🤖 AI

SCAN: A Decision-Making Framework for Effective Task Allocation with Generative AI

This paper introduces SCAN, a human-centric framework grounded in Vygotsky's Zone of Proximal Development and metacognition that guides learners and knowledge workers in effectively allocating tasks to Generative AI through four distinct sub-zones (Substitute, Complement, Aid, and Non-negotiable) to foster hybrid intelligence and lifelong learning.

Original authors: Fendi Tsim, Alina Gutoreva

Published 2026-06-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Fendi Tsim, Alina Gutoreva

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a chef trying to cook a complex new dish. You have a powerful, super-smart kitchen assistant (Generative AI) who can chop vegetables, mix sauces, and even suggest recipes instantly. But here's the problem: if you let the assistant do everything, you never learn to cook. If you refuse to let the assistant help at all, you might burn the dinner or take three hours to make a simple salad.

This paper introduces SCAN, a simple mental checklist to help you decide exactly how to use your AI assistant so you get the job done and keep getting smarter.

Think of SCAN as a "Cognitive GPS" that helps you navigate four different zones of work. It's based on an old idea from a psychologist named Vygotsky, who said we learn best when we do things that are just a little bit too hard for us to do alone, but easy enough to do with a little help.

Here is how the four zones of SCAN work, using the Chef and the Kitchen Assistant analogy:

1. The "Substitute" Zone (S) – The "Let the Robot Do It" Zone

The Situation: You have no idea how to make this dish. You've never seen the ingredients, and you don't know the steps.
The AI Role: The AI takes over completely. It cooks the whole meal.
The Risk: If you do this too often, you become a "passenger" in your own kitchen. You aren't learning; you are just watching. The paper warns that if you use AI here without understanding the basics, you might start believing whatever the AI says (even if it's wrong) because you don't have the knowledge to check it.
When to use it: Only for tasks you truly cannot do at all, and only if you accept that you aren't learning the skill right now.

2. The "Aid" Zone (A) – The "Training Wheels" Zone

The Situation: You know the basics of cooking, but this specific recipe is tricky. You know how to chop an onion, but you don't know how to balance the spices for this specific sauce.
The AI Role: The AI acts as a scaffold (like training wheels). It gives you hints, suggests the spice ratios, or shows you a step-by-step guide. You still do the chopping and the stirring.
The Benefit: This is the "sweet spot" for learning. The AI fills in the gaps so you don't get overwhelmed, but you are still the one doing the work. You are stretching your skills just enough to grow.
When to use it: When you are learning something new and need a little push to get over the hump.

3. The "Complement" Zone (C) – The "Power Team" Zone

The Situation: You are an expert chef. You know exactly how to make this dish. You could do it in your sleep.
The AI Role: The AI acts as a super-speed assistant. It might chop the vegetables in seconds or find a rare ingredient for you. You check its work, but you are the one in charge.
The Benefit: This is pure efficiency. You and the AI work together as a team. You save time, and because you are an expert, you can spot if the AI makes a silly mistake immediately.
When to use it: When you already know the skill well and just want to get the job done faster.

4. The "Non-negotiable" Zone (N) – The "Human Only" Zone

The Situation: The task requires deep human judgment, empathy, or ethical responsibility. For example, deciding who gets a promotion, comforting a crying child, or making a final ethical call on a medical diagnosis.
The AI Role: The AI stays out of the kitchen.
The Reason: Some things require a human heart and a human conscience. An AI can't truly understand the weight of a decision or the nuance of a human relationship.
When to use it: Always. These are the tasks that must remain between humans (like a mentor and a student) because they build your character and professional identity.

How to Use SCAN (The "Scan" Part)

The paper suggests that before you start a task, you should pause and "scan" your own brain. Ask yourself two simple questions:

Do I know how to do this?
Do I need help, and if so, what kind?

If you know nothing: Substitute (but be careful not to get lazy).
If you know a little but need a boost: Aid (this is where the magic learning happens).
If you are an expert: Complement (let the AI speed you up).
If it's a human-to-human moment: Non-negotiable (keep the AI out).

Why This Matters (The "Upskilling" vs. "Deskilling" Trap)

The paper warns about a trap called Deskilling. If you always put tasks in the "Substitute" zone (letting AI do everything), you stop learning. Eventually, you forget how to cook at all.

However, if you use the Aid and Complement zones correctly, you get Upskilling. You move from needing help to becoming an expert. Over time, a task that used to be in the "Aid" zone (hard for you) moves to the "Complement" zone (easy for you).

The ultimate goal of SCAN is Hybrid Intelligence. This doesn't mean AI replacing humans. It means humans and AI working together so well that the combination is smarter than either one alone. But for this to work, the human must stay in the loop, constantly checking their own knowledge and making sure they are learning, not just copying.

In short: SCAN is a rulebook to make sure you use AI as a tool to grow your brain, not as a crutch to replace it. It helps you decide when to let the AI drive, when to let it steer, and when to keep your hands firmly on the wheel.

Technical Summary: SCAN – A Decision-Making Framework for Effective Task Allocation with Generative AI

Problem Statement

The rapid integration of Generative AI (GenAI) into knowledge work and education has introduced a critical challenge: ineffective task allocation. While GenAI offers significant potential as a scaffold to augment human capabilities, empirical evidence indicates that without proper self-assessment, users fall into patterns of over-reliance (automation bias), under-utilization, or "cognitive offloading" that erodes intrinsic motivation and metacognitive abilities. The core problem is not GenAI's capability, but the individual's inability to assess their own knowledge state relative to a task. Consequently, learners cannot determine where their competence ends and where AI assistance genuinely adds leverage, leading to arbitrary delegation that may result in deskilling rather than upskilling. Existing heuristics often describe the outcomes of task assignment (e.g., "automated" vs. "delegated") rather than providing a systematic procedure rooted in the learner's real-time knowledge state.

Methodology and Theoretical Framework

The paper introduces SCAN, a human-centric decision-making framework designed to systematize AI-human interaction. SCAN is grounded in two psychological constructs:

Vygotsky's Zone of Proximal Development (ZPD): Specifically, the distinction between what a learner can do independently (Zone of Actual Development), what they can do with assistance (ZPD), and what is currently beyond their reach (Zone of Potential Development).
Metacognition: Defined as the awareness and regulation of one's own cognitive processes (knowledge, monitoring, and control).

The SCAN Task Paradigm
The framework extends the traditional ZPD by introducing a fourth dimension: "Known to GenAI." This addition creates a Venn diagram-based task paradigm comprising four distinct sub-zones for task allocation:

Substitute (S): The intersection of "Unknown to Learner" and "Known to GenAI." Here, the learner lacks task-specific knowledge entirely. GenAI performs the task via automation. This zone carries high cognitive risk, as the learner cannot verify outputs, leading to potential cognitive offloading and automation bias.
Complement (C): The intersection of "Known to Learner" and "Known to GenAI." The learner possesses sufficient task-specific knowledge to complete the task independently and verify outputs. GenAI is used to reduce time and allocate cognitive resources to higher-order tasks, fostering human-AI collaboration.
Aid (A): The intersection of "ZPD" (Some but insufficient knowledge) and "Known to GenAI." The learner has partial knowledge but requires GenAI as a scaffold (e.g., providing general knowledge, structure, or hints) to bridge the gap. This represents human-GenAI augmentation.
Non-negotiable (N): Tasks located in the original ZPD that require a "More Knowledgeable Other" (MKO) who is human, not AI. These tasks involve tacit knowledge, ethical judgment, accountability, or relational attunement that GenAI cannot replicate.

Metacognitive Mechanism
SCAN operationalizes metacognition as a control layer. It requires learners to perform a "cognitive scan" involving:

Metacognitive Knowledge: Assessing task demands and personal competence before starting.
Metacognitive Monitoring: Real-time evaluation of task execution and AI output reliability to detect misalignments (e.g., fluency-accuracy dissociation).
Metacognitive Control: Strategic adjustment of delegation (e.g., shifting from automation to augmentation) based on performance feedback.

Key Contributions

Formalization of Task Allocation: SCAN provides a structured, four-zone taxonomy (Substitute, Complement, Aid, Non-negotiable) that moves beyond binary "human vs. AI" distinctions to a nuanced spectrum of interaction modes.
Integration of ZPD and GenAI: The paper theoretically extends Vygotsky's ZPD by explicitly mapping GenAI's capabilities against human knowledge states, distinguishing between "digital scaffolds" (GenAI) and human MKOs.
Differentiation of Learning Trajectories: The framework distinguishes between two developmental paths:
- Internalization Trajectory (Human-Human): Occurring in the Non-negotiable zone, leading to autonomous self-regulation and professional identity.
- Symbiosis Trajectory (Human-AI): Occurring in the S, A, and C zones, where the learner and GenAI form a symbiotic cognitive unit, aiming for "Hybrid Intelligence" rather than full independence from the tool.
Cognitive Load Analysis: The paper maps the four sub-zones to Cognitive Load Theory (CLT), demonstrating how different allocations affect intrinsic, extraneous, and germane load. For instance, the Aid zone is identified as pedagogically optimal because it reduces extraneous load while activating germane processing.
Upskilling and Deskilling Dynamics: SCAN offers a mathematical illustration of how task migration across sub-zones over time ( $S \to A \to C$ ) leads to upskilling, while regression ( $C \to A$ ) or static over-reliance in $S$ leads to deskilling.

Results and Application Scenarios

The paper does not present empirical experimental results but rather a theoretical demonstration of the framework's applicability in two primary scenarios:

Knowledge Workers: SCAN serves as a decision protocol where workers classify tasks based on domain competence and AI capability. It advocates for "reversible delegation," ensuring that automation (Substitute) is only used when capability thresholds are met and human oversight (in the loop) is maintained to prevent skill erosion.
Education: SCAN operationalizes ZPD by requiring students to classify sub-tasks, justify their choices via metacognitive notes, and maintain an audit trail (prompts, drafts, revisions). This shifts assessment focus from the quality of the AI artifact to the student's ability to unassistedly perform the task after the learning cycle.

The framework also addresses the "Augmentation-Automation Paradox" and the "Upskilling-Deskilling Paradox," suggesting that these phenomena are driven by the dynamic migration of tasks between sub-zones based on the learner's evolving metacognitive calibration.

Significance and Claims

The authors claim that SCAN offers a "great starting point" for discussing whether GenAI complements or replaces human abilities. Its primary significance lies in:

Sustaining Lifelong Learning: By ensuring that task allocation preserves the conditions for growth (e.g., keeping tasks in the Aid or Complement zones rather than defaulting to Substitute), SCAN aims to prevent deskilling.
Achieving Hybrid Intelligence: The framework proposes a path toward hybrid intelligence, where humans and AI collaborate effectively, with the human retaining agency and epistemic responsibility.
Human-in-the-Loop: SCAN reinforces the necessity of human presence in the decision-making loop, ensuring that learners actively evaluate AI outputs rather than passively accepting them.
Future-Proofing: The authors posit that SCAN is applicable not only to current GenAI but also to future advanced forms of AI, such as Artificial General Intelligence (AGI), by grounding interaction in fundamental cognitive and developmental principles.

The paper concludes modestly, acknowledging that SCAN requires a level of metacognitive capacity that novices may lack, suggesting that educators or supervisors must initially guide the classification process. It identifies future research directions, including mathematical formalization of the framework and empirical comparisons between human-centric and AI-centric task allocation approaches.