A comprehensive survey finds that the promise of large foundation models for human-AI collaboration depends less on model scale and more on human-centered design, preference-driven objective shaping, and governance; the literature is expanding quickly but remains non-systematic and highlights open challenges in bias, evaluation, and socio-economic effects.

Main Finding

Large Foundation Models (LFMs) substantially expand the potential of Human-AI (HAI) collaboration, but better models alone do not guarantee effective, safe, or equitable partnerships. Real-world HAI systems require careful human-centered design across data curation, objective shaping, interaction design, evaluation, and governance. Human roles are shifting—from mass annotators toward targeted teachers, evaluators, and overseers—while methods like RLHF/DPO and HITL active learning are central to aligning LFMs with human values. The survey emphasizes opportunities (productivity, richer multimodal collaboration, new applications) and risks (bias reinforcement, misaligned incentives, labor impacts, privacy, governance gaps), and identifies open technical and socio-economic research challenges.

Key Points

• Scope and approach

  • Survey of literature (2014–2024) on HAI collaboration, with attention to LFMs (LLMs, LVMs).
  • Non-systematic Google Scholar search using targeted keywords; includes published papers and arXiv preprints.
  • Authors used an LLM (ChatGPT-4) to assist in drafting and revision.

• Human-guided model development

  • Human-in-the-loop (HITL) and active learning: humans now act as targeted teachers for LFMs, focusing labeling and correction where models are uncertain or biased.
  • Objective shaping: preference-based methods (RLHF) and gradient-based alternatives (DPO, Constitutional AI) convert human judgments into model objectives; multimodal feedback is emerging.
  • Evaluation: human experts and end-users remain essential for measuring alignment, safety, trustworthiness, and usability.

• Design principles for HAI systems

  • Interaction and interface design: UI/UX, explainability, and feedback channels determine effective human-AI teaming.
  • Collaboration patterns & role allocation: research explores task decomposition, dynamic role switching, and allocation of responsibilities between humans and LFMs.
  • System integration and evaluation: end-to-end evaluation metrics and deployment strategies are needed to validate HAI effectiveness.

• Ethics, governance, and societal concerns

  • Fairness and bias: LFMs can perpetuate or amplify biases; bias-aware active learning and evaluation methods are open needs.
  • Labor & autonomy: HAI affects job content, autonomy, and worker well-being (augmentation vs. displacement trade-offs).
  • Privacy, security, and trust: data governance, adversarial risks, and accountability frameworks are required.
  • Policy and regulation: legal, regulatory, and governance frameworks lag technical advances.

• Applications

  • Broad sectors covered: healthcare, autonomous vehicles, surveillance/security, games, education, accessibility.
  • Domain-specific challenges (safety, liability, interpretability) are particularly acute in high-stakes settings.

• Open technical challenges

  • Scalable, faithful preference elicitation for RLHF; multimodal human feedback; bias-aware sampling in active learning; standardization of HAI evaluation metrics.
  • Socio-technical research: measuring real-world impacts, adoption barriers, and long-term equilibrium effects.

• Limitations of the survey

  • Non-systematic selection and reliance on visible literature; includes preprints; rapidly evolving field—findings are provisional.

Data & Methods

• Type of paper: literature survey (30 pages; arXiv preprint).
• Search procedure: Google Scholar searches covering 2014–2024 with keywords including “Human-AI”, “Human-AI collaboration”, “large models”, “foundation models”, “large language models”, and “large vision models”.
• Organization: papers were filtered and grouped into four thematic areas—(1) Human-guided AI development, (2) Design principles for HAI collaboration, (3) Ethics and governance, (4) Applications.
• Analytical framing:
  • Adopted a three-phase model (adapted from Maadi et al.): data control → model optimization → evaluation.
  • Synthesizes methods (HITL, active learning, RLHF/DPO, constitutional approaches), interface/role design literature, ethics and governance work, and applied case studies.
• Empirical content: survey compiles and synthesizes prior empirical and methodological work rather than presenting new field experiments or econometric analysis.
• Note on reproducibility: selection was not systematic; the survey maps visible research trends rather than providing a comprehensive meta-analysis.

Implications for AI Economics

• Task reallocation and labor market effects

  • Complementarity vs substitution: LFMs plus HAI design change task boundaries—routine/automatable subtasks are likely to be automated, while demand rises for tasks requiring oversight, judgment, labeling, and model steering (AI-literate, higher-skill roles).
  • Role shift: human workers move from bulk annotation to higher-value activities (curation, feedback, evaluation), altering skill demands and wage premiums.
  • Crowdwork & compensation design: platforms that monetize human feedback (for RLHF, evaluation) will shape bargaining power, compensation norms, and the microeconomics of data labor.

• Productivity, value capture, and industry structure

  • Productivity gains: properly designed HAI can raise productivity in many sectors (medical diagnostics, design, coding), but realized gains depend on integration, retraining, and complementary investments.
  • Value capture: platform owners and model proprietors may capture disproportionate returns (control over LFMs, fine-tuning datasets, deployment pipelines), potentially increasing market concentration.
  • Measurement challenges: traditional productivity statistics may understate or mis-attribute HAI-driven gains—new metrics needed to capture human-AI joint output and reallocation effects.

• Wages, inequality, and distributional outcomes

  • Skill-biased demand: increased returns to AI oversight and model engineering skills could widen wage dispersion absent compensating policies.
  • Displacement risk: some intermediate-skilled roles could shrink, pushing workers into lower-paid tasks or requiring transition support.

• Markets for data, feedback, and governance

  • Markets and incentives for human feedback: RLHF and other preference-data markets create new economic goods (quality-rated human judgments); designing incentives (monetary, reputation, regulatory) affects data quality and fairness.
  • Data ownership and privacy externalities: governance of training/evaluation data affects firms’ costs, consumer welfare, and potential liabilities—policy choices alter competitive dynamics.

• Policy and regulation implications

  • Labor policy: policies for retraining, bargaining rights in platform-mediated feedback markets, and safety nets will shape transition costs.
  • Competition policy: concentration around LFMs and feedback/data ecosystems calls for antitrust scrutiny and possibly data-portability rules.
  • Standards and certification: sector-specific certification for HAI systems (e.g., healthcare, autonomy) can influence adoption rates and liability allocation.
  • Measurement & monitoring: regulators need new tools (audits, disclosure regimes, standardized HAI evaluation) to oversee economic and societal impacts.

• Research agenda for economists

  • Causal studies: randomized trials or natural experiments evaluating productivity and employment effects of HAI deployments.
  • Microdata on feedback labor: datasets on who provides RLHF/evaluation labels, compensation structures, and quality outcomes.
  • Equilibrium models: task-allocation models that incorporate human feedback costs, model-improvement dynamics, and monopoly rents.
  • Metrics development: frameworks to measure joint human-AI output, complementarity indices, and welfare impacts.

Summary implication: LFMs magnify opportunities for productivity and new services, but they also reshape labor demand, value capture, and regulatory needs. Economic outcomes will hinge on how HAI systems are designed, who supplies and gets paid for human feedback, and the policy frameworks that govern data, labor, and market power.