literature without new derivation.
  • ad hoc to paper A consensus ranking by three independent raters along a sparse-to-busy continuum produces a valid perceptual centroid and variance manipulation for poster designs.
    Method section: continuum construction and validation; disagreements resolved by discussion. This ordering defines high vs low variance sets.
  • domain assumption Grid-based simultaneous presentation of eight designs approximates the interaction structure of commercial image-generation interfaces sufficiently for the bias to be relevant.
    Stated in Method and Discussion; underpins the claim about AI tool implications.

pith-pipeline@v1.1.0-grok45 · 12499 in / 2243 out tokens · 25392 ms · 2026-07-13T00:57:48.703404+00:00 · methodology

0 comments
read the original abstract

Image-generation AI systems increasingly support creative work by producing multiple design variations for users to evaluate and select. In such human-AI co-creation workflows, selection becomes a critical stage where human judgment guides AI-generated possibilities toward final outcomes. While presenting multiple alternatives is intended to encourage exploration, the simultaneous multi-option presentation may introduce systematic biases in human decision making. Drawing on ensemble perception theory, we investigate whether these interfaces induce central tendency bias-the tendency to favor options closer to the center of a design set. We conducted a controlled experiment manipulating the variance of design sets (high vs. low) and measured participants' selections in both aesthetic preference and representativeness tasks. Results show that higher variance increases the selection of center-proximal designs across both tasks. These findings suggest that multi-variation interfaces in image-generation AI systems may constrain selection diversity, revealing a potential tension between diversity in generated outputs and diversity in human selection outcomes.

Figures

Figures reproduced from arXiv: 2607.09018 by Huiyang Chen, Keqing Jiao.

Figure 1
Figure 1. Figure 1: Example stimulus sets used in the experiment. Top row shows a low [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Manipulation Check: perceived variance ratings for high- and low [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

18 extracted references · 11 canonical work pages

  1. [1]

    Seeing the mean: Ensemble coding for sets of faces,

    J. Haberman and D. Whitney, “Seeing the mean: Ensemble coding for sets of faces,”Journal of Experimental Psychology: Human Per- ception and Performance, vol. 35, no. 3, pp. 718–734, 2009. doi: 10.1037/a0013899

    work page doi:10.1037/a0013899 2009

  2. [2]

    Ensemble perception,

    D. Whitney and A. Yamanashi Leib, “Ensemble perception,”Annual Review of Psychology, vol. 69, pp. 105–129, 2018. doi: 10.1146/annurev- psych-010416-044232

    doi:10.1146/annurev- 2018

  3. [3]

    Representing multiple objects as an ensemble enhances visual cognition,

    G. A. Alvarez, “Representing multiple objects as an ensemble enhances visual cognition,”Trends in Cognitive Sciences, vol. 15, no. 3, pp. 122–131, 2011. doi: 10.1016/j.tics.2011.01.003

    work page doi:10.1016/j.tics.2011.01.003 2011

  4. [4]

    Hierarchical encoding in visual working memory: Ensemble statistics bias memory for individual items,

    T. F. Brady and G. A. Alvarez, “Hierarchical encoding in visual working memory: Ensemble statistics bias memory for individual items,”Psychological Science, vol. 22, no. 3, pp. 384–392, 2011. doi: 10.1177/0956797610397956

    work page doi:10.1177/0956797610397956 2011

  5. [5]

    An after- effect of adaptation to mean size,

    J. E. Corbett, N. Wurnitsch, A. Schwartz, and D. Whitney, “An after- effect of adaptation to mean size,”Visual Cognition, vol. 20, no. 2, pp. 211–231, 2012. doi: 10.1080/13506285.2012.657261

    work page doi:10.1080/13506285.2012.657261 2012

  6. [6]

    Averaging facial expression over time,

    J. Haberman, T. Harp, and D. Whitney, “Averaging facial expression over time,”Journal of Vision, vol. 9, no. 11, pp. 1–13, 2009. doi: 10.1167/9.11.1

    work page doi:10.1167/9.11.1 2009

  7. [7]

    Limited memory for ensemble statistics in visual change detection,

    W. J. Harrison, J. M. V . McMaster, and P. M. Bays, “Limited memory for ensemble statistics in visual change detection,”Cognition, vol. 214, Art. no. 104763, 2021. doi: 10.1016/j.cognition.2021.104763

    doi:10.1016/j.cognition.2021.104763 2021

  8. [8]

    Investigating how experienced UX designers effectively work with machine learning,

    Q. Yang, A. Scuito, J. Zimmerman, J. Forlizzi, and A. Steinfeld, “Investigating how experienced UX designers effectively work with machine learning,” inProceedings of the 2018 Designing Interactive Systems Conference (DIS ’18), Hong Kong, China, 2018, pp. 585–596. doi: 10.1145/3196709.3196730

    doi:10.1145/3196709.3196730 2018

  9. [9]

    I lead, you help but only with enough details: Understanding user experience of co-creation with artificial intelligence,

    C. Oh, J. Song, J. Choi, S. Kim, S. Lee, and B. Suh, “I lead, you help but only with enough details: Understanding user experience of co-creation with artificial intelligence,” inProceedings of the 2018 CHI Conference on Human Factors in Computing Systems (CHI ’18), Montreal, QC, Canada, 2018, pp. 1–13. doi: 10.1145/3173574.3174223

    doi:10.1145/3173574.3174223 2018

  10. [10]

    Generative AI-enhanced human-AI collaborative conceptual design: A systematic literature review,

    C. Fang, Y . Zhu, L. Fang, Y . Long, H. Lin, Y . Cong, and S. J. Wang, “Generative AI-enhanced human-AI collaborative conceptual design: A systematic literature review,”Design Studies, vol. 97, p. 101300, 2025. doi: 10.1016/j.destud.2025.101300

    work page doi:10.1016/j.destud.2025.101300 2025

  11. [11]

    Exploring creativ- ity in human–AI co-creation: A comparative study across design experience,

    N. Wang, H. Kim, J. Peng, and J. Wang, “Exploring creativ- ity in human–AI co-creation: A comparative study across design experience,”Frontiers in Computer Science, vol. 7, 2025. doi: 10.3389/fcomp.2025.1672735

    work page doi:10.3389/fcomp.2025.1672735 2025

  12. [12]

    Co-Designer: A human-AI collaborative framework for deriving physical product concept design opportunities from user-generated content (UGC) in man- ufacturing,

    R. Cai, M. Guo, S. Yang, R. Li, X. Wu, and J. Li, “Co-Designer: A human-AI collaborative framework for deriving physical product concept design opportunities from user-generated content (UGC) in man- ufacturing,” inProceedings of the 7th ACM International Conference on Multimedia in Asia, 2025, Art. no. 154. doi: 10.1145/3743093.3771654

    work page doi:10.1145/3743093.3771654 2025

  13. [13]

    Framing effects influence interface feature decisions,

    A. Cockburn, B. Lewis, P. Quinn, and C. Gutwin, “Framing effects influence interface feature decisions,” inProceedings of the 2020 CHI Conference on Human Factors in Computing Systems (CHI ’20), New York, NY , USA: ACM, 2020, pp. 1–11. doi: 10.1145/3313831.3376496

    doi:10.1145/3313831.3376496 2020

  14. [14]

    Choice architecture for human-computer interaction,

    A. Jameson, B. Berendt, S. Gabrielli, F. Cena, C. Gena, F. Vernero, and K. Reinecke, “Choice architecture for human-computer interaction,” F oundations and Trends in Human–Computer Interaction, vol. 7, no. 1–2, pp. 1–235, Oct. 2014. doi: 10.1561/1100000028

    work page doi:10.1561/1100000028 2014

  15. [15]

    Mapping the landscape of creativity support tools in HCI,

    J. Frich, L. MacDonald Vermeulen, C. Remy, M. M. Biskjaer, and P. Dalsgaard, “Mapping the landscape of creativity support tools in HCI,” inProceedings of the 2019 CHI Conference on Human Factors in Computing Systems (CHI ’19), New York, NY , USA: ACM, 2019, pp. 1–18. doi: 10.1145/3290605.3300619

    doi:10.1145/3290605.3300619 2019

  16. [16]

    The effects of generative AI on design fixation and divergent thinking,

    S. Wadinambiarachchi, R. M. Kelly, S. Pareek, Q. Zhou, and E. Velloso, “The effects of generative AI on design fixation and divergent thinking,” inProceedings of the 2024 CHI Conference on Human Factors in Computing Systems (CHI ’24), Honolulu, HI, USA: ACM, 2024, pp. 1–18. doi: 10.1145/3613904.3642919

    doi:10.1145/3613904.3642919 2024

  17. [17]

    Extremeness aversion and attribute-balance effects in choice,

    A. Chernev, “Extremeness aversion and attribute-balance effects in choice,”Journal of Consumer Research, vol. 31, no. 2, pp. 249–263, Sep. 2004. doi: 10.1086/422105

    work page doi:10.1086/422105 2004

  18. [18]

    Choice in context: Tradeoff contrast and extremeness aversion,

    I. Simonson and A. Tversky, “Choice in context: Tradeoff contrast and extremeness aversion,”Journal of Marketing Research, vol. 29, no. 3, pp. 281–295, 1992. doi: 10.1177/002224379202900301

    work page doi:10.1177/002224379202900301 1992