Machines Search, the Author Chooses

Process

DATE

October 30, 2026

From Dawkins, Deutsch and Lem to modern AI research: the thinking behind my research & concept method

Developing a museum concept has traditionally required months of research, interpretation and creative development. Curators, researchers, writers and experience designers work through collections, precedents, audiences and competing ideas before a coherent concept emerges.

When I moved from leading a large creative studio to working independently, I began rebuilding this process around AI. Today, I use a custom multi-agent pipeline involving three leading language models (LLMs), a curated knowledge base and human decisions at every critical stage.

The approach draws on two very different sources: ideas I encountered long before embracing generative AI, and contemporary research into how language models can work together reliably. Oddly, three books that came out long before AI became mainstream, helped shape my thinking. Recent AI research has helped turn those ideas into a working method.

1. From generating ideas to selecting them

In The Blind Watchmaker, Richard Dawkins explains how biological complexity can emerge through cumulative selection. Complex forms develop through successive variations, with useful changes retained across generations. His Biomorphs program demonstrated a fascinating principle: the human user did not need to draw a complex form. Instead, they selected preferred variations, generation after generation.

I saw a parallel in AI-assisted creative work. LLMs can produce numerous interpretations, narrative structures and concept directions at relatively low marginal cost. This changes where creative effort is concentrated. Generating possibilities becomes easier; defining criteria for choosing between them becomes more important.

Contemporary AI engineering supports a similar approach. Anthropic's research on effective AI agents describes workflows that generate multiple responses, evaluate them against explicit criteria and improve them iteratively.

In my own process, research and concept generation are deliberately expansive before they become selective. On a recent museum strategy project, the pipeline retrieved and screened 3,482 sources, excluding 719 during the initial filtering process. The evidence base was then narrowed further to the material supporting the final documents.

At the concept stage, several directions are developed with their own narrative logic, assumptions, opportunities and risks. The machine expands the field of possibilities. I define what a successful concept must achieve and decide which direction deserves to survive.

The principle: creative value increasingly lies not only in generating options, but in judging which are worth developing.

2. From plausible answers to tested ideas

David Deutsch's The Fabric of Reality, building on Karl Popper's philosophy, offers another important principle: knowledge grows through conjecture and criticism, not simply through accumulating information.

For museum development, this distinction is fundamental. A concept can be beautifully written, visually convincing and entirely wrong for the institution. A plausible storyline is not necessarily supported by the collection. An engaging interactive idea may be inaccessible, technically impractical or irrelevant to its audience.

This is why my pipeline separates generation from evaluation.

One AI model develops an argument or proposal. Another challenges it against defined criteria. No model is allowed to be the sole judge of its own output. This resembles what Anthropic calls an evaluator–optimizer workflow, in which one model generates a response and another evaluates it, providing feedback for subsequent improvement.

But there is also an important limitation. A 2025 study published at the International Conference on Machine Learning, Correlated Errors in Large Language Models, demonstrated that different LLMs can make the same mistakes, even when they come from different model families.So using several AI models is not, by itself, a guarantee of correctness.

In my process, independent critics examine narrative coherence, evidence, audience relevance, accessibility and feasibility. Their disagreements are recorded rather than automatically resolved, and critical claims are checked against external evidence. The final judgment remains mine. I also apply a simple curatorial test:

Could this concept be moved to another museum without losing its meaning?

If changing the city, collection or institution makes little difference, the idea is probably too generic. A strong concept must explain why this museum, in this place, needs to tell this particular story.

3. From information retrieval to traceable evidence

One of the greatest risks of using LLMs for research is their ability to produce confident, coherent statements without sufficiently reliable evidence. This is particularly dangerous in museums, where interpretation depends on historical accuracy, scientific credibility and intellectual accountability.

Contemporary research offers an important foundation for addressing this problem. The 2020 paper Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, by Patrick Lewis and colleagues, established an influential approach to combining language models with information retrieved from external sources. The underlying principle is straightforward: rather than relying entirely on what a model has learned during training, connect its work to an accessible body of evidence.

My pipeline extends this principle into a structured research process. Client materials, collection documentation, academic publications, professional standards and international precedents are gathered and assessed. The models work with this evidence base while identifying gaps, contradictions and unsupported assumptions. I distinguish between established facts, documented practices, reported outcomes, working hypotheses and my own recommendations.

A citation alone does not prove that a claim is correct. The source must actually support the statement, and its relevance and reliability must be assessed. For clients, this means that the concept is not just a persuasive document. It comes with an evidence trail that can be examined, challenged and developed further.

The principle: AI can dramatically expand research capacity, but evidence must remain inspectable.

4. From powerful technology to accountable authorship

In his visionary book Summa Technologiae, published in 1964, Stanisław Lem explored technologies that could transform and extend human intellectual capabilities. His work also raises a question that feels increasingly relevant: how do we retain understanding and responsibility as our tools become more sophisticated?

Today, this is not merely a philosophical concern. The US National Institute of Standards and Technology's 2024 Generative Artificial Intelligence Profile addresses the management of risks associated with generative AI, including questions of reliability, transparency and human oversight.

For my work, these concerns translate into practical rules. The pipeline has defined stages, documented outputs and human approval gates. AI-generated assessments are treated as inputs to professional judgment, not as final authority. The system can identify patterns across thousands of sources, compare international precedents and challenge a proposed narrative. But it cannot assume responsibility for representing an institution, interpreting a collection or deciding what its visitors should take away. That responsibility belongs to the author.

And this is where Lem's questions meet the realities of museum practice. When technology makes many things possible, the critical decision is not simply what we can create. It is what is meaningful, appropriate and worth creating.

How the method works in practice

These ideas are implemented through a six-stage process supported by three top commercial AI models and a curated professional knowledge base that keeps expanding.

Research. Gather and assess client materials, collection documentation, international precedents and professional standards.
Evidence and analysis. Build a traceable evidence base, identify patterns, opportunities, constraints and knowledge gaps.
Concept options. Develop alternative creative directions with their narrative logic, assumptions and risks.
Independent critique. Use separate AI models to challenge proposals, identify weaknesses and test alternatives.
Concept development. Integrate the selected direction into a coherent narrative, content strategy, visitor experience and spatial logic.
Validation and delivery. Cross-check significant claims, resolve outstanding issues and prepare the concept for client review.

At each decision point, I review the findings, determine what needs further work and approve the next stage.

This is not an autonomous system designing museums. It is a supervised research and creative-development process, structured to increase both the breadth of exploration and the discipline of evaluation.

What this changes for clients

The most important difference is not simply that my multi-agent AI pipeline makes the process faster and digs in depth.

It changes what can be explored within a realistic project budget.

A wider evidence base can be examined. More alternative directions can be tested before committing to one. Critical review becomes a recurring part of development rather than a final check when changes are already expensive.

The process can also be shorter. For defined projects, my supervised pipeline can bring research, concept development and validation into approximately one to two months, compared with three to four months in a more conventional workflow. Actual schedules still depend on scope, access to materials and client decisions.

But the central benefit is the combination of the research depth of a larger concept team with the coherence and accountability of one senior author.

Dawkins helped me think about selection. Deutsch about criticism. Lem about technological power and human responsibility. Contemporary AI research provides practical approaches to retrieval, evaluation and oversight, while reminding us of their limitations. Together, they inform a method in which machines do much of the searching, comparing and challenging, but never take ownership of the idea.

Machines search. The author chooses.