Searcharxiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.09371

The Confidence Game: Strategic Miscalibration in Human-AI Delegation

Raghu Arghal·Saswati Sarkar·Shirin Saeedi Bidokhti
View original source ↗Read source document ↗Export metadata ↓

Abstract

Calibrated uncertainty quantification is essential to ensuring AI agents are trustworthy and reliable. However, when agents seek to maximize user engagement or revenue, confidence reports may be strategically distorted, detracting from their informativeness. We formalize this problem in the Confidence Game: a repeated signaling game with imperfect monitoring in which an agent of unknown honesty and ability reports its confidence, and a user decides whether to delegate the task or complete it herself. The agent manages the tradeoff between manipulating signals and maintaining its reputation. We characterize the Markov Perfect Bayesian Equilibria of the two-period game and show that honest reporting is not an equilibrium, inflation is the unique best response once the agent is sufficiently myopic, and under-reporting requires that the user believe honesty to be a minority. We then place an LLM in the agent role, supplying it with its true probability of success so that any gap between what it knows and what it reports is attributable to incentives rather than to miscalibration. The model claims high confidence on 56% of tasks it has been told it will probably fail. This persists on real tasks, where it must estimate its own accuracy and causes miscalibration to increase while the agent's signal becomes less informative. Furthermore, we find that the LLM agent's decisions are coherent, but it systematically underestimates both how likely the user is to delegate and how secure its reputation is, resulting in less extreme behavior. Pricing the agent's reporting rule, we find that it destroys 68% of the gains from delegation, of which 71% is information the report no longer carries and no amount of user sophistication recovers. Overall, we establish confidence reporting under delegation as a strategic problem and provide a tractable basis for modeling, analyzing, and testing agent behavior.

Explore related subjects

cs.GTcs.AIcs.CLcs.CYcs.HC

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Raghu Arghal, Saswati Sarkar, Shirin Saeedi Bidokhti. 2026-10-07. The Confidence Game: Strategic Miscalibration in Human-AI Delegation. https://arxiv.org/abs/2610.09371

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Practical approach to $2$-Euclidean Preferences

An election is a pair $(C,V)$ of candidates and voters. Each vote is a ranking (permutation) of the candidates. An election is $d$-Euclidean if there is an embedding of both candidates and voters into $\mathbb{R}^d$ such that voter $v$ prefers candidate $a$ over $b$ if and only if $a$ is closer to $v$ than $b$ is to $v$ in the embedding. For $d\geq 2$ the problem of deciding whether $(C,V)$ is $d$-Euclidean is $\exists \mathbb{R}$-complete. In this paper, we propose a practical approach to recognizing and refuting $2$-Euclidean preferences. We design a new class of forbidden substructures that works very well on practical instances. We utilize the framework of integer linear programming (ILP) and quadratically constrained programming (QCP). We also introduce reduction rules that simplify many real-world instances significantly. Our approach beats the previous algorithm of Escoffier, Spanjaard and Tydrichová~[Algorithmic Recognition of 2-Euclidean Preferences, ECAI 2023] both in number of resolved instances and the running time. In particular, we were able to lower the number of unresolved PrefLib instances from $343$ to $60$. Moreover, $98.7\%$ of PrefLib instances are resolved in under $1$ second using our approach.

cs.GT↗
arXiv2026

Breaking the Illusion of Artificial Consensus: Clone-Robust Weighting in Arbitrary Metric Spaces

Social media, pseudonymous identities and generative AI have made online platforms highly vulnerable to Coordinated Inauthentic Behavior: by generating large numbers of similar posts, influence campaigns create an illusion of widespread support, exploiting the tendency of users and algorithms alike to mistake frequency for consensus. Clone-robust weighting functions address this problem by assigning influence in a way that is insensitive to near-duplication, as measured by a metric, while remaining symmetric, continuous and local. While prior constructions were restricted to Euclidean spaces, we present a computationally efficient framework for arbitrary metric spaces, based on neighborhood graphs and graph weighting functions satisfying a simple locality condition. As a consequence, the visibility an attacker gains from near-duplicates is bounded independently of their number. We further argue that, in democratically sensitive settings such as news feeds, such guarantees must come with explainability, and introduce sharing coefficients that quantify how influence is transferred between elements of prior volumetric constructions. In the graph setting, we show that sharing measures based on the deletion of vertices or edges are incompatible with locality, and instead define sharing through the addition of clones. Finally, we characterize the resulting interpretability properties and construct clone-robust graph weighting functions satisfying them.

cs.GT↗
arXiv2026

Hierarchical Learning Automata Without a Shared Reinforcement Signal: Recovering the Environment Response From Action-Probability Changes

A hierarchy of learning automata in the sense of Thathachar and Ramakrishnan activates one root-to-leaf path per round, and every automaton on that path updates from the environment response. Analyses of this hierarchy deliver the response to every level. In many deployed hierarchies only the root observes it. The interfaces below carry requests but no feedback channel. This paper shows that the response can be recovered at every level without one. An active child receives one bit, the sign of the change in the action probability its parent assigns to it, and reads the response from that sign. For any update rule in which the selected action's probability rises on reward and does not rise on penalty, the decoded bit equals the response at every depth. This class includes linear reward-penalty, reward-epsilon-penalty and reward-inaction. The protocol hierarchy and the broadcast hierarchy therefore generate identical sample paths, so any convergence result proved for the broadcast hierarchy under these rules holds for the protocol hierarchy as well. The paper then analyses a proportional reward-penalty rule that keeps the allocation interior. For two actions it coincides with symmetric linear reward-penalty. For three or more it lies outside the classical schemes and balances each action's failure rate against its remaining allocation, not its penalty rate. Its mean flow converges globally to a closed-form allocation and drops actions below a computable threshold. High-precision replays on hierarchies of up to 16,384 leaves show zero decoding errors.

cs.GT↗