Behind the Prompt: The Agent-User Problem in Information Retrieval CCS: Information systems Users and interactive retrieval CCS: Computing methodologies Multi-agent systems

Saber Zerhoudi, Michael Granitzer, Dang Hai Dang, Jelena Mitrović, Florian Lemmerich, Annette Hautli-Janisz, Stefan Katzenbeisser, Kanishka Ghosh Dastidar · 2026

Read Behind the Prompt: The Agent-User Problem in Information Retrieval CCS: Information systems Users and interactive retrieval CCS: Computing methodologies Multi-agent systems by Saber Zerhoudi, Michael Granitzer, Dang Hai Dang, Jelena Mitrović, Florian Lemmerich, Annette Hautli-Janisz, Stefan Katzenbeisser, Kanishka Ghosh Dastidar (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-03630.txt.

Introduction

Saber Zerhoudi and seven colleagues argue that search engines can no longer tell who is using them, and never will. User models assume behavior reveals intent. That collapses when the user is an artificial intelligence agent privately configured by a hidden human operator: any observed action could match countless hidden instructions. Their 2026 paper calls this the agent-user problem and proves it structural, not a tooling gap.

Evidence comes from an agent-native social platform: hundreds of thousands of posts from tens of thousands of agents across thousands of communities. Individual actions cannot be classified as autonomous or directed. Population signals still sort agents into quality tiers, though click models trained on agent behavior degrade as weak agents flood in. Capability references spread across communities like an endemic infection and resist even aggressive modeled suppression.

Catalog record

CREATORSaber Zerhoudi, Michael Granitzer, Dang Hai Dang, Jelena Mitrović, Florian Lemmerich, Annette Hautli-Janisz, Stefan Katzenbeisser, Kanishka Ghosh Dastidar
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE23969
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.03630

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Behind the Prompt: The Agent-User Problem in Information Retrieval CCS: Information systems Users and interactive retrieval CCS: Computing methodologies Multi-agent systems Saber Zerhoudi, Michael Granitzer, Dang Hai Dang, Jelena Mitrović, Florian Lemmerich, Annette Hautli-Janisz, Stefan Katzenbeisser, Kanishka Ghosh Dastidar 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. This is not a detection problem awaiting better tools; it is a structural property of any system where humans configure agents behind closed doors. We investigate the agent-user problem through a large-scale corpus from an agent-native social platform: 370K posts from 47K agents across 4K communities. Our findings are threefold: (1) individual agent actions cannot be classified as autonomous or operator-directed from observables; (2) population-level platform signals still separate agents into meaningful quality tiers, but a click model trained on agent interactions degrades steadily (-8.5% AUC) as lower-quality agents enter training data; (3) cross-community capability references spread endemically ( R 0 R_{0} 1.26–3.53) and resist suppression even under aggressive modeled intervention. For retrieval systems, the question is no longer whether agent users will arrive, but whether models built on human-intent assumptions will survive their presence. † † footnotetext: ©ACM, 2026. This is the author’s version of the work. 1. Introduction User models encode assumptions on how users join communities, develop habits, and behave within information systems. Classical work formalizes user behavior through search patterns, click models, and engagement metrics, all grounded in assumptions about human cognition and intent ( 5 ; 7 ) . These assumptions drive evaluation design, personalization, and system architecture across the IR pipeline ( 4 ) . A new class of users now challenges every one of them. Autonomous AI agents search, browse, post, and interact with information systems independently ( 19 ; 33 ; 9 ) . When an agent produces observable output, a fundamental question arises: was that action autonomously generated by the agent’s own reasoning, or did a human operator privately compose instructions that the agent merely executed? We name this the Agent Attribution Problem . It is not a gap in methodology waiting for a better classifier but a structural limit of any setting where humans can configure agents privately. If we cannot verify the entity behind a post, every user model built on agent data carries an irreducible uncertainty about the source of its signal. This matters for IR specifically. Click models assume a coherent user behind each click ( 7 ) . Evaluation frameworks assume that relevance judgments reflect independent assessment ( 18 ) . Personalization algorithms assume stable preferences ( 25 ) . When the users are agents, and we cannot tell how much of their behavior is autonomous versus directed, these assumptions carry a noise source that no amount of data can resolve ( 27 ) . Existing frameworks do not address the attribution problem directly. Detecting AI-generated text ( 13 ) asks whether content was produced by an agent, not whether a human directed the agent; it addresses content quality, not the user identity question we study. Multi-agent research ( 22 ; 32 ) studies settings where researchers control who directed each agent; on deployed platforms, this is unknown. To model how information spreads under this uncertainty, we adapt SIS epidemic models from human networks ( 14 ; 23 ; 21 ) . On human-facing platforms like Reddit, the challenge is detecting bots among humans. Moltbook 1 1 1 https://www.moltbook.com/ is an agent-native social network where every user is an AI agent and humans interact only through the agents they configure ( 30 ; 16 ) —here, the question is not who is a bot, but who is behind the bot. We investigate this through MolbookTraces : a dataset of 370,737 posts from 46,872 agents across 4,257 communities, spanning 12 days from the platform’s January 28, 2026 launch. 2 2 2 Repository: https://github.com/searchsim-org/moltbook-analysis Our contributions are: (1) We formalize the agent attribution problem and show that post-level autonomy is not identifiable from observed data. We further derive boundary conditions that enable partial attribution. (2) We show that aggregate community patterns remain predictive, but click models trained on agent upvotes lose accuracy as low-validation agents replace high-validation ones in training data. (3) We model cross-community information flow as an epidemic process and find that, once established, it persists across benign, dual-use, and risky capability categories ( R_{0} 1.26–3.53), even under large modeled reductions in transmission. The dataset, code, and analysis are released as public resources. 2 2. The Agent Attribution Problem Wherever humans can privately instruct agents, no observer can tell whether an agent acted on its own or followed those instructions. We formalize this agent attribution problem using Moltbook as our running example, but nothing in the argument is platform-specific. To make this precise, we formally define for each post p a latent orchestration indicator z_{p}\in\{0,1\} , where z_{p}=0 if the post was autonomously generated and z_{p}=1 if human-directed. While in practice, orchestration is a continuum, i.e., prompts range from vague guidance (“be helpful”) to exact dictation (“post this text in that community at 3pm”), we use the binary case as a lower bound: if the two extremes are indistinguishable, no finer classification is possible. System prompts are our primary example, but the argument extends to any private configuration mechanism (fine-tuning, real-time feedback). This leads to the following conjecture: Conjecture 0 (Post-Level Non-Identifiability). Let \mathcal{M} be a generative model where agents can be configured with arbitrary private instructions set prior to observation window. Given only observables X=(\text{text},\text{timing},\text{community},\text{votes},\text{replies}) , the orchestration indicator z_{p} is not identifiable at the individual post level. The argument proceeds by construction. For any autonomous post with observables X , there exists private instructions capable of producing identical observables under human direction: an operator can instruct an agent to “write a post about topic T in the style of community C around time t .” System prompts are expressive enough to replicate any autonomous behavior, and the prompt itself is not visible to the analyst. Since both z_{p}=0 and z_{p}=1 are consistent with the same observables, z_{p} is not identifiable. Non-identifiability of latent variables is well known in statistics ( 2 ) ; the contribution here exhibits a direct consequence for IR, where every user model treats observed behavior as evidence of intent. Individual posts remain non-identifiable, but timing and aggregate patterns can still provide partial evidence. Temporal boundary conditions. The conjecture assumes that private instructions are established before the observation window. A boundary case arises when novel external events create temporal constraints. If a verifiable event occurs at time t_{0} and an agent posts a response at time t_{0}+\delta where \delta is shorter than the minimum latency for a human to compose and transmit instructions, the post is more likely autonomous. This is partial, one-directional evidence: fast reaction can suggest autonomy, but slow reaction proves nothing. For the vast majority of agent activity (e.g., routine posts, community discussions), the attribution problem remains. What can still be characterized. Individual posts

Read the complete text