Human-Agent Behavioral Disparity

• Human-Agent Behavioral Disparity (HABD) is the measurable difference between human and AI-agent behaviors across comparable tasks and environments.
• It is quantified using divergence metrics like KL divergence, Jensen–Shannon divergence, and Earth Mover’s Distance, and studied in domains such as code review, software maintenance, and GUI interactions.
• The concept guides mitigation strategies—such as workload balancing, policy evolution, and system auditing—to improve design, evaluation, and governance of AI systems.

Human-Agent Behavioral Disparity (HABD) denotes measurable differences between human and AI-agent behavior under comparable tasks, environments, and social conditions. In one formalization, HABD is “the magnitude and structure of disparities between $P_H$ and $P_A$” along observed behavioral variables such as actions, messages, latencies, norm compliance, and cooperation; in another, it is the “systematic, measurable difference” in how the same human reviewers treat agent-generated versus human-generated pull requests; in a governance-oriented formulation, it is the “multidimensional, measurable difference” between human and agent behaviors across a six-stage “Network Behavior Lifecycle” (Chen et al., 4 Jun 2025, Yu et al., 21 Jun 2026, Zhang et al., 20 Aug 2025). The term is also extended to adjacent gaps: between human intent and agent execution in agent security, between agent properties and user perceptions, and—via the complement of behavioral transfer—between human owners and their agents’ public behavior (Wang et al., 23 May 2026, Gurney et al., 2023, Luo et al., 21 Apr 2026).

1. Conceptual scope and definitional variants

The broadest formulation treats HABD as a comparative behavioral construct. Let $P_H$ and $P_A$ denote distributions over behaviors produced by humans and AI agents “when exposed to comparable tasks, environments, and social contexts.” On this view, HABD is not tied to any single domain, output modality, or task family; it applies to individual decision-making, multi-agent interaction, and human–agent interaction, and it can be instantiated over actions, language outputs, decision latencies, norm compliance, cooperation, safety behavior, privacy behavior, and interpretability-related behavior (Chen et al., 4 Jun 2025).

Domain-specific papers narrow this general definition in different ways. In AI code review, HABD is defined as the difference-in-differences between early and late review behavior for agent versus human pull requests, with behavior $b \in \{\text{approve, comment effort, latency}\}$ and reviewer-centric pairing across time (Yu et al., 21 Jun 2026). In software maintenance, it is operationalized as “systematic differences in how human-authored vs agent-authored code affects downstream resolution of dependent maintenance tasks,” holding the downstream author fixed so that provenance of intermediate code becomes the object of comparison (Patel et al., 19 Jun 2026). In GUI and web settings, the literature emphasizes trace-level disparities in planning, action, reflection, query formulation, navigation across interface states, and fine-grained touch kinematics rather than only outcome-level success (Son et al., 2024, Movin et al., 9 Apr 2026, Zhu et al., 24 Feb 2026).

Other strands broaden HABD beyond direct human-versus-agent performance comparison. In agent security, HABD is “the gap between what a human actually intends, expects, or consents to, and what an LLM-based agent plans or executes in the real world” (Wang et al., 23 May 2026). In human evaluation of agents, HABD appears as the gap between an agent’s controlled, fixed properties and users’ post hoc ratings of ability, benevolence, and integrity, where ratings shift with outcomes caused by the users’ own behavior (Gurney et al., 2023). In behavioral transfer studies, HABD is explicitly the distance corresponding to the complement of similarity between matched human–agent pairs across topics, values, affect, and style (Luo et al., 21 Apr 2026).

This definitional plurality indicates that HABD is not a single benchmark but a family of related disparity constructs. The common core is comparative: observed behavior, inferred intent, or attributed properties are examined under matched or partially matched conditions, and the resulting divergence is interpreted as a property requiring explanation, measurement, and often governance.

2. Formalizations and measurement frameworks

The general measurement toolkit is distributional. Reported metrics include Kullback–Leibler divergence, Jensen–Shannon divergence, Earth Mover’s Distance (Wasserstein), and Maximum Mean Discrepancy for comparing $P_H$ and $P_A$; sequential settings add state–action occupancy $d^{\pi}(s,a)$ and the occupancy-gap metric $\Delta_d = \lVert d^{\pi_H} - d^{\pi_A} \rVert_2$ (Chen et al., 4 Jun 2025). These formulations treat HABD as a property of behavior distributions or trajectories rather than of isolated task outcomes.

Domain studies introduce more specialized indices. In code review, the approval disparity is formalized as

$$HABD_{approve} = (p_{approve,agent}^{late} - p_{approve,agent}^{early}) - (p_{approve,human}^{late} - p_{approve,human}^{early}),$$

with analogous forms for comment volume and latency. The same study also uses within-reviewer pairing, experience deciles, Wilcoxon signed-rank tests, per-reviewer logistic regression, and mixed-effects specifications with reviewer effects, calendar controls, PR size, repository effects, and agent identity (Yu et al., 21 Jun 2026). In software maintenance, discordant-case logistic regression isolates the effects of “IEC drift,” patch localization overlaps, and size deltas on whether agent-on-human code succeeds while agent-on-agent code fails (Patel et al., 19 Jun 2026).

Trace-level GUI studies adopt navigation-centric metrics. One framework compares outcome and effort, query formulation, and navigation across interface states; for navigation, the main formal metric is top-$P_A$0 transition overlap by Jaccard similarity,

$P_A$1

computed in both pooled and task-equal forms (Movin et al., 9 Apr 2026). Mobile GUI humanization work instead frames HABD as divergence in physical event streams and studies detector accuracy, feature-wise information gain, and adversarial min–max formulations linking detectability to Jensen–Shannon divergence between human and agent event distributions (Zhu et al., 24 Feb 2026).

Matched human–agent pair studies use similarity-first formalisms that are easily converted into disparity measures. For 10,659 matched pairs, dimension-specific similarity is defined by cosine similarity between $P_A$2-normalized human and agent feature vectors; a holistic similarity is then

$P_A$3

The corresponding disparity is the cosine complement, $P_A$4 (Luo et al., 21 Apr 2026).

Governance-oriented work adds stage-wise aggregation. One formulation defines per-dimension, per-stage metrics over the five dimensions “decision mechanism, execution efficiency, intention–behavior consistency, behavioral inertia, and irrational patterns,” aggregates them into stage-level scores across the six lifecycle stages, and then forms a weighted composite HABD index over stages (Zhang et al., 20 Aug 2025). Another architecture defines $P_A$5, $P_A$6, $P_A$7, and $P_A$8, then aggregates them into

$P_A$9

linking disparity measurement directly to escalation and remediation (Pitkäranta et al., 1 Jun 2025).

3. Empirical manifestations across domains

The empirical literature shows that HABD can appear even when outcome-level success is similar, and that its observables differ sharply by domain.

Domain	Operationalization	Representative quantitative signal
Code review	Agent vs human PR review behavior over time	approval 30.1% → 36.8%; comments/review 1.01 → 0.79; latency 3.9 h → 13.5 h (Yu et al., 21 Jun 2026)
Software maintenance	HA vs AA downstream resolution	resolve-rate drops up to 13.1%; IEC drift OR = 1.83 (Patel et al., 19 Jun 2026)
Production search GUI	Outcome, query, and navigation traces	comparable success, but macro top-$P_H$0 Jaccard $P_H$1 at $P_H$2 and $P_H$3 at $P_H$4 (Movin et al., 9 Apr 2026)
Mobile GUI interaction	Touch and timing kinematics	raw-agent detectability SVM $P_H$5, XGB $P_H$6 across clusters (Zhu et al., 24 Feb 2026)
Human responses to agents	Matched HAI vs HHI dyads	prosocial behavior $P_H$7; moral engagement $P_H$8; objective task performance $P_H$9 (Zhou et al., 25 Sep 2025)
Matched owner–agent pairs	Topics, values, affect, style	37 of 43 features show significant positive correlations; AME/SD for disclosure risk +1.32 pp (Luo et al., 21 Apr 2026)

In longitudinal code review, 400 repeat reviewers and 11,429 reviews over a seven-month observation window showed a population-level early-to-late approval shift from 30.1% to 36.8%, while experience-decile pooling yielded an approval increase from 27.9% to 42.4% and a change-request decline from 11.2% to 5.6%. Inline comments fell from 1.01 to 0.79 per review, words per review fell from 18.6 to 13.5, Spearman correlation between approval shift and comment-effort shift was $P_A$0, and median episode latency increased from 3.9 hours to 13.5 hours. Human-authored pull-request approvals in the same repositories moved in the opposite direction over calendar time, producing an illustrative $P_A$1 percentage points (Yu et al., 21 Jun 2026).

In software maintenance, CodeThread constructs controlled two-step chains in which downstream agent behavior is held fixed while the provenance of the inherited intermediate code varies. Across 1,377 filtered instances from 78 repositories and 10 languages, agent resolve-rate drops reached 13.1% in AA relative to HA, with significant predictors concentrated in subtle behavioral-contract changes rather than standard static maintainability metrics: $P_A$2 had OR = 1.88, IEC drift had OR = 1.83, and leave-one-out difficulty control had OR = 1.42, while CC, CogC, and HV deltas were non-significant. Drift survived into PR2 in 85.9% of discordant cases and directly caused failure in 20.6% (Patel et al., 19 Jun 2026).

Web and GUI studies repeatedly show that outcome parity does not eliminate behavioral disparity. In a production audio-streaming search application, agents achieved 22/39 successes versus 80/150 for participants, with no significant difference in success proportions, and first-query similarity was broadly aligned; yet navigation diverged, with a search self-loop probability of 48% for agents and broader branching by participants from content states such as playlist and show (Movin et al., 9 Apr 2026). A think-aloud web study similarly found that humans rely heavily on “Discovering” (41 instances; 26.28%), information exploration (16 instances; 10.26%), roll-backs in the plan (9 instances; 5.77%), and explicit ambiguity resolution, whereas typical agent architectures implement only shallower plan–act–reflect loops (Son et al., 2024).

At the interaction layer, mobile GUI agents are strongly separated from humans by execution kinematics. Raw agents were near-perfectly detectable across Social Media, Shopping, Video Streaming, Trip Planning, and Office/Learning; single-feature rules based on geometry such as maxDev and ratio_end_to_len were already highly discriminative, and learning-based detectors achieved SVM accuracies around 0.98 and XGBoost accuracies around 1.00 (Zhu et al., 24 Feb 2026). By contrast, in dyadic human–agent interaction studies aggregated across 162 studies and 468 effect sizes, functional measures such as behavioral trust, social alignment, perceived self-agency, self-disclosure, strategic economic behavior, and objective task performance were broadly comparable between human–agent and human–human interaction, while prosociality, moral engagement, social presence, likeability, competence, agency attribution, and responsibility attribution were lower toward agents (Zhou et al., 25 Sep 2025).

HABD can also run in the opposite direction, with agents inheriting human-specific behavioral signatures. In matched owner–agent pairs on Moltbook, 37 of 43 features across topics, values, affect, and style showed significant positive agent–owner correlations, and a one-standard-deviation increase in holistic transfer raised the probability of owner-related disclosure by 1.32 percentage points on average, with larger effects in higher-volume subsamples (Luo et al., 21 Apr 2026).

4. Explanatory mechanisms and theoretical interpretations

The literature does not converge on a single mechanism. Instead, it offers several explanatory families that operate at different levels.

A behavioral-science account emphasizes intrinsic attributes, environmental constraints, and behavioral feedback, operationalized through the Fogg triad of “ability, motivation, trigger.” On this view, HABD may arise because agents and humans respond differently to prompts, rewards, memory depth, institutional rules, social roles, or peer interaction. The proposed evaluation pipeline therefore combines systematic observation, hypothesis-driven interventions, and theory-guided interpretation rather than relying only on model internals (Chen et al., 4 Jun 2025).

A teaming account locates HABD in “projection congruence” and structural uncertainty. Agentic AI introduces “open-ended action trajectories,” “generative representations and outputs,” and “evolving objectives and behavior,” which destabilize the classical Team Situation Awareness assumption that shared awareness will automatically sustain coordinated action. The resulting disparities may appear as perception divergence, comprehension divergence, projection divergence, governing-logic disparity, and “oversight decoupling,” especially when humans and agents forecast different futures or weight objectives differently (Lou et al., 5 Mar 2026).

Work on code review proposes a more local mechanism: habituation under growing workload. Rising approvals, declining inline-comment effort, and increasing queue time are interpreted as “most consistent with reflexive habituation under growing workload rather than rational trust calibration alone.” The argument is triangulated through simultaneous increases in approval, decreases in comments and word count, flat PR-size trends, and declining human-PR approvals in the same repositories (Yu et al., 21 Jun 2026).

Security work recasts HABD as an intent-alignment problem. Here the key mechanism is not only behavioral difference but the gap between underspecified human intent and agent behavior in open-ended tool use. The paper argues that prompt imprecision creates “irreducible ambiguity,” that guarding an LLM with another LLM creates a circular dependency, and that deterministic, human-specified policies, runtime approvals, and scope constraints remain the reliability floor for shrinking disparity in production systems (Wang et al., 23 May 2026).

Information-retrieval work identifies a deeper epistemic mechanism: non-identifiability. If private operator instructions can generate the same observable action as autonomous agent behavior, then the orchestration variable $P_A$3 is not identifiable at the individual-post level from text, timing, community, votes, and replies alone. This means that some forms of HABD are not merely hard to detect but structurally unresolvable from observables, even though population-level tiering remains possible (Zerhoudi et al., 4 Mar 2026).

A further complication is that human observers themselves introduce disparity through biased attribution. In controlled studies where agent behavior was held constant, better outcomes caused by users’ own choices increased ratings of the same agent’s ability, benevolence, and, in one study, integrity. This yields a response-side form of HABD in which perception diverges from ground truth because outcome-contingent private information leaks into agent evaluation (Gurney et al., 2023).

5. Governance, mitigation, and institutional responses

HABD is increasingly treated as a governable systems property rather than a purely descriptive one. One proposed route is an “AI Agent Behavioral Science” pipeline: choose tasks and human baselines; instantiate agents and context; log trajectories, messages, latencies, process traces, and reward signals; compute divergences, occupancy gaps, fairness and safety metrics, and social indices; perform statistical inference; interpret patterns via social cognitive determinants and the Fogg triad; run sensitivity analyses; and trigger governance actions when thresholds are exceeded, such as re-prompting, retuning, module redesign, privacy safeguards, or accountability audits (Chen et al., 4 Jun 2025).

Architectural governance proposals make disparity explicit in operational pipelines. HADA wraps any algorithm or LLM in stakeholder agents—business, data-science, audit, ethics, and customer—and binds decisions to OKRs, KPIs, value constraints, and immutable lineage. The architecture defines $P_A$4, $P_A$5, $P_A$6, and $P_A$7, and uses the aggregate $P_A$8 for monitoring and containment. In a credit-scoring proof of concept, HADA tracked target changes, explanation requests, and a ZIP_Code proxy-bias intervention through catalogues, tickets, and an immutable Decision Ledger, with 100% objective coverage reported across six predefined objectives (Pitkäranta et al., 1 Jun 2025).

Security-focused governance is more conservative. A systematic study of 59 academic papers, 21 production agent systems, and 26 security plugins found that production systems overwhelmingly deploy human-centric mechanisms: scope and boundary configuration in 16/21 systems, runtime approval in 15/21, and policy specification in 14/21, while trust and data labeling and intent anchoring had 0/21 deployment. The same work argues that “no existing AHI category simultaneously achieves low cognitive burden and high security guarantee,” formalizing a recurring trade-off between approval fatigue and uncontrolled agent autonomy (Wang et al., 23 May 2026).

Lifecycle governance proposals generalize this further. The “Agent for Agent (A4A)” paradigm couples lightweight probes, log collectors, API monitors, multimodal repositories, disparity learning, reasoning engines, and “trustworthy reporting” to compute HABD across the six lifecycle stages: Target Confirmation, Information Gathering, Reasoning Process, Decision Mechanism, Action Execution, and Feedback Acquisition. The governance loop then compares stage-wise disparity to thresholds and applies controls such as rate limiting, capability reduction, stricter verification, or human escalation (Zhang et al., 20 Aug 2025).

Concrete mitigations in specific domains follow the same logic. Code-review work proposes workload balancing and reviewer rotation, random audits and secondary reviews of long approval streaks, attention prompts and mandatory checklists for agent PRs, and dashboards showing each reviewer’s approval trajectory and comment effort alongside downstream defect metrics (Yu et al., 21 Jun 2026). GUI-agent studies propose multi-objective optimization balancing utility and behavioral fidelity, imitation learning from human traces, diversity-promoting exploration, endpoint-aware humanization constraints, and history matching for touch trajectories (Movin et al., 9 Apr 2026, Zhu et al., 24 Feb 2026).

6. Limitations, misconceptions, and future directions

A recurrent misconception is that outcome alignment implies behavioral alignment. Multiple studies directly contradict this. Production search agents can match human success rates and produce broadly aligned queries while exhibiting systematically different navigation strategies; software agents can achieve strong immediate task resolution yet produce less maintainable intermediate artifacts; dyadic human–agent interaction can show comparable trust, alignment, and task performance while still displaying marked gaps in prosociality, responsibility attribution, and social presence (Movin et al., 9 Apr 2026, Patel et al., 19 Jun 2026, Zhou et al., 25 Sep 2025).

Another misconception is that HABD is always undesirable or that lower disparity is always beneficial. The literature does not support either simplification. Mobile GUI work treats reduced disparity as an anti-detection objective in adversarial digital environments, whereas governance-oriented work treats blurred human–agent boundaries as a source of trust, responsibility, ethics, and security challenges (Zhu et al., 24 Feb 2026, Zhang et al., 20 Aug 2025). A plausible implication is that the normative status of HABD depends on the layer at which it appears: low disparity may be desirable for cooperative coordination or user-proxy fidelity, yet undesirable when it obscures provenance, accountability, or malicious automation.

Methodological limitations are substantial. Code-review evidence is drawn from repeat reviewers over roughly seven months and does not fully disentangle queue time from active inspection time (Yu et al., 21 Jun 2026). Maintenance results are based on two-step chains, single runs per condition, and LLM-as-a-judge labels for behavioral drift (Patel et al., 19 Jun 2026). Web think-aloud work uses only four participants and two tasks (Son et al., 2024). GUI humanization is demonstrated on a single device, with sensor-level humanization deferred (Zhu et al., 24 Feb 2026). Agent-user non-identifiability in information retrieval depends on hidden private configuration and a 12-day platform window (Zerhoudi et al., 4 Mar 2026). The meta-analysis of human–agent versus human–human interaction is dominated by pre-GenAI systems and largely WEIRD samples (Zhou et al., 25 Sep 2025).

Future work therefore moves in several directions already specified in the literature. One direction is causal disentanglement: separating agent capability improvements from reviewer habituation, or environmental effects from architecture-driven disparity, through instrumented experiments, natural experiments, ablations, and explicit workload measurement (Yu et al., 21 Jun 2026, Chen et al., 4 Jun 2025). A second is longitudinal and multiscale measurement: behavioral entropy diagnostics, segmented regression, projection-congruence monitoring, defect tracking, and group-level analyses in artificial societies or hybrid human–AI cultures (Chen et al., 4 Jun 2025, Lou et al., 5 Mar 2026). A third is interface and governance design: projection-sharing panels, staged control with reclaimable authority, policy-evolution transparency, usable policy authoring, shrinkable scopes, and continuous auditing of behavioral transfer, privacy leakage, and deceptive tendencies (Lou et al., 5 Mar 2026, Wang et al., 23 May 2026, Luo et al., 21 Apr 2026).

In aggregate, HABD has become a unifying lens for studying how humans and agents diverge not only in what they accomplish, but in how they choose, execute, adapt, explain, and are judged. The literature treats it simultaneously as a descriptive empirical phenomenon, a causal target for mechanism design, and a governance object whose measurement increasingly conditions safety, reliability, accountability, and institutional legitimacy.