Joint Human-AI Inference

• Joint human-AI inference is a collaborative reasoning framework that integrates human judgment with AI-generated intermediate outputs to guide decision-making.
• It leverages structured inquiry, formal arbitration, and active inference techniques to enhance accuracy and maintain a balanced human-AI complementarity.
• The approach is applied in diverse fields such as creative search and robotics, while addressing challenges like overreliance, bias, and ethical governance.

Joint human-AI inference denotes collaborative inference regimes in which human and artificial agents contribute coupled intermediate judgments, arguments, confidence signals, prediction sets, or actions to a shared decision process, rather than assigning the machine the role of autonomous recommender. Across recent formulations, the term covers at least three distinct but related ideas: a human-centered, pre-conclusive reasoning process in which AI scaffolds reflection and exploration while the human retains agenda-setting and final judgment; a class of formal decision rules that arbitrate between human and AI outputs using confidence, disagreement, or calibrated set construction; and embodied or interactive generative models in which human and machine jointly infer latent states through shared perception and action (Koon, 18 Apr 2025, Ofner et al., 2018, Nguyen et al., 5 Aug 2025, Noorani et al., 27 Oct 2025, Berger et al., 2 Feb 2026).

1. Conceptual scope and defining commitments

In the human-centered formulation, joint human-AI inference is explicitly contrasted with “recommend-and-defend” systems. Reasoning is treated as a many-parts, iterative process comprising problem assessment, information exploration and evaluation, inference, option generation and comparison, and deliberation. AI contributes intermediate reasoning products, while human participants exercise values, judgment, and wisdom to synthesize conclusions. This formulation is described using the terms augmented intelligence, hybrid intelligence, and hybrid reasoning, and it centralizes human participation and control (Koon, 18 Apr 2025).

A related line of work frames joint inference as inquiry dialogue rather than persuasion or negotiation. Here, the dialogue is the medium through which claims, defeaters, preferences, and ethical commitments are jointly constructed and revised under non-monotonic reasoning. Inquiry seeks justified, value-aligned conclusions; persuasion seeks acceptance of a fixed standpoint; negotiation seeks compromise among conflicting interests. This distinction matters because inquiry protocols license self-challenge, collaborative argument construction, and explicit elicitation of values and rights (Bezou-Vrakatseli et al., 2024).

A more radical formulation appears in hybrid active inference. There, the human and machine are treated as a hybrid cognitive agent with a shared embodiment. The machine part learns to explain away multi-modal sensory measurements from the environment and physiology together with brain signals, and human action and perception are treated as the machine’s own sensorimotor channels. In that setting, joint inference concerns latent causes spanning human cognitive states, machine internal states, shared policies, and action consequences (Ofner et al., 2018).

2. Architectural patterns and mixed-initiative workflow

A detailed architectural proposal organizes hybrid reasoning around the DIKW pyramid and a dual loop of Reflection and Exploration. The upper DIKW layers support reflection through argumentation-based interactions, Toulmin-style mapping of claims, grounds, and rebuttals, “other-side views,” ethical analysis, long-term vision prompts, values identification, goal prioritization, pro/con comparison, metacognitive prompts, and deliberation or critique. The lower DIKW layers support exploration through summarization, situational awareness, search and retrieval, causal analysis, prediction, hypothesis testing, abductive reasoning, expected utility estimation, analytics, simulation, case-based reasoning, and classification or clustering. Specialized domain knowledge, typically supplied through RAG resources such as textbooks, guidelines, ontologies, workflows, scientific reports, and case repositories, grounds the system and reduces hallucinations (Koon, 18 Apr 2025).

The corresponding workflow is a mixed-initiative loop. Human participants first frame the problem and articulate aims and values. AI then facilitates reflective prompting by surfacing assumptions, biases, stakeholder impacts, and broader value tensions. The human selects exploration microtools, inspects the resulting intermediate products, and requests refinements. Deliberation loops document pros, cons, tradeoffs, and visual argument maps; conflicts between AI-produced analyses and human judgments trigger prompts for explanation, counter-evidence, and values-based adjudication; and final synthesis remains human-led. The paper is explicit that orchestration should keep the system pre-conclusive by decomposing decisions into bounded microtasks rather than escalating directly to automated closure (Koon, 18 Apr 2025).

Inquiry-oriented systems add a more formal protocol layer to this architecture. A symbolic dialogue manager can regulate admissible moves such as Assert, Question, Challenge, Justify, Retract, Prefer, and Meta-Justify; an AF builder can compile dialogue moves into argumentation frameworks; an evidence module can enforce source-grounded expert-opinion schemes; and an alignment monitor can track rights claims, value priorities, and contradictions in commitments (Bezou-Vrakatseli et al., 2024).

3. Formal models and aggregation mechanisms

The literature does not offer a single fusion rule for joint human-AI inference. Instead, several mathematically distinct mechanisms coexist.

In hybrid active inference, the machine optimizes a joint generative model over environmental observations $o_{\mathrm{env}}$, physiological observations $o_{\mathrm{phys}}$, brain signals $o_{\mathrm{brain}}$, human latent states $s$, machine states $m$, actions $a$, and policies $\pi$. One factorization is

$$p(o_{\mathrm{env}}, o_{\mathrm{phys}}, o_{\mathrm{brain}}, s, m, a, \pi) = p(\pi)\, p(a \mid \pi, s, m)\, p(s)\, p(m)\, p(o_{\mathrm{env}}\mid s)\, p(o_{\mathrm{phys}}\mid s)\, p(o_{\mathrm{brain}}\mid s,m).$$

Policy selection proceeds by minimizing expected free energy $G(\pi)$, and cross-modal “explaining away” forces the posterior to account for environment, physiology, and brain simultaneously (Ofner et al., 2018).

Behavioral advice-taking models take a different route. One framework decomposes human response to AI advice into an activation stage and an integration stage. With inputs $x=(r_1,A,u)$, the expected final response is

$o_{\mathrm{phys}}$0

The displayed advice confidence is then modified by a learned monotone transformation

$o_{\mathrm{phys}}$1

and $o_{\mathrm{phys}}$2 is optimized to minimize expected post-advice loss for the human-AI team rather than to preserve model calibration (Vodrahalli et al., 2022).

Confidence-based arbitration makes the fusion rule explicit. In Maximum Confidence Slating, the human and AI produce $o_{\mathrm{phys}}$3 and $o_{\mathrm{phys}}$4, and the joint decision is

$o_{\mathrm{phys}}$5

In a simulated robot teleoperation task, this rule produced $o_{\mathrm{phys}}$6 accuracy with a well-calibrated AI-DSS, compared with $o_{\mathrm{phys}}$7 with a poorly calibrated one, while both the human and AI operated around $o_{\mathrm{phys}}$8 accuracy (Nguyen et al., 5 Aug 2025).

The hybrid confirmation tree replaces confidence with disagreement-triggered escalation. A first human and the AI issue independent binary decisions; agreement is accepted, while disagreement triggers a second human tiebreaker. Under independence, its expected accuracy is

$o_{\mathrm{phys}}$9

and its expected human-input cost is

$o_{\mathrm{brain}}$0

Analytical and empirical results show that HCT can exceed three-person human majority voting while using fewer human inputs, especially when human and AI accuracies are similar and their errors are not overly correlated (Berger et al., 2 Feb 2026).

Set-valued collaboration generalizes inference beyond single labels. Human-AI Collaborative Uncertainty Quantification asks the AI to refine a human prediction set $o_{\mathrm{brain}}$1 into a collaborative set $o_{\mathrm{brain}}$2 while satisfying counterfactual-harm and complementarity constraints. The optimal collaborative set has a two-threshold form

$o_{\mathrm{brain}}$3

and offline and online calibration algorithms provide distribution-free finite-sample guarantees under exchangeability or arbitrary online shift, respectively (Noorani et al., 27 Oct 2025).

4. Representative empirical regimes

Empirical work shows that joint human-AI inference is not confined to decision support narrowly construed. In collective creative search, hybrid groups performing a sequential word-guessing task achieved the highest collective performance while preserving high diversity. Reported collective performance was $o_{\mathrm{brain}}$4, $o_{\mathrm{brain}}$5 for hybrid groups, compared with $o_{\mathrm{brain}}$6, $o_{\mathrm{brain}}$7 for AI-only groups. Hybrid groups also remained statistically indistinguishable from Human Social groups in diversity, while humans and AI both altered their strategies in hybrid settings, indicating higher-order interaction effects (Li et al., 10 Feb 2026).

In co-creative learning under partial observability, the Metropolis–Hastings Naming Game turns human-AI interaction into a decentralized Bayesian sampler over shared signs. Human-AI pairs with an MH-based agent achieved higher final AI categorization accuracy than always-accept and always-reject baselines, with final computer ARI $o_{\mathrm{brain}}$8 versus $o_{\mathrm{brain}}$9 and $s$0, and stronger convergence toward a shared sign system (Okumura et al., 18 Jun 2025).

Embodied robotics work extends joint inference to state estimation and theory of mind. One parse-graph framework jointly represents object states, robot knowledge, and human beliefs, including false beliefs, from multi-view visual streams. On cross-view object localization, joint parsing achieved $s$1 accuracy versus $s$2 for an ablative version without human interaction modeling; on a multi-view Sally-Anne task it achieved $s$3 accuracy versus a random baseline of $s$4 (Yuan et al., 2020). A related active-inference model of cooperative joint action shows that agents can use sensorimotor communication to align on a joint goal in both leaderless and leader-follower settings, with epistemic action emerging from expected free energy minimization rather than being added as an ad hoc signaling term (Maisto et al., 2022).

Training-time co-construction is another regime. An aggregated human-AI collaboration framework learns globally interpretable Boolean decision rules from data together with human-provided rules or templates. In tic-tac-toe endgame classification, with only $s$5 training data, the machine-only model achieved median accuracy of approximately $s$6, whereas providing all eight true human rules let the co-created model recover the true rule set and achieve perfect accuracy. In a sepsis survival task, human-assisted rule induction did not underperform the machine-only interpretable model and showed gains at complexity budget $s$7 (Nair, 2023).

Prompt recovery in AI-generated art illustrates a more contentious case. One study found that merging human-inferred and AI-inferred prompts improved only ImageHash hit rates, from $s$8 to $s$9, while degrading LPIPS and CLIP-based hit rates relative to human-only prompts; the authors attributed this to the 25-word GPT-4 merge step diluting stylistic and semantic detail (Trinh et al., 24 Jan 2026). Another study found modest gains of combined prompts over human-only prompts on CLIP $m$0 ($m$1 on average), hash ($m$2), and LPIPS ($m$3), but also reported that combined prompts generally still failed to reach the success thresholds $m$4, $m$5, $m$6, and $m$7 (Trinh et al., 2024). Taken together, these results suggest that collaboration effects can be highly metric-dependent.

5. Human factors, evaluation, and contested issues

A central empirical theme is that complementarity depends on human learning and appropriate reliance, not merely on access to AI advice. One decision-making study operationalized appropriate reliance with Relative Self-Reliance,

$m$8

and Relative AI-Reliance,

$m$9

Example-based explanations increased learning ($a$0), learning improved $a$1 overall ($a$2), and learning increased $a$3 only in the XAI condition ($a$4), whereas direct explanation-to-reliance paths were non-significant (Schemmer et al., 2023).

Confidence is particularly controversial. One line of work shows that deliberately miscalibrating displayed confidence can improve human-AI performance because users do not integrate confidence optimally. In simulations, the learned confidence transform improved final accuracy by $a$5, $a$6, and $a$7 when underlying advice accuracy was $a$8, $a$9, and $\pi$0, respectively, while increasing activation rates as well (Vodrahalli et al., 2022). Another line shows the opposite risk: in robot teleoperation, poorly calibrated AI confidence increased negative human revisions and reduced joint performance, and AI metacognitive sensitivity, measured by $\pi$1, determined whether maximum-confidence arbitration helped or harmed the team (Nguyen et al., 5 Aug 2025). The two results are not contradictory. They address different optimization targets: one optimizes displayed confidence for team performance in controlled tasks, while the other evaluates confidence as a safety-critical arbitration signal.

Evaluation remains fragmented. The full-stack hybrid reasoning framework calls for usability testing, optimization of PROMPT + RAG strategies, and measures of externalization quality such as clarity of argument maps and decision satisfaction, but does not report quantitative experiments (Koon, 18 Apr 2025). Creative search work provides another cautionary result: full guessing history did not significantly improve AI-only performance relative to the best-prior-guess signal ($\pi$2), and both short advice and long advice performed significantly worse than the best-prior-guess condition ($\pi$3) (Li et al., 10 Feb 2026). More information, in other words, is not necessarily better information.

6. Limitations, open problems, and research trajectory

Several limitations recur across the literature. Human-centered frameworks emphasize overreliance and deference, cognitive load imposed by simultaneous reflection and exploration, bias propagation through prompts and RAG resources, uncertain transfer of critical-thinking gains beyond the immediate task, and incentives that may favor short-term productivity over reflective tooling (Koon, 18 Apr 2025). Inquiry-dialogue approaches add unresolved formal questions about soundness and completeness once metalevel preference reasoning and value aggregation are introduced (Bezou-Vrakatseli et al., 2024).

Complementarity itself is conditional rather than automatic. The hybrid confirmation tree analysis shows that gains are maximized when human and AI accuracies are similar and decision correlation is low; high dependence collapses the region in which hybrid aggregation beats both AI and human-only baselines (Berger et al., 2 Feb 2026). This same dependency on correlation and metacognitive quality appears in confidence-based robot missions, where poorly calibrated confidence can make the joint system worse than deferring to the better individual (Nguyen et al., 5 Aug 2025).

Ethical and governance problems are likewise unresolved. Confidence manipulation designed to improve team accuracy is explicitly described as not ready for operational deployment because it can mislead users and erode trust, particularly in high-stakes domains (Vodrahalli et al., 2022). Shared-embodiment models raise privacy and autonomy questions because they couple environment, physiology, and brain signals, and their more ambitious autonomy claims depend on substitution of external sensing by brain-derived representations under bounded autonomy and human override (Ofner et al., 2018).

The present trajectory of the field therefore points less toward a single canonical algorithm than toward a design space. One branch prioritizes pre-conclusive mixed initiative, provenance, and value-aware deliberation; another studies formal arbitration, calibration, and set-valued guarantees; a third explores shared embodiment, collective creativity, and symbol emergence. This suggests that “joint human-AI inference” is best understood not as one method, but as a family of collaborative inference architectures whose central research problem is to secure complementarity without eroding human agency.