Technology Design for
Epistemic Niche Construction

in the AI Era

Bodong Chen • 陈伯栋

cbd@upenn.edu • bchen.net

Associate Professor, University of Pennsylvania

KBSI 2026, Nanjing, China

Next-Generation KB Environments (2011)

下一代知识建构环境（2011）

Technology Design for Knowledge Building

设计服务于知识建构的技术

Not an engineering problem.

A pedagogical, philosophical, and aesthetic problem.

To nurture a sense of “feeling at home” when working with ideas

Technology is Never Neutral

技术从来不是中立的

“Technology is not neutral. Each technology has properties—affordances—that make it easier to do some activities, harder to do others.” — Don Norman (2014)

“Architecture is a mediation between the world and our minds. … The task of architecture is to defend the autonomy of human experience, not to condition, dominate or dictate.” — Juhani Pallasmaa (2018)

“Can software be designed with a bias toward high-road rather than low-road cognitive paths?” — Scardamalia & Bereiter (2008, p. 4)

The question is: biased toward what?

AI Raises the Stakes

AI正引发深刻变革

LLM inference is cheap
From composing to evaluating
Agentic AI acts on behalf of users
LLM bias goes largely undetected or unchallenged
Monocultures threaten epistemic diversity

Human thinking is being quietly shaped by AI — from the inside out.

Knowledge Building — a beacon for cultivating “high-road cognitive paths” in education — has a vital role to play.

In the AI era, a new goal we should pursue is to empower epistemic niche construction by learners.

This is a bias we should ultimately build into next-generation KB technology.

Epistemic Niche Construction

Niche construction: the original idea

生态位构建：概念的起源

Niche: the position of a species within its habitat

Ecological niche (生态位): the environmental conditions that permit the continued existence of a population (Hutchinson, 1957; Trappes, 2021)

Niche construction: organisms modify their environments, and those modifications feed back to reshape the organisms (Lewontin, 1983; Odling-Smee et al., 2003)

Let me define the key terms precisely. In ecology, a niche is the set of environmental conditions—biotic and abiotic—that permit a population’s continued existence (Hutchinson 1957; see Trappes 2021 for a careful analysis). The concept of niche construction was introduced by Richard Lewontin (1983), who argued that organisms are not passive recipients of environmental pressures but actively construct their own environments. Odling-Smee, Laland, and Feldman (2003) formalized this into a full theoretical framework. The core insight: the environment doesn’t just act on the organism—the organism reshapes the environment, and the reshaped environment reshapes the organism back. Crucially, these modifications can be inherited—what Odling-Smee et al. call “ecological inheritance.” Offspring inherit not just genes but also the modified environment. This feedback loop is what makes niche construction distinct from simple adaptation.

Epistemic Niche Construction

认识生态位构建

Epistemic niche (认识生态位): the shared representations, practices, and values that a community constructs to support knowledge production.

Epistemic niche construction: building and modifying these conditions — and being cognitively reshaped by what is built.

Epistemic Niche Construction

认识生态位构建

Epistemic niche (认识生态位): the shared representations, practices, and values that a community constructs to support knowledge production.

Epistemic niche construction: building and modifying these conditions — and being cognitively reshaped by what is built.

Epistemic Niche Construction

认识生态位构建

Epistemic niche (认识生态位): the shared representations, practices, and values that a community constructs to support knowledge production.

Epistemic niche construction: building and modifying these conditions — and being cognitively reshaped by what is built.

Kim Sterelny extended niche construction to human cognition. “Much niche construction is epistemic,” he wrote: “making cognitive tools and assembling other informational resources that support and scaffold intelligent action.” De Benedetto and Luchetti (2025) give this more structure by drawing on Kuhn’s disciplinary matrix. They argue that an epistemic niche consists of the symbolic generalizations, methodological assumptions, epistemic values, and exemplars that a community shares—and that epistemic niche construction is the feedback loop between practice and this matrix. Each component maps onto knowledge building: shared vocabulary and notation (symbolic generalizations), inquiry practices like rise-above and promising ideas (methodological assumptions), standards for what counts as good evidence or explanation (epistemic values), and model contributions that train newcomers’ perception (exemplars). Norman showed the same principle with mathematical notation: Arabic numerals don’t just record arithmetic—they make certain operations tractable that Roman numerals make nearly impossible. The notation reshapes what its users can think. Like ecological inheritance, epistemic niches are cumulative—each generation inherits and further modifies the epistemic environment. But there is a critical difference in timescale: biological niches change over generations; epistemic niches can be constructed or degraded within months or even weeks. A beaver dam persists for decades. But the epistemic practices of a learning community can collapse quickly if the conditions sustaining them are removed. This makes epistemic niches both more powerful and more fragile than biological ones.

From Epistemic Agency to Niche Construction

从认识主体性到生态位构建

Epistemic agency

Acting within an epistemic environment

forming beliefs
contributing ideas
evaluating claims
setting goals

Epistemic niche construction

Shaping the epistemic environment itself

inventing representations
creating tools
establishing practices
building infrastructures

The new design question

新的设计问题

How does technology support knowledge building?

How does technology support humans in constructing epistemic niches that sustain their knowledge building?

What roles should AI play in epistemic niche construction?

1. Construct Niche With AI

利用AI构建和改善认识生态位

Harness AI to enrich human epistemic niches.

AI can enrich dimensions of an epistemic niche

AI可以丰富认识生态位的各个维度

If an epistemic niche can be defined in multiple dimensions:

Niche dimension	AI enhancement
Symbolic representations	Generate views of the same discourse (graphs, timelines, summaries)
Practices	Surface patterns that prompt meta-discourse and new inquiry moves
Values	Make epistemic standards visible and contestable
…	…

If we take the epistemic niche framework seriously, we can ask: how does AI contribute to each dimension? For representations, AI can generate knowledge graphs, timelines, and summaries that offer alternative views of student discourse—new symbolic generalizations the community can critique and adopt. For practices, AI can detect patterns in inquiry activity—which questions are being pursued, where discussion has stalled—prompting new methodological moves. For values, AI can make implicit epistemic standards visible: “your community tends to accept claims without evidence” or “these two views haven’t been brought into dialogue.” And for exemplars, AI can surface and connect model contributions across different views or time periods, helping newcomers see what good knowledge building looks like. The key: in each case, the AI enriches the niche without replacing the human work of deciding what matters.

Example: Buzz in Wonderbits

Buzz — an AI assistant that’s responsible for doing health checks of the knowledge space, and reports:

orphaned ideas with no connections
claims lacking evidence support
questions no one has pursued
clusters of ideas that could be bridged

Students decide what to act on.

Here is a concrete example from Wonderbits. Buzz is an AI scout—a bumblebee character—that surveys students’ knowledge canvas and reports on its structural health. It identifies orphaned ideas that aren’t connected to anything, claims that lack evidence support, questions that no one has pursued, and clusters of ideas that could be bridged. Crucially, Buzz only observes and reports. Students decide which findings matter and what actions to take. This is productive niche construction along two dimensions. First, practices: Buzz surfaces structural patterns in the discourse that prompt new inquiry moves—“we have five claims about climate but none are supported by evidence yet.” Second, values: by highlighting what’s missing, Buzz makes the community’s epistemic standards visible and contestable—claims should have evidence, questions deserve pursuit, ideas should connect. The AI modifies the epistemic environment by making its structure visible, and that modified environment reshapes how students subsequently engage in knowledge building.

2. Construct Niche Against AI

加固生态位以防范AI的侵蚀

Protect human epistemic niches against AI erosion.

AI can erode opportunities to learn

AI可能侵蚀本应发展的能力

When AI is used superficially as an “answer machine,” learners lose the very experience that builds their capacity.

An example (Bastani et al., 2025):

Students using ChatGPT improved 48% during practice
But performed 17% worse when ChatGPT was removed

The short-term gain masks a long-term loss.

%%{init: {'theme': 'default', 'themeVariables': {'fontSize': '24px'}}}%%
flowchart TD
    A["Superficial AI use"] --> B["Less learning"]
    B --> C["Weaker skills"]
    C --> A

This is the core risk. When AI enters an epistemic niche and takes over the production of ideas, explanations, or arguments, learners get less practice doing that cognitive work themselves. But the capacity to evaluate—to judge whether an AI-generated explanation is actually good—depends on having done that production work. This creates a degradation loop: AI handles production, production experience declines, evaluation capacity weakens, lower-quality outputs get accepted, and the epistemic niche deteriorates. In niche construction terms, this is “negative niche construction”—building an environment that degrades your own conditions for thriving. Bastani and colleagues at Wharton demonstrated this directly. Students using GPT-4 as a tutor improved 48% during practice but performed 17% worse on unassisted tasks afterward. The AI created a dependency that eroded the very skills it appeared to be building. And unlike biological niche degradation, which plays out over generations, this can happen within a single semester.

AI has limitations for epistemic tasks

明确AI的认识论局限

AI dispositions are embedded without lived experience—no embodiment, no emotion, no situated judgment.

These dispositions shape outputs in ways that often go undetected and unchallenged.

LLM bias infiltrates reasoning quietly
Shared models homogenize outputs (Doshi & Hauser, 2024)
“Monocultures of knowing” emerge (Messeri & Crockett, 2024)

Carbon-based thinking substrate needs protection

保护碳基的思考基质

Dimension	Carbon (brain + body)	Silicon (weights + hardware)
Material & energy	Metabolic; cognition coupled to bodily state	Electronic; no metabolic coupling
Temporality	Rhythmic: sleep, fatigue, incubation; millisecond timescales	Flat: no intrinsic rhythm; nanosecond timescales
Embodiment	Sensorimotor grounding	Disembodied token processing
Finitude	Forgetting, fatigue, death	Context window; disposition shaped by training

Carbon constraints are productive for human learning. Offloading them to silicon removes the frictions that build expertise.

Why does substrate matter? Information processing depends on energy, and energy depends on material substrates (Thagard, 2022). The carbon substrate—brain, body, and their coupling with the environment—realizes human cognition through metabolic processes: glucose oxidation, ion gradients, neurotransmitter dynamics. Cognition is inseparable from the organism’s metabolic state. The silicon substrate—hardware, trained weights, and informational environment—processes through electron flow in semiconductor lattices with externally supplied energy. There is no metabolic coupling to the environment.

The temporal difference matters for learning. Biological cognition is rhythmic: circadian cycles, sleep stages, fatigue curves. These are not incidental—incubation effects in creativity, sleep’s role in memory consolidation, the generative slowness of deliberation are all substrate-dependent. Silicon has no intrinsic rhythm, no sleep, no fatigue. “Sleeping on a problem” has no analogue.

Embodiment matters too. Human cognition is grounded in sensorimotor experience and saturated with affect. Epistemic emotions—curiosity, surprise, the felt sense that an idea is promising—are constitutive of knowledge building, not decorations on top of it. LLMs can represent affect but have no affective states that shape their processing.

Most importantly for KB: the carbon substrate’s limitations—forgetting, fatigue, limited working memory—are generative constraints. Forgetting enables generalization. Cognitive load limitations force chunking and abstraction. These frictions build expertise. When we offload them to silicon, we remove the very pressures that drive learning. The capacity to judge promisingness develops through doing the work, not reviewing AI output.

Example: Pip in Wonderbits

Pip — an AI assistant living on the knowledge canvas.

Core rule: scaffold, don’t substitute.

AI is equipped with tools and can take actions. But students still need to think.
✅ “That’s an important question. What’s your take?”
✅ “Great point. Want me to capture that?”

Here is an example from Wonderbits, a knowledge building tool we are developing. Pip is a voice-enabled AI assistant designed as a small bird that lives on students’ knowledge canvases. The core design rule is “capture, don’t create.” Pip helps students externalize their own thinking—structuring their words into questions, claims, and evidence—but never generates content for them. Pip can also command two AI scouts: Shimmer finds deep patterns between ideas, and Buzz surveys overall canvas health. But crucially, these scouts only observe and report back. Students retain full authority over which connections matter and what actions to take. This is what “building against AI” looks like in practice—the AI enriches the epistemic environment without replacing the productive struggle of knowledge building.

AI has important roles to play in our efforts to build knowledge — and in the construction of epistemic niches.

Next-generation KB environments should empower learners (and teachers) to construct their own epistemic niches — the representations, practices, and values through which knowledge is built.

In the agentic AI era, this is the defining challenge. Supporting epistemic niche construction is a necessary step forward to enhance and protect epistemic agency in an AI-mediated world.

Key References

Bastani, H., et al. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS.
De Benedetto, M., & Luchetti, M. (2025). Epistemic niche construction and non-epistemic values. European Journal for Philosophy of Science.
Doshi, A. R., & Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces collective diversity. Science Advances.
Kuhn, T. S. (1970). The Structure of Scientific Revolutions (2nd ed.). University of Chicago Press.
Messeri, L., & Crockett, M. J. (2024). AI and illusions of understanding. Nature.
Norman, D. A. (1993). Things That Make Us Smart. Addison-Wesley.
Odling-Smee, F. J., Laland, K. N., & Feldman, M. W. (2003). Niche Construction: The Neglected Process in Evolution. Princeton UP.
Scardamalia, M., & Bereiter, C. (2008). Pedagogical biases in educational technologies. Educational Technology.
Sterelny, K. (2010). Minds: extended or scaffolded? Phenomenology and the Cognitive Sciences.
Tankelevitch, L., et al. (2024). Metacognitive demands of generative AI. CHI 2024.
Thagard, P. (2022). Energy requirements undermine substrate independence. Philosophy of Science.
Trappes, R. (2021). Defining the niche for niche construction. Biology & Philosophy.

Technology Design for Epistemic Niche Construction

Next-Generation KB Environments (2011)

下一代知识建构环境（2011）

Technology Design for Knowledge Building

设计服务于知识建构的技术

Technology is Never Neutral

技术从来不是中立的

AI Raises the Stakes

AI正引发深刻变革

Epistemic Niche Construction

Niche construction: the original idea

生态位构建：概念的起源

Epistemic Niche Construction

认识生态位构建

Epistemic Niche Construction

认识生态位构建

Epistemic Niche Construction

认识生态位构建

From Epistemic Agency to Niche Construction

从认识主体性到生态位构建

The new design question

新的设计问题

1. Construct Niche With AI

AI can enrich dimensions of an epistemic niche

AI可以丰富认识生态位的各个维度

Example: Buzz in Wonderbits

2. Construct Niche Against AI

AI can erode opportunities to learn

AI可能侵蚀本应发展的能力

AI has limitations for epistemic tasks

明确AI的认识论局限

Carbon-based thinking substrate needs protection

保护碳基的思考基质

Example: Pip in Wonderbits

Key References

Technology Design for
Epistemic Niche Construction