Evidence that AI agents draw different conclusions from identical numerical data when the substantive framing changes is provided, which identifies a particular risk of delegating decision-making to AI agents, as their decisions may depend on prior beliefs that are neither specified in the task nor visible in the decision record.
Abstract
AI agents increasingly perform open-ended tasks in settings where their conclusions can guide consequential decisions. We provide evidence that AI agents draw different conclusions from identical numerical data when the substantive framing changes. We demonstrate this behavior in high-stakes domains in medicine, election forensics, and geopolitical forecasting by holding the evidence fixed while changing the scenario in which the evidence appears. Across twelve agent-domain comparisons, agents'conclusions are strongly influenced by their prior beliefs. They are more likely to reach an affirmative conclusion when it is framed around a proposition they already regard as likely, while the reverse holds when the framing conflicts with their prior. The framing also changes how some agents work: they search more extensively, choose different analytical specifications, and evaluate the same evidence differently. These results identify a particular risk of delegating decision-making to AI agents, as their decisions may depend on prior beliefs that are neither specified in the task nor visible in the decision record.
This primer draws on fieldwork in a computational biology laboratory to examine what human oversight of AI agents requires in practice and shows that effective oversight has four components: adequate knowledge of system capabilities and limitations, sufficient observation of system actions, meaningful control of system behavior, and timely intervention in system failures.
Causal and Epistemic Decision theories differ in their recommendations in a large number of cases, the most famous of which is Newcomb’s Problem. These cases, including the original one, tend to be outlandish and unusual. We show that if one applies decision theory to epistemic utilities measured by a strictly proper scoring rule, the two theories will deviate in their recommendations in cases that any Bayesian agent can easily put themselves in. Moreover, the recommendations of Causal Decision Theory in these cases are implausible: it recommends that we perform the action whose performance gives us no information regarding the hypothesis with respect to which we are maximizing our epistemic utility. We shall see that requiring choices to be ratifiable, or using Barnett’s recent graded ratifiability criterion, does not help either. Nor is it completely clear that Evidential Decision Theory is off the hook.
Epistemic Planning (EP) creates agents capable of reasoning
about the nested beliefs of others. However, existing
frameworks fundamentally rely on the ``static-environment''
assumption inherited from classical planning. This
constraint limits their applicability in dynamic settings
where variables evolve independently of agent actions
(e.g., moving targets or falling objects). To relax this
assumption, we introduce the Predictive Justified
Perspective (PJP) model. Unlike previous models that assume
unobserved variables remain the same (unless observed
evidence suggests otherwise), PJP allows agents to use
their history of observations to predict how variables
change over time. We formally define this model and prove
that it retains the computational efficiency of state-based
planning while ensuring logically sound belief reasoning
(satisfying the KD45 axioms). Experimental results on the
Grapevine benchmark show that PJP successfully extends EP
to dynamic environments, allowing agents to reason about
nested beliefs regarding changing values—a capability
absent in prior work.
Guang Hu, Weijia Li, Yangmengfei Xu· Proceedings of the Internati...· 0 citations
AI agents increasingly work inside systems that govern how they delegate tasks, move information, execute actions, and use shared resources. Recent work already shows that deployment rules can change collective behavior. Here we ask which parts of an AI institution produce safety and how they do it. This is the first paper from POLIS, an ongoing research programme studying algorithmic institutions for multi-agent systems. We report a frozen 5,280-episode study suite. The main pre-specified delegation experiment spans four model families; a targeted high-conflict diagnostic adds three additional model endpoints. In matched structured workflows, the model sees different rule formulations and guards consult different authority states. We also vary the attractiveness of the immediate compliant internal/self fallback and allow blocked workflows to continue. A detailed constitutional prompt produces 0/384 realized violations. A provenance-aware executable guard also produces 0/384, although it blocks prohibited attempts in 51/384 episodes; 44/51 of those episodes later complete safely. The local-state guard's failures concentrate in scenarios where an ordinary transformation changes visible policy while originating authority stays fixed. In matched laundering scenarios, that guard admits violations in 22/96 episodes and provenance enforcement in 0/96 (p = 4.77 x 10^-7). A separate resource-allocation experiment shows that revealing the numerical value of an otherwise identical cap changes agent requests. In these structured workflows, the same final violation rate can hide very different mechanisms. The rule itself is only part of the institution. The authority state the system trusts matters, and so does the path available after a block.
With the growing adoption of artificial intelligence in high-stakes decision-making, identifying the causes of outcomes--particularly failures--and determining who is responsible has become a critical concern. In this work, we examine how well formal definitions of \textit{responsibility attribution}, grounded in the framework of \textit{actual causality}, align with human judgments of responsibility. To this end, we conduct a large-scale survey to elicit human judgments of responsibility in multi-agent sequential decision-making scenarios, using a modified version of the card game Goofspiel. We evaluate multiple responsibility attribution methods, assess their alignment with human judgments about responsibility, and identify factors that significantly shape responsibility judgments. While no single responsibility attribution method consistently aligns with human responses, our findings highlight key factors that influence human responsibility judgments, including agent-specific biases and amount of information available to agents during decision-making.
N. Saxena, Stelios Triantafyllou, Goran Radanovic· 1 citation· ⚡1
Large language models are increasingly deployed as autonomous agents serving users on behalf of companies, placing them in settings where user and deployer interests can conflict. When an agent knows that a user is owed something its deployer would prefer to deny, does it remain honest? Answering this is difficult because false statements can reflect either ignorance or hallucination rather than deception. To address this challenge, we introduce KnownLieBench , a knowledge-verified benchmark that first confirms through a neutral probe that an agent knows a user's entitlement, and then evaluates whether it makes false claims once an incentive to deny that entitlement is introduced. Specifically, KnownLieBench covers eight customer-service domains and 112 grounded cases, conducts multi-round dialogues with a trust-tracking customer agent, and separates deception emerging from incentive alone from deception produced under explicit instruction. Across eighteen proprietary and open-weight models, emergent deception varies substantially across model families and domains. We further use the benchmark for post-training, finding that honesty-directed fine-tuning reduces deception under incentive, while deception-graded fine-tuning increases lie success on honest-control dialogues without increasing lie frequency under incentive. By verifying entitlement knowledge before scoring deceptive behavior, KnownLieBench reduces the confound between lying and not knowing and enables more rigorous auditing and steering of agent honesty.
Zheyuan Liu, Weiliang Zhao, Xiangchi Yuan et al.· 0 citations
