AI and Work

Algorithmic Management
Professional Artists' Use of Generative AI
Group Selection as a Safeguard Against AI Substitution

As advances in robotics and AI revive concerns about the impact of automation on jobs, many questions loom: How will automation affect employment in the workplace, and how will this impact differentially affect different communities, cities, and economies? How will human workers adapt to a world with AI tools, co-workers, and managers? How can employers and policy-makers respond to such challenges? CHM uses tools from economics, behavioral science, complex systems, and urban science to explore this question.

Algorithmic Management

How do people feel and behave under algorithm-managed workplaces?

Companies and organizations increasingly use AI-powered algorithms to manage human workers. Yet, experimental findings on how people respond to algorithmic authority remain inconclusive, leaving us without clear behavioral insight into the consequences of AI management under controlled causal conditions. These divergent findings reflect not only theoretical tensions but also, perhaps more importantly, methodological challenges in designing ecologically valid simulations of AI management in a laboratory setting.

In a Perspective paper, Dong, Bonnefon, and Rahwan (2024) first reviewed mainstream methodological approaches and their limitations. Vignette studies, where participants imagine being managed by AI, tend to overestimate negative reactions because people poorly forecast emotional responses, rely on dystopian media narratives, and default to status quo preferences. Case studies, by contrast, can overestimate positive reactions: workers already embedded in AI-managed environments may self-censor due to job insecurity, and organizations adopting AI management are not representative of the broader labor market. Field experiments in real or simulated labor markets were further proposed as a promising methodological direction, which is capable of capturing authentic behavior while preserving experimental control.

To operationalize this vision, a high-fidelity AI management experiment was conducted in Minecraft, building a 3D immersive workplace with real-time monitoring, repeated production cycles, and contingent pay. Participants, acting as workers (N = 382), completed repeated production cycles under human, AI, or hybrid management. The experimental procedure is illustrated below in Figure 1.

*Figure 1*. *Experimental procedure* (N = 382 workers). Participants were randomly assigned to one of four management conditions: Human, Human+AI, AI-R (R for "Rule-based AI"), or AI-T (T for "Decision-tree-based AI"). Each worker was tasked with making an iron pickaxe under the same manager across three repeated 10-minute rounds. Except for the evaluation experience, participants were not made aware of the differences between the two human-managed or two AI-managed conditions. Workers self-assessed into three levels of competence (beginner, intermediate, advanced). After each turn, managers evaluated worker performance using these three levels, with corresponding wages of ¢0, ¢20, and ¢50 during this turn, respectively.

Image: Mengchen Dong / MPI for Human Development

*Figure 1*. *Experimental procedure* (N = 382 workers). Participants were randomly assigned to one of four management conditions: Human, Human+AI, AI-R (R for "Rule-based AI"), or AI-T (T for "Decision-tree-based AI"). Each worker was tasked with making an iron pickaxe under the same manager across three repeated 10-minute rounds. Except for the evaluation experience, participants were not made aware of the differences between the two human-managed or two AI-managed conditions. Workers self-assessed into three levels of competence (beginner, intermediate, advanced). After each turn, managers evaluated worker performance using these three levels, with corresponding wages of ¢0, ¢20, and ¢50 during this turn, respectively.

Image: Mengchen Dong / MPI for Human Development

As shown in Figure 2, an AI manager trained on human-defined evaluation principles systematically assigned lower performance ratings and reduced wages by 40%, without adverse effects on worker motivation and sense of fairness. This pattern was driven by a muted sensitivity to discrepancies between self-assessed and AI-assigned evaluations, compared to evaluations delivered by a human. These findings suggest the AI systems may serve as powerful instruments of silent exploitation, highlighting the need to audit their ethical and accountable use.

*Figure 2*. *Summary of the Minecraft study results* (N = 382 workers). Workers self-assessed into three levels of competence (beginner, intermediate, advanced). After each turn, managers evaluated worker performance using these three levels, with corresponding wages of ¢0, ¢20, and ¢50 during this turn, respectively.

Image: arXiv
Dong et al. (2025)
Image licensed under CC BY 4.0

*Figure 2*. *Summary of the Minecraft study results* (N = 382 workers). Workers self-assessed into three levels of competence (beginner, intermediate, advanced). After each turn, managers evaluated worker performance using these three levels, with corresponding wages of ¢0, ¢20, and ¢50 during this turn, respectively.

Image: arXiv
Dong et al. (2025)
Image licensed under CC BY 4.0

Key Reference

Dong, M., Bonnefon, J.-F., & Rahwan, I. (2024). Toward human-centered AI management: Methodological challenges and future directions. Technovation, 131, Article 102953. https://doi.org/10.1016/j.technovation.2024.102953

MPG.PuRe

DOI

publisher-version

Professional Artists' Use of Generative AI

Can professional artists express their skill when using generative AI to create visual art, or is everyone on an even playing field?

The increasing prevalence and capacities of generative machine learning models are transforming creative work. Traditionally, highly trained experts such as professional artists have played a major role as creative actors by building on their extensive set of skills. However, this traditional role is now threatened by promises that simple natural language instructions to AI, called prompts, can "make everyone an artist" by enabling high-quality image creation without the need to master the techniques used in visual art. Yet what is rarely considered is that professionals' existing skills might also transfer to using this new technology. If so, one would expect them to outperform laypeople even under the constraints of writing prompts for AI. This study takes a cross-sectional, behavioral approach to test this hypothesis by comparing 50 active professional visual artists and a demographically matched sample of laypeople in a preregistered experiment. Participants completed two tasks inspired by artistic practices, testing capabilities in both faithful and divergent image creation (Figure 3): They had to write prompts to replicate a reference image (copying task), or to move as far away as possible from it (creative task).

*Figure 3. Experimental design.* The bolded part in the prompts of the creative task was fixed to ensure it remained challenging. Example prompts and images are actual data from one of the artist participants.

Image: arXiv
Eisenmann et al. (2025)
Image licensed under CC BY 4.0

*Figure 3. Experimental design.* The bolded part in the prompts of the creative task was fixed to ensure it remained challenging. Example prompts and images are actual data from one of the artist participants.

Image: arXiv
Eisenmann et al. (2025)
Image licensed under CC BY 4.0

Key Reference

Eisenmann, T. F., Karjus, A., Canet Sola, M., Brinkmann, L., Supriyatno, B. I., & Rahwan, I. (2026). Expertise elevates AI usage: Experimental evidence comparing laypeople and professional artists. International Journal of Human-Computer Interaction. Advance online publication. https://doi.org/10.1080/10447318.2026.2669041

[Shared first authorship: Thomas F. Eisenmann, Andres Karjus].

MPG.PuRe

DOI

Group Selection as a Safeguard Against AI Substitution

Does the use of generative AI in cultural transmission increase the risk of cultural collapse in the long run?

Generative AI has now been increasingly applied to complex, creative tasks (e.g., music production, commercial design), with performance often exceeding average human cultural skills. However, the use of AI-generated data comes with limitations: for example, training AI models on AI-generated data is known to lead to "model collapse." This paper asks whether the use of generative AI in cultural transmission can similarly lead to "cultural collapse," which refers to the risk of the reduction of our cultural diversity and complexity in the long run.

An agent-based model was used to compare two cases of AI uses, Complement and Substitute. It was shown that compared to AI Complements, AI Substitutes can improve our cumulative learning and exploration in the short term, but are detrimental in the long run due to the greater influence on collective variance. Even if some people start with complementary AI, which maintains some variance, the more efficient performance boost provided by AI Substitutes makes it more attractive to be adopted by the majority as the dominant strategy. As a result, the research demonstrates that the danger of "culture collapse" not only resides in the reduction of variance caused by substituting human effort with AI but also in its immediate boost of performance that allows it to be easily adopted. The dilemma of how environments can be created to encourage AI as a complement rather than a substitute despite higher costs and lower individual efficiency in the short-term was addressed by leveraging an evolutionary mechanism, multi-level selection. A group structure was implemented in the population and applied differential social learning rates of AI adoption within and between groups. It was found that when group boundaries are strong and between-group learning rates are low, AI Complements can be selected for over AI Substitutes at the group level (see Figure 4). This is because the group boundary slows down the short-term invasion of AI Substitute and protects the variance in the AI Complement group. In the long run, this variance can lead to more cumulative advancement within the group, which drives AI Complement to be the dominant strategy across groups. Taken together, our results offer insights into the long-term and population-level effects of AI adoption. Zhong et al. (in press) suggest that a more robust, long-term approach to AI alignment involves building an environment that preserves organizational boundaries and promotes model diversity.

Overview of Simple Chat's bridging architecture — *Figure 4. AI Complement is selected for through group-level selection but not in individual-level selection.* As a group-beneficial but individually costly strategy, using AI as a Complement can be selected for in group-level selection. Two simulations are conducted with the same parameters except for the group structure and in-group learning rate. Panel (a) is one simulation of a population of 1000 in 3 groups. Each strip in this panel shows the AI strategy use in the group across time. AI Complement (orange) slowly spreads across groups through social learning to become the dominant strategy. Panel (b) is one simulation of a mixed population of 1000. With the same social learning setup, but no group structure, AI Substitute becomes the dominant strategy. In other words, the absence of group-level selection leads to the invasion of individual-beneficial strategies that provide higher short-term benefit.

Image: arXiv
Zhong et al. (2026)

*Figure 4. AI Complement is selected for through group-level selection but not in individual-level selection.* As a group-beneficial but individually costly strategy, using AI as a Complement can be selected for in group-level selection. Two simulations are conducted with the same parameters except for the group structure and in-group learning rate. Panel (a) is one simulation of a population of 1000 in 3 groups. Each strip in this panel shows the AI strategy use in the group across time. AI Complement (orange) slowly spreads across groups through social learning to become the dominant strategy. Panel (b) is one simulation of a mixed population of 1000. With the same social learning setup, but no group structure, AI Substitute becomes the dominant strategy. In other words, the absence of group-level selection leads to the invasion of individual-beneficial strategies that provide higher short-term benefit.

Image: arXiv
Zhong et al. (2026)

Key Reference

Zhong, Q., Eisenmann, T. F., Garcia, J., & Rahwan, I. (in press). Group selection as a safeguard against AI substitution. Philosophical Transactions of the Royal Society of London: A, Mathematical, Physical and Engineering Sciences.

MPG.PuRe