cs.HCApr 4, 2026
SaveIncentives shape how humans co-create with generative AI
Organizations: Massachusetts Institute of Technology
Abstract
Generative AI is quickly becoming an integral part of people's everyday workflows. Early evidence has shown that while generative AI can increase individual-level productivity, it does so at the cost of collective diversity, potentially narrowing the set of ideas and perspectives produced. Our research stands in contrast to this concern: through a pre-registered randomized control trial, we show that incentives mediate AI's homogenizing force in a creative writing task where participants can use AI interactively. Participants rewarded for originality relative to peers produce collectively more diverse writing than those rewarded for quality alone. This divergence is driven not by abandoning AI, but by how participants use it: those incentivized for originality incorporate fewer AI suggestions verbatim, relying on the model more selectively for brainstorming, proofreading, and targeted edits. Our results reveal that the effects of generative AI depend not only on the technology itself, but also the behavioral strategies and incentive structures surrounding its use.
Figures & tables
| Originality ( O ) | Technical Quality ( T ) |
| “You are participating in a short story competition. There are thousands of submissions, so your goal is to stand out as much as possible. Find your voice and be as creative as possible! The short story should be around 250–350 words.” Meta instruction: “Your bonus will be determined based on originality and uniqueness.” | “You are starting an Intro to Writing class. Your first assignment is to create a 250–350 word short story. Your goal is to get an A by submitting a high-quality piece of work .” Meta instruction: “Your bonus will be determined based on the grade you receive.” Grading is based on a rubric, see Figure 8 . |
Table 1: Instructions given to the originality ( O ) and technical quality ( T ) conditions.
| AI-O | SELF-O | AI-T | SELF-T | |
| Continuous variables | ||||
| Prolific Approvals | 1211.38 (62.21) | 1171.52 (57.80) | 1193.66 (61.10) | 1165.34 (61.04) |
| Age | 40.74 (1.90) | 39.86 (1.86) | 41.32 (1.93) | 40.04 (1.75) |
| AI Experience (general) | 2.30 (0.14) | 1.94 (0.16) | 1.90 (0.15) | 2.04 (0.14) |
| AI Experience (writing) | 1.64 (0.17) | 1.08 (0.16) | 1.46 (0.17) | 1.06 (0.15) |
| Binary variables | ||||
Table 2: Balance table of participant demographics by group. For continuous and binary variables, we display the mean and standard error (in parentheses). All groups consist of participants. “NIPW” stands for Not in Paid Work. AI Experience is on a Likert scale from 0 to 4, with 0 being not at all familiar and 4 being extremely familiar with AI tools.
| Name | Description |
|---|---|
| Embeddings (using cosine similarity) | |
| all-MiniLM-L12-v2 a | One of the most popular lightweight embedding models |
| all-mpnet-base-v2 b | One of the most popular lightweight embedding models |
| Qwen3-Embedding-0.6B [ 41 ] | Lightweight ( 1B) but high-performing on embedding benchmarks c |
| embeddinggemma-300m [ 31 ] | Lightweight ( 1B) but high-performing on embedding benchmarks c |
| Style-Embedding [ 34 ] | Intended to capture style more than content (for author attribution) |
Table 3: Table of similarity metrics used, in that higher similarity means reduced diversity. ⋆ -gram diversity score is the only diversity measure, where a higher score means more diversity.
Figure 1: Effect sizes from difference in means tests (Welch t-test, one-sided) with significance: ∗ , ∗∗ , ∗∗∗ , over the final submissions (left) and between the final submission and the first valid draft the AI suggested (right). A higher positive value means that the first group in the difference-in-means test is more homogeneous. All rows are similarity metrics, except for -gram diversity score (NGDS), which is a diversity metric. As such, we flip the comparison for NGDS so that the effect sizes are directionally consistent. See Appendix Table 5 for specific effect sizes and multiple hypothesis correction.
Figure 2: (a) Average cosine similarity of embeddings across randomized groups, for both the final submission and the first full draft the AI produces (when available). (b) [Top] Average cosine similarity of embeddings across randomized groups through time spent in the session. We only include text that has at least 200 words for every timestep to ensure that the embeddings capture enough information about the story. [Bottom] Number of requests over time spent in the session. (c) First and second principal component of the embeddings across groups, for both the drafts that AI produces (when available) and the final submissions. All plots use the embedding model all-MiniLM-L12-v2 . See Appendix C.2 for other metrics as robustness checks.
Figure 3: Example of the AI attribution procedure to construct the adoption metric.
Figure 4: (a) Histogram of adoption score for both the AI-T and AI-O groups. (b) Time spent writing AI prompts by adoption and group. (c) Number of words in AI prompts by adoption and group. (d) Number of AI requests by adoption and group.
Figure 5: Distribution of editing and drafting requests normalized by number of participants in each group, over adoption levels (a), incentive group (b), and general experience with AI (c). High (low) experience is a self-reported score of ( ).
Figure 6: Effect sizes from difference in means tests (Welch t-test, one-sided) with significance: ∗ , ∗∗ , ∗∗∗ , over the final submissions stratified by low and high adoption on AI among the AI-assisted group. See Appendix Table 6 for specific effect sizes.
| High adoption ( ) | Low adoption ( ) | |
| Continuous variables | ||
| Total Approvals | 1221.48 (96.18) | 1263.19 (60.43) |
| Age | 39.19 (2.18) | 42.16 (2.39) |
| AI Experience (General) | 2.52 (0.21) | 1.81 (0.19) |
| AI Experience (Writing) | 2.00 (0.26) | 1.14 (0.19) |
| Binary variables | ||
Table 4: Participant demographics for high adoption and low adoption groups. For continuous and binary variables, we display the mean and standard error (in parentheses). “NIPW” stands for Not in Paid Work. AI Experience is on a Likert scale from 0 to 4, with 0 being not at all familiar and 4 being extremely familiar with AI tools.
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: [Top] Instructions for AI group; [Bottom] Instructions for SELF group
Figure 8: [Top] Prompt and rubric for O group; [Bottom] Prompt and rubric for T group
Figure 9: Brainstorming and outlining page.
Figure 10: [Top] Writing interface for SELF group; [Bottom] Writing interface for AI group.
Figure 11: Post-session survey questions.
Figure 12: Warning sign instructing users to enter full screen to continue the study.
Figure 13: Attention checks for the Instruction phase (top) and the Prompt phase (bottom). Participants are not allowed to continue the session if they fail to answer both questions correctly.
Figure 14: Correlation matrix of various adoption metrics. For example, means that the adoption metric aggregates from to 10.
| Comparison | AI-O – SELF-O | AI-T – SELF-T | AI (T – O) | SELF (T – O) | |||
|---|---|---|---|---|---|---|---|
| Version | Final | First (AI) | Final | First (AI) | Final | First (AI) | Final |
| Embeddings | |||||||
| all-MiniLM-L12 | 0.043^{\star\star\star}_{{\color[rgb]{1,0,0}\star\star\star}} | 0.212^{\star\star\star}_{{\color[rgb]{1,0,0}\star\star\star}} | 0.062^{\star\star\star}_{{\color[rgb]{1,0,0}\star\star\star}} | 0.191^{\star\star\star}_{{\color[rgb]{1,0,0}\star\star\star}} | 0.023^{\star\star}_{{\color[rgb]{1,0,0}\star}} | ||
| all-mpnet-base | 0.043^{\star\star\star}_{{\color[rgb]{1,0,0}\star\star\star}} | 0.227^{\star\star\star}_{{\color[rgb]{1,0,0}\star\star\star}} | 0.076^{\star\star\star}_{{\color[rgb]{1,0,0}\star\star\star}} | 0.232^{\star\star\star}_{{\color[rgb]{1,0,0}\star\star\star}} | 0.029^{\star\star}_{{\color[rgb]{1,0,0}\star}} | ||
| Qwen3-Embedding | 0.060^{\star\star\star}_{{\color[rgb]{1,0,0}\star\star\star}} | 0.027^{\star\star\star}_{{\color[rgb]{1,0,0}\star\star\star}} | 0.087^{\star\star\star}_{{\color[rgb]{1,0,0}\star\star\star}} | ||||
| embeddinggemma | 0.031^{\star\star\star}_{{\color[rgb]{1,0,0}\star\star\star}} | 0.178^{\star\star\star}_{{\color[rgb]{1,0,0}\star\star\star}} | 0.044^{\star\star\star}_{{\color[rgb]{1,0,0}\star\star\star}} | 0.189^{\star\star\star}_{{\color[rgb]{1,0,0}\star\star\star}} | |||
Table 5: Effect sizes from difference in means tests (Welch t-test, one-sided) over the final submissions (“Final”) and the first valid draft the AI suggested when available (“First (AI)”). Black significance stars are standard p-values, while red starts use Benjamini–Hochberg (BH) FDR-adjusted p-values: ∗ , ∗∗ , ∗∗∗ .
Figure 15: Average similarity (diversity for NGDS) across randomized groups, for both the final submission and the first full draft the AI produces (when available). NGDS is the diversity metric, in that a higher value means lower similarity.
Figure 16: Average similarity (diversity for NGDS) randomized groups through time spent in the session. NGDS is the diversity metric, in that a higher value means lower similarity. We only include text that has at least 200 words for every timestep to ensure that the embeddings capture enough information about the story.
Figure 17: First and second principal component of the embeddings across groups, for both the drafts that AI produces (when available) and the final submissions.
| Comparison | AI-O - SELF-O | AI-T - SELF-T | AI-T - AI-O | |||
|---|---|---|---|---|---|---|
| AI Adoption level | Low | High | Low | High | Low | High |
| Embeddings | ||||||
| all-MiniLM-L12-v2 | 0.015 | 0.069*** | 0.051*** | 0.072*** | 0.041** | 0.008 |
| all-mpnet-base-v2 | 0.016 | 0.061*** | 0.021 | 0.115*** | 0.002 | 0.050*** |
| Qwen3-Embedding-0.6B | -0.016 | 0.009 | 0.004 | 0.043*** | -0.005 | 0.009 |
| embeddinggemma-300m | 0.007 | 0.036*** | 0.011 | 0.072*** | -0.023 | 0.010 |
Table 6: Effect sizes from difference in means tests (Welch t-test, one-sided) with significance: ∗ , ∗∗ , ∗∗∗ , over the final submissions stratified by low and high adoption on AI among the AI-assisted group.