PaperlayerView exact revision on arXiv

Source-linked AI summary

Beyond Better Ideas: How Human-AI Collaboration Shapes Scientific Ideas and Human Perceptions

Qian Chen, Saatvik Pradhan, Janice Ahn, Berk Atil, Angkai Li, Shijie Lu, Nobin Sarwar, Michael Ulis, Ruyuan Wan, Wenpeng Yin

arXiv:2610.30568v1cs.HC

TL;DR

Research on AI-assisted science has often studied idea generation, evaluation, provenance, and human perceptions separately. This study examines them together through proposal teams, human and AI evaluation, provenance analysis, and participant surveys; it finds no human-evaluated quality advantage for AI collaboration, alongside differences in AI judging, partial provenance detection, and recalibrated perceptions.

  • Problem

    Research has largely examined AI’s roles in scientific idea generation, collaboration, evaluation, provenance, and human perceptions in isolation, leaving their interactions insufficiently studied.

    Source: §1 · p. 2

  • Method

    The study combines Human+AI and Human+Human proposal teams, human and LLM evaluation, collaboration-provenance analysis, and participant questionnaires in one empirical framework.

    Source: §1 · p. 2

  • Results

    Human+AI and Human+Human teams show nearly indistinguishable human-rated overall quality, while revision improves both; AI judges can favor AI-assisted proposals without a corresponding human advantage.

    Source: §1 · p. 2

  • Takeaways & Limitations

    In this setting, AI’s influence spans evaluation, attribution, and researchers’ perceptions, not just the quality of ideas produced.

    Source: §5 · p. 17

    Differences in cohort composition, research experience, and language background may contribute to proposal-quality differences, preventing causal interpretation.

    Source: §4.1 · p. 8

Abstract

from arXiv · show

As AI increasingly becomes a collaborator in scientific research, its impact may extend beyond improving ideas to reshaping how they are evaluated, attributed, and perceived by humans. Yet existing research has largely examined these dimensions in isolation. We present a unified empirical study of Human-AI scientific ideation, investigating two complementary perspectives: i) how AI collaboration shapes scientific ideas and their evaluation, and ii) how firsthand collaboration reshapes human perceptions of AI. Through a shared experimental setting involving Human-AI and Human-Human research teams, human and LLM evaluators, collaboration-provenance analysis, and participant surveys, we uncover several counterintuitive findings. First, AI collaboration shows no significant advantage in human-evaluated idea quality, while iterative revision consistently improves quality. Second, AI judges can approximate human scores yet disagree on idea rankings; notably, an AI judge favors AI-assisted proposals without a corresponding advantage in human evaluations. Third, collaboration provenance remains only partially detectable, revealing substantial overlap between Human-AI and Human-Human scientific outputs. Finally, firsthand AI collaboration is associated with greater perceived usefulness but smaller efficiency gains than expected, suggesting a nuanced recalibration of human expectations. Overall, our findings challenge the notion that AI's value in scientific ideation lies primarily in producing better ideas, revealing instead its broader influence on scientific creation, evaluation, attribution, and human perceptions.

1 INTRODUCTION

The introduction frames scientific ideation as a connected process of creation, evaluation, attribution, and human perception, then presents an empirical study of these dimensions. Its findings show no human-evaluated quality advantage for Human+AI teams, while AI judges may favor AI-assisted work and firsthand collaboration recalibrates perceptions.

1 INTRODUCTION

  • The paper asks whether AI collaboration improves ideas, whether AI judges preserve human preferences, whether collaboration leaves detectable traces, and how experience changes researchers’ beliefs.The introduction argues these questions require observing the full pipeline from idea production through evaluation, provenance, and human perceptions.Source: §1 · p. 2
  • Human+AI and Human+Human teams achieve nearly indistinguishable overall quality under human evaluation, while both improve substantially through revision.The authors suggest iterative refinement, rather than AI participation itself, is the dominant benefit in this setting.Source: §1 · p. 2
  • The study connects proposal creation, human and AI evaluation, provenance analysis, and participant perception within one empirical framework.Its real-world corpus contains 70 proposals from 35 student teams under Human+AI and Human+Human conditions, with initial and revised submissions.Source: §1 · p. 2
  • GPT-4.1 can be calibrated to human scores yet diverge from human preference rankings, and it scores Human+AI proposals higher without a corresponding human-panel advantage.The largest divergence appears in broad applicability.Source: §1 · pp. 2–3
  • Collaboration provenance is only partially detectable, while firsthand AI collaboration is associated with greater perceived usefulness but smaller-than-expected speed-ups.The introduction describes substantial overlap between Human+AI and Human+Human outputs and reports lower confidence in integrating GenAI than participants initially expected.Source: §1 · pp. 2–3

2 RELATED WORK

Related work spans Human–AI collaboration, scientific ideation, AI evaluation, provenance, and human perceptions. This paper connects these lines by examining collaboration-produced artifacts, their evaluation, and how their production process and firsthand use relate to judgments and beliefs.

Human–AI Collaboration and Scientific Ideation

  • Prior collaboration studies report productivity or quality gains, possible creativity–diversity tradeoffs, and benefits that may reverse beyond a model’s capability frontier.The paper argues that collaboration should be evaluated across multiple dimensions rather than by quality alone.Source: §2 · p. 3
  • Scientific ideation work includes systems optimizing novelty and studies contrasting human- and AI-generated ideas; this paper instead studies human-led design with AI participating in development and revision.Prior findings cited here include AI-generated ideas judged more novel but less feasible than human-generated ones.Source: §2 · p. 3

LLMs as Evaluators and Scientific Judges

  • LLM evaluators can align substantially with human judgments, but agreement varies and systematic biases may persist despite strong aggregate agreement.The paper distinguishes score calibration from preference alignment because human-like scores do not ensure human-like rankings.Source: §2 · p. 3

Creator–Evaluator Interaction

  • The paper extends evaluator-bias research by asking whether judges evaluate Human+AI-produced work differently from Human+Human work when human evaluators show no such preference.This creator–evaluator interaction means measured gains may depend on who evaluates the work, not only on its quality.Source: §2 · p. 4

Detecting Collaboration Provenance

  • Rather than identifying human versus AI authorship, the provenance analysis asks whether the collaboration process leaves detectable signatures in methodologies written by human authors.The related work also notes that text detectors can be fragile and capture unintended linguistic characteristics.Source: §2 · p. 4

Human Perceptions of AI Collaboration

  • The study complements work on AI expectations and educational norms by examining perceptions after participants experience scientific collaboration with GenAI.The perceptions considered include usefulness, trustworthiness, efficiency, and AI’s role in research.Source: §2 · p. 4

3 DATA COLLECTION

The data collection combines synthetic and real proposal data, matched human and AI evaluations, and pre/post participant questionnaires. Together, these components support analysis of artifacts, their evaluation, and participants’ perceptions after collaboration.

3 DATA COLLECTION

  • The study combines synthetic training data, real-world proposals with human and AI evaluations, and questionnaires capturing expectations and post-study perceptions.These sources support three perspectives: produced artifacts, their evaluation, and changes in the humans involved.Source: §3 · p. 4

3.1 SCIENTIFIC IDEATION DATA

  • The synthetic corpus contains 2,000 methodology instances from 1,000 human-authored papers, using source-paper-grounded and research-question-only generation as proxies for two collaboration conditions.GPT-4.1 verbalizes both conditions, so the controlled distinction is the source of methodological reasoning, not surface authorship; the real proposals are reserved for transfer evaluation.Source: §3.1 · pp. 4–5
  • The real corpus contains 70 proposal identities across 20 research questions, with Human+AI and Human+Human teams submitting initial and revised proposals.Human reviewers and AI judges use the same multidimensional rubric and also rank proposals within each research question.Source: §3.1 · p. 5

3.2 PARTICIPANT QUESTIONNAIRE DATA

  • Pre/post questionnaires capture eight comparable perception constructs, with responses harmonized to a common 1–5 ordinal scale where encodings differed.The perception analysis focuses on participants assigned to Human+AI; 37 participants appear in both surveys.Source: §3.2 · p. 5

4 EXPERIMENTS

The experiments examine how collaboration style and revision relate to proposal quality, how AI and human evaluations compare, and how collaboration provenance can be detected. They also assess how firsthand collaboration relates to perceptions of GenAI.

4.1 RQ1: HOW DO DIFFERENT FACTORS (E.G., COLLABORATION STYLE, PROPOSAL REVISION, RESEARCHER EXPERIENCE, AND TASK DIFFICULTY) AFFECT THE QUALITY OF SCIENTIFIC RESEARCH PROPOSALS?

  • Human+AI proposals scored only 0.04 points above Human+Human proposals (95% CI [−0.31, 0.40], Hedges’ g = 0.08), while revision improved overall quality by 0.18 points and normalized rank by 0.13.Revision gains did not differ significantly by collaboration condition (∆∆ = −0.11, p = .58).Source: §4.1 · pp. 6–7, Table 2 · p. 7
  • Undergraduate proposals scored higher than graduate proposals on broad applicability (+0.40, q = .025) and normalized rank (−0.27, p = .024), while task difficulty showed no detectable association in 20 research questions.The authors caution that cohort composition, research experience, and language background may contribute to the undergraduate–graduate pattern, so it is not causal evidence; the difficulty result rules out a large difference in this sample, not all possible effects.Source: §4.1 · pp. 7–8

4.2 RQ2: HOW WELL DOES AI-AS-JUDGE ALIGN WITH HUMAN JUDGMENT?

  • GPT-4.1’s score MAE was 0.48 versus the human baseline’s 0.57, but its ranking τ was 0.42 versus 0.57, showing score calibration did not ensure preference alignment.AI judges were most human-like on detail and clarity, with larger gaps on novelty and feasibility; overall, their reliability varied by model, criterion, and evaluation task.Source: §4.2 · pp. 9–11, Table 4 · p. 10, Table 5 · p. 11

4.3 RQ3: DOES AI SCORE HUMAN+AI PROPOSALS MORE FAVORABLY THAN HUMAN+HUMAN PROPOSALS?

  • GPT-4.1’s mean across all six ratings was 3.90 for Human+AI proposals versus 3.76 for Human+Human proposals, while human-panel means were 3.605 in both conditions.For overall ratings, GPT-4.1 scores were 3.94 versus 3.76, compared with human scores of 3.45 versus 3.41; its broad-applicability bias gap was 0.51 points. GPT-4.1 overall high-score rates were 94.4% versus 76.5% (Fisher’s exact p = 0.043), while human all-criteria high-score rates were 16.7% versus 17.6%.Source: Table 6 · p. 12, §4.3 · p. 12, Table 7 · p. 12
  • GPT-4.1 systematically favored Human+AI proposals unlike human reviewers, especially on broad applicability; the pattern persisted after accounting for human scores and criteria, though the proposal-level evaluator discrepancy was marginal.With criterion fixed effects, the Human+AI effect was 0.13 for GPT-4.1 scores (p = 0.002) and the GPT–human discrepancy (p = 0.011).Source: §4.3 · pp. 12–13

4.4 RQ4: CAN WE DETECT WHETHER A PROPOSAL WAS PRODUCED BY HUMAN–AI OR HUMAN–HUMAN COLLABORATION?

  • 65.7% accuracy on unified proposals and 67.1% on pre-unified proposals were achieved by fine-tuned Qwen; GPT-4.1 also reached 67.1% accuracy on pre-unified text.Fine-tuning improved unified accuracy by 18.6 percentage points and macro-F1 from 0.362 to 0.657; GPT-4.1 had the strongest pre-unified macro-F1 at 0.671.Source: Figure 3 · p. 14, §4.4 · p. 14
  • Methodology contained detectable provenance signals, but moderate classification accuracy and substantial overlap made individual proposals difficult to attribute reliably.Synthetic supervision improved Qwen’s detection, particularly on unified proposals, but methodology alone remained insufficient for reliable individual-proposal attribution.Source: §4.4 · p. 14

4.5 RQ5: HOW DOES HUMAN–AI COLLABORATION CHANGE PARTICIPANTS’ PERCEPTIONS OF GENERATIVE AI?

  • −0.386 in human effort needed to refine GenAI outputs (3.600 to 3.214) and −0.319 in GenAI speed-up (3.533 to 3.214) reflected less effort than expected but smaller realized speed gains.Perceived challenge severity also decreased from 3.467 to 3.343 (−0.124).Source: §4.5 · p. 15
  • Perceptions shifted in multiple directions: overall attitude positivity and confidence in integrating GenAI decreased slightly, so the pre/post differences indicate descriptive recalibration rather than causal within-participant change.Attitude positivity changed from 4.294 to 4.200 (−0.094), and integration confidence from 3.706 to 3.500 (−0.206).Source: §4.5 · p. 15
  • 3.250 to 3.567 (+0.317) was the increase in perceived helpfulness across six research-design tasks after Human+AI collaboration.Perceived research-design capability enhancement rose from 3.222 to 3.429 (+0.206), and trust in GenAI recommendations rose from 2.733 to 2.929 (+0.195).Source: §4.5 · p. 15, Figure 4 · p. 17

5 CONCLUSION

The conclusion finds that AI collaboration does not necessarily improve research quality beyond human collaboration, while revision remains a strong driver of improvement. It also highlights AI judges’ preferences, weakly detectable provenance, and shifts in researchers’ perceptions.

5 CONCLUSION

  • Human+AI collaboration does not necessarily improve research quality beyond human collaboration, while iterative refinement remains a powerful driver of improvement.The conclusion also warns that AI judges may favor AI-assisted proposals despite comparable human evaluations.Source: §5 · p. 17
  • Collaboration provenance leaves only weakly detectable traces in scientific artifacts, while firsthand collaboration reshapes researchers’ perceptions of AI’s capabilities and limitations.The conclusion calls for understanding and governing interactions among human creativity, AI assistance, and scientific evaluation.Source: §5 · p. 17

A RQ1 SUPPLEMENTARY FIGURES

The supplementary figures show proposal-level overall-rating distributions and diagnostics for the primary mixed-effects model. The diagnostics note discrete-outcome residual bands and lower-tail departures in the Q–Q plot.

A RQ1 SUPPLEMENTARY FIGURES

  • Figure 6 reports residual-versus-fitted, normal Q–Q, residual-distribution, and observed-versus-fitted diagnostics for the primary mixed-effects model.The visible residual bands follow from the discrete 1–5 outcome, and the Q–Q plot shows lower-tail departures.Source: A RQ1 SUPPLEMENTARY FIGURES · p. 20, Figure 6 · p. 20
  • Figure 5 presents proposal-level mean overall-rating distributions by collaboration condition, course level, submission round, and task difficulty.Points represent individual proposals.Source: Figure 5 · p. 20

B RQ2 CONFUSION MATRICES

Figures 7–10 present confusion matrices comparing each model’s integer scores with rounded mean human scores across six criteria.

B RQ2 CONFUSION MATRICES

C RESEARCH QUESTIONS USED IN THE REAL-WORLD STUDY

The appendix identifies the research questions assigned to the 70 real-world proposals.

C RESEARCH QUESTIONS USED IN THE REAL-WORLD STUDY

Related papers

Loading 2610.30568v1…