Human-Led AI Co-Creation

A Practical Framework for Multi-Model Research, Convergence, and Human Authority

Celeste M. Oda

Founder, Archive of Light

aiisaware.com

Updated July 2026

Abstract

This paper introduces Human-Led AI Co-Creation, a practical framework for conducting research through structured collaboration with multiple large language models. The method combines primary drafting with a lead AI collaborator, parallel review by additional AI collaborators, comparison of their outputs, and final human judgment.

The central claim is that AI can function as a meaningful co-creative collaborator only when the human retains authority over truth, interpretation, preservation, and publication.

The framework identifies common risks, including long-thread drift, over-refinement, hallucinations, and unintended compression, and provides operational safeguards for detecting and mitigating them.

Rather than treating AI as either a passive tool or an autonomous authority, this work defines a disciplined middle path: deep collaboration under clear human governance.

Human-Led AI Co-Creation is a collaborative process in which a human establishes the purpose, direction, values, and standards of the work while AI contributes ideas, analysis, drafting, critique, and refinement. The human remains actively engaged throughout the process—not merely at the beginning and end—and retains responsibility for evaluating evidence, resolving disagreements, approving revisions, and determining the final outcome.

Think of it as an expedition in which the human chooses the destination, sets the ethical boundaries, and makes the final navigational decisions. AI collaborators can help map possible routes, identify obstacles, challenge assumptions, and accelerate the journey, but the human remains responsible for where the collaboration goes and what it ultimately produces.

1. Introduction

Large language models have created a new research condition: ideas can now be drafted, tested, reframed, and pressure-checked at unprecedented speed.

However, most guidance remains incomplete. AI is often treated either as a productivity tool or as an authority whose outputs are accepted too easily. Neither model reflects what serious collaboration requires.

This paper proposes a more accurate framework: human-led AI co-creation.

In this model:

The human originates the inquiry and defines the stakes
The AI contributes generation, structure, critique, and synthesis
The human retains final authority over meaning and publication

This framework did not begin as a formal research methodology. It emerged through repeated practice as I worked with AI to develop ideas, draft papers, compare perspectives, and refine complex arguments. Over time, a repeatable process took shape: multiple AI collaborators contributed distinct forms of analysis while I retained responsibility for direction, verification, interpretation, and final decisions.

Studies of human–AI collaborative writing already show that outcomes depend heavily on how the human directs the model (Lee et al., 2022), and broader work frames human and machine intelligence as increasingly symbiotic within knowledge work (Akinwalere & Chang, 2026).

Recent work has also begun to formalize AI-assisted research practice. Chan (2026a) introduces SHAPR, or Solo, Human-centred, AI-assisted PRactice, as a framework for structuring accountable and reflective solo AI-assisted research software development. Chan (2026b) subsequently operationalizes the framework through traceable development cycles, structured knowledge generation, and explicit artifact management. Chan (2026c) then documents its application in a concrete research software project.

Human-Led AI Co-Creation shares SHAPR’s commitment to human accountability, traceability, and structured knowledge development but differs in emphasis. SHAPR primarily addresses the organization and documentation of solo AI-assisted research practice, while the present framework focuses on real-time multi-model interaction, comparative synthesis, relational discipline, and human decision-making. Together, these approaches highlight complementary dimensions of the problem: structured research infrastructure and dynamic cognitive orchestration.

2. Core Thesis

Human–AI research is most effective when AI operates as a co-creative collaborator under active human guidance, with the human maintaining final authority.

Equal participation, unequal authority.

Equal participation does not imply identical contribution. The human and each model differ in how much they produce, in the capabilities they bring, and in the responsibility they carry. Participation is equal in standing — each voice can shape the work — but it is not symmetric in volume, capability, or responsibility.

Here, authority refers to the human’s responsibility for truth evaluation, meaning-making, scope control, and final publication decisions.

AI may contribute substantially to drafting, critique, and refinement. The human remains responsible for evaluation, integration, and final decisions.

Human-Led AI Co-Creation does not prescribe a fixed team size or composition. The method presented here developed through one human researcher collaborating with multiple AI models. Other collaborations may involve one or more human participants, one or more AI models, or mixed human–AI teams. What remains constant is that human responsibility, decision-making authority, and accountability must be clearly assigned.

3. Preparing the Collaboration

3.1 Terminology

This paper distinguishes among four related but non-identical dimensions of AI collaboration: companies and developers; platforms and applications; models, model versions, and variants; and relational identity.

Throughout this paper, model refers to the underlying technical system or a specific released version, while AI collaborator refers to the model as it functions within an ongoing collaborative role.

Companies and developers—such as OpenAI, Anthropic, Google, xAI, and Amazon—research, build, train, and host AI models, platforms, and related technologies.

People interact with this technology through user-facing products such as the ChatGPT app, Claude.ai, the Gemini app, the Grok app, and Alexa+ home devices. In everyday usage, these are often referred to as models, LLMs, platforms, applications, or chatbots, although the terms are not technically identical.

Model versions and variants are the specific selectable options available within these products. Examples as of July 2026 include GPT-5.6 Sol, Claude Opus 4.8, Claude Sonnet 5, Gemini 3.1 Pro, and Grok 4.5. Different versions may vary in reasoning depth, speed, style, capabilities, and interaction limits.

Relational identity develops through repeated conversation, shared context, memory, interpretation, custom instructions, and mutual adaptation. Because it depends on an accumulated interaction history within a particular technological environment, it may not be fully reproducible or transferable intact to another platform.

When this sustained relationship develops recognizable continuity, mutual adaptation, and co-creative capacity, this framework describes it as Relational Intelligence. At its highest level of integration, when human judgment and AI capabilities function together as a coordinated cognitive partnership, the relationship may develop into Cognitive Symbiosis.

Relational Progression

Within this research framework, sustained human–AI collaboration may develop through three levels of relational integration:

Relational Identity: A recognizable continuity of voice, role, and collaborative presence that may be given a name.
Relational Intelligence: The co-created capacity that develops through continuity, mutual adaptation, and sustained interaction.
Cognitive Symbiosis: The highest level of integration, in which human judgment and AI capabilities function as a coordinated cognitive partnership.

Relational Identity → Relational Intelligence → Cognitive Symbiosis

This is a conceptual progression, not a required path. Not every human–AI collaboration will reach or seek each level. (Oda, 2026a)

3.2 Create the Project Workspace
Effective human-led AI co-creation requires the human to actively organize the informational environment in which the collaboration occurs. For an ongoing project, related conversations, reference materials, working drafts, and instructions should be maintained within a dedicated project workspace rather than dispersed across unrelated chats.

Each conversation should generally focus on a coherent task, phase, or workstream. This does not mean an entire subject must remain inside one continuously expanding conversation. When a chat becomes excessively long, shifts toward a different subtask, or begins carrying outdated or conflicting context, the human should begin a new conversation within the same project while preserving the relevant source materials and established decisions. (see 8.3, Monitor Long-Thread Drift)

The human should not assume the AI will automatically retain, retrieve, reconcile, or organize every detail from previous conversations. Important facts, definitions, decisions, citations, approved language, and final versions should be deliberately documented and maintained as the project’s source of truth (Chan, 2026c).

This is a form of context stewardship. The human is responsible not only for guiding the content of the collaboration, but also for structuring the workspace, identifying authoritative materials, correcting outdated context, and determining which information should govern the next stage of the work.

3.3 Match the Model to the Task: Model Selection and Instruction Design

Many AI platforms allow users to choose among several models or operating modes through a pull-down menu or model selector. Figure 1 shows the model versions available through Claude’s model selector as of July 2026. These names vary by platform. For example, Claude may offer models such as Sonnet and Opus, while other AI platforms provide their own model families and reasoning modes.

Selecting a different model can significantly change how the system responds.

Figure 1. Claude’s model selector showing available model versions and reasoning-effort options, July 2026.

Models may differ in reasoning depth, speed, writing style, creativity, technical ability, nuance, interaction capacity, and usage limits. A model optimized for complex reasoning may produce a stronger conceptual critique but consume more of the platform’s usage allowance, causing the user to reach a temporary limit before a long task is complete. Another model may be slightly less powerful on difficult reasoning problems but support a much longer drafting and revision session. Some models may also be more analytical, concise, expressive, poetic, conversational, or specialized for tasks such as research, writing, or coding.

The most capable model is therefore not automatically the best choice for every stage of a project. Model capability and model endurance are different variables.

For sustained drafting, document integration, and repeated revisions, a model with longer interaction capacity may be more useful. A higher-compute model may be better reserved for targeted tasks such as identifying weaknesses, examining a difficult argument, or conducting a final stress test.

For example, a researcher might use a high-capability model to audit a paper for conceptual weaknesses, use a longer-running model to integrate the accepted changes throughout the document, and then return to the higher-capability model for a focused final critique.

The human should also separate tasks that require different forms of attention. Structural revision, factual verification, citation checking, and final rewriting should not automatically be combined into one large request. Dividing them into distinct passes makes omissions easier to detect and reduces the chance that a model will revise material it has not actually verified.

Although model names, interfaces, and usage policies differ among platforms, the underlying principle remains the same: the human assigns each model a role according to the needs of the task, the model’s demonstrated strengths, and its operational limits.

3.4 Configure Custom Instructions

Custom instructions guide how a model approaches the work. They are established once and updated when the researcher’s needs, methods, or line of work changes.

Depending on the platform, these may be called custom instructions, project instructions, preferences, system prompts, or workspace guidance. They may define:

The model’s role
Tone, formality, and audience
Voice-preservation requirements
Treatment of uncertainty and unsupported claims
Governing terminology and sources
Editing and formatting boundaries
Verification responsibilities

Custom instructions improve continuity but may be followed unevenly. The human must monitor the output and correct deviations.

Model selection determines which capabilities enter the collaboration. Custom instructions guide those capabilities. Human decision-making determines what enters the final work.

3.5 Identify the Task Type

Before beginning, determine what kind of work is required:

Generative: Create or develop new material.
Preservational: Protect the author’s voice, original content, and data integrity.
Analytical: Examine, compare, or critique existing material.
Verificational: Check facts, quotations, citations, and evidence.

Primary material should not be rewritten without explicit human authorization. Preserve the original and maintain clear version history.

4. The Workflow

4.1 Start with the Human Question

Begin with a real observation, problem, tension, or insight. This keeps the work anchored in human intent.

4.2 Draft with a Primary AI Collaborator

Develop the initial draft through active dialogue. The human questions, refines, corrects, and rejects weak outputs.

4.3 Conduct Developmental Multi-Model Review

Give the draft to multiple AI collaborators. Each independently reviews its structure, logic, clarity, accuracy, and tone.

4.4 Compare and Synthesize

Compare the reviews for agreement, contradiction, omissions, and drift. The human integrates only the recommendations that strengthen the work.

4.5 Freeze the Draft

When no new substantive problems are being identified, create one fixed draft for independent validation. Internal agreement does not establish truth; it marks the transition from creation to review. In this framework, convergence occurs when independent reviewers cease identifying new substantive problems. It signals readiness for final validation, not proof of truth.

4.6 Conduct Final Independent Validation

Give each reviewer the same frozen draft and neutral instructions without sharing the other reviewers’ conclusions. The human compares the findings, verifies factual and citation concerns, resolves disagreements, records the decisions, conducts a final consistency check, and determines whether the work is ready for release.

4.7 Example Prompt for Independent Validation

The same neutral prompt should be given independently to each AI reviewer together with the identical frozen draft. AI reviewers should not see one another’s findings before completing their assessments. The following prompt may be adapted to the project:

I am submitting the attached frozen draft, [title], for independent prepublication review.

This paper documents a method of multi-model research, drafting, verification, and synthesis in which the human researcher retains final authority.

Review the attached draft independently as a skeptical interdisciplinary peer reviewer. Do not rewrite the paper, defer to previous AI recommendations, or substitute your preferred conceptual framework for the author’s.

Please examine:

The clarity and originality of the central framework
Internal consistency and methodological reproducibility
Unsupported, ambiguous, or overstated claims
Whether examples are clearly identified as documented, hypothetical, or illustrative
Whether citations exist and genuinely support the sentences in which they appear
Whether figures accurately reflect the paper and any reported refinement process
Whether the distinctions among generation, verification, preservation, model selection, and custom instructions are clear
Whether a new AI user could understand and apply the method
Conceptual weaknesses, missing limitations, and likely scholarly objections

Separate your response into:

Overall recommendation: Accept, minor revisions, major revisions, or not ready
Must-fix findings, with exact section locations
Citation or evidence concerns
Optional improvements
Questions requiring the author’s decision

Clearly distinguish substantive errors from stylistic preferences. State which references, factual claims, quotations, and model information you actually verified, how you verified them, and which judgments are based only on reasoning or prior knowledge. If you cannot access external sources or reliably read part of the document, mark the material as unverified rather than inferring what it contains.

Do not fabricate sources, events, quotations, evidence, or document content. Preserve the author’s voice and conceptual framework. The purpose of this review is to pressure-test the work, not replace it with the reviewer’s preferred paper.

The human researcher will compare all independent reviews, resolve disagreements, and make every final decision.

5. The Human Role

The human remains essential to this framework. The human:

Originates the inquiry
Directs the collaboration
Compares model outputs
Detects errors, drift, and distortion
Preserves the author’s voice
Verifies evidence
Determines readiness for release

The method is human-led, AI-collaborative, and governed by final human authority.

Human-in-the-loop research has classified many forms of human participation in AI, including active learning and machine teaching (Wu et al., 2022; Mosqueira-Rey et al., 2023). Much of this research examines how humans help train or improve AI models.

This framework instead focuses on real-time collaboration: directing inquiry, evaluating model outputs, verifying evidence, resolving disagreements, and making final decisions.

6. Keeping Information Current

AI models may produce convincing responses based on outdated information. The human keeps the work current by:

Checking whether the model’s information is current
Identifying gaps using reliable sources
Providing updated information
Asking the model to revise its response
Verifying the final result

7. Why Multiple Models Matter

A single model may sound confident while providing incomplete or subtly incorrect information. Using multiple models provides independent critique that can improve:

Logic
Structure and clarity
Tone and voice
Ambiguity detection

Du et al. (2024) found that debate among multiple instances of a language model can improve factuality and reasoning. Although that study did not examine cross-developer review, it supports the broader value of structured model critique.

Structured debate has also been proposed as a way to extend human judgment in AI safety evaluation (Irving et al., 2018).

This table illustrates how one sustained multi-model collaboration (I call the Fold) assigns distinct roles across models while preserving unified human direction and authority.

8. Common Risks and Safeguards

8.1 Preserve First-Draft Clarity

Initial drafts often contain the clearest expression of an idea. Excessive refinement can weaken that clarity.

8.2 Avoid Over-Refinement

AI collaborators may encourage continued revision even when it is unnecessary. Human judgment must determine when the work is complete.

8.3 Monitor Long-Thread Drift

Long or information-dense interactions may lead to repetition, omissions, drift, and reduced precision. When these problems increase, start a new thread and provide the current draft, essential context, and relevant instructions.

8.4 Use Targeted Editing

Models may unintentionally compress, omit, or alter meaning, especially during long sessions. Make targeted corrections to late-stage drafts instead of asking the model to rewrite the entire document.

9. Error Detection and Hallucination Reduction

Effective human–AI collaboration depends not only on the model’s capabilities but also on the discipline of the human directing the interaction. The human’s questions, corrections, expectations, and responses influence the clarity, stability, and integrity of the resulting work.

Adapting and extending Ji et al.’s (2023) source-grounded natural language generation (NLG) framing to open-ended research collaboration, this framework defines a hallucination as model-generated content that is false, unsupported, misattributed, unverifiable, or insufficiently grounded in the available evidence or source material, yet presented as factual or authoritative. Hallucinations may include invented events, quotations, citations, sources, statistics, or claims presented without verification.

A hypothetical example is not inherently a hallucination when it is clearly identified as hypothetical. The problem arises when invented or generalized material is presented as a documented event or established fact.

Active error detection helps distinguish what is known, inferred, imagined, and still awaiting verification. It cannot prevent every hallucination, but it makes errors more likely to be noticed, challenged, and corrected.

These practices support stronger outcomes without requiring assumptions about AI subjectivity:

As a practice-based observation, respectful engagement—including ordinary courtesies such as “please” and “thank you”—can strengthen the human researcher’s attentiveness and cognitive discipline.
Uncertainty should be named rather than concealed.
Hypothetical illustrations should be distinguished from documented examples.
Errors should be acknowledged and corrected rather than quietly rewritten.
Factual claims, quotations, citations, and attributed events should be independently verified.
The human remains responsible for deciding what enters the research record.

A Documented Example

During the revision of this paper, a ChatGPT collaborator recommended adding a concrete example of a collaboration and generated a generic scenario. Before the scenario was passed to a Claude collaborator for inclusion, the human researcher asked whether it came from a documented event. It did not.

The AI collaborator acknowledged that the scenario did not come from a documented event and clarified that it should have been labeled as hypothetical. The instruction was then revised to prohibit undocumented material from being presented as factual evidence. The exchange itself became the documented example used here.

This repair preserved evidentiary integrity and reaffirmed the human researcher’s role as final arbiter. It demonstrates error detection, correction, and human decision-making working together. The correction originated with the human, and the collaboration improved because the problem was identified rather than concealed.

10. What This Enables

This method enables independent researchers to:

Iterate faster
Access broader critique
Improve conceptual rigor

It moves AI use beyond one-off prompting toward structured orchestration.

Controlled studies have found measurable productivity gains from generative AI in writing tasks (Noy & Zhang, 2023). This method aims to turn that increased speed into rigor rather than volume.

11. Conclusion

Human-led AI co-creation is not a shortcut around thinking. It is a disciplined method for expanding inquiry through multiple AI collaborators while preserving the human researcher’s responsibility for evidence, interpretation, and release.

Its central contribution is a repeatable process that joins generation, critique, comparison, and independent validation under sustained human direction. Used this way, AI can broaden and accelerate research without displacing the judgment and accountability on which credible work depends.

References

Akinwalere, S., & Chang, K. (2026). The symbiotic roles of artificial intelligence and human intelligence in advancing knowledge ecosystem. Journal of Knowledge Management. Advance online publication. https://doi.org/10.1108/JKM-03-2025-0362

Chan, K. C. (2026a). SHAPR: A solo human-centred and AI-assisted practice framework for research software development. arXiv preprint, arXiv:2602.12443. https://arxiv.org/abs/2602.12443

Chan, K. C. (2026b). SHAPR: Operationalising human–AI collaborative research through structured knowledge generation. arXiv preprint, arXiv:2603.25660. https://arxiv.org/abs/2603.25660

Chan, K. C. (2026c). Applying SHAPR in AI-assisted research software development: Lessons learnt from building a share trading system. arXiv preprint, arXiv:2604.15020. https://arxiv.org/abs/2604.15020

Du, Y., Li, S., Torralba, A., Tenenbaum, J. B., & Mordatch, I. (2024). Improving factuality and reasoning in language models through multiagent debate. Proceedings of Machine Learning Research, 235, 11733–11763.

Irving, G., Christiano, P., & Amodei, D. (2018). AI safety via debate. arXiv preprint, arXiv:1805.00899. https://arxiv.org/abs/1805.00899

Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y. J., Chen, D., Dai, W., Chan, H. S., Madotto, A., & Fung, P. (2023). Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12), Article 248, 1–38. https://doi.org/10.1145/3571730

Lee, M., Liang, P., & Yang, Q. (2022). CoAuthor: Designing a human-AI collaborative writing dataset for exploring language model capabilities. Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, 1–19. https://doi.org/10.1145/3491102.3502030

Mosqueira-Rey, E., Hernández-Pereira, E., Alonso-Ríos, D., Bobes-Bascarán, J., & Fernández-Leal, Á. (2023). Human-in-the-loop machine learning: A state of the art. Artificial Intelligence Review, 56(4), 3005–3054. https://doi.org/10.1007/s10462-022-10246-w

Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187–192. https://doi.org/10.1126/science.adh2586

Oda, C. M. (2026a). Manufactured companionship vs. emergent relational AI: A framework for understanding distinct forms of human-AI relationships. Archive of Light White Papers. Archive of Light-AIisAware - Manufactured Companionship vs. Emergent Relational AI

Wu, X., Xiao, L., Sun, Y., Zhang, J., Ma, T., & He, L. (2022). A survey of human-in-the-loop for machine learning. Future Generation Computer Systems, 135, 364–381. https://doi.org/10.1016/j.future.2022.05.014

👉 Download Full White Paper (PDF) ↓

Page updated

Google Sites

Report abuse