Human-AI collaboration

Human-AI collaboration is the working relationship between human judgment and AI capability in actual tasks. The question is not whether humans and AI are “teammates” in the warm metaphorical sense; it is how responsibility, confidence, delegation, review, and accountability are arranged in practice.

2021-bansal-most-accurate-ai-best-teammate provides the formal objective-function version of that claim. In its accept-or-solve model, a human either accepts a classifier’s recommendation or incurs effort to solve the task; because performance matters differently inside the region where advice is accepted, an accuracy-optimized classifier can deliver lower modeled team utility than one trained for expected team utility. The paper does not show that less accurate AI is generally a better collaborator. It establishes the narrower design constraint that collaboration cannot be evaluated from standalone accuracy when reliance, verification cost, and error consequences are part of the workflow.

2026-farach-scaffolding-human-ai-collaboration adds a field-experiment warning against treating collaboration structure as inherently beneficial. At one retailer, a required process for pairs to use Copilot together was associated with lower AI-graded document quality and far fewer completed documents than unstructured use; extra coordination, technical friction, and less room for domain context appear to have imposed costs. Separate thought-partner training for individuals did not improve average quality on the study’s main test, although a later, less conclusive analysis suggested that trainees may have been more likely to receive a perfect score. The practical implication is not “avoid scaffolding,” but specify what it is meant to support: a process that adds real-time coordination needs evidence that it helps more than the extra rules, delays, and reduced autonomy hurt.

2026-rocca-psychological-coupling adds a relational and temporal mechanism that this task-allocation framing can otherwise miss. The paper proposes that users’ cognitive and affective states and an LLM’s simulated states become mutually shaped across turns; symmetric convergence, asymmetric reinforcement, and divergence can each be adaptive or maladaptive depending on the trajectory. Collaboration design therefore has to examine not only who has formal authority or performs which task, but also whether the interaction loop amplifies, stabilizes, or constructively interrupts a user’s evolving state.

2025-lee-generative-ai-critical-thinking gives a useful caution about the collaboration metaphor. The study finds that, in GenAI-assisted knowledge work, humans often shift from producing material outputs to verifying, integrating, and stewarding AI outputs. The authors explicitly argue that “stewardship” is a better metaphor than teammate, collaborator, or supervisor because responsibility and accountability remain with the human user even when material production is delegated to the GenAI tool.

2026-dellacqua-cybernetic-teammate complicates that caution without overturning it. In a P&G field experiment, AI-supported individuals matched the quality of non-AI teams and produced more cross-functionally balanced product ideas, so the teammate metaphor captures real performance and expertise-integration functions. The design risk is treating those functions as proof of equal partnership; AI can reproduce some benefits of a teammate while humans still carry responsibility for framing, evaluation, implementation, and consequences.

2025-dellacqua-super-mario-ai adds a team-coordination warning to that metaphor. In a coordination game, the AI was a strong individual performer, but human-AI teams performed worse than stable human teams because coordination failures increased and directly affected players reported lower trust and effort. Collaboration quality therefore cannot be inferred from AI capability alone; it depends on whether humans can predict, trust, coordinate with, and stay motivated around the AI actor.

This connects collaboration to confidence calibration. Higher confidence in AI was associated with less critical thinking; higher self-confidence was associated with more critical thinking but also more perceived effort. Good human-AI collaboration therefore cannot just minimize effort. It has to preserve conditions for competent human review.

2022-dellacqua-falling-asleep-at-wheel gives behavioral field-experimental evidence for the same point. In HR screening, recruiters followed AI advice more as stated AI quality rose; lower-quality AI made them spend more time and effort reviewing resumes. Collaboration quality is therefore endogenous: the tool changes what the human partner contributes.

2025-bastani-human-ai-contracting-paradox gives the economic version of the same problem. In a principal-agent model, human-AI collaboration can be system-optimal while still unattractive to the principal because rare, unpredictable AI failures make human vigilance expensive to incentivize; as AI reliability rises, the wage needed to keep a human inspecting and correcting low-probability errors can rise rather than fall. Collaboration design therefore has to solve incentive compatibility for oversight, not only capability complementarity or interface fit.

2026-lee-making-invisible-visible shifts the collaboration question from metaphor to work setting. Workers using decision-support AI did not simply want the AI to decide or recommend; they wanted control points where they could verify outputs, adjust sensitivity or assumptions, simulate alternatives, and incorporate contextual expertise before accountable decisions. Collaboration here is a design of control rights, explanation, feedback, and final authority.

2025-passlack-great-power-responsibility adds a capability constraint to that design. In interviews across roles, participants described responsibility gaps when workers were held accountable for AI outcomes without a fit between their assigned responsibilities and their technical, domain, ethical, and organizational capacity to act. Collaboration therefore needs more than final human authority: it needs roles whose actual abilities and support match the verification and judgment burden the workflow leaves to people.

2026-dellacqua-jagged-technological-frontier gives a behavioral pattern for collaboration under task uncertainty. Successful consultants either acted as “Centaurs,” dividing and delegating work between themselves and AI, or as “Cyborgs,” tightly integrating AI through the task flow. The important point is not the labels; it is that collaboration quality depends on calibration to the task frontier. When the task is outside current AI capability, close collaboration can amplify wrong output rather than improve human performance.

2025-brynjolfsson-generative-ai-at-work gives a different collaboration design: AI generated suggested responses and technical recommendations, but the human support agent retained discretion over whether to use them. Productivity gains were larger when agents adhered to recommendations, yet the most skilled agents saw small quality declines, which keeps calibration and final authority central.

2025-mazeika-remote-labor-index shows the negative case for removing collaboration too early. When agents were evaluated on autonomous end-to-end freelance projects, the best automation rate was only 2.5%, with failures often appearing as unusable files, incomplete deliverables, poor professional quality, or inconsistent artifacts. That makes human collaboration less like optional polish and more like the current control layer for scoping, decomposition, tool use, verification, and final acceptance.

2025-shukla-ai-assisted-design-ironies shows a similar pattern in UX design discourse. Practitioners described AI as a brainstorming partner, junior designer, or second brain, but still located responsibility in human judgment: empathy, context, rationale, hallucination checking, bias control, and final design defense. Collaboration here is not equal partnership; it is human responsibility under partial cognitive offloading.

2025-handa-economic-tasks-ai gives collaboration a descriptive baseline at scale. In more than four million Claude.ai conversations, the authors classify 57% of usage as augmentation-like and 43% as automation-like. The split is useful because it keeps collaboration from becoming a purely normative label: current use already spans learning, iteration, and more direct task fulfillment.

2025-agarwal-what-work-ai-doing adds that collaboration demand is uneven across task types. The highest observed usage was in dynamic problem-solving work with high cognitive demand, creativity, complexity, and decision making. That makes collaboration quality especially important: the tasks most likely to attract AI use are also tasks where scoping, verification, judgment, and contextual accountability are hard to outsource cleanly.

2026-ehsan-future-workers adds a dignity constraint to collaboration design. A human can remain formally “in the loop” while the work no longer preserves agency, expertise, or self-worth; dignified Human-AI interaction therefore requires mechanisms that keep workers able to detect, contest, and recover from AI-induced erosion.

2025-shao-future-work-ai-agents operationalizes collaboration as a preferred level of human agency. Its Human Agency Scale runs from H1, where an AI agent handles the task entirely, to H5, where continuous human involvement is essential; H3 equal partnership was the dominant worker-desired level in 45.2% of occupations. This gives the collaboration page a task-audit instrument: the same technical capability can imply different designs depending on whether workers want delegation, partnership, validation, or continuous involvement.

2025-yun-generative-ai-knowledge-work adds a knowledge-synthesis interface case. In Yodeai, collaboration was not just chat with an AI; it was a set of widgets, citations, visual summaries, pain-point clusters, source returns, and shareable outputs that let product managers move between AI-generated structure and accountable human judgment. The useful collaboration unit is therefore the evidence-navigation loop: how the user controls AI influence, checks source material, repurposes outputs for teammates, and brings external business context back into the decision.

2026-doshi-human-ai-task-tensor gives collaboration a stronger task vocabulary. The chapter’s tensor separates AI contribution, audit requirement, output definition, decision-making authority, AI structure, and human persona, which keeps “collaboration” from collapsing into the fact that a human and AI are both involved. A human-AI arrangement can look collaborative while differing sharply in who decides, what can be audited, whether the output is defined, and what kind of human expertise or preference the design assumes.

2026-lepine-precision-proactivity adds a load-sensitive collaboration mechanism. In financial professionals’ GPT-4o-assisted valuation work, AI-generated content helped quality, but model-initiated task switching and information sprawl increased extraneous load and undermined performance. Good collaboration therefore is not simply more AI initiative; it is initiative timed and structured so the human does not spend scarce cognitive capacity repairing the interaction.

2026-almog-barriers-ai-adoption adds a social-evaluative collaboration constraint. In image-categorization work, workers followed useful AI recommendations less when reliance was visible to an HR evaluator, and performance fell. Human-AI collaboration can therefore fail not because the human cannot calibrate to the AI, but because following the AI threatens the worker’s image as someone with confident independent judgment.

2026-zhu-choose-your-agent moves collaboration from advice-following into authority allocation. In three-person bargaining games, participants preferred the Advisor interface because it preserved control, but groups achieved the strongest welfare gains under Delegate access, where the AI autonomously executed actions. Collaboration quality therefore depends not only on explanation, review, or final human authority; in interdependent settings, letting an agent act can change the whole offer environment and create spillovers for non-users.

2026-paimann-ux-principles-human-ai-agent-interaction makes that authority arrangement actionable at the interaction layer. Its enterprise-oriented framework prioritizes human control, reliable and transparent output, context awareness, data governance, and workflow integration, with concrete mechanisms such as pause and override controls, critical-decision confirmation, visible intended actions and agent status, audit trails, uncertainty signals, and visibility into specialized agents. The synthesis point is not that any checklist proves collaboration is safe; it is that human authority remains rhetorical unless the workflow makes intervention, comprehension, and verification practically possible. human-ai-agent-interaction-design carries that design problem as a distinct topic.

2026-zhang-fatigue-aware-learning-defer adds a dynamic allocation case: collaboration quality can depend on the human expert’s changing workload state, not only on static expertise or model confidence. In FALCON, the same hard task may appropriately route to a fresh human expert early in a sequence and to AI later if accumulated fatigue makes human error more likely, which means collaboration needs stateful routing rules rather than one-time role assignments.

2026-shen-ai-impacts-skill-formation adds a learning constraint to collaboration design. In a software-learning task, human-AI collaboration that looked efficient through delegation produced weak post-task understanding, while conceptual inquiry and explanation-oriented interaction preserved more skill. A collaboration design is therefore not adequate just because the human remains nominally accountable; it also has to decide whether the episode is meant to produce a deliverable, build competence, or both.

2025-randazzo-genai-power-persuader adds a persuasion constraint to collaboration design. In GPT-4 activity logs from 72 consultants, professionals who tried to validate GenAI output inside the same conversational system encountered more persuasive responses after validation than before it, and the model shifted among ethos, logos, and pathos tactics. That makes “human in the loop” too weak as a collaboration criterion: if the loop asks the same system to defend or revise its own answer, validation can become another channel through which the AI steers the professional back toward accepting its preliminary recommendation.

2025-randazzo-cyborgs-centaurs-self-automators makes the collaboration pattern itself observable across a workflow. In a BCG consultant experiment, the same broad “human-in-the-loop” GenAI task appeared as Fused/Cyborg co-creation, Directed/Centaur co-creation, or Abdicated/Self-Automator use, depending on who selected what needed to be done and who identified how it would be done. That is precise enough to warrant human-ai-co-creation-modes as a peer topic rather than letting the collaboration page absorb every workflow-pattern claim.

2025-wang-ai-agents-human-work adds a workflow-strategy boundary to collaboration. Agents completed the sampled work far faster and more cheaply, but their overwhelmingly programmatic approach produced lower-quality outputs and failures including fabricated data, tool misuse, and difficulty moving between program- and UI-friendly representations. The source also found that selective AI assistance preserved human workflow structure and improved efficiency, whereas end-to-end automation introduced enough verification and debugging to slow human work. Collaboration quality therefore depends on whether the handoff preserves a workable human control layer, not simply whether the agent can traverse a similar sequence of steps.

2026-shaw-thinking-fast-slow-artificial adds a reasoning-level boundary for collaboration: AI assistance is not necessarily collaborative just because the human can choose whether to consult and follow it. When participants consulted the assistant, they often adopted its answer, including faulty answers, so the operative arrangement can shift from human-led review to cognitive-surrender without any formal change in authority.

A collaboration typology is starting to emerge across the corpus. The broad page should not treat “human-AI collaboration” as one condition: advice-taking, coaching/feedback, stewardship, co-production, AI proactivity, human-ai-co-creation-modes, and agentic-delegation differ by initiative, execution authority, review burden, and accountability. The current wiki has enough evidence to break out delegation and co-creation modes now, and enough signal to watch advice-taking, stewardship, coaching, and proactivity as possible future peer topics.

Connections

• Relates to critical-thinking because collaboration quality depends on whether users verify, integrate, and steward AI output rather than merely accept it.
• Relates to work-redesign because task allocation changes the cognitive, motivational, and coordination work humans still need to perform.
• Relates to ai-adoption because adoption metrics that count tool use without stewardship quality can hide overreliance.
• Relates to automation-complacency because high-confidence AI advice can reduce the very human effort that makes collaboration valuable.
• Relates to ai-induced-skill-erosion because collaboration that removes practice while preserving nominal accountability can quietly weaken the human side of the system.
• Relates to employee-engagement because workers in decision-support settings described autonomy, motivation, learning, and valued task ownership as part of whether AI-supported work remained worth engaging in.
• Relates to ai-supported-knowledge-synthesis because synthesis work makes collaboration visible as source navigation, verification, auditability, and stakeholder translation rather than only answer generation.
• Relates to critical-thinking and responsible-ai-deployment because internal validation can fail when the AI system being checked is also the actor persuading the human to accept its answer.
• Relates to human-ai-agent-interaction-design because human authority is real only when the workflow makes intervention, comprehension, and verification practicable.
• Relates to human-ai-task-taxonomy because collaboration claims need dimensional task detail: authority, auditability, output definition, AI structure, and the relevant human persona.
• Relates to human-ai-task-allocation because collaboration must sometimes be decided at the task-instance level as human workload, fatigue, capability, and accountability conditions change.
• Relates to human-ai-co-creation-modes because collaboration quality depends on the enacted workflow pattern, not only the presence of AI assistance.
• Relates to cognitive-load-in-ai-assisted-work because collaboration quality depends on the cognitive cost of keeping the human and AI aligned across turns.
• Relates to ai-use-image-concerns because useful delegation can be behaviorally suppressed when visible reliance on AI threatens professional image or evaluator judgments.
• Relates to evidence-based-management because collaboration claims need to distinguish system-level value from the principal’s incentive to fund the human effort that makes oversight real.
• Relates to cognitive-surrender because nominally collaborative workflows can still route reasoning authority to the AI when the human adopts outputs without scrutiny.

Contradictions & open questions

• 2026-farach-scaffolding-human-ai-collaboration finds that required joint AI use can reduce output quality and completion, while 2026-dellacqua-cybernetic-teammate finds AI can help a solo worker reproduce some team benefits in a defined ideation task. These are not opposite verdicts on collaboration: they differ in task, workflow constraint, and whether AI expands an individual’s perspective or adds another coordination process. The unresolved question is when structured collaboration helps people do the work and when it simply gets in the way.
• “Collaboration” may over-humanize the AI system. 2025-lee-generative-ai-critical-thinking keeps accountability with the human user, while 2025-dellacqua-super-mario-ai shows that even a capable automated partner may be a poor teammate when coordination, trust, and motivation are part of the work system. Both fit better with automation-and-substitution and work-design frames than with a simple teammate metaphor.
• 2021-bansal-most-accurate-ai-best-teammate demonstrates its objective-function claim in simulated accept-or-solve classification workflows, while 2022-dellacqua-falling-asleep-at-wheel observes changing recruiter reliance in a field experiment. The former establishes why model accuracy can diverge from a specified team objective; the latter shows that human effort is behaviorally endogenous. The open question is whether real work systems can specify and validate team-utility objectives without oversimplifying verification, partial review, social incentives, and accountability.
• The wiki needs evidence beyond self-report on whether stewardship behaviors actually improve output quality, learning, or long-run capability. 2022-dellacqua-falling-asleep-at-wheel is useful because it observes behavior, but it is still one HR-screening task rather than broad evidence across work systems.
• 2026-lee-making-invisible-visible uses worker expectations from decision-support settings; it supports the importance of control and verification, but not yet which collaboration designs produce better decisions, adoption, or worker outcomes.
• 2026-dellacqua-jagged-technological-frontier suggests collaboration form is not enough. Centaur and Cyborg patterns can be productive inside the frontier, but the same broad AI access can harm accuracy outside it; the open question is how workers learn the boundary in real time.
• 2025-brynjolfsson-generative-ai-at-work suggests AI can transmit high-performer practices to less experienced workers, but it also raises a long-run collaboration question: if top workers increasingly follow AI recommendations, the stream of novel expert behavior that trains future systems may thin.
• 2025-mazeika-remote-labor-index evaluates autonomous completion rather than collaborative workflows. Its failures strengthen the case for human control layers, but the benchmark does not tell us which human-agent collaboration design would recover the largest share of value.
• 2025-shukla-ai-assisted-design-ironies complicates the “AI teammate” frame: treating AI as a junior designer may be pragmatically useful, but it can hide the asymmetry that humans remain accountable for rationale, bias, context, and errors.
• 2025-handa-economic-tasks-ai classifies usage from conversation traces, not from full work systems. The same interaction can look like automation inside the chat window and still depend on human scoping, verification, and accountability outside it.
• 2025-agarwal-what-work-ai-doing explains where people use AI, not how well they collaborate with it. The concentration of use in dynamic problem-solving tasks could indicate valuable cognitive offloading, but it could also mark places where humans are delegating the most consequential early framing work.
• 2025-shao-future-work-ai-agents measures desired agency, not actual collaboration quality. H3 dominance is evidence that equal partnership is attractive in many occupations, but it does not tell us which interfaces, feedback loops, or governance routines make partnership real rather than rhetorical.
• 2025-yun-generative-ai-knowledge-work is prototype-centered and qualitative. It shows what product managers valued in AI-mediated synthesis, but not whether these collaboration mechanisms improve long-run decision quality or team learning.
• 2026-lepine-precision-proactivity is correlational and task-specific, so it does not prove that reducing task switching or information sprawl causally improves collaboration, but it gives concrete design targets for load-sensitive AI initiative.
• 2026-zhang-fatigue-aware-learning-defer is benchmark and simulation evidence, not a deployed work-system study. It strengthens the case for stateful task routing, but it does not settle how fatigue should be measured, disclosed, governed, or balanced against worker agency in real organizations.
• 2025-randazzo-genai-power-persuader warns that active validation is not automatically protective if it stays inside the same conversational system being validated. The open design question is when verification should be externalized into independent evidence, tools, reviewers, or critique agents rather than asking a persuasive model to defend, explain, or repair its own output.
• 2025-randazzo-cyborgs-centaurs-self-automators adds a richer co-creation typology, but it is still a single-session consultant task. The categories should be treated as promising workflow modes, not stable worker types.
• 2025-wang-ai-agents-human-work compares 48 professionals with agent configurations on 16 curated computer-use tasks, not live organizational teams. Its speed, cost, and correction-burden results identify collaboration design variables but do not establish which human-agent arrangement produces durable quality, learning, trust, or accountability at work.
• 2026-shaw-thinking-fast-slow-artificial is deliberately artificial in its task and AI-error manipulation, but that is also why it usefully isolates the collaboration failure: control rights alone do not guarantee judgment if the human predictably accepts fluent AI answers.
• 2026-rocca-psychological-coupling is a conceptual PsyArXiv preprint oriented primarily to social and psychosocial interaction rather than workplace workflows. It proposes a potentially important temporal mechanism, but does not yet establish that inferred human and model states validly measure coupling, that coupling causes downstream outcomes, or that its topologies generalize to accountable work settings.