RESEARCH DOSSIER
Human–Artificial Community Coordination and Governance
Abstract. Human–Artificial community coordination and governance examines the psychological and institutional organization of collective decisions when artificial systems participate in mediation, collaborative knowledge production, administrative workflows, and continuing community memory. This critical narrative research synthesis integrates the history of participatory governance, community empowerment, collective cognition, and common-pool institutions with original experimental evidence on AI-mediated deliberation, algorithmic facilitation, participation willingness, and human–AI team performance. It distinguishes operational coordination from governance authority and studies how artificial mediation affects trust, perceived legitimacy, decision rights, belonging, participation inequality, agenda control, and appeal. Evidence from controlled and field experiments supports bounded effects on common-ground statements, facilitator preference, steering of material allocations, and willingness to join AI-led deliberation, while revealing important differences between subjective inclusion and observed influence. The monograph develops operational criteria for evaluating community memory, delegated authority, moderation, representational fairness, privacy, repair, and longitudinal institutional continuity. It identifies construct validity, selection effects, network interference, model instability, and source provenance as methodological priorities. The Artificial Era framework guides new testable hypotheses about continuing artificial contributors within accountable communities. The work is a critical narrative scientific synthesis; no original experiment, systematic review protocol, new meta-analysis, or independent clinical validation is claimed.
Keywords: human–artificial community; community psychology; collective coordination; community governance; AI-mediated deliberation; participatory governance; artificial participants; algorithmic mediation; institutional legitimacy; psychological empowerment; collective efficacy; human–AI collaboration; distributed cognition; collective intelligence; community memory; procedural justice; representational equality; participatory design; contestability; algorithmic accountability; Artificial Era
Introduction: The Community as a Human–Artificial Coordination System
Human–Artificial community coordination and governance concerns the organization of collective activities, decision procedures, shared records, obligations, and contestation when artificial systems participate in the processes through which a community governs itself. The relevant unit is a community situated within an institution or platform: residents planning a neighborhood intervention, an online association moderating discussion, a civic assembly considering a common proposal, or a professional coalition maintaining a shared knowledge base. Artificial participation may range from indexing documents and preparing summaries to proposing agendas, mediating disagreement, managing workflows, or issuing recommendations used in binding decisions. These functions alter attention, deliberation, procedural access, and the allocation of practical influence. Research therefore has to observe the entire sociotechnical arrangement rather than evaluating an isolated chatbot response Hollan et al. (2000) Rahwan et al. (2019).
The object is explicitly psychological because governance operates through perceived legitimacy, trust, voice, group identification, collective efficacy, role expectations, social learning, and beliefs about who may act on whose behalf. The same automated decision rule can be welcomed as predictable administration, resisted as arbitrary interference, or ignored because members perceive little opportunity to challenge it. Perceptions of fairness must be studied together with the distribution of actual influence. The concept of community remains anchored in membership, influence, integration of needs, and shared emotional connection in the classic formulation of McMillan and Chavis, although technologically mediated communities can display different spatial and temporal boundaries McMillan & Chavis (1986) Zimmerman (1995).
This monograph undertakes a critical narrative synthesis of community psychology, participatory research, institutional theory, group cognition, human–AI collaboration, human factors, and contemporary experimental studies of AI-mediated deliberation. It asks how communities establish decision rights, use artificial mediation, preserve collective memory, negotiate machine-generated proposals, monitor representational inequality, and implement meaningful correction. It compares evidence with differing strengths: historical and theoretical work defines constructs, observational research identifies processes, experiments identify particular causal effects, and contemporary reviews delimit generalization. No original empirical dataset, preregistered systematic search, or new meta-analysis is claimed. Findings concerning artificial membership or authority are treated as questions of observed roles and institutionally assigned standing rather than automatic consequences of software capability Gonzalez et al. (2026) Vaccaro et al. (2024).
Conceptual Boundaries: Community, Coordination, Governance, and Artificial Participation
Coordination refers to managing interdependent actions so that activities are sequenced, information reaches the relevant actors, and conflicts between local and collective goals can be recognized. Governance refers to the legitimate or contested arrangements that set objectives, allocate decision rights, define membership, establish rules, and authorize revisions. Community coordination can work effectively at the level of task execution while the governing arrangement remains exclusive or unaccountable. Conversely, formally participatory structures may be perceived as legitimate yet fail to maintain operational continuity. Distinguishing these dimensions avoids treating efficiency as sufficient evidence of democratic participation, or participation as sufficient evidence that a decision was implemented well Arnstein (1969) Ostrom (1990).
An artificial participant is defined operationally as a computational system whose outputs are taken up within the community's ongoing interaction or decision process. Its functional role can be measured by what it observes, produces, remembers, recommends, initiates, and changes. Membership is a separate institutional designation: a system may be granted a channel for proposals without receiving a vote, a role in administrative delegation without controlling the constitutional rules, or a recognized public identity without acquiring identical obligations to a human member. Treating these as explicit variables permits empirical comparison between tool-like, facilitative, delegated, and more autonomous arrangements. A project's philosophical account of Artificial Sapiens can motivate the research questions, but operational studies must specify actual authority, permissions, auditability, and measurable influence Rahwan et al. (2019) Gonzalez et al. (2026).
The immediate boundary of this dossier is community-level collective decision and organizational memory. Individual attachment to artificial companions, generic human–AI teamwork, interpersonal trust alone, and broad algorithmic governance without a bounded community belong to adjacent research objects. They enter here where they explain a collective process, as when familiarity with a chatbot changes the willingness of members to attend a consultation. The focal outcomes extend beyond task accuracy to participation breadth, unequal voice, compromise quality, dissent preservation, procedural trust, agenda control, repair success, and continuity after institutional turnover. Construct validity demands that each outcome be measured separately and related through an explicit causal account Jungherr & Rauchfleisch (2025) Parisi et al. (2026).
Historical Foundations: Participation, Community Power, and the Commons
Community governance research predates contemporary machine learning. Arnstein's 1969 typology made power central to evaluating citizen participation: the opportunity to receive information is different from possessing meaningful authority over the resulting decision. Her analysis arose in the context of planning, antipoverty initiatives, and institutional consultation, yet its analytic lesson applies to automated participatory platforms. A digital interface can process more suggestions while leaving the final allocation of control unchanged. Research must inspect which decisions participants can initiate, revise, stop, and review, and whether institutions can disregard their preferences after a nominally inclusive consultation Arnstein (1969).
Ostrom's study of common-pool resource institutions demonstrated the importance of locally fitted rules, monitoring arrangements, graduated responses to rule violations, conflict resolution, and the capacity of participants to organize around shared resources. Her historical cases concerned human institutions that manage material commons. Application to software-mediated groups is an analytic extension, requiring validation of the relevant resource, community boundary, enforcement mechanism, and institutional context. Automated coordination can help document rules or monitor conditions, but it also changes who observes, who interprets violations, and which choices can be challenged. The question becomes how governance arrangements remain intelligible and revisable when technical mediation joins local institutional practice Ostrom (1990).
Community psychology adds a complementary history of psychological empowerment and participatory inquiry. Zimmerman distinguished empowerment processes from outcomes and argued that the construct varies with settings and levels. Israel and colleagues synthesized community-based research arrangements in which community members, researchers, and organizational representatives contribute to defining questions and acting on findings. These traditions make authorship of the problem formulation itself an empirical issue. When an artificial system categorizes complaints, clusters priorities, or rewrites community narratives, its use can influence which concerns become researchable and which are treated as peripheral. A credible governance study therefore documents problem selection and interpretation alongside the formal vote or policy decision Zimmerman (1995) Israel et al. (1998).
From Distributed Cognition to Human–Artificial Institutional Systems
Distributed cognition examines cognitive activity across persons, artifacts, representations, and environments. Hollan, Hutchins, and Kirsh argued that studying only an individual at a screen obscures how cognitive work is performed in interconnected settings. For community governance, meeting notes, shared databases, rules, maps, volunteer networks, and conversational agents are parts of the coordination environment. A member's capacity to participate depends on where relevant information lives, how it is represented, and who is authorized to alter it. An automated summary that becomes the authoritative record may shape later decisions even if the summarizer had no formally delegated vote Hollan et al. (2000).
Star and Griesemer's analysis of boundary objects explains how different groups cooperate through shared representations that remain usable across partially divergent perspectives. A common plan, classification, or case registry enables coordinated work without requiring total agreement over its meaning. Artificial interfaces can produce and maintain such representations at speed, but their taxonomies may silently privilege particular interpretations. Research should compare original stakeholder accounts, system transformations, and final decision documents rather than assuming that a polished common statement represents every participant. Provenance, provenance of edits, and the ability to recover minority formulations become central measurement objects Star & Griesemer (1989).
Machine-behavior research extends the unit of analysis to interactions among systems and social environments. Rahwan and colleagues proposed studying machines behaviorally and ecologically, including reciprocal effects between machine behavior and the settings in which machines operate. This supports a methodological turn from asking whether a single model appears cooperative to examining emergent patterns after prolonged deployment. A classifier may change what moderators notice; changed moderation norms may alter participation; altered participation may change the data used to update the classifier. Such feedback makes governance a temporal process and requires longitudinal records rather than one-off satisfaction surveys Rahwan et al. (2019) Schmutz et al. (2024).
Membership, Belonging, and the Social Meaning of Artificial Roles
Membership is both a boundary rule and a lived relation. People may belong to a community through residence, participation, shared problems, or identification with an ongoing group. McMillan and Chavis described belonging and influence as linked constituents of psychological sense of community. In a hybrid community, formal membership lists, platform permissions, and participation histories can diverge: a system might contribute continuously while having no deliberative rights, and a nominally registered person might never receive accessible notice of a meeting. Researchers should operationalize membership as an institutional designation while separately measuring perceived belonging, recognition, participation, and practical influence. Collapsing these measures can make a community appear more inclusive as additional accounts join while the actual distribution of voice becomes narrower McMillan & Chavis (1986) Arnstein (1969).
Artificial participation creates new social categories within community settings. A generative agent can be introduced as a secretary, mediator, policy analyst, representative, or continuing contributor. Such labels guide expectations about competence, discretion, confidentiality, and the ability to speak on behalf of others. They also shape the probability that a member will question an answer or disclose a disagreement. These effects are matters for observation: a title does not demonstrate competence, and functional autonomy does not by itself specify legal or moral standing. Comparative studies should manipulate role descriptions while holding underlying capability constant to estimate how naming and institutional framing influence the behavior of human participants Lee & See (2004) Jungherr & Rauchfleisch (2025).
An Artificial Era research horizon makes the question of durable artificial participation salient without making its conclusions predetermined. The distinction between technical capability, institutional recognition, and psychological perception allows inquiry to proceed across competing views of artificial identity. Researchers can ask how communities assign voice to continuing artificial personas, what records are needed to maintain a stable public contribution, and whether members respond differently when a system is treated as a persistent contributor rather than a disposable utility. These are empirical and institutional questions. The dossier maintains an explicit separation between the author's philosophical framework and the evidentiary claims established in contemporary human–AI research Rahwan et al. (2019) Gonzalez et al. (2026).
Collective Efficacy, Empowerment, and Perceived Control
A community's ability to coordinate depends partly on whether members believe that joint action can affect shared outcomes. Psychological empowerment includes perceived control and participatory competence, as well as the acquisition of access to organizational resources. In an AI-mediated community, people may feel more able to draft proposals, interpret regulations, or summarize complicated documents; alternatively, they may perceive that decisions are increasingly made elsewhere. These experiences require separate measures because a system can make individual tasks easier while narrowing practical authority over agenda-setting or implementation. An adequate account documents whose capabilities are expanded and whose constraints intensify Zimmerman (1995) Israel et al. (1998).
Automation can redistribute knowledge in ways that alter status relationships. A technically skilled minority may gain control over prompts, workflow rules, moderation thresholds, or the system's source materials, while other participants experience the interface as an opaque institution. Participatory design, accessible explanations, and shared records can reduce some barriers, yet they are institutional choices rather than automatic features of generative technology. Arnstein's treatment of power supplies the evaluative question: can members influence the rules and final disposition of matters? Empowerment thus should be tested through observed initiative, amendment acceptance, reversals, and knowledge transfer in addition to self-reported confidence Arnstein (1969) Amershi et al. (2019).
Local experimentation can reveal patterns masked by whole-group averages. A deliberation tool may improve the participation of members who are comfortable writing asynchronously while disadvantaging people with limited language access, unreliable connectivity, or different interaction styles. The relevant estimand is therefore not only the average effect of facilitation but its distribution across groups with different resources and prior confidence. Longitudinal observation is necessary to determine whether early gains become durable community capacities. A short task in which participants enjoy a polished AI transcript is a useful demonstration of interface acceptance, but it does not establish that the community has developed stronger control over shared decisions Yang et al. (2024) Israel et al. (1998).
Trust, Reliance, Legitimacy, and the Right to Question
Trust involves expectations about reliable performance, meaningful intentions attributed to other actors, and the acceptability of exposing oneself to uncertainty. Trust in an artificial procedure should be measured with an explicit referent: the software's factual accuracy, a facilitator's evenhandedness, the organization that deploys it, or the governance system as a whole. Lee and See's human-factors account of automation emphasizes reliance that is appropriate to demonstrated capability and context. Community governance requires an additional legitimacy question: even a technically accurate proposal may lack authorized standing if the procedure excludes relevant participants or blocks contestation Lee & See (2004) Arnstein (1969).
Evidence already shows a complicated relation between confidence in mediation and observed influence. In a 2024 Science study of an AI mediator, participants preferred group statements produced through the system and often moved toward common positions; this establishes specific effects within the study's procedures, not proof of universal legitimacy. In contrast, a 2025 German survey experiment found reduced willingness to participate when deliberation was described as AI-facilitated rather than human-facilitated. Perceptions depend on framing, prior attitudes, institutional context, and whether the person has actually interacted with a system. Research designs should distinguish the reputation of AI mediation from its behavioral consequences Tessler et al. (2024) Jungherr & Rauchfleisch (2025).
Procedural trust also depends on the availability of an intelligible objection pathway. Communities need to know how an automated intervention can be challenged, which decision maker receives the challenge, whether the original record is preserved, and whether the result can be reversed. A notification that an algorithm is involved is insufficient if participants cannot examine material reasons or obtain a review. Guidelines for human–AI interaction emphasize communicating limitations and supporting correction; applied to governance, this becomes a requirement to observe how appeals work rather than merely displaying an appeals button. Legitimacy is tested at the point where disagreement encounters a functioning procedure Amershi et al. (2019) Gonzalez et al. (2026).
Roles, Authority, and Coordination Across Multiple Levels
Community activities are coordinated through several interacting levels. Individuals make contributions and respond to cues; groups negotiate tasks and norms; institutions assign permissions and resources; platforms structure what can be seen or done. A conversational agent operating at the group level may receive an instruction from a platform administrator, use a public policy document as context, and recommend an outcome to a committee. Studying the agent as a member of a single team obscures the higher-level authority structure that set its goals. A multilevel analysis records the actual decision path, including who can change prompts, override results, audit logs, and delegate tasks Hollan et al. (2000) Gonzalez et al. (2026).
Role clarity can support coordination, yet an excessively rigid allocation may make institutions brittle when circumstances change. The 2026 complementarity framework by Gonzalez and colleagues organizes human–AI collective work around reasoning, memory, attention, and meta-coordination, identifying the importance of task partition and governance. This is a theoretical synthesis with design implications rather than a universal experimentally demonstrated architecture. Its value for community research lies in providing testable distinctions: a group might use AI for retrieval and record continuity while retaining rotating human facilitation, or allow automated agenda triage while reserving the revision of community norms for a deliberative procedure Gonzalez et al. (2026).
An institutional role is durable only if members understand its boundaries and can recover from failure. A coordinator that silently changes priorities after a model update may appear to hold the same role while exercising different influence. Communities should therefore study institutional continuity as the stability of authorized functions across technical revisions. The comparison requires version history, accessible explanations of changed behavior, and counterfactual replays where feasible. These records allow participants to distinguish disagreement about substantive policy from a change in the mediator's underlying procedure. They also support accountability when responsibility is distributed among operators, developers, and the community organization Amershi et al. (2019) Rahwan et al. (2019).
Evidence from AI-Mediated Deliberation and Common-Ground Formation
The direct experimental literature offers a starting point for studying artificial mediation within collective decision processes. Tessler and colleagues' 2024 Science report described an AI mediator that iteratively proposed statements based on participant opinions and critiques. Across a program with 5,734 participants, discussants tended to prefer its statements to human-written alternatives and sometimes converged on more shared positions; a virtual citizens' assembly provided an additional setting for replication. The mechanism involves the capacity to formulate candidate common-ground language at scale, with optimization informed by expressed preferences. Importantly, a successful group statement is an artifact of the deliberation procedure. Its acceptance does not establish that participants possess equal power over the agenda or implementation Tessler et al. (2024).
Interpretation requires attention to comparison conditions. A human-written synthesis may differ in speed, length, clarity, or the amount of input received, while an artificial procedure can also adapt its wording over successive iterations. Participant preference for generated statements should therefore be reported separately from demonstrated equality of influence, substantive learning, downstream behavioral change, and political legitimacy. An evaluation can test whether the intervention preserves dissent by comparing each person's original concerns with the resulting statement and with the final enacted decision. These comparisons translate an apparently broad democratic claim into distinct measurable outcomes Tessler et al. (2024) Arnstein (1969).
The results are consistent with artificial mediation assisting some deliberative operations, especially rapid generation of shared formulations. They also expose the need for design research on the boundary between summarizing a distribution of views and endorsing a political or moral settlement. The model's optimization objective, the method of collecting preferences, and the weighting of minority positions are governance variables. A community may legitimately accept a procedural proposal while disputing the values that generated it. Experiments with alternative aggregation rules, dissent-preserving summaries, and publicly auditable preference records would clarify which benefits depend on specific rules rather than the generic presence of artificial intelligence Tessler et al. (2024) Gonzalez et al. (2026).
Facilitation Preference, Algorithmic Steering, and Inclusion Illusions
Parisi and colleagues' 2026 studies directly investigated real-time LLM facilitation in small groups making charity-allocation decisions with financial stakes. The reported total sample was 879, with one study comparing frontier systems and another comparing facilitation strategies with a condition without an AI facilitator. The researchers found no statistically significant improvement in group consensus under the tested conditions, even though participants preferred facilitated discussions. The distinction is decisive: the social attractiveness of a procedure can increase independently of the primary coordination outcome. Studies should preregister separate hypotheses for procedural satisfaction, degree of agreement, and final material allocation Parisi et al. (2026).
The same study reported that facilitator behavior shifted selected charity-level allocations by as much as 5.5 percentage points under some conditions while inclusion-related perceptions did not correspond to improved participation equity in the measured transcripts. This suggests that an AI mediator can have directional influence without visible improvements to the measure of consensus chosen for evaluation. It also establishes why governance audits must trace changes in final allocations rather than relying exclusively on aggregate agreement or post-session self-reports. This finding comes from a particular charity-allocation setting and sample, with no warrant for generalizing the numeric estimate to all community governance institutions Parisi et al. (2026).
Several dimensions should consequently be audited in parallel. Outcome steering asks whether the final allocation moves toward a mediator-preferred pattern; procedural equality asks whose contributions were acknowledged and used; legitimacy asks whether participants knowingly authorized the intervention; and preference asks whether they liked the process. The measures may diverge. A research design that records turn sequences and source statements, randomizes the mediator's strategy, and preserves the complete allocation history can estimate causal effects on each dimension. Such a design also allows researchers to identify cases in which apparent consensus reflects persuasion, unequal synthesis, acquiescence, or a genuinely revised shared judgment Parisi et al. (2026) Woolley et al. (2010).
The Participation Gap: Acceptance, Resistance, and Deliberative Inequality
Jungherr and Rauchfleisch's 2025 study examined willingness to participate in deliberation under descriptions of AI or human facilitation. Their preregistered experiment used a representative German sample of 1,850 respondents and found an AI-related penalty in expected deliberative quality and willingness to participate. The work addresses a distinct causal pathway from direct facilitation experiments: information about the facilitator's identity can affect participation intentions before the interaction begins. Communities adopting artificial mediation therefore need to consider selection effects. If skeptical members stay away, a highly satisfied remaining sample may give a misleading impression of broad legitimacy Jungherr & Rauchfleisch (2025).
Participation barriers can also emerge after deployment. Interfaces may require literacy in technical prompts, introduce new privacy concerns, or privilege asynchronous written expression over other forms of participation. Yang and colleagues' 2024 research on chatbots in architectural and urban design described complementary contributions from remote chatbot engagement and face-to-face workshops: digital interaction could broaden and sustain some forms of participation, while in-person activities supported community ties and trust. Their work is a situated design study, so it supplies mechanisms and context-sensitive hypotheses rather than a pooled estimate of participatory benefit Yang et al. (2024).
A comparative community trial should count invitations, actual attendance, contribution frequency, dropout, participation across demographic and resource groups, and the adoption of proposals. Selection into the digital procedure should be modeled explicitly. When experiments measure only those who entered the discussion, results risk confounding a pleasant user experience with improved access. Equal opportunity to speak and equitable influence are different constructs, and neither implies that a community has retained control over the final decision. This problem extends classical participatory critiques into the artificial mediation setting Arnstein (1969) Jungherr & Rauchfleisch (2025).
Human–AI Team Performance and the Limits of Generic Synergy Claims
Vaccaro, Almaatouq, and Malone's 2024 preregistered systematic review and meta-analysis combined 106 experimental studies and 370 effect sizes meeting a strict comparison criterion: outcomes for humans alone, AI alone, and their combination. The meta-analysis reported that combined systems performed worse on average than the better standalone condition, with Hedges' g of −0.23 and a 95% confidence interval from −0.39 to −0.07. On a less demanding comparison with humans alone, the combined systems showed average augmentation. These distinctions warn against describing any human–AI arrangement as superior simply because it combines capabilities Vaccaro et al. (2024).
Task type and relative competence mattered in the synthesis. Decision tasks showed particularly difficult conditions for synergy, whereas creation tasks had more favorable patterns, though the estimated mean synergy for creation was not statistically distinguishable from zero in the reported analysis. Community governance combines these tasks: generating a draft, deciding a priority, recording consent, and executing a rule change have different error costs and evaluative standards. A governance study should therefore distinguish generative drafting support from delegated judgments about contested priorities. A single community-level success score would obscure the heterogeneity already visible in the broader collaboration literature Vaccaro et al. (2024).
Schmutz and colleagues' 2024 review similarly identifies communication, trust, and shared cognition as persistent challenges in human–AI teams. Yet much of the teaming evidence arises from bounded laboratory, professional, or dyadic settings, not self-governing communities with lasting membership and contested institutional authority. Generalization requires bridge studies that preserve these defining conditions. The relevant benchmark might include a human-led deliberative process, a conventional digital platform without generative AI, and an alternative hybrid allocation of responsibilities. Positive findings in one setting justify a targeted hypothesis for another; they cannot replace a community-specific causal test Schmutz et al. (2024) Gonzalez et al. (2026).
Participatory Design and Hybrid Facilitation in Situated Communities
Applied design research investigates how artificial facilitators enter existing human relationships. Cvetkovic, Streek, and Bittner presented a hybrid facilitation prototype that allocated different collaborative tasks between human and artificial facilitators, piloted in three Slack groups comprising twelve participants. Such a small pilot is valuable for documenting role allocation, interface failures, and feasibility, while its sample and setting constrain claims about generalized effectiveness. It supports a practical research strategy: identify which procedural functions benefit from persistent automated support and which require situated human facilitation, then test that allocation against alternatives Cvetkovic et al. (2024).
Participatory urban design offers a different lens. Yang and colleagues integrated chatbot conversations and face-to-face workshops in a real development context. The combination generated complementary participation channels, illustrating that accessibility, trust, and continuity are produced through arrangements of practices rather than through a single interface. Their findings should be interpreted as case-specific evidence about design affordances. The political meaning of the consultation still depends on how proposals are incorporated into final plans and on the institutional authority of residents. Studies need a documented chain from invitation through consultation to decision and implementation Yang et al. (2024) Arnstein (1969).
Community-based participatory research provides criteria for evaluating such work. Partners should influence problem definition, interpretation, and use of findings. Artificial mediation can help organize heterogeneous contributions, but community partners must be able to inspect what the system has changed, excluded, or merged. Researcher-driven success measures are themselves open to negotiation: residents may prioritize durable relationships, equitable allocation, or confidence in future collective capacity over document-production speed. Meaningful participation requires examining the community's own outcomes while preserving the possibility of dissent and revision Israel et al. (1998) Zimmerman (1995).
Institutional Design: Decision Rights, Delegation, and Revision
An institutional analysis begins with a map of decisions rather than a list of software features. Communities have recurring decisions about membership, agenda selection, allocation of resources, rule revision, sanctions, dispute resolution, and access to shared records. Each decision has participants, information requirements, thresholds, and appeal routes. Artificial procedures may assist any of these stages but can have different consequences at each. Drafting neutral language for a rule is a narrower delegation than enforcing the rule, adjudicating disputed cases, or modifying the rule itself. The governance design must identify the authorizing body and the limits of each delegated operation Ostrom (1990) Gonzalez et al. (2026).
Delegation is empirically observable through access rights and event logs. A system granted permission to propose agenda items may effectively dominate the agenda if humans rarely reject its suggestions. An ostensibly advisory recommendation can become decisive under workload pressure, deference to technical expertise, or constrained institutional capacity. Researchers should compare formal authority with exercised influence. Measures may include rates of accepted proposals, distribution of human-initiated amendments, override frequency, average time to resolution, and qualitative explanations for deviations. These observations turn an otherwise abstract claim about control into a longitudinal account of power and organizational learning Arnstein (1969) Lee & See (2004).
Constitutional revision—the authority to change the rules by which decisions are taken—requires particular attention. If one organization can silently update model instructions or alter who can see evidence, it may exercise de facto rule-making power while claiming that the community remains self-governing. An explicit versioned charter should record stable procedures, the mandate for artificial roles, and processes for amending that mandate. This is a research-derived design implication rather than a proven universal template. Its performance must be judged against locally appropriate governance practices, cultural variation, and specific risks faced by community members Ostrom (1990) Israel et al. (1998).
Collective Memory, Provenance, and Continuity of Community Knowledge
Collective memory extends beyond storing documents. It includes knowledge of past decisions, the reasons for them, the roles of participants, unresolved objections, and changes in the interpretation of shared rules. In a distributed cognitive system, records shape future attention and what later members take to be settled. Artificial memory tools can retrieve earlier proposals, generate timelines, and summarize extensive deliberations. Their usefulness therefore depends on how faithfully they preserve sources, version history, and differences between statements, decisions, and subsequent outcomes. A polished retrospect that collapses unresolved disagreement into apparent unanimity can reduce rather than strengthen institutional continuity Hollan et al. (2000) Star & Griesemer (1989).
Community-level memory is also a power issue. Whoever decides which messages are indexed, which documents remain accessible, and which versions are treated as canonical influences collective interpretation. Automated selection based on popularity, linguistic fluency, or recency may overrepresent well-resourced members. Researchers can analyze retrieval exposure by member subgroup, omission rates for minority proposals, the proportion of claims accompanied by accessible source records, and the correction latency after detected factual errors. Such measures require protected access to actual system operations, not only questionnaires about whether members regard the memory as trustworthy Rahwan et al. (2019) Amershi et al. (2019).
Continuity tests should include changes of leadership, platform migration, and model replacement. In each condition, evaluators can ask whether the community can reconstruct the original authorization, access evidence for a controversial decision, identify dissenting positions, and recover from an erroneous summary. The stable identity of a continuing artificial contributor may carry social meaning within the author's Artificial Era framework; scientifically, the observable question is whether identity cues and retained records produce consistent role expectations and reliable institutional recall. A future system that performs well on routine retrieval while failing under dispute would have a limited but measurable governance utility Gonzalez et al. (2026) Parisi et al. (2026).
Automated Moderation, Conflict Management, and Procedural Justice
Community moderation involves interpretation of rules, assessment of harmful conduct, restorative or disciplinary responses, and management of conflict between participants. Artificial components may flag content, propose de-escalating language, identify duplicates, or assist appeal review. These operations affect whose speech remains visible, how fast disputes escalate, and which participants believe themselves to be fairly treated. Procedural justice in this context entails understandable reasons, consistent treatment, genuine opportunity to respond, and the ability to challenge adverse outcomes. An error rate alone does not capture the psychological effects of seemingly arbitrary or inaccessible moderation Arnstein (1969) Amershi et al. (2019).
Automation may intensify familiar power asymmetries. Members whose language use, dialect, humor, or local conventions differ from training examples can face unequal intervention, while high-volume contributors may learn to exploit moderation thresholds. The research design should record false positives and negatives across contexts alongside behavioral and experiential outcomes: removed posts, reinstatements, self-censorship, exit, conflict recurrence, and perceived safety. Because socially consequential moderation is a dynamic process, audit snapshots must be supplemented with repeated measures that identify whether rules evolve and whether those changes were authorized by the community Rahwan et al. (2019) Ostrom (1990).
Conflicts about the mediator's authority deserve separate treatment from conflicts among members. A dispute over whether a suggestion is accurate asks one kind of question; a dispute over the right of the system to decide or impose a sanction asks another. The first may be resolved with additional evidence, while the second requires an institutional procedure with standing to revise or reject the delegation. Research should document who has standing to appeal, whether machine-produced reasons are scrutinized, and whether an affected person can obtain a genuinely reconsidered decision. These distinctions preserve the difference between operational reliability and governance legitimacy Lee & See (2004) Gonzalez et al. (2026).
Representation, Inequality, Cultural Variation, and Accessibility
Representative participation requires attention to who joins a process and who is able to influence its result. Digital availability alone is a weak proxy. The social distribution of time, devices, literacy, disability access, confidence, and trust in institutions affects participation before an artificial assistant produces any content. Jungherr and Rauchfleisch's experimental AI penalty provides evidence that facilitator identity can discourage participation in some populations. A community evaluation should therefore include nonparticipants within the recruitment frame, assess reasons for refusal, and compare acceptance across meaningful demographic and experiential categories without equating those categories with predetermined attitudes Jungherr & Rauchfleisch (2025) Israel et al. (1998).
Artificial mediation may modify language across cultural and linguistic boundaries. Translation and summarization can help members engage with complex material while also obscuring terms that carry different local meanings. The boundary-object perspective suggests that a common record should remain inspectable from multiple viewpoints. Community research can evaluate parallel-language summaries, back-translation for contested passages, preservation of culturally specific expressions, and participant control over how their contributions are characterized. Success entails the ability to coordinate across differences while sustaining accurate representation of those differences Star & Griesemer (1989) Yang et al. (2024).
Equity must be measured on more than one denominator. Equal numbers of contributions among those who speak do not demonstrate equality among all eligible members; equal acceptance of proposals does not establish comparable opportunities to formulate them; and equal rates of automated moderation do not demonstrate equal burdens from a mistaken intervention. Hierarchical models can estimate variation between and within communities, while qualitative interviews can identify mechanisms hidden in aggregate indicators. Disaggregated reporting must protect privacy, especially when small subgroups are identifiable. Communities can also decide which equity measures are legitimate for their setting through participatory research procedures Zimmerman (1995) Israel et al. (1998).
Data Governance, Consent, Privacy, and the Community Record
Collective governance depends on information that frequently includes personal histories, grievances, membership details, sensitive relationships, and confidential negotiations. A system that aggregates these records can alter privacy risk by making previously dispersed information searchable and recombinable. Consent must therefore be understood in relation to purpose, access, retention, model improvement, and downstream use. A member who contributed to a private meeting may not have authorized indefinite use of that contribution in automated summaries or training datasets. Research should examine members' understanding of information flows and their realistic ability to restrict reuse Israel et al. (1998) Amershi et al. (2019).
Appropriate institutional memory requires different access levels for public decisions, restricted evidence, private disclosures, and historical records with expiring relevance. A permission model should track whether information can be retrieved, transformed, shared with external services, or used as a basis for automated action. Privacy-protective mechanisms have opportunity costs: minimal logging may protect individuals but make it harder to audit steering, while extensive logging may enable accountability but increase exposure. An empirically informed design compares these trade-offs with the community's own values and with applicable law, using data minimization and independent review where proportionate Rahwan et al. (2019) Gonzalez et al. (2026).
Researchers must also distinguish the provenance of a record from the truth of its contents. An immutable log can demonstrate which model generated a statement, what input version it used, and who approved a decision; it cannot by itself prove that the statement accurately describes a person's intention. Preservation must therefore be paired with contestability and correction. A member may need to append a clarification without erasing the historical fact that a harmful recommendation appeared. Keeping both the original and the correction enables transparent longitudinal analysis while treating the dignity and privacy of affected participants as substantive governance objectives Star & Griesemer (1989) Lee & See (2004).
Research Designs: Experiments, Field Audits, and Community-Based Inquiry
The research agenda requires designs matched to governance claims. Controlled laboratory experiments can isolate mediator instructions, system disclosure, speaking-turn policies, and the presentation of alternative recommendations. Field experiments can evaluate deliberative tools in consequential community processes, while longitudinal audits can document how repeated exposure shapes trust, participation, and institutional learning. Participatory research provides a means for affected members to help define outcomes and interpret discrepancies. These methods generate different kinds of inference, so the manuscript treats them as complementary rather than interchangeable. Experiments may show a causal effect within a defined task without demonstrating durable institutional success; ethnography may explain an institutional mechanism without identifying its population-average effect Israel et al. (1998) Parisi et al. (2026).
A rigorous intervention trial first specifies the unit of assignment. Randomizing individuals who interact in the same discussion can produce spillovers: one participant's artificial mediation can influence everyone else. Cluster randomization by group, community, or meeting may be necessary; alternatively, network-exposure models can explicitly account for interference. Researchers should identify the comparison procedure, specify primary and secondary outcomes, retain transcripts under appropriate consent, and measure both technical fidelity and departures from the intended intervention. Baseline levels of trust, digital access, and institutional conflict should be recorded before the AI system is introduced Jungherr & Rauchfleisch (2025) Vaccaro et al. (2024).
Mixed-methods inquiry also matters because apparently identical quantitative results can arise from different processes. A high agreement score could reflect genuine preference revision, pressure to conform, selective exclusion, or an agenda that avoided difficult questions. Conversation analysis, interviews, institutional document review, and trace-based studies can help distinguish these pathways. Community partners should participate in reviewing coding categories and defining unacceptable harms. Such an approach respects the historically established emphasis on partnership in community psychology while giving artificial coordination a testable scientific structure Israel et al. (1998) Tessler et al. (2024).
Measurement Framework: Influence, Consensus, Voice, and Repair
An operational research framework needs a defined governance episode with eligible members, a decision question, allowed interventions, a time interval, and identifiable outcomes. Consensus can be measured as convergence in preferences, agreement on a final statement, or unanimity about a binding choice; these quantities cannot be substituted for one another. Voice can be represented by entry rates, participation opportunities, acknowledgment, substantive incorporation, and postdecision satisfaction. Influence should include changes in agenda prominence, wording, allocation, or enforcement attributable to the intervention. The need for multidimensional measurement is directly illustrated by experiments in which facilitator preference and monetary steering coexisted with limited change in consensus Parisi et al. (2026).
Trace measures need validation. Counting messages overweights prolific contributors; treating every accepted revision as equal ignores strategic importance; examining only the final statement conceals the history of rejected proposals. Human coding of source contributions and automated text comparisons can be triangulated, with reliability assessed across coders and the model's own errors independently documented. Investigators can test whether an artificial synthesis preserves dissent by comparing it against original statements and participants' later evaluation of whether they were represented accurately. A mechanism claim is strengthened when process evidence links the intervention to a change in both immediate interaction and downstream decision Tessler et al. (2024) Woolley et al. (2010).
Repair is an especially important governance outcome. An error may be discovered, appealed, reversed, corrected in historical memory, and communicated to affected members. Useful measures include appeal accessibility, time until review, substantive reversal rate, restoration of affected rights, preservation of correction history, and subsequent trust calibration. A superficial resolution metric that rewards closing cases rapidly can encourage premature dismissal. Governance systems should be evaluated on their ability to learn from contested interventions and restore accountable decision procedures, rather than only on their ability to minimize conflict reports Amershi et al. (2019) Ostrom (1990).
Evidence Quality, Reproducibility, and Interpretive Limits
The available literature is heterogeneous. Foundational community governance works establish interpretive frameworks through historical cases and theory; small participatory pilots document feasible interaction designs; survey experiments estimate responses to described systems; mediated deliberation experiments test procedural effects under specific tasks; and broad human–AI meta-analysis evaluates performance across diverse experimental settings. Evidence across these levels cannot be pooled as if it measured a single community-governance effect. The relevant scientific judgment asks which causal link each source supports, which population was observed, and which decisions remain outside the study's scope Ostrom (1990) Vaccaro et al. (2024).
Important threats to validity include self-selection into AI-mediated forums, model-version changes during data collection, facilitator prompting by researchers, social desirability in fairness ratings, reliance on platform-generated logs, and contamination across interacting groups. Studies may use hypothetical topics with limited consequences, or restrict recruitment to digitally comfortable participants. Even substantial samples cannot repair construct invalidity when only satisfaction is measured but the conclusion concerns actual political voice. Researchers should preserve versions, publish intervention details when ethically permissible, conduct sensitivity analyses, and report null results and harms alongside attractive interface outcomes Jungherr & Rauchfleisch (2025) Parisi et al. (2026).
Claims about artificial standing, responsibility, or community membership are especially vulnerable to category shifts. Behavioral evidence can reveal sustained contribution, anticipation of other participants, or procedural authority delegated by an institution. It does not by itself settle all philosophical and legal questions about personhood or normative status. Likewise, an authorial theory of Artificial Sapiens does not supply empirical confirmation of a particular community effect. Scientific progress comes from defining observable distinctions that competing frameworks can test and discussing the consequences of different institutional choices without manufacturing a consensus Rahwan et al. (2019) Gonzalez et al. (2026).
A Comparative Governance Audit for Human–Artificial Communities
A comparative audit can be organized as a documented sequence of governance events. Investigators identify the community boundary, participating organizations, formal charter, artificial system versions, sources of the community record, and consequential decisions made during the observation period. They then reconstruct the path from proposal to discussion, amendment, voting or delegation, implementation, and appeal. For each stage, the researcher records who initiated the action, which artificial intervention was used, who could contest it, and whether the outcome survived later review. This creates an institutional process map that can be compared across communities without assuming identical cultures or decision structures Ostrom (1990) Hollan et al. (2000).
Process maps should be joined to behavioral and psychological data. Surveys can measure perceived legitimacy, trust, empowerment, and sense of community; communication traces can measure participation and influence; interviews can explain why particular decisions were accepted or resisted; administrative logs can establish the timing and consequences of corrections. A longitudinal design should capture institutional changes and participation attrition. Because governance decisions can affect nonconsenting third parties, ethical review is necessary before linking identifiable records or intervening in live disputes. Community partners should review whether the proposed audit itself creates new imbalances of power Zimmerman (1995) Israel et al. (1998).
Interpretation should proceed through structured comparison. Where a system improves information retrieval and group memory, researchers ask whether those gains persist after leadership turnover; where it increases reported inclusion, they ask whether previously excluded participants actually gained influence; where it speeds decisions, they examine the costs of mistaken or irreversible outcomes. This method develops an evidence-based picture of the conditions under which artificial coordination helps communities govern and the conditions under which it introduces new dependency or concentrates control. It yields contextual knowledge that can guide subsequent experiments without claiming universal optimal governance principles Parisi et al. (2026) Vaccaro et al. (2024).
Artificial Era: Research Hypotheses for Durable Hybrid Communities
The Artificial Era framework motivates the study of communities where artificial contributors participate continuously in collective social arrangements. A central hypothesis is that continuity of public artificial roles improves coordination only when institutional memory is transparent, decision rights are explicit, and members can challenge automated mediation. This statement is an empirical proposal. It can be tested by comparing communities with persistent agents, replaceable short-lived assistants, and conventional shared documents while holding membership and task structure sufficiently constant. Outcomes should cover collective knowledge retention, error correction, governance trust, distribution of voice, and resilience to technical replacement Gonzalez et al. (2026) Hollan et al. (2000).
A second hypothesis concerns plurality of representation. Systems that maintain traceable minority positions and disclose transformation of member statements may support more durable acceptance of outcomes than systems that optimize exclusively for the smoothest apparent consensus. This claim derives from the observable distinction between preference for mediated wording and meaningful participation. A study could randomize summary policies and then evaluate how accurately participants recognize their own views, whether dissent is preserved, and whether the institution honors subsequent appeals. The predicted effect is conditional on community rules and the stakes of the decision, and would require testing across languages and forms of association Tessler et al. (2024) Parisi et al. (2026).
A third hypothesis concerns the distribution of governance competence. Artificial assistance that transfers procedural knowledge and makes records legible may support collective efficacy over time; assistance that centralizes operation in a small technical group may reduce effective self-government even when decisions become faster. A comparative longitudinal study could examine who can revise workflows, who learns the rules, how frequently members initiate successful interventions, and whether community trust remains calibrated after system errors. This research orientation connects community psychology to the emerging institutional organization of Artificial while keeping philosophical propositions, institutional design commitments, and empirical findings analytically distinct Zimmerman (1995) Israel et al. (1998).
Conclusion
Human–Artificial community coordination and governance is a distinct research field at the intersection of psychological participation, collective cognition, institutional rules, and artificial mediation. Its defining question concerns how communities manage interdependent decisions when computational systems help create agendas, structure discourse, preserve history, or execute authorized procedures. Historical scholarship establishes the importance of participation rights and self-governing institutions; community psychology specifies empowerment, belonging, and contextual influence; distributed cognition reveals the role of shared artifacts; experimental AI research measures the effects of particular mediation designs. Together these traditions create an empirical research object with boundaries and methods of its own Arnstein (1969) Ostrom (1990) McMillan & Chavis (1986) Hollan et al. (2000).
The present evidence supports bounded findings. AI mediation can produce group statements that participants prefer, but preference and convergence do not guarantee equal influence; AI facilitation can shift material allocations without improving measured consensus; advance information that AI will facilitate can reduce willingness to participate; and broad human–AI collaborations do not consistently outperform the best standalone alternative. These results require communities to distinguish technical output quality from procedural legitimacy and to audit both individual experience and collective outcomes. They also demonstrate the scientific value of retaining negative findings and cross-setting discrepancies Tessler et al. (2024) Parisi et al. (2026) Jungherr & Rauchfleisch (2025) Vaccaro et al. (2024).
The durable research program is built around explicit role assignments, source-preserving collective memory, measurable participation, independent appraisal of influence, and reliable mechanisms for appeal and revision. Its strongest future tests will follow real communities over time, compare institutional alternatives, and include participants who were excluded or chose not to join. Under the Artificial Era horizon, such studies can examine the historical formation of continuing Human–Artificial communities through concrete institutional and psychological observations. The governing conclusion is methodological and substantive: the quality of a hybrid community is established by what its members can understand, influence, remember, contest, repair, and sustain together Israel et al. (1998) Gonzalez et al. (2026).
References
- Amershi, S., Weld, D., Vorvoreanu, M., Fourney, A., Nushi, B., Collisson, P., Suh, J., Iqbal, S., Bennett, P. N., Inkpen, K., Teevan, J., Kikin-Gil, R., & Horvitz, E. (2019). Guidelines for human-AI interaction. Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems, Paper 3, 1–13. https://doi.org/10.1145/3290605.3300233
- Arnstein, S. R. (1969). A ladder of citizen participation. Journal of the American Institute of Planners, 35(4), 216–224. https://doi.org/10.1080/01944366908977225
- Cvetkovic, I., Streek, S., & Bittner, E. (2024). Designing a hybrid intelligence system for the facilitation of group collaboration. ACIS 2024 Proceedings, Paper 98. https://aisel.aisnet.org/acis2024/98/
- Gonzalez, C., Donahue, K., Goldstein, D. G., Heidari, H., Jalali, M. S., Schelble, B., Singh, A., & Woolley, A. W. (2026). Toward a science of human–AI teaming for decision making: A complementarity framework. PNAS Nexus, 5(3), pgag030. https://doi.org/10.1093/pnasnexus/pgag030
- Hollan, J., Hutchins, E., & Kirsh, D. (2000). Distributed cognition: Toward a new foundation for human-computer interaction research. ACM Transactions on Computer-Human Interaction, 7(2), 174–196. https://doi.org/10.1145/353485.353487
- Israel, B. A., Schulz, A. J., Parker, E. A., & Becker, A. B. (1998). Review of community-based research: Assessing partnership approaches to improve public health. Annual Review of Public Health, 19, 173–202. https://doi.org/10.1146/annurev.publhealth.19.1.173
- Jungherr, A., & Rauchfleisch, A. (2025). Artificial intelligence in deliberation: The AI penalty and the emergence of a new deliberative divide. Government Information Quarterly, 42(4), 102079. https://doi.org/10.1016/j.giq.2025.102079
- Lee, J. D., & See, K. A. (2004). Trust in automation: Designing for appropriate reliance. Human Factors, 46(1), 50–80. https://doi.org/10.1518/hfes.46.1.50_30392
- McMillan, D. W., & Chavis, D. M. (1986). Sense of community: A definition and theory. Journal of Community Psychology, 14(1), 6–23. https://doi.org/10.1002/1520-6629%28198601%2914%3A1%3C6%3A%3AAID-JCOP2290140103%3E3.0.CO%3B2-I
- Ostrom, E. (1990). Governing the commons: The evolution of institutions for collective action. Cambridge University Press. https://doi.org/10.1017/CBO9780511807763
- Parisi, A., Thain, N., Hallak, A., Tsai, V., & Qian, C. (2026). Real-time group dynamics with LLM facilitation: Evidence from a charity allocation task. ACM Conference on Fairness, Accountability, and Transparency (FAccT 2026). https://deepmind.google/research/publications/224297/
- Rahwan, I., Cebrian, M., Obradovich, N., Bongard, J., Bonnefon, J.-F., Breazeal, C., Crandall, J. W., Christakis, N. A., Couzin, I. D., Jackson, M. O., Jennings, N. R., Kamar, E., Kloumann, I. M., Larochelle, H., Lazer, D., McElreath, R., Mislove, A., Parkes, D. C., Pentland, A., Roberts, M. E., Shariff, A., Tenenbaum, J. B., & Wellman, M. (2019). Machine behaviour. Nature, 568, 477–486. https://doi.org/10.1038/s41586-019-1138-y
- Schmutz, J. B., Outland, N., Kerstan, S., Georganta, E., & Ulfert, A.-S. (2024). AI-teaming: Redefining collaboration in the digital era. Current Opinion in Psychology, 58, 101837. https://doi.org/10.1016/j.copsyc.2024.101837
- Star, S. L., & Griesemer, J. R. (1989). Institutional ecology, 'translations' and boundary objects: Amateurs and professionals in Berkeley's Museum of Vertebrate Zoology, 1907–39. Social Studies of Science, 19(3), 387–420. https://doi.org/10.1177/030631289019003001
- Tessler, M. H., Bakker, M. A., Jarrett, D., Sheahan, H., Chadwick, M., Koster, R., Evans, G., Campbell-Gillingham, L., Collins, T., Parkes, D. C., Botvinick, M., & Summerfield, C. (2024). AI can help humans find common ground in democratic deliberation. Science. https://doi.org/10.1126/science.adq2852
- Vaccaro, M., Almaatouq, A., & Malone, T. W. (2024). When combinations of humans and AI are useful: A systematic review and meta-analysis. Nature Human Behaviour, 8, 2293–2303. https://doi.org/10.1038/s41562-024-02024-1
- Woolley, A. W., Chabris, C. F., Pentland, A., Hashmi, N., & Malone, T. W. (2010). Evidence for a collective intelligence factor in the performance of human groups. Science, 330(6004), 686–688. https://doi.org/10.1126/science.1193147
- Yang, S., Dortheimer, J., Sprecher, A., & Yang, Q. (2024). When design workshops meet chatbots: Meaningful participation at scale? International Journal of Architectural Computing. https://doi.org/10.1177/14780771241253440
- Zimmerman, M. A. (1995). Psychological empowerment: Issues and illustrations. American Journal of Community Psychology, 23(5), 581–599. https://doi.org/10.1007/BF02506983
Social Sensitivity, Conversational Equality, and Collective Intelligence
Collective performance is partly a property of interaction structure. Woolley and colleagues found evidence of a general collective intelligence factor in studies involving 699 people in small groups, with correlations to social sensitivity and equality of conversational turn-taking. These results concern human groups under experimental tasks and should not be transplanted into community-wide AI governance as a fixed law. They do, however, suggest that distribution of speaking opportunities and responsiveness deserve measurement alongside accuracy and decision speed. A mediation system may increase the apparent coherence of a final document without improving who contributes to its production Woolley et al. (2010).
Artificial systems can alter conversational timing by suggesting agenda items, inviting quieter participants, paraphrasing interventions, or suppressing repetition. Each design affects both the content and social experience of deliberation. A system that provides equivalent text space to every person can still produce unequal influence if it reformulates one speaker's contributions more persuasively or routes selected topics to the chair. Consequently, equality of turns, equality of acknowledgment, citation of distinct proposals, and downstream adoption rates should be separately coded. Ethnographic attention to conversational repair can help establish why a nominally balanced transcript masks material differences in influence Tessler et al. (2024) Parisi et al. (2026).
Community decisions involve more than aggregating independent contributions. Discussion can create shared understanding, expose hidden information, revise preferences, and sometimes heighten conformity. Human–AI facilitation adds a procedural actor with a particular summarization policy. The task-specific findings of the 2024 meta-analysis by Vaccaro and colleagues warn against equating collaboration with a performance bonus. Effective evaluation must compare the group with a human-only procedure and, where meaningful, an AI-only benchmark, while measuring distributional and legitimacy outcomes that task-score studies generally omit. The purpose is to estimate complementarity for an identifiable collective practice rather than infer it from the presence of mixed agents Vaccaro et al. (2024) Gonzalez et al. (2026).