🧬 biology

Competitive and Complementary Tools

This paper models the co-evolution of human competence and tool reliance as a bistable dynamical system, demonstrating that excessive tool availability can trigger an irreversible collapse of human agency into dependency due to hysteresis, a phenomenon validated across domains from navigation to language models.

Original authors: David C. Krakauer

Published 2026-07-22

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: David C. Krakauer

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Great Tool Trap: Why Your Brain Might Be Getting Lazy

Imagine your brain is a gym, and every time you solve a problem, you're lifting a mental weight. For thousands of years, humans have built tools to help us lift heavier loads, from the first stone axe to the modern smartphone. But there's a tricky question that scientists have been wrestling with: Do these tools make us stronger, or do they turn our muscles into jelly? This field of study looks at "cognitive artifacts"—tools that help us think, remember, or calculate. The big idea is that when we use a tool, we aren't just borrowing its power; we are changing how our own brain works. Sometimes, a tool helps us build a better internal map of the world (like learning to read a map helps you remember the streets). Other times, it might trick us into thinking we know the answer when we've just outsourced the work to a machine. Understanding this difference is crucial because we are currently surrounded by powerful new tools, like artificial intelligence, and we need to know if they are making us smarter or just making us dependent.

The Tug-of-War Between Skill and Convenience

In this paper, David Krakauer sets up a fascinating mental experiment to figure out exactly how tools change us. He treats a person, a tool, and a task as a single, living system where two things are constantly fighting each other: Competence (what you can still do on your own) and Reliance (how much you let the tool do for you).

Think of it like learning to ride a bike with training wheels. If you use the training wheels just enough to stay upright while you learn to balance, you eventually build strong leg muscles and a great sense of balance. That's a complementary tool. But imagine if the training wheels were so magical that they did all the balancing for you, and you never had to wobble or adjust your weight. You'd get really good at sitting on the bike, but the moment someone took the wheels away, you'd crash. That's a competitive tool.

Krakauer's model shows that this isn't just a slow slide; it's a cliff. There is a "tipping point." As long as you have some skill and the tool is somewhat transparent (meaning you can see how it works), you stay in the "Competent" zone. But if the tool is too easy to use and too mysterious inside, and you use it too much, your brain suddenly snaps into a "Dependent" zone. Once you fall into this zone, your ability to do the task without the tool collapses toward zero.

The Trap of History: Why Two People Can Be Opposites

Here is the most surprising part of the story: History matters more than the present.

Imagine two people, Alex and Jordan, standing in front of the exact same GPS device with the exact same signal strength.

Alex learned to navigate using paper maps first. They built a strong mental map of the city. When they start using the GPS, they use it as a helper, but they still know the streets. They stay in the "Competent" zone.
Jordan never learned the streets. They started using the GPS immediately. Because they didn't have a mental map to begin with, the GPS did all the work. Their brain stopped building the map entirely. They fall into the "Dependent" zone.

Even though they have the same tool right now, they are in two completely different worlds. If you take the GPS away from Alex, they can still find their way. If you take it away from Jordan, they are lost.

The paper suggests that this state is bistable, which is a fancy way of saying the system has two stable "parking spots." Once you park in the "Dependent" spot, it's incredibly hard to get out. You can't just stop using the tool and expect your skills to magically return. You have to go back to a much lower level of tool use to even start rebuilding your skills. It's like a ratchet: you can click forward easily, but clicking backward is a struggle.

The Secret Ingredient: Transparency

Why do some tools (like a map) make us smarter, while others (like a GPS) make us dumber? The paper points to a key feature called Transparency.

Transparent Tools: Think of an abacus or a paper map. You can see the gears turning, or the lines connecting the dots. You can look at the tool and say, "Oh, I see how that works!" You can rebuild that logic inside your own head. These tools are complementary. They help you build a mental model that stays with you even when the tool is gone.
Opaque Tools: Think of a GPS or a complex AI. It gives you an answer, but the "how" is hidden inside a black box. You can't see the math or the logic. Because you can't rebuild the logic in your head, your brain has no choice but to stop trying. These tools are competitive. They eat your skills.

The paper argues that if a tool is too opaque, or if the task is too huge for your brain to hold, you will lose your ability to do the task, no matter how smart you are. It's not about your intelligence; it's about the tool's design.

The Big Warning: The Order of Learning

The most practical takeaway from this research is about timing. The paper suggests that the order in which we learn matters more than we think.

If you teach someone a skill (like math or coding) before giving them the powerful tool (like a calculator or an AI), they build a strong foundation. They can then use the tool to go faster and do harder things, but they keep their core skills. They land in the "Competent" attractor.

But if you give the tool first, before they have built that foundation, they will likely skip the hard work of learning the basics. They will rely on the tool so much that their brain never builds the necessary skills. They land in the "Dependent" attractor.

The paper warns that this isn't just a theory; it matches what we see in real life. Pilots who rely too much on autopilot lose their manual flying skills. People who use GPS every day lose their ability to remember routes. And students who use AI to write essays without learning how to write first might lose their ability to think critically.

The Future: Building a Tool-Proof Education

So, what does this mean for us? The paper suggests that we need to be careful about how we introduce new technology. We shouldn't just hand out powerful, opaque tools to beginners. Instead, we should make sure people build their "mental muscles" first.

Education needs to be "tool-resistant" in the beginning. This means teaching the hard stuff without the crutches, so that when the crutches (the advanced tools) are finally introduced, the student is strong enough to use them as a partner rather than a master. If we get the order wrong, we risk a generation that is incredibly capable with their tools, but completely helpless without them.

In the end, the paper paints a picture of a world where our tools are not just passive helpers, but active partners in a dance. If we lead the dance, we get stronger. If we let the tool lead, we might forget how to dance at all. The choice of which tool to use, and when to use it, determines whether we remain the masters of our own minds or become the instruments of our own machines.

Technical Summary: Competitive and Complementary Tools

Problem Statement
The paper addresses a fundamental question in cognitive science and human-computer interaction: under what conditions does the externalization of thought onto tools (cognitive artifacts) enhance human intelligence, and when does it erode it? While tools have historically extended human capability, there is a growing concern that modern technologies, particularly opaque systems like large language models (LLMs) and satellite navigation, may induce a state of dependency where human competence collapses. The core problem is to mathematically characterize the dynamics between an agent, a tool, and a task to determine why some tools leave users more capable after withdrawal (complementary) while others leave them less capable (competitive).

Methodology
The author models the agent, the tool, and the environment as a single open dynamical system composed of coupled transducers. The system is defined by three internal registers: memory ( $M$ ), operations ( $O$ ), and policy ( $\Pi$ ). The agent holds the policy register and sets the goal.

The model tracks two primary state variables:

Competence ( $\iota \in [0, 1]$ ): The fidelity of the agent's internal model of the task, representing the performance remaining when the tool is withdrawn.
Reliance ( $x \in [0, 1]$ ): The probability that the agent offloads a control step to the tool.

The dynamics are governed by a pair of differential equations (Eqs. 1–2) describing the mean-field limit of a stochastic loop. Key parameters include:

Tool Availability ( $a$ ): The control parameter representing the probability of tool access.
Transparency ( $\tau$ ): Defined as the ceiling of competence ( $\iota_{max}$ ), representing the fraction of the tool's function the user can internally reconstruct. Transparent tools (e.g., maps, abacuses) have $\tau \approx 1$ ; opaque tools (e.g., GPS, complex LLMs) have low $\tau$ .
Atrophy Rate ( $\mu$ ): The rate at which competence decays when steps are offloaded.
Acquisition Rate ( $r$ ): The rate at which competence grows when the agent performs steps independently.

The system incorporates a "scale gate" function, $g(m^*/B)$ , which accounts for the irreducible size of a tool's model ( $m^*$ ) relative to the agent's acquisition budget ( $B$ ). If the model is too large to internalize, the effective ceiling drops to zero.

Key Contributions

Formal Definition of Tool Types: The paper provides a rigorous mathematical distinction between complementary tools (which scaffold internal representations and persist after withdrawal) and competitive tools (which increase reliance and erode internal competence).
Bistability and Hysteresis: The model demonstrates that the coupled dynamics of competence and reliance are bistable. For a range of tool availability, two stable states exist: a "competent" state and a "dependent" state (where competence collapses toward a low floor). The system exhibits hysteresis; once competence collapses due to high availability, reducing availability does not immediately restore competence. A much lower threshold is required to recover the competent state.
The Collapse Threshold: The paper identifies a critical threshold of tool availability ( $a^*$ $a^{*}$ ) that depends jointly on the user's prior competence ( $\iota_0$ $ι_{0}$ ) and the tool's transparency ( $\tau$ $τ$ ).
- If an agent has high prior competence and uses a transparent tool, they can withstand high availability without collapse.
- If an agent has low prior competence or uses an opaque tool, competence collapses even at low levels of availability.
Agentic Reversal: The model extends to goal uncertainty. If a tool's policy register (goal-setting capability) is larger than the human's and the goal is uncertain, the tool effectively becomes the agent, and the human is reduced to an "agentic-instrument." This reversal is irreversible if the tool's model is too large to internalize.
Population Lock-in: The dynamics scale to populations, showing that societies can become trapped in a "collective dependence" state if the initial adoption of a tool occurs before a critical mass of prior competence is established.

Results

Divergence of Capability and Competence: Above a critical availability threshold (the upper fold $a^+$ ), the system enters a regime where "tool-present capability" remains high, but "tool-absent competence" collapses to the dependent floor. This explains the phenomenon where users perform well with a tool but lose the ability to perform the task without it.
Path Dependence: Two users with identical current access to a tool can occupy opposite, lasting states (one competent, one dependent) solely based on the order of learning. A user who builds competence before tool introduction lands in the augmenting attractor; a user introduced to the tool before building competence lands in the collapsed state.
Transparency as a Boundary: The paper maps tools onto a phase plane defined by transparency and task load.
- Augmenting Region: Tools with high transparency (maps, abacuses) allow competence retention.
- Diminishing Region: Tools with low transparency (satellite navigation) or tasks exceeding the "capacity wall" (where the internal surrogate overflows) lead to competence loss.
- LLMs: The model predicts that LLMs can be augmenting if their output is reconstructed and verified (acting as a coach), but diminishing if accepted uncritically or if the model is too large to internalize (acting as a player-agent).
Empirical Validation: The model's predictions are tested against independent datasets, including:
- GPS vs. Maps: GPS users show spatial memory decline, while map users retain navigational knowledge.
- Abacus Experts: Experts retain a mental surrogate (high transparency) despite heavy tool use.
- Pilot Automation: Pilots with low manual practice show a decline in hand-flying skills, exhibiting hysteresis.
- Skill Decay: Cognitive skills decay approximately twice as fast as physical skills when not practiced, supporting the atrophy parameter $\mu$ .

Significance and Claims
The paper claims to reframe the debate on technology and human agency from a qualitative worry into a quantifiable dynamical system. Its primary significance lies in identifying prior competence and tool transparency as the two critical variables determining whether a tool strengthens or weakens a user.

The author argues that the "collapse" of competence is not a failure of the tool but a geometric property of the system's dynamics. Consequently, the paper suggests that:

Education Sequencing: Competence must be built before reliance is introduced. Tools should be introduced only after the user has established a baseline of internal competence to ensure they land in the augmenting attractor.
Tool Design and Selection: Tools should be designed or selected for transparency. Opaque tools (or those with models too large to internalize) remove the possibility of a retained competent state, effectively ceding agency to the tool.
AI Deployment: The widespread deployment of opaque AI tools to users with limited prior competence places society in a regime predicted to result in competence collapse and irreversible agency transfer.

The paper concludes that preserving human agency in the age of AI requires an educational approach that builds competence first and restricts the use of opaque tools until that competence is secure, ensuring that humans remain the policy-setters rather than becoming instruments of computational agents.