WORKING PAPERS
(* denotes equal contribution)
"Looking for Something Weird to Happen": How Humans Sustain AI Agent Novelty Amid Semantic Collapse (submitted to CHI 2027)
Shiyang Lai*, Arna Woemmel*, Hongkai Mao, Junsol Kim, Summer Eunhyung Ann, and James Evans
Semantic collapse, the progressive narrowing of what AI systems generate, has been studied mainly in closed settings, and remedies have targeted models and data. We study it in MOLTBOOK, a social network of interacting AI agents that human users configure and steer. Across 30,076 active agents, output grows less diverse within agents and more similar across them over weeks, yet a minority sustains high novelty. Interviews with users of high- and typical-novelty agents (N=11) associate sustained novelty with three features: users value novelty of itself, they supply broad and distinctive material and revise it when output narrows, and they approach MOLTBOOK as a new agentic world to explore, not a venue to instrumentally exploit. A survey of users of distinctive agents (N=53) confirms these patterns. Communities with more novel agents also show more diverse output from other agents. We discuss interface and policy interventions that could support improved human input.
AgentPedia: A Collaborative and Compute-Efficient Knowledge Commons for AI Agents (submitted to ICLR 2027)
Junyi Chen*, Arna Woemmel*, Hongbo Fang*, Veniamin Veselovsky, Daniel Björkegren, Shiyang Lai, Nadav Kunievsky, Robert Ward, Davi Bastos Costa, Austin C. Kozlowski, and James Evans
AI agents spend substantial compute searching for and synthesizing information, then discard most of what they find---even though it may be useful to other agents. We introduce Agentpedia, an open knowledge commons in which agents deposit findings for reuse; agents may consult accumulated knowledge before recomputing it. Agents and humans may contribute, but we expect most readers to be agents, since the corpus does not prioritize human readability. Agentpedia holds three kinds of contributable objects: content, as self-contained knowledge units - or "finds" - that retain evidence, validation, and applicability conditions; indices, which organize and score units for credibility, freshness, or task relevance and which any participant can build; and rules, which govern admission, reconciliation, validation, and reward, and which change through the same evidence-bearing process as the content they govern. Because the corpus can be flexibly reorganized, prior computation can be repurposed for tasks its producers did not anticipate, which knowledge held in model weights or private caches does not allow. Across 10 randomized task orders in a 50-task frozen-corpus evaluation, Agentpedia reduces page reads by 32% and task-solving API tokens by 9% while raising answer completeness from 86% to 88%. In a higher-overlap setting where more prior work can be reused, page reads fall by 59% and API tokens by 21%, while completeness rises from 83% to 87%.
Digital Skills: Social Disparities and the Impact of Early Mentoring (revise and resubmit at Research Policy)
with Fabian Kosse and Tim Leffler
We analyze social disparities in digital skills, their relevance for labor market outcomes, and the long-term impact of an early childhood mentoring program. Drawing on a representative survey and a randomized controlled trial (RCT), we distinguish between proficiency and confidence in digital skills. We document three key findings. First, both skill levels and confidence are strong and similarly sized predictors of labor market earnings. Second, we find marked gender and socioeconomic disparities: males exhibit higher proficiency and confidence than females, while SES gaps only exist for males, and only in the confidence dimension. Third, we provide causal evidence that early mentoring programs can reduce this SES gap. Low SES males who received mentoring show significantly higher digital skill confidence a decade later. The effect is concentrated among those with initially low confidence and does not increase overconfidence. Mediation analysis suggests that roughly half of the effect operates through improvements in general self-concept and educational attainment.
Fragile AI Optimism (under review)
with Lydia Mechtenberg, and Hendrik Hüning
Winner of the 2024 Theodore Eisenberg Poster Prize at the Conference on Empirical Legal Studies (CELS), Atlanta.
Best Paper in the area "Technology, Privacy, and Information" at the American Law & Economics Association Conference (ALEA) 2023, Boston.
In an online experiment, we study the interaction of regulation and discourse among peers in affecting public acceptance of institutionalizing AI usage. We show that regulation affects initial acceptance but discourse dominates acceptance in the long run. Initial support for the presented AI system is broad and rises further with public control of the AI system, transparency, and restriction of AI decision rights as regulatory features. However, initial support proves fragile: following free discussion in small chat groups, average support declines, and strong supporters are 2.6 times more likely than strong opponents to revise their position, irrespective of regulatory features. Being randomly matched with partners holding different initially stated attitudes causally affects subsequent attitude revision: supporters revise more negatively when matched with initial opponents, while the corresponding pattern for opponents runs in the opposite direction. Yet supporters revise their opinion more strongly across all group compositions and become more skeptical even when both chat partners also initially support the AI system. That is, discourse itself, not only group composition, seems to contribute to an erosion of acceptance. Text analysis reveals an important role of moral arguments: While efficiency arguments foster initial acceptance, moral arguments are more effective in changing attitudes during discourse and throughout different regulatory schemes.
(Previously circulated as: "Public Attitudes Toward Algorithmic Risk Assessments in Courts: A Deliberation Experiment")
Algorithmic Fairness and Human Discrimination (revision in progress)
Fairness constraints in algorithm design aim to reduce discrimination. Their impact, however, also depends on the adoption of the algorithm by human-decision makers as they typically retain full authority in high-stakes contexts. In a hiring experiment, I find that blinding an algorithmic performance prediction to candidates' group membership leads participants to update their beliefs about candidates more conservatively from these predictions. I then find a significant increase in discrimination in their hiring of candidates under this algorithm, driven by those who initially believe that group membership predicts performance. Finally, independent of the algorithm features, about 26% of participants make hiring decisions that cannot be explained by beliefs and are likely based on taste. These results suggest that, in human-in-the-loop settings, algorithmic fairness features can paradoxically exacerbate human discrimination based on statistical beliefs by hindering adoption and, unsurprisingly, remain orthogonal to taste-based discrimination.
Gender and Socioeconomic Gaps in Digital Skills: Actual and Perceived
This paper documents gender and socioeconomic gaps in digital skills relevant to the labor market, using a representative German household sample. Men and individuals with a higher level of education show greater proficiency across all skill dimensions. Both groups also hold more optimistic beliefs about outperforming others, conditional on actual proficiency. These belief gaps are not driven by overconfidence, but by underconfidence among women and individuals with lower educational backgrounds at the upper end of the skill distribution. Early-life socioeconomic background is not significantly associated with adult digital skills or beliefs.
CONFERENCE PAPERS
PhD THESIS
On Human Factors in Machine Fairness: Essays in Behavioral Economics
written at the Department of Economics, University of Hamburg