Analysis of "Agentic Flooding" in Public Administration

Authors: Chris Schmitz, Lewis Hammond, and Alan Chan
Paper ID: arXiv:2608.16603
Venue: To be presented at the 9th AAAI Conference on AI, Ethics, and Society (AIES), October 12-14, 2026.

---

Executive Summary

The integration of AI agents into the public sphere is a double-edged sword. While these tools empower citizens to navigate bureaucratic hurdles, they simultaneously threaten to overwhelm the systems they interact with.

Core Thesis: The authors introduce the concept of "agentic flooding"—a phenomenon where Large Language Models (LLMs) drastically lower the cost of generating complex administrative requests, leading to surges in demand that can paralyze government infrastructure.

The Mechanism of Flooding

The process can be visualized as a pipeline where AI removes the "cognitive friction" previously required to interact with the state:

---

Key Research Contributions

The paper provides three primary insights into the current state of public service vulnerability:

1. Empirical Evidence: The researchers analyzed a dataset consisting of 84 potential instances of flooding across 11 different jurisdictions. This suggests that flooding is not a theoretical risk but a current reality driven by the cheap production of high-quality text.
2. Vulnerability Assessment: The authors developed a framework to determine which services are most likely to be targeted.
3. Policy Recommendations: They explore the tension between stopping flooding and maintaining equitable access.

The Risk Matrix

The likelihood of a service experiencing agentic flooding is determined by two primary variables: Financial Incentive and Procedural Complexity.

The highest risk is found in services that are financially attractive but difficult to navigate without assistance, as LLMs provide the exact "bridge" needed to exploit these systems.

---

The Mathematical Perspective of Strain

The authors imply that the strain on a service ($S$) is a function of the volume of requests ($V$) and the processing capacity ($C$), where $V$ is now amplified by the efficiency of AI ($\alpha$):

$S = \frac{V \cdot \alpha}{C}$

Where:

• $\alpha > 1$ represents the multiplier effect of LLM-assisted filing.
• If $S > 1$, the service enters a state of flooding.

---

Mitigation and the Equity Trade-off

Governments often react to surges by introducing "friction." However, this creates a systemic conflict:

• Friction-inducing measures: (e.g., introducing application fees) $\rightarrow$ $\text{Effective at stopping bots}$ $\approx$ $\text{Harmful to low-income humans}$.

Easy Access for All $\leftrightarrow$ System Stability

To avoid this trade-off, the authors suggest a set of strategic actions:

• Develop non-financial friction (e.g., identity verification).
• Implement AI-detection layers for incoming appeals.
• Scale backend processing using the same AI technology causing the flood.

Implementation Example

If a government were to track these requests in a database, the "flooding" signature might look like this in a JSON log:

{
  "request_id": "gov_99283",
  "timestamp": "2026-08-19T10:00:01Z",
  "content_analysis": {
    "perplexity": "low",
    "burstiness": "low",
    "llm_probability": 0.98
  },
  "status": "flagged_for_review"
}

---

Metadata & Licensing

• Submission Date: August 17, 2026 (v1)
• Last Revised: August 19, 2026 (v2)
• Subject Category: cs.CY (Computers and Society)
• DOI: https://doi.org/10.48550/arXiv.2608.16603