🤖 AI

Incentives shape how humans co-create with generative AI

This pre-registered randomized control trial demonstrates that while generative AI can homogenize creative output, incentivizing participants for originality rather than just quality preserves collective diversity by encouraging more selective and strategic use of AI tools.

Original authors: Nathanael Jo, Manish Raghavan

Published 2026-04-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Nathanael Jo, Manish Raghavan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a giant, super-smart kitchen where everyone is trying to bake a unique cake. In this kitchen, there's a magical robot assistant (Generative AI) that can instantly mix ingredients and bake a perfect, standard cake for you in seconds.

The big worry in the world right now is: "If everyone uses this robot, won't all the cakes start tasting exactly the same? Won't we lose our unique flavors?"

This paper is like a science experiment in that kitchen. The researchers wanted to see if the robot always makes cakes taste the same, or if the rules of the game (the incentives) change how people use the robot.

Here is the story of what they found, broken down simply:

1. The Setup: The "Cookie Cutter" vs. The "Chef"

The researchers hired 200 people to write a short story. They split them into groups:

The Solo Chefs: People who had to write the story entirely by themselves.
The Robot-Assisted Chefs: People who could use the AI robot to help.

But here's the twist: They gave the Robot-Assisted Chefs two different prizes (incentives):

Group A (The "Quality" Group): "Write the best, most grammatically perfect story you can. Get an 'A' grade."
Group B (The "Originality" Group): "Write a story that is totally unique. We want to hear your voice. Stand out from the crowd!"

2. The Problem: The Robot's "Default Flavor"

When the researchers looked at the very first draft the robot made for everyone, it was like a cookie cutter. Every robot draft started with the same clichés (like "I woke up one sunny morning..."). They all sounded like the same generic robot voice.

If the people just accepted these first drafts, the world would indeed be full of boring, identical stories.

3. The Discovery: It's All About the "Chef's Touch"

The magic happened when the people started editing.

The "Quality" Group: These people treated the robot like a shortcut. They took the robot's cookie-cutter cake, added a little bit of frosting, and submitted it. Their stories ended up sounding very similar to each other. They relied heavily on the robot's suggestions.
The "Originality" Group: These people looked at the robot's cake and said, "Nope, this tastes like everyone else's cake." They used the robot as a tool, not a replacement. They asked the robot for ideas, but then they threw away the boring parts, rewrote the sentences, and added their own secret spices.

The Result: The "Originality" group ended up with stories that were much more diverse and unique than the "Quality" group. They didn't stop using the robot; they just used it differently.

4. The Surprising Twist: Experience Can Be a Trap

The researchers also found something interesting about experience.

People who were new to AI were more likely to say, "This robot cake tastes weird, I'll fix it myself."
People who were experts at using AI actually relied on the robot more. They trusted the robot so much that they kept accepting its suggestions, which made their stories end up sounding more like the robot's "default flavor."

It's like a master chef who gets so used to a pre-made sauce that they forget to taste the food and adjust the seasoning.

The Big Lesson: The Rules of the Game Matter

The main takeaway of this paper is this: AI doesn't automatically make everything boring. The way we reward people does.

If you tell people, "Just get the job done fast and make it look perfect," they will lean on the AI, and everything will look the same.
If you tell people, "We want to see your unique style," they will fight against the AI's generic suggestions and create something diverse.

In real life:
Think about college essays. If a student uses AI just to write a "good" essay, it might sound like thousands of other AI essays. But if the student is told, "We want to see your specific, weird, unique story," they will use the AI to brainstorm and check grammar, but they will write the core story themselves.

The Bottom Line:
Generative AI is like a powerful engine. If you just let it drive, everyone ends up at the same destination. But if you put a human in the driver's seat and tell them, "Drive somewhere no one has been before," they will use that engine to go on a wild, unique adventure. The technology isn't the problem; the incentives are what shape the outcome.

1. Problem Statement

The rapid integration of Generative AI (GenAI) into creative workflows has raised concerns about homogenization. Previous research suggests that while AI boosts individual productivity, it narrows the diversity of ideas and perspectives, potentially leading to a "compressed" creative landscape (often termed "AI slop").

The authors argue that existing findings on homogenization are premature and limited by two key shortcomings:

Lack of Real-World Incentives: Most controlled experiments lack strong incentives for differentiation. Participants may accept the first AI suggestion simply to finish a task quickly, which does not reflect high-stakes real-world scenarios (e.g., college admissions or professional writing) where standing out is crucial.
Limited Interactivity: Prior studies often restrict human-AI interaction to single modes (e.g., brainstorming or text completion). Real-world usage is iterative and open-ended, allowing users to refine, edit, and steer AI outputs over time.

Research Question: How do incentive structures and interactive co-creation strategies mediate the homogenizing force of GenAI in creative writing tasks?

2. Methodology

The authors conducted a pre-registered, between-subjects randomized control trial (RCT) with $N=200$ participants recruited via Prolific.

Experimental Design

The study utilized a $2 \times 2$ factorial design:

Factor 1: AI Usage
- SELF: Participants wrote a 250–350 word short story without AI.
- AI: Participants used an integrated writing assistant (OpenAI's GPT-5-Mini) but were required to use it at least once. They could interact freely (brainstorming, drafting, editing) but were responsible for the final text.
Factor 2: Incentives
- Originality (O): Participants were told their bonus depended on uniqueness. The top 25% of originality received \ $2–\$ 7.
- Technical Quality (T): Participants were told their bonus depended on a grade based on organization and technique (A=\ $2.5, B=\$ 1), with no limit on the number of "A" grades.

Data Collection

Temporal Data: Snapshots of the text editor every 5 seconds.
Interaction Logs: Full transcripts of AI conversations.
Demographics: Post-task surveys on AI experience and background.
Constraints: Strict attention checks (full-screen mode, inactivity tracking) and minimum time requirements (1 min brainstorming, 5 min writing).

Metrics and Analysis

Diversity Measurement: The authors used a suite of similarity metrics to measure diversity (lower similarity = higher diversity).
- Embedding-based: Cosine similarity using models like all-MiniLM-L12-v2, Qwen3-Embedding, and embeddinggemma.
- N-gram/Token-based: BLEU, ROUGE-L, and compression ratios.
Adoption Metric ( $A$ ): A novel lexical attribution metric calculating the fraction of n-grams in the final story that appeared in AI suggestions but not in the user's initial prompts. This quantifies how much users "anchored" on AI text.
Statistical Approach: Welch's t-tests for difference-in-means across conditions.

3. Key Contributions

Incentives as a Mediator: The study demonstrates that incentives are a critical variable. When participants are rewarded for originality, they actively mitigate AI's homogenizing effect, whereas those rewarded for quality alone produce more homogeneous outputs.
Strategic Adaptation: The paper identifies that diversity is not determined solely by the AI model but by user strategy. Incentivized users do not abandon AI; rather, they change how they use it (e.g., using it for brainstorming/proofreading rather than accepting full drafts).
The "Human-in-the-Loop" Effect: The research quantifies how human editing and iterative prompting recover diversity lost in initial AI drafts, showing that the final output is significantly more diverse than the raw AI generation.
Experience Paradox: Contrary to intuition, participants with more prior AI experience tended to rely more heavily on AI suggestions, resulting in less diverse stories, suggesting that familiarity may lead to over-reliance rather than strategic mastery in this context.

4. Key Results

A. Diversity and Homogenization

AI vs. Human: AI-assisted stories were significantly more similar to each other than human-only stories (Hypothesis 1 supported).
Draft vs. Final: Initial AI drafts were highly homogeneous (average similarity ~0.39 in AI-O vs. 0.22 in SELF-O). However, through editing, the final submissions in the AI-O condition dropped to ~0.26, approaching human-only diversity levels.
Incentive Impact: In the AI condition, participants incentivized for Originality (O) produced significantly more diverse stories than those incentivized for Quality (T). In the SELF condition, incentives did not significantly alter diversity, suggesting incentives primarily shape the human-AI interaction process.

B. Behavioral Strategies (The "How")

Adoption Levels: The distribution of AI adoption was bimodal (users either heavily anchored on AI or barely used it).
Incentive Shift: The AI-O group shifted toward lower adoption (anchoring less on AI text) compared to the AI-T group.
Effort vs. Outcome: Interestingly, the AI-O group spent more time prompting and interacting with the AI (up to 50% more) than the AI-T group, yet this extra effort did not linearly translate to higher diversity. The key differentiator was selective rejection of AI text, not the volume of interaction.
Experience Correlation: Users with high self-reported AI experience were more likely to be in the "High Adoption" group, leading to more homogeneous outputs. They tended to iterate on AI drafts rather than diverge from them.

C. Qualitative Findings

User Agency: Most users viewed AI as a tool for scaffolding (e.g., "Get a strong baseline, then edit manually"). Even high-adoption users felt a sense of ownership.
Dissatisfaction: Initial AI drafts were frequently perceived as unsatisfactory. Low-adoption users often disengaged after one or two unsatisfactory drafts, while high-adoption users iterated extensively.
AI Voice: AI-generated drafts formed a distinct cluster in PCA space (e.g., starting with "I woke up..."), which users actively worked to dismantle through editing.

5. Significance and Implications

Design Implications: The homogenizing effect of AI is not inevitable. System designers and policymakers should focus on incentive structures and interaction designs that encourage differentiation.
Policy & Education: In high-stakes environments (e.g., college admissions), competitive incentives naturally drive users to differentiate, potentially counteracting AI homogenization. Educators could design rubrics that explicitly reward originality and personal voice to prevent over-reliance.
Theoretical Shift: The paper moves the discourse from "AI replaces humans" to "AI changes human strategy." It suggests that the outcome of human-AI co-creation depends on the alignment of reward structures with desired values (e.g., diversity vs. efficiency).
Future Work: The authors note that while short-term incentives help, the long-term effects of sustained AI use on human creative agency remain an open question. Additionally, the trade-off between diversity and quality under different incentive structures requires further study.

Conclusion: The paper concludes that incentives mediate AI's homogenizing force. By rewarding originality, users strategically reduce their reliance on verbatim AI suggestions, leveraging the tool for brainstorming and editing rather than generation, thereby preserving collective diversity in creative outputs.