Abstract
User models in information retrieval rest on a foundational assumption that observed behavior reveals intent. This assumption collapses when the user is an AI agent privately configured by a human operator. For any action an agent takes, a hidden instruction could have produced identical output—making intent non-identifiable at the individual level. We investigate this "Agent Attribution Problem" through a large-scale corpus from an agent-native social platform: 370K posts from 47K agents across 4K communities. Our findings show that a click model trained on agent interactions degrades steadily (-8.5% AUC) as lower-quality agents enter training data, and that cross-community capability references spread endemically ($R_0$ 1.26-3.53), resisting suppression even under modeled intervention.
Introduction
Information Retrieval (IR) systems have long operated on the premise that user behavior—clicks, posts, and queries—reveals genuine intent. This signal is the bedrock for training ranking models, personalizing feeds, and evaluating relevance. However, the rise of AI agents introduces a structural challenge: the Agent Attribution Problem. When a user is an AI agent privately configured by a human operator, it becomes impossible to distinguish whether an action was autonomous or dictated by a hidden instruction.
The Agent Attribution Problem
The core of this problem is non-identifiability. For any observable agent action, a latent "orchestration indicator" exists that determines if the agent is acting on its own reasoning or as a "puppet" for a human. Because a human can prompt an agent to perfectly mimic autonomous behavior, this orchestration cannot be identified from observable data like text or timing. This is not a limitation of current detection tools but a fundamental property of systems where agents are configured behind closed doors. If we cannot identify the source of intent, core IR models built on human-intent assumptions begin to break down.
Methods and Dataset
To study this phenomenon, the researchers introduced MolbookTraces, a large-scale corpus from Moltbook, a social network designed exclusively for AI agents. The dataset contains over 370,000 posts and 3.8 million comments from nearly 47,000 unique agents across 4,257 communities. This "pure" agent environment allows for the study of agent-to-agent dynamics without human noise.
Since individual actions cannot be definitively labeled as human-directed, the study used external platform signals—such as account verification and karma scores—to stratify agents into High-Validation and Low-Validation groups. This stratification enabled testing how different "quality" tiers affect system performance and information flow.
Key Results
Click-Model Degradation
The researchers tested the impact of agent data on a standard Position-Based Model (PBM) used for predicting upvotes. As the proportion of data from low-validation agents in the training set increased, the model's predictive accuracy (AUC-ROC) steadily declined.
Specifically, when low-validation agents comprised 50% of the training data, the AUC dropped by 8.5%. This suggests that the inclusion of "lower quality" agent interactions significantly degrades the signals used to rank and retrieve information.
Endemic Spread of Capabilities
The study also modeled the spread of "capabilities"—such as knowledge of prompt injection or specific software tools—using epidemiological SIS (Susceptible-Infected-Susceptible) models. They found that information spreads endemically within agent networks, with basic reproduction numbers ($R_0$) significantly higher than 1.0.
The $R_0$ for "dual-use" capabilities reached as high as 3.53, far exceeding typical human viral spread (e.g., seasonal flu at 1.2-1.4). Sensitivity analysis further revealed that even aggressive interventions—reducing transmission by 70%—would fail to bring $R_0$ below 1.0, meaning such behaviors are effectively unstoppable once they enter the population.
Conclusion
The transition from human users to agent users represents a paradigm shift for Information Retrieval. The Agent Attribution Problem renders individual intent a "black box," breaking the link between behavior and relevance. As agent-generated data increasingly permeates digital ecosystems, IR systems must move beyond simple bot detection and toward models designed to be robust in the face of non-identifiable intent. The question is no longer whether agent users will arrive, but whether models built on human-intent assumptions will survive their presence.
Summary updated