From Gap-Filling to Possibility Expansion: AI and the Cyclical Process of Knowledge Discovery

Executive Summary

This article presents a knowledge-centric framework for understanding how AI creates value in knowledge work. Rather than acting in a single, static way, AI alternates between two complementary modes in a continuous Knowledge Discovery Process:

Gap-Filling — Closing immediate, explicit knowledge gaps so a task can proceed.
Amplification — Expanding the space of possibilities for innovation by recombining existing human and AI knowledge into new ideas, perspectives, and solutions.

These modes are interdependent. Gap-Filling enables Amplification by supplying the missing pieces needed to explore new possibilities. Amplification, in turn, generates novel directions that create fresh gaps, restarting the cycle. This dynamic mirrors findings from novelty research (Loreto et al., 2016; Di Bona et al., 2023) on how exploration expands the possibilities and drives the emergence of new ideas.

The article resolves common misconceptions — that AI is only useful to novices, or that it fails in complex contexts — by showing how AI’s marginal value changes across the two modes. Gap-Filling value decreases as user expertise rises, while Amplification value increases, making AI relevant at all levels of expertise. That just means the smarter you are, the smarter the AI is.

The key implications for organizations are:

Measure AI’s impact across the full discovery cycle, not just in one mode.
Design tools that switch fluidly between Gap-Filling and Amplification.
Invest in explicit, machine-readable knowledge to strengthen Amplification.
Harness the cycle to drive both productivity and innovation over time.

By accelerating the alternation between filling gaps and expanding possibilities, AI can increase not only the speed of task completion but also the pace at which new opportunities and innovations emerge.

1. Introduction — AI as a Knowledge Source

Viewed through a knowledge-centric lens, AI is not a replacement for human intelligence but one more source of knowledge in a worker’s toolkit — alongside search engines, technical documentation, books, and the insights of experienced colleagues. In software development, for example, AI now competes with Stack Overflow for quick answers, and with internal wikis for codebase-specific guidance.

This raises a practical question: When and how does AI create the most value for human work? The answer is not as simple as “whenever the human knows less than the AI.” Nor is it simply “when the human knows more and can guide the AI.” In practice, AI’s usefulness is not static — it operates in two alternating modes that feed each other in an ongoing process of knowledge discovery.

In one mode, AI acts as a substitute, filling explicit gaps in the human’s current knowledge so they can complete the task at hand. In the other, it acts as an amplifier, helping the human combine what they already know into new perspectives, possibilities, and solutions. These modes are not mutually exclusive; rather, they form a cycle in which each creates the conditions for the other.

This dynamic mirrors findings from novelty research in complex systems. Loreto et al. (2016) describe innovation as movement in an expanding possibility space, where each novelty reshapes what is possible next. Di Bona et al. (2023) extend this to higher-order novelties, where new combinations of existing elements open further opportunities. In the same way, AI-assisted work alternates between filling the gaps in today’s space of possibilities and expanding that space to create the challenges and opportunities of tomorrow.

2. Two Modes of AI Use

AI does not provide value in a single, uniform way. In knowledge work, its contribution follows two distinct but interdependent modes: Gap-Filling and Amplification. Each mode operates differently, benefits from different conditions, and aligns with different strands of cognitive and educational research.

Gap-Filling — Filling Immediate Knowledge Gaps

In gap-filling mode, AI acts as a stand-in for missing, explicit knowledge needed to complete a task. The developer (or other knowledge worker) is already competent in their domain but encounters a specific gap: a fact they don’t recall, a syntax they’ve never used, or a configuration they’ve never encountered.

Example: An expert programmer working in a familiar codebase needs the exact flags for a rarely used cloud SDK function. They understand the architecture, coding standards, and implications of using the SDK, but this specific detail is missing. AI can retrieve it in seconds.
This mode parallels the Expertise Reversal Effect in educational psychology: support is most useful when knowledge is lacking, but its value declines — and can even turn into friction — when the user already knows what they need.
Limitation: AI cannot substitute for tacit or undocumented knowledge embedded in human experience or in the unrecorded conventions of a mature codebase.

Amplification — Expanding the Space of Possibilities for Innovation

In amplification mode, AI helps humans combine human knowledge — found in mental models, design documents, or requirements — with AI’s own knowledge from its training data and contextual inputs. The result is a shared reasoning space that can generate new perspectives, combine known elements in novel ways, and propose solutions that might not have been considered by either side alone.

Example: A programmer who has just mastered the use of a new SDK function now asks an AI coding agent for architectural patterns that integrate it with existing modules. Drawing on the design documentation (human knowledge) and its learned patterns from thousands of prior systems (AI knowledge), the agent proposes several combinations. The programmer’s expertise allows them to quickly assess trade-offs, discard unworkable ideas, and select the most promising path. The developer then instructs the agent to make the necessary code changes, which it implements according to the agreed design.
In this way, amplification mode expands the space of possibilities by merging human and AI knowledge, while leveraging human expertise to navigate and narrow it effectively.
Analogy: In the game 20 Questions, the more concepts a player knows, the better they can frame each question to reduce the search space. Without prior knowledge, the game stalls; with it, every answer accelerates convergence.
This mode aligns with the concept of Cognitive Scaffolding in human–AI interaction: the AI extends and accelerates reasoning, but only when there is enough prior knowledge to guide its queries and interpret its outputs. That just means the smarter you are, the smarter the AI is.
Limitation: Without a coherent mental framework, amplification produces noise or overwhelming results — like staring through a telescope without knowing where to point it.

These two modes are not independent silos. In practice, they alternate and feed each other: gap-filling fills gaps in the current task space; amplification uses that fuller knowledge to generate new possibilities; those possibilities create new gaps, and the cycle repeats.

3. The Continuous Knowledge Discovery Process

In real-world work, AI’s two modes — gap-filling and amplification — do not occur in isolation. They form a continuous cycle of knowledge discovery, in which each mode naturally triggers the other.

Gap-Filling: Fill immediate gaps in the current space of possibilities.
- The human encounters a missing fact, unfamiliar method, or unclear requirement.
- AI provides the explicit information needed to proceed.
- This is an exploitation phase: drawing on AI knowledge to address the present need.
Amplification: Use AI’s fuller knowledge base to generate novel combinations and perspectives.
- With the gap closed, the human can explore beyond the immediate problem.
- AI suggests new arrangements, patterns, or strategies that build on what is already known.
- This is an exploration phase: expanding the boundaries of what is possible.
- Crucially, expertise plays a dual role here — guiding the generation of meaningful alternatives and selecting between them.
Gap-Filling again: Address the new gaps created by the expansion.
- Every new possibility comes with its own missing pieces — a library you’ve never used, a dependency you don’t know how to configure, a concept you haven’t studied.
- AI re-enters gap-filling mode to fill these gaps.
Repeat: The cycle continues as long as new possibilities emerge.

This pattern maps directly to models of innovation in complex systems:

In Loreto et al.’s framework, gap-filling corresponds to exploring the actual — the current reachable set of knowledge and possibilities. Amplification corresponds to expanding the possible — reshaping the conceptual space so new elements become reachable.
Di Bona et al.’s concept of higher-order novelties describes how new combinations of existing elements create further opportunities for exploration. In our cycle, amplification generates these higher-order novelties; gap-filling then explores them in detail.

AI’s role in the cycle

AI can support both phases, but in different ways:

In gap-filling mode, it is a retrieval and precision tool — quickly supplying explicit, factual knowledge.
In amplification mode, it is a catalyst and combinator — surfacing unexpected patterns, alternative framings, and novel arrangements of what is already known.

The power of AI-assisted work comes not from excelling at one phase alone, but from accelerating the alternation between phases. The faster we can fill current gaps and open new possibilities, the faster the overall pace of discovery.

4. Implications for AI Adoption in Knowledge Work

Recognizing that AI operates in an alternating cycle of gap-filling and amplification has practical consequences for how organizations evaluate, design, and integrate AI into knowledge-intensive workflows.

Measure Impact Across the Full Cycle

Most current productivity metrics focus on isolated effects — for example, how quickly AI helps a user produce a piece of code (gap-filling) or how many new ideas it can suggest (amplification). These partial measurements can be misleading. True impact should be assessed across the full knowledge discovery cycle:

How effectively does AI fill immediate gaps?
How often does it open genuinely new possibilities?
How well do those new possibilities translate into valuable outcomes?

Design for Smooth Transitions Between Modes

Knowledge work rarely stays in one mode for long. A single task might shift from gap-filling to amplification and back several times in minutes. AI tools should support this fluidity:

In gap-filling mode, prioritize accuracy, clarity, and minimal friction.
In amplification mode, prioritize breadth, combinational creativity, and exploratory prompts.
Switching between these should be seamless — without forcing the user into a rigid workflow.

Invest in Knowledge Capture and Documentation

Amplification mode depends on the availability of explicit, machine-readable knowledge. Tacit, undocumented expertise cannot be recombined by AI. This means:

Documenting architecture patterns, decision rationales, and domain constraints increases AI’s amplification power.
Treating internal wikis, code comments, and structured data as strategic assets maximizes return on AI investment.
As explicit knowledge grows, amplification can produce higher-order novelties that further expand the possibility space.

Leverage the Cycle for Productivity and Innovation

The alternation between gap-filling and amplification is not just a way to solve today’s problems — it’s a mechanism for long-term growth:

Productivity improves as gaps are closed faster and new possibilities are explored more effectively.
Innovation emerges from repeated cycles, as each amplification phase reshapes what is possible and each gap-filling phase equips people to act within the expanded space.

Organizations that understand and harness this cycle position themselves to move faster, adapt more easily, and innovate continuously.

5. Real-World Examples of the Cycle

Expert Programmer in a Mature Codebase

Gap-Filling: While implementing a feature, an expert programmer needs to integrate with a cloud SDK they rarely use. They understand the codebase and the architectural context but don’t recall the exact parameters for a specific SDK call. AI quickly retrieves the correct syntax and options.
The benefit here is real but marginal — AI is replacing only the small portion of knowledge the programmer lacks. If the programmer had already known the API call, AI would add nothing unique and might even slow them down with irrelevant or slightly incorrect suggestions. This reflects a Heaps’ Law–like pattern: as the programmer’s prior knowledge about the task increases, the marginal value of AI in gap-filling mode decreases, eventually turning negative when knowledge is complete.
Amplification: With this knowledge in place, the programmer asks AI to suggest alternative architectural patterns for integrating the SDK with existing services. AI proposes several combinations. Drawing on their expertise, the programmer evaluates trade-offs and selects the most promising approach. AI implements the needed code changes and produces new source code.
Gap-Filling again: The chosen design requires configuring a new build tool unfamiliar to the programmer. AI supplies the commands, configuration templates, and troubleshooting advice.
Repeat: Equipped with the new tool, the programmer can now consider optimizations to the build process — triggering another amplification phase.

The 20 Questions Game

The game "20 questions" is an old game that gained popularity in the late 1940s when it was used as the format for a successful weekly radio quiz program. In the traditional game, the inquirer leaves the room while the remaining people agree on an object - a person, place, or thing. The inquirer then comes back and has to guess what the object is by asking successively questions that can be answered with a simple "yes" or "no". If the inquirer cannot guess the object after asking 20 questions, the respondents have stumped the inquirer.

There are many ways to acquire the same knowledge. Each depends on the strategy and level of subject matter expertise.
Let's suppose the object to be guessed is "Abraham Lincoln's stove pipe hat" from Box, G. E. et. all. (2005) . The initial clue is "sugarloaf with animal associations".

Gap-Filling: The inquirer begins with only the clue “sugarloaf with animal associations.” AI suggests broad, high-yield opening questions (“Is it living?” “Is it human?”) that help fill in basic facts. At this stage, the inquirer’s prior knowledge about the object is low, so AI’s contribution is substantial. But as more answers accumulate, the remaining unknowns shrink. Eventually, the AI is only supplying small factual fills — and if the inquirer already knew those facts, AI’s marginal value here would be negligible or even negative if it distracted with irrelevant leads. This follows a Heaps’ Law–like pattern: as the inquirer’s knowledge of the target increases, the marginal benefit of gap-filling decreases.
Amplification: Once the inquirer knows the object is a famous 19th-century hat, AI can propose targeted, high-information questions (“Was it worn by a political leader?” “Is it connected to Abraham Lincoln?”). Here, the inquirer’s background knowledge is what makes AI’s suggestions powerful: they can recognize the significance of the clues and quickly converge on “Abraham Lincoln’s stovepipe hat.” In this mode, the more the inquirer knows, the more AI can help — the “telescope effect.”
Gap-Filling again: A specific answer (“Abraham Lincoln”) may still leave a gap — the inquirer may need to recall or learn that Lincoln was known for a distinctive stovepipe hat. AI can supply this missing fact, closing the loop.
Repeat: The newly confirmed answer enriches the inquirer’s mental library, enabling faster and more strategic questioning in future games.

If Lincoln’s hat’s cultural significance were missing from AI’s training data, no combination of clues would lead it to propose the correct answer. If the information exists in explicit sources, AI can connect the dots and recombine the known facts. This illustrates why AI thrives with explicit, documented knowledge but struggles with tacit, undocumented insights.

In both cases, gap-filling addresses immediate, explicit gaps, while amplification leverages existing knowledge to explore new possibilities and select among them. Each amplification phase expands the possibility space, and each expansion generates fresh opportunities for gap-filling — an alternating rhythm that can, in principle, continue indefinitely.

These examples show that AI’s usefulness is not a simple function of how much the human already knows. In gap-filling mode, AI bridges specific, explicit gaps; in amplification mode, it thrives when the human can guide, evaluate, and integrate its output. In both modes, AI’s success depends heavily on whether the underlying knowledge is explicit or tacit.

6. Conclusion

AI’s role in knowledge work cannot be reduced to a single pattern of use. Sometimes it serves as a replacement for missing, explicit knowledge, with value that diminishes as the user’s mastery of the task increases. Other times it acts as a force multiplier, becoming more valuable the more expertise the user brings. In practice, these modes are not isolated — they alternate in a continuous Knowledge Discovery Process.

In this cycle:

Gap-Filling closes the gaps in the current space of possibilities.
Amplification expands that space by merging existing human and AI knowledge into novel combinations and perspectives.
Expansion creates new gaps, triggering another round of gap-filling.

This process mirrors the actual vs. possible dynamic in Loreto et al.’s (2016) work on the emergence of novelties and the higher-order novelties described by Di Bona et al. (2023). Each iteration not only solves the problem at hand but also reshapes what can be imagined and achieved.

For individuals and organizations, the implication is clear: the goal is not merely to speed up output in one mode, but to accelerate the entire cycle. That means designing AI systems that move fluidly between gap-filling and amplification, investing in explicit knowledge capture to power amplification, and measuring success across the full loop.

The real opportunity is not just faster answers — it’s a sustained increase in the pace at which we discover what’s possible.

Reference

1. Box, G. E. P., Hunter, J. S., & Hunter, W. G. (2005). Statistics for Experimenters: Design, Innovation, and Discovery, 2nd Edition (2nd edition). Wiley-Interscience.

2. Di Bona, G., Bellina, A., De Marzo, G. et al. The dynamics of higher-order novelties. Nat Commun 16, 393 (2025). https://doi.org/10.1038/s41467-024-55115-y

3. Loreto, V., Servedio, V.D.P., Strogatz, S.H., Tria, F. (2016). Dynamics on Expanding Spaces: Modeling the Emergence of Novelties. In: Degli Esposti, M., Altmann, E., Pachet, F. (eds) Creativity and Universality in Language. Lecture Notes in Morphogenesis. Springer, Cham. https://doi.org/10.1007/978-3-319-24403-7_5

From Gap-Filling to Possibility Expansion

Related Articles