A Survey on Human-AI Collaboration with Large Foundation Models

In plain words

Large AI models, such as chatbots, are capable, but they work best when people stay involved. This survey looks at how people and these models can work together well. It covers four areas: how people shape a model while it is being built, by choosing its data, giving feedback on its answers and judging its results; how the person and the model share roles once the model is in use; the ethical and governance questions such partnerships raise, such as fairness, privacy and who is responsible when things go wrong; and how this plays out in fields such as healthcare, driving, security, games, education and accessibility.

The survey's main message is that a good partnership comes from careful, human-centred design, not from a stronger model alone. It ends with open problems. It reviews other people's work and does not run experiments of its own.

TL;DR

A survey, not a method. It maps human-AI collaboration with large foundation models (LFMs) into a build-time loop (data curation, preference-based optimisation, human evaluation), five run-time collaboration patterns, ethics and governance concerns, and six application areas. It proposes no algorithm, reports no experiments of its own, and gives no thresholds or metrics to implement. Use it as a map of the field and of its open problems.

Problem

The survey argues that purely automated AI has known limits, and that safety, fairness and control of large models need human involvement §1. It also argues that a model's performance alone does not make a human-AI team work: interfaces, role allocation, training and evaluation decide the outcome §3. Its stated gap is that model development has moved faster than governance and than the methods for judging collaboration quality §4 §6.

The survey's working definition is a cooperative arrangement where people and AI cover for each other's weak spots, share the flow of information and, at times, of control, and work toward one target §1.

Method

The survey is a literature review. It is organised as follows §1 fig. 3:

1. Build-time human guidance (§2). A three-phase view adapted from Maadi et al.: data control, then model optimisation, then evaluation, with feedback from evaluation returning to data control §2 fig. 4. The three sub-areas are human-in-the-loop data curation and active learning (§2.1) fig. 5, human-guided objective shaping with RLHF and DPO (§2.2) fig. 6, and human evaluation and alignment metrics (§2.3).
2. Run-time design principles (§3). Interaction and interface design (§3.1), collaboration patterns and role allocation (§3.2), and system integration and evaluation (§3.3). §3.2 names five patterns, summarised in the table below tab. 2 fig. 7.
3. Ethics, societal and governance aspects (§4). Fairness, autonomy and well-being, labour dynamics, data privacy and security, trust and shared accountability, and policy and regulation fig. 8.
4. Applications (§5). Healthcare (§5.1), autonomous vehicles (§5.2), surveillance and security (§5.3), games (§5.4), education (§5.5) and accessibility (§5.6) fig. 9.
5. Open challenges (§6). Scalable and diverse human guidance; fluid but controllable interaction; accountability and governance; contextual grounding in high-stakes domains.

The five run-time patterns in §3.2, as the survey tabulates them tab. 2:

The survey's build-time loop and its active-learning round, as described in the text. It gives no stopping rule for the overall loop and no threshold for the active-learning round §2 §2.1.

# Build-time loop: data control, optimization, evaluation  [§2] [fig. 4]
# Input:  foundation model; data pool; human experts
# Output: model after the last iteration

def build_time_loop(model, data_pool, experts):
    # Iteration count is not stated  [§2]
    for iteration in range(num_iterations):
        # Humans curate, label or filter examples  [§2.1]
        training_data = curate(data_pool, experts)

        # Fit the model, then shape its objective with preferences  [§2.2]
        model = fine_tune(model, training_data)
        model = optimize_with_preferences(model, experts)

        # Judge outputs for quality, trust, safety and fairness  [§2.3]
        # Findings steer the next round of data control  [§2.3]
        findings = evaluate(model, experts)
        data_pool = update_with_findings(data_pool, findings)

    return model

# Active-learning round  [§2.1] [fig. 5]
# Input:  pre-trained model; unlabeled pool; labeled set; threshold
# Output: model retrained on human-labeled uncertain samples
# Threshold value is not stated  [§2.1]

def active_learning_round(model, unlabeled, labeled, threshold):
    # Predict a label and a confidence for each unlabeled sample  [§2.1]
    predictions = model.predict_with_confidence(unlabeled)

    # Send low-confidence samples to human annotators  [§2.1]
    uncertain = [s for s, conf in predictions if conf < threshold]
    labeled = labeled + annotate(uncertain)

    # Retrain on the enlarged labeled set; repeat until confident  [§2.1]
    return retrain(model, labeled)

Hyperparameters

Not applicable. The survey specifies no training, model or evaluation settings. Its only numeric content is in the figures and results it quotes (see Results). The active-learning threshold is not stated §2.1; the number of loop iterations is not stated §2; no learning rate, budget or rubric size is given anywhere in the text.

Results

Not applicable as results of this survey: it runs no experiments. The numbers below are figures the survey quotes from cited studies. They are author-reported by those studies, and this card has not checked them against the originals.

Paper vs code

No official code. The survey is a literature review; it has no method or model to implement. The LaTeX source, bibliography and arXiv metadata contain no code repository for this work. No GitHub search was run for it.

Pitfalls

• Quoted numbers are not the survey's. Every figure in Results comes from a cited paper that this card has not read. Check the original before citing a number.
• Search and selection are thin. Articles were collected from Google Scholar for 2014 to 2025 using keyword searches, then sorted into themes by the authors. The 2025 to early 2026 agent literature was added as a "targeted update" with no stated selection rule §1, Scope of the Survey. No screening counts or inclusion criteria are given.
• Disclosed AI assistance. The authors state that ChatGPT gave feedback on language flow and citation placement §1, Scope of the Survey. Treat the text as human-reviewed, but that is the authors' own account.
• Organising labels are not operational definitions. "Supervisory orchestration", "control handoff" and "mixed-initiative" are described in prose. The safeguards in the table are suggestions the survey does not test tab. 2 §3.2.
• No metrics, despite calls for them. The survey asks for measures of trust, workload, resilience and collaboration quality, but names no metric, scale or threshold §3.3 §6.
• Patterns overlap. The survey acknowledges that its categories are not mutually exclusive: studies may belong to several themes, and the scope figure shows overlaps fig. 2.
• Figure and table numbers are from source order. The cards cite figures and tables as numbered in the LaTeX source (for example fig. 4 for the three-phase view). The compiled PDF was not read; check the numbering there before quoting it externally.
• Commented-out drafts exist in the source. Parts of §3.2 are commented out in the source (for example a draft on supervisory orchestration). This card uses only the active text.

Lineage

• Builds on: the three-phase view of model development (data control, model optimisation, evaluation) from Maadi et al., which the survey adapts §2 fig. 4.
• Compared against: not stated; the survey does not benchmark other surveys.
• Superseded by: not known from content in this wiki.