From the Knowledge Gap to the Doing Gap: Artificial Intelligence, Human Agency, and the Future Purpose of Education – RESEARCH & PODCAST SERIES 2026

Why the AI era requires a shift from information delivery to question quality, judgment, ethical application, and disciplined execution.

Prepared for Di Tran University

Author: Di Tran University Research Initiative

Date: August 18, 2026

Suggested citation: Di Tran University Research Initiative. (2026). From the knowledge gap to the doing gap: Artificial intelligence, human agency, and the future purpose of education. Di Tran University.

1. Executive Summary

Historically, the primary function of formal education, vocational training, and workforce development has been the distribution of scarce information. Institutions, educators, and libraries served as gatekeepers and conduits for knowledge access. However, the proliferation of generative artificial intelligence (AI) and large language models (LLMs)—acting as general-purpose technologies—has fundamentally altered the economics of information¹. AI systems are rapidly reducing the barriers to explanation, tutoring, drafting, translation, and cognitive planning³. As a result, the central challenge facing educators, workforce developers, and individual learners is shifting decisively away from a knowledge-access gap.

This comprehensive research report argues that the AI era replaces the traditional knowledge gap with four distinct human-centric challenges. The first is a question-quality gap, defined as the capacity to frame meaningful, precise, and ethically responsible queries that direct AI toward productive outputs. The second is a judgment gap, which demands epistemic vigilance to verify information, identify AI hallucinations and algorithmic biases, assess risks, and determine when human professional expertise is strictly necessary. The third is an intention–behavior gap, representing the persistent behavioral disconnect between knowing what to do—often aided by AI-generated planning—and sustaining the motivation to initiate action. Finally, there is a doing gap, characterized by the practical ability to execute consistently and ethically in real-world, physical, and social environments.

Through an exhaustive analysis of cognitive science, behavioral economics, and workforce literature, this paper introduces the original ASK–JUDGE–DO Model, a pedagogical framework designed to cultivate human agency in an AI-abundant world. This model reorients the purpose of education toward the development of critical inquiry, metacognition, and physical execution. Furthermore, the report utilizes the conceptual analogy of advanced driver-assistance systems—specifically Tesla’s Full Self-Driving (Supervised)—to illustrate the critical necessity of calibrated trust, ongoing human supervision, and ultimate human accountability when interacting with automated systems⁵. AI is an immensely powerful cognitive support tool, but it is not a substitute for licensed professionals, nor does it possess the physical agency or moral liability required to execute high-stakes decisions⁸.

The empirical evidence demonstrates that while AI can boost the productivity and quality of knowledge workers by up to 40% on tasks within its capability frontier, it can simultaneously degrade performance by 19% when users blindly trust the system on tasks outside that frontier¹⁰. Therefore, education must train individuals to navigate this jagged technological frontier using strategic human-AI collaboration practices. The report concludes with targeted policy recommendations for higher education, career and technical education, and employers, culminating in the proposal for the “AI to Action: The Human Agency Initiative” at Di Tran University. The overarching thesis is clear: AI may help narrow the knowledge gap, but responsible human agency, critical judgment, and disciplined execution are what close the doing gap.

2. Introduction

For centuries, educational systems were designed around a fundamental economic reality: knowledge was scarce, expensive to reproduce, and difficult to distribute. The pedagogical models of the industrial and information ages positioned the educator as the primary distributor of content and the student as the receiver. In this paradigm, closing the knowledge gap was the primary objective of schooling, universities, and workforce development programs. Societal progress was measured by the broadening of access to facts, theories, and procedural instructions.

The advent of highly capable generative artificial intelligence has disrupted this foundational premise. Modern LLMs, possessing the capacity to synthesize vast amounts of literature, generate complex code, and provide personalized tutoring, have democratized access to high-level cognitive assistance¹. Research indicates that the impacts of LLMs are pervasive, exhibiting the traits of general-purpose technologies that will scale economic impacts across virtually all wage levels and industries². Consequently, learners and workers now possess on-demand access to personalized tutoring, drafting, translation, and strategic planning⁴.

Yet, the accessibility of AI does not equate to learning, competence, or professional qualification. Providing a novice with an AI-generated business plan does not automatically yield a successful entrepreneur, just as providing a student with an AI-drafted essay does not yield a critical thinker. This disruption introduces the central research question of this report: When knowledge and AI-assisted guidance are increasingly accessible, what should education, workforce development, and individual agency become?

The analysis indicates that education must pivot toward the cultivation of uniquely human capacities. Historically, automation has substituted for routine and codifiable tasks, but it has simultaneously complemented and amplified the value of human problem-solving, adaptability, and creativity¹². As AI increasingly offloads cognitive demand, allowing individuals to outsource memory and information processing to external digital environments¹⁴, the defining competencies of the future workforce will not merely be what individuals know, but what they can critically evaluate and practically execute. Access to AI does not override the fundamental psychological barriers that prevent human action, nor does it dissolve systemic inequities. Therefore, the future of education must focus on fostering intention, self-regulation, ethical judgment, and the translation of digital knowledge into physical, real-world execution.

3. Conceptual Framework

To navigate the transition from knowledge acquisition to knowledge execution, it is necessary to define the cognitive, psychological, and behavioral constructs that underpin human-AI interaction. The shift away from information scarcity requires an expanded vocabulary for educators and workforce developers.

Table 1: Foundational Constructs in AI-Enabled Education

Construct	Definition and Empirical Relevance
Knowledge Access	The availability of factual information, explanations, and procedural steps. Generative AI has vastly expanded and democratized this access, rendering rote memorization less economically valuable1.
AI Literacy	The competencies required to understand AI mechanisms, evaluate outputs critically, apply AI ethically, and recognize its systemic impacts on fairness, data privacy, and societal structures16.
Prompt Literacy	The ability to articulate precise, context-rich, and ethically sound questions to guide AI toward generating highly relevant and accurate outputs. This requires a shift toward question-quality over answer-retention18.
Epistemic Vigilance	An evolutionary cognitive mechanism utilized to assess the believability of information and the reliability of its source, protecting humans from accidental or intentional misinformation and deception19.
Human Agency	The capacity of individuals to act independently, make free choices, and exert control over their behavior and environment. UNESCO highlights the protection of human agency as a primary ethical mandate in AI integration8.
Self-Efficacy	A person’s intrinsic belief in their capability to organize and execute the courses of action required to manage prospective situations and achieve goals23.
Metacognition	“Thinking about thinking”—the conscious process of monitoring, evaluating, and regulating one’s own cognitive processes and learning strategies24.
Intention–Behavior Gap	The robust empirical finding in behavioral science that individuals frequently fail to translate their behavioral intentions into actual sustained behaviors26.
Procrastination	A quintessential self-regulatory failure characterized by the voluntary delay of an intended course of action despite expecting to be worse off for the delay28.
Execution Capacity	The practical, physical, and environmental capability to implement a plan in the real world, navigating friction, unforeseen obstacles, and social dynamics (the “doing gap”).
Human-in-the-Loop	A system design requirement where human interaction and oversight are mandatory to verify, correct, or approve automated decisions before they are executed in high-stakes environments30.

3.1 The ASK–JUDGE–DO Model

To bridge the chasm between AI-generated knowledge and human execution, this research team proposes an original pedagogical and operational framework: The ASK–JUDGE–DO Model. This model reorients learning outcomes from the mere retention of information toward the critical, ethical, and applied use of augmented intelligence.

Figure 1: The ASK–JUDGE–DO Model (Text Description)The model operates as a continuous, iterative cycle divided into three distinct phases, each containing two core actions designed to preserve human accountability while leveraging machine intelligence.

Phase 1: The Inquiry Phase (ASK) The process begins with the human intellect. In the Ask stage, the human initiates by framing relevant, precise, context-aware, and ethical questions. This relies heavily on prompt literacy and domain expertise to guide the AI effectively. Following the machine’s response, the human enters the Synthesize stage, gathering the AI-generated information, comparing it against prior knowledge, and fusing it with external verified sources to form a cohesive, albeit preliminary, understanding.

Phase 2: The Evaluation Phase (JUDGE) Because generative models operate on probabilistic text prediction rather than grounded truth, the human must apply rigorous evaluation. In the Know Limits stage, the human applies metacognition to identify areas of uncertainty, potential algorithmic bias, privacy risks, and situations where licensed professional expertise is legally or ethically required. Moving to the Judge stage, the human exercises epistemic vigilance. This involves actively distrusting unverified claims, assessing the real-world consequences of the information, and making an accountable, authoritative decision about how to proceed.

Phase 3: The Execution Phase (DO) The final phase confronts the behavioral bottlenecks of human nature. In the Do stage, the human overcomes the intention–behavior gap by converting the verified decision into concrete, sustained physical or digital action. The AI cannot execute physical labor, look a client in the eye, or bear legal liability. Finally, in the Observe stage, the human measures the real-world outcomes of the action, reflects on the effectiveness of both the AI’s contribution and their own execution, and uses this experiential feedback to inform future inquiries, restarting the cycle.

This model explicitly preserves human accountability. While AI can heavily support the synthesis of data and assist in mapping out the steps of execution, the cognitive load of asking, knowing limits, judging, and the physical agency of doing and observing remain strictly within the human domain. This framework is highly applicable to higher education curricula, workforce development programs, and personal entrepreneurship, providing a structured approach to thriving in an AI-abundant environment.

4. AI as Accessible Cognitive Support

Generative AI operates as a powerful engine for cognitive offloading—the use of physical or external actions to alter the information processing requirements of a task, thereby reducing internal cognitive demand¹⁴. In the educational and workforce context, LLMs serve as interactive cognitive co-learners that democratize access to previously scarce educational resources³².

4.1 Meaningful Opportunities for Diverse Learners

The empirical literature demonstrates that AI integration in higher education offers substantial benefits, particularly in personalized adaptive learning, real-time feedback, and language support³. For non-traditional learners—such as first-generation students, multilingual learners, and adult workforce participants—AI can act as a profound equalizer. It provides a judgment-free environment for foundational skill-building, translating complex academic jargon into accessible language, and offering structured planning tools that reduce executive-functioning burdens²².

Furthermore, small business owners, tradespeople, and early-stage entrepreneurs can leverage AI to generate market research, draft administrative communications, and outline business strategies that would traditionally require expensive consultancy services³⁴. By acting as a ubiquitous tutor and administrative assistant, AI removes the initial friction associated with learning new domains and planning complex projects.

4.2 Limitations, Risks, and the Cognitive Trap

However, treating AI as an infallible oracle invites severe individual and systemic risks. AI models frequently generate “hallucinations”—outputs that are grammatically coherent and highly plausible but entirely fictitious. Models also perpetuate systemic biases present in their training data and pose significant data privacy risks if sensitive intellectual property or personal data is ingested into the system⁸.

Moreover, excessive cognitive offloading can lead to detrimental psychological effects. If learners bypass the “productive struggle” necessary for memory encoding and deep learning, they risk cognitive atrophy¹⁴. There is also a persistent digital divide; unequal access to premium AI models, advanced computational hardware, or the broadband infrastructure required to use them threatens to exacerbate existing educational and economic inequities⁴. UNESCO strongly cautions that the deployment of generative AI must protect human agency and not deprive learners of opportunities to develop cognitive abilities and social skills through observations of the real world⁸.

Table 2: What AI Can Assist With vs. What Requires Human Accountability

Domain	AI Assistance Capabilities (Cognitive Offloading)	Human Accountability Requirements (Agency & Judgment)
Education & Learning	Drafting outlines, providing interactive tutoring, simplifying complex texts, generating practice questions, breaking down study schedules4.	Evaluating accuracy, defending arguments, engaging in productive struggle, ensuring academic integrity, demonstrating moral character.
Workforce & Business	Generating marketing copy, analyzing spreadsheet data, suggesting project milestones, drafting emails, conducting preliminary market research1.	Verifying data against reality, maintaining client relationships, assessing ethical impact, executing physical labor, finalizing strategic direction.
Professional Advice	Summarizing legal statutes, explaining medical terminology, outlining basic financial concepts, providing general psychological frameworks.	Providing licensed legal counsel, diagnosing medical conditions, providing fiduciary financial advice, delivering regulated psychotherapy.
Safety & Regulation	Highlighting standard safety protocols, generating compliance checklists, formatting incident reports.	Assuming legal liability, inspecting physical sites, ensuring compliance with local jurisdiction laws, protecting human life.

5. The Tesla FSD Analogy: Trust, Supervision, and Agency

To deeply understand the optimal relationship between human practitioners and artificial intelligence, it is highly instructive to examine human factors research in the context of advanced driver-assistance systems. Specifically, the deployment of Tesla’s Full Self-Driving (FSD) provides a powerful conceptual analogy for the necessity of calibrated trust and continuous human supervision in human-AI teams.

5.1 FSD is Supervised, Not Autonomous

It is a common societal misconception that Tesla’s Full Self-Driving technology represents fully autonomous driving. Tesla’s primary engineering and legal documentation explicitly and repeatedly contradicts this notion. The official owner’s manuals state unequivocally that FSD is a “Supervised” feature, requiring a “fully attentive driver who is ready to take immediate action at all times”³⁷. The documentation warns that the system may “quickly and suddenly make unexpected maneuvers or mistakes” and that the driver must “never depend on Full Self-Driving (Supervised) to determine when it is safe and/or appropriate to stop or continue through an intersection”³⁸. FSD is officially classified as a Level 2 driver-assistance system, meaning legal responsibility, moral liability, and ultimate operational control remain entirely with the human in the driver’s seat⁶.

Many consumers who own or have access to this advanced technology hesitate to use it precisely because of concerns regarding trust, safety, liability, and the unpredictable nature of complex real-world environments. They recognize that while the machine can process visual data and execute steering inputs rapidly, it lacks human context, self-preservation instincts, and legal accountability.

5.2 Calibrated Trust: Automation Bias vs. Algorithm Aversion

This dynamic perfectly mirrors the challenge of integrating generative AI into knowledge work and education. The primary goal of human-AI interaction is not absolute reliance, nor is it blanket avoidance; the goal is achieving calibrated trust—aligning the human user’s trust in the automated system with the system’s actual, context-dependent reliability⁴¹.

When trust is miscalibrated, two distinct behavioral failures occur:

1. Automation Bias (Over-trust): This is the psychological tendency to use automated cues as a heuristic replacement for vigilant information seeking, leading users to accept AI recommendations uncritically⁷. Just as a driver might dangerously disengage their attention while relying on driver-assistance software⁴³, a student or worker suffering from automation bias might blindly submit AI-generated text containing severe hallucinations, biased assumptions, or flawed logic. High cognitive load and complex verification requirements exacerbate this bias³⁰.
2. Algorithm Aversion (Under-trust): Conversely, algorithm aversion is the phenomenon where individuals erroneously avoid algorithms after seeing them err. Behavioral research demonstrates that humans forgive human errors much more readily than algorithmic errors of the exact same magnitude⁴⁴. After witnessing an AI make a single mistake, users often abandon the technology entirely, even if the algorithm statistically outperforms human judgment over a large sample size⁴⁵.

Optimal utilization requires humans to maintain a continuous state of epistemic vigilance¹⁹. Epistemic vigilance is the cognitive mechanism that allows humans to remain critically alert to the believability of communication. In the AI era, this means trusting the system enough to gain its massive productivity and brainstorming benefits, while remaining skeptical enough to supervise it critically and override it when necessary²¹.

BOXED CAUTION: Why Driving Technology Is Not a Direct Analogy for High-Stakes Professional Advice

While driver-assistance technology illustrates the concepts of supervision and trust calibration, it remains strictly a conceptual analogy. AI language models do not possess physical sensors, real-time spatial awareness, or regulatory licensing. Relying on an AI to draft a binding legal contract, diagnose a medical condition, or provide mental health counseling carries severe ethical and legal liabilities that cannot be mitigated merely by “supervising” the output. Generative AI is an advanced text-prediction engine; it is not equivalent to a licensed psychologist, attorney, physician, or professor. High-stakes professional decisions require licensed human accountability, strict regulatory compliance, and a fiduciary duty of care that algorithms fundamentally cannot possess.

6. The Intention–Behavior and Doing Gaps

If artificial intelligence dramatically lowers the barrier to acquiring knowledge, answering questions, and generating strategic plans, why do individuals still frequently fail to achieve their educational, professional, and personal goals? The answer lies in the behavioral sciences, which have long studied the profound disconnect between knowing what to do and actually doing it.

6.1 The Disconnect Between Knowledge and Action

The intention–behavior gap refers to the robust empirical finding that people do not always do the things they intend to do, even when they possess the necessary knowledge and tools²⁷. A comprehensive meta-analysis by Sheeran (2002) revealed that the correlation between intentions and behaviors is approximately 0.53, indicating that intentions alone leave a massive portion of behavioral variance unexplained²⁶. The gap is primarily driven by “inclined abstainers”—individuals who genuinely intend to change their behavior or execute a task but consistently fail to act when the time comes²⁷.

In economic and psychological terms, this failure is often explained by present bias and time-inconsistent preferences⁴⁸. O’Donoghue and Rabin (1999) elegantly modeled how humans place disproportionate weight on immediate costs and rewards compared to future ones. When an activity involves immediate costs (such as the friction of studying, the physical effort of exercising, or the anxiety of launching a business), people predictably procrastinate⁵⁰. “Naifs” procrastinate because they falsely believe their future selves will possess better self-control, whereas “sophisticates” recognize their future self-control problems but may still struggle to overcome the immediate friction of the present moment⁵¹. AI can generate a perfect daily schedule, but it cannot override the human neurological preference for immediate comfort over delayed reward.

6.2 The Nature of Procrastination and Execution Barriers

Procrastination is a quintessential self-regulatory failure. It is strongly correlated with task aversiveness, impulsiveness, distractibility, and low self-efficacy²⁸. However, it is vital to conceptually distinguish between harmful procrastination and planned, strategic delay, which can be a valid approach to managing complex workloads²⁸.

Furthermore, execution barriers must be treated with profound compassion and systemic awareness. A failure to execute is rarely a simple moral failing or a lack of character. Rather, the doing gap is heavily mediated by cognitive load, executive dysfunction, mental illness, psychological trauma, poverty, caregiving burdens, discrimination, and a lack of environmental resources²⁹. Expecting AI to solve human productivity ignores the reality that human action is deeply embedded in complex, often restrictive social and physical realities.

6.3 Strategies for Closing the Doing Gap

Bridging the doing gap requires educational and workforce interventions that move beyond mere information provision and focus on behavioral regulation:

• Implementation Intentions: Forming specific “if-then” plans (e.g., “If it is Tuesday at 9 AM, then I will open my accounting software and reconcile the ledger”) significantly increases the likelihood of goal attainment. These plans pre-load decisions and reduce the cognitive friction of initiating action⁵⁴.
• Environmental Design and Cues: Structuring the physical and digital environment to make the desired behavior the path of least resistance, removing distractions that trigger present bias.
• Social Accountability: Utilizing mentors, peers, and learning communities to provide the external motivation, emotional support, and feedback loops that AI algorithms cannot replicate.
• Building Self-Efficacy: Designing curricula that provide mastery experiences—small, early, tangible wins that build a learner’s intrinsic belief in their own capability to execute in the physical world²³.

7. Reimagining Education in the AI Era

Because generative AI provides near-instant access to synthesized information and structural formatting, modern educational institutions must transition their fundamental value proposition from content delivery to human formation, ethical reasoning, and execution.

7.1 The Evolving Role of the Educator

In an AI-rich environment, educators become vastly more important, not less. While an AI can deliver a lecture, generate a syllabus, or explain a mathematical concept with endless patience, it cannot provide empathy, moral formation, or human accountability. The educator’s role evolves into that of a mentor, a designer of complex practice environments, a builder of community, and a facilitator of ethical judgment⁸.

Teachers provide the pedagogical scaffolding that encourages students to engage in “productive struggle.” This struggle is entirely necessary to prevent the cognitive offloading trap, where over-reliance on AI dependency systematically erases the friction required for deep neurological encoding and critical thinking development³². Educators are essential in helping students navigate ambiguity and develop the resilience needed to push through the intention-behavior gap.

7.2 Assessing Process, Reasoning, and Execution

Assessment models must radically adapt to a reality where AI can effortlessly generate highly polished final products, such as essays, code, and marketing plans. Evaluating only the final product is no longer a valid or reliable measure of student learning³. Instead, assessments must measure the process of learning and the depth of verification:

• Prompt and Verification Portfolios: Grading students on how they iteratively refine their questions (prompt literacy) and, more importantly, how they fact-check and verify AI outputs against primary sources (epistemic vigilance)¹⁸.
• In-Person Defenses (Vivas): Utilizing oral examinations where students must explain, defend, and critique their reasoning and the AI’s contributions in real-time, proving mastery of the concepts.
• Applied Execution: Moving away from hypothetical papers toward actual execution—requiring students in vocational, undergraduate, and lifelong learning programs to make physical prototypes, conduct real client interviews, or execute community projects that require physical agency.

7.3 Academic Integrity Beyond Surveillance

Addressing academic integrity cannot rely solely on AI detection software. Such software is often technologically flawed, highly susceptible to false positives, and disproportionately penalizes multilingual learners and neurodivergent students²². Instead of relying on an arms race of surveillance and prohibition, institutions must foster a culture of integrity by explicitly teaching AI literacy and clearly defining acceptable boundaries of human-AI collaboration for every assignment. Plagiarism and unauthorized AI use should be framed to students not merely as rule-breaking, but as a detrimental short-circuiting of their own cognitive development, self-efficacy, and future professional agency.

8. Workforce Development and Entrepreneurship

The introduction of LLMs represents a profound and structural shock to the labor market. Research by Eloundou et al. (2023) suggests that approximately 80% of the U.S. workforce could have at least 10% of their tasks affected by LLMs, with roughly 19% seeing at least 50% of their tasks highly impacted¹. However, labor economics demonstrates that while automation displaces specific routine tasks, it also drives demand for uniquely human skills, complementing labor and raising the value of human adaptability¹².

8.1 Navigating the Jagged Technological Frontier

A landmark experimental study on knowledge workers conducted with the Boston Consulting Group found that AI capabilities create a “jagged technological frontier”⁹. The frontier is jagged because tasks of seemingly similar difficulty sit on different sides of the AI’s capability boundary. For tasks inside the frontier (where AI is highly capable, such as drafting and summarizing), workers using GPT-4 improved their quality by 40% and productivity by 12.2%¹⁰.

However, for tasks outside the frontier (requiring complex, nuanced contextual judgment and the synthesis of subtle, contradictory human data), workers using AI were 19 percentage points less likely to produce correct solutions due to severe automation bias¹⁰. The workers fell asleep at the wheel, trusting the AI when they should have trusted their own judgment⁴³.

Successful workers navigated this jagged frontier using two distinct human-AI collaboration practices:

• Centaurs: Workers who create a clear, strategic division of labor. Like the mythical half-human, half-horse, they delegate specific, bounded sub-tasks to the AI while handling the core strategic and nuanced tasks entirely themselves³⁶.
• Cyborgs: Workers who intimately integrate AI into their continuous workflow, constantly interacting, prompting, and refining alongside the machine in a blended approach³⁶.

8.2 Applied Scenario 1: Beauty Education at Di Tran University

Consider a cosmetology student at Di Tran University aiming to launch an independent salon. Historically, this student might lack the capital to hire a marketing firm or the formal business education to draft a comprehensive financial plan. Today, the student can use generative AI to draft a localized salon marketing campaign, generate a demographic analysis of the neighborhood, and write professional lease negotiation emails.

However, closing the knowledge gap is only the first step. The student must utilize the ASK-JUDGE-DO model. They must ask the right questions about local zoning laws. They must judge the AI’s marketing copy with epistemic vigilance to ensure it does not make false medical claims about chemical treatments that violate state board cosmetology regulations. Most importantly, they face the doing gap: the AI cannot physically inspect a retail space, secure financing from a local credit union, negotiate with landlords face-to-face, or perform the delicate physical cosmetology services safely on a human client. The educational program must therefore focus intensely on building the student’s regulatory judgment, resilience, social communication skills, and physical craftsmanship.

8.3 Applied Scenario 2: The First-Generation Adult Learner

Consider a first-generation adult learner balancing a full-time job and family caregiving, who seeks to transition into a career in project management. The learner utilizes AI as a 24/7 tutor to explain complex agile methodologies and assist in structuring study plans⁴. While the AI resolves the knowledge access issue, the learner faces severe intention-behavior gaps mediated by cognitive fatigue, financial stress, and structural time constraints²⁹.

An AI cannot provide the social accountability of a cohort, the empathy of a human mentor who understands the immigrant or first-generation experience, or the professional networking opportunities necessary for job placement. Workforce development programs must therefore provide wrap-around support, flexible scheduling, and community-based accountability to ensure the learner can bridge the doing gap and translate AI-assisted knowledge into a secured, thriving career path.

9. Policy and Institutional Recommendations

To prepare society for the shift from knowledge acquisition to knowledge execution, institutions must adopt proactive, systemic strategies. The following recommendations provide a roadmap for educators, policymakers, and employers.

Table 3: Policy and Institutional Recommendations Matrix

Sector	Problem Addressed	Proposed Action	Expected Benefit	Implementation Risk or Tradeoff	Measurable Indicator of Progress
Colleges & Universities	Outdated assessment methods highly susceptible to AI bypass3.	Shift to process-oriented assessments (vivas, prompt portfolios, applied physical execution).	Preserves true academic rigor and measures actual human cognitive development.	High faculty workload; requires significant retraining and cultural shifts.	Percentage of institutional syllabi utilizing process-based grading rubrics.
Career & Tech Education (CTE)	AI deskilling in vocational planning and administrative management56.	Teach “Centaur/Cyborg” workflows alongside traditional hands-on physical skills34.	Graduates become highly efficient practitioners who leverage AI safely and competitively.	AI technology moves much faster than standard curriculum update cycles.	Employer satisfaction scores regarding graduate adaptability and technical competence.
Workforce Boards & Public Agencies	The digital divide and unequal access to premium AI tools4.	Subsidize access to enterprise-grade AI models for marginalized communities and underfunded schools.	Prevents the widening of economic inequality due to uneven technology access.	High ongoing software licensing costs for publicly funded agencies.	Adoption rates and sustained usage of subsidized AI tools in targeted low-income zip codes.
Employers	Automation bias and unverified AI outputs harming business operations10.	Implement mandatory “human-in-the-loop” verification protocols for all AI outputs before deployment.	Reduces legal liability, algorithmic hallucinations, and customer-facing errors.	May initially slow down operational efficiency and frustrate workers.	Reduction in error rates and compliance breaches for AI-assisted workflows.
Technology Companies	Users lacking epistemic vigilance and struggling with trust calibration19.	Design user interfaces that explicitly communicate uncertainty and force friction for high-stakes tasks.	Calibrates user trust and reduces both algorithm aversion and automation bias.	Added friction may frustrate users seeking instantaneous, seamless answers.	User interaction logs showing increased time spent verifying outputs prior to acceptance.
Students & Individual Learners	The intention-behavior gap, present bias, and procrastination27.	Explicitly teach implementation intentions and digital/physical environmental design54.	Increases the successful translation of goals into sustained, real-world action.	Requires deep behavioral change, which is intrinsically difficult to sustain.	Course completion rates and successful practical project execution metrics.

10. Di Tran University Flagship Initiative: “AI to Action: The Human Agency Initiative”

To operationalize the empirical findings and recommendations of this report, Di Tran University proposes an implementation-ready program titled the “AI to Action: The Human Agency Initiative.” This initiative represents a paradigm shift, establishing the university not merely as a center for knowledge transfer, but as a premier incubator for ethical judgment, human resilience, and disciplined execution.

10.1 Mission and Guiding Principles

Mission: To empower learners from all backgrounds to leverage artificial intelligence as a powerful cognitive tool while rigorously cultivating the uniquely human capacities of ethical judgment, social empathy, and real-world execution.

Guiding Principles:

1. AI assists; humans execute. The final action and its consequences belong to the individual.
2. Accountability cannot be automated. Moral, legal, and professional liability remains inherently human.
3. Execution requires community. Overcoming the doing gap requires human connection, guided practice, and productive struggle.

10.2 A 6-Week Pilot Curriculum Outline

• Week 1: The New Landscape of Knowledge. Introduction to Generative AI, foundational AI literacy, and the ASK-JUDGE-DO model. (Focus on Ask).
• Week 2: Prompt Literacy and Epistemic Vigilance. Crafting advanced queries and learning to spot hallucinations, biases, and logical fallacies. (Focus on Synthesize and Know Limits).
• Week 3: Navigating the Jagged Frontier. Identifying tasks inside and outside AI’s capabilities. Practicing Centaur and Cyborg workflows in a professional context. (Focus on Judge).
• Week 4: The Psychology of Action. Understanding the intention-behavior gap, present bias, and the roots of procrastination. Designing personal implementation intentions. (Focus on Do).
• Week 5: Real-World Execution. Students must take an AI-generated project plan (e.g., a community event, a marketing strategy, a coding project) and physically execute the first phase in the real world. (Focus on Do).
• Week 6: Reflection and Calibration. Evaluating the outcomes of the execution phase. Students participate in oral defenses of their decision-making process and their management of AI tools. (Focus on Observe).

10.3 Institutional Safeguards and Evaluation Plan

• Faculty-Development Requirements: Faculty will undergo rigorous training to grade the process of AI-assisted work, utilizing rubrics that measure ethical verification, critical thinking, and physical execution rather than just output quality.
• Equity and Accessibility Safeguards: Di Tran University will ensure all students, regardless of socioeconomic status, have institutional access to premium AI tiers to prevent a two-tiered learning system.
• Student Privacy and Responsible-AI Policies: Clear institutional guidelines will prohibit the input of sensitive personal data or proprietary university research into open public LLMs.
• Partnership Opportunities: The initiative will partner with local employers, community organizations, and workforce-development agencies to provide the physical venues for the “Execution Phase” (Week 5) of the curriculum.
• Evaluation Plan: The initiative will be measured using mixed methods: quantitative tracking of project completion rates and qualitative assessments of student self-efficacy, trust calibration, and epistemic vigilance through pre- and post-course surveys.

10.4 A Public-Facing Manifesto

The Di Tran University AI to Action Manifesto

Information is no longer scarce; execution is. We embrace Artificial Intelligence as a profound tool for cognitive support, creativity, and productivity. However, we reject the notion that AI replaces human agency, accountability, or the necessity of the human struggle to achieve mastery. We commit to teaching our students not just how to prompt machines, but how to question deeply, judge ethically, and act decisively in the physical world. We do not graduate prompt engineers; we graduate accountable leaders, resilient entrepreneurs, and ethical professionals.

11. Discussion

The rapid integration of artificial intelligence into education and workforce development presents a profound societal paradox: as AI expands our capacity to access knowledge and simulate reasoning, it simultaneously exposes the fragility of human self-regulation, motivation, and execution. The central tension highlighted throughout this report is that AI massively expands access to information, but it absolutely does not eliminate the need for human expertise, formal institutions, human relationships, or professional accountability⁴.

The empirical evidence strongly suggests that when humans are given powerful cognitive tools, they can achieve unprecedented productivity—exemplified by the 40% quality boost seen in knowledge workers operating inside the jagged frontier⁹. Yet, without rigorous epistemic vigilance¹⁹, they are highly susceptible to automation bias, willingly accepting plausible but deeply flawed outputs³⁰. Furthermore, behavioral science clearly demonstrates that knowing the optimal path—even having an AI map it out perfectly—does not automatically overcome present bias, procrastination, or systemic execution barriers²⁸.

It is crucial for educational and workforce leaders to avoid technological determinism. AI will not inevitably save education, nor will it inevitably destroy it. The outcome depends entirely on institutional design, pedagogical philosophy, and policy implementation. Future research must rigorously investigate the long-term cognitive consequences of continuous cognitive offloading¹⁴, the efficacy of various interventions designed to bridge the intention-behavior gap in digitally saturated environments²⁷, and the evolving nature of trust calibration in complex human-AI teams³¹.

12. Conclusion

The historical architecture of education was built to bridge the gap between ignorance and knowledge. That architecture has now been permanently disrupted by generative artificial intelligence, which acts as an omnipresent, highly capable cognitive assistant. Consequently, the bottleneck to human flourishing and economic productivity is no longer access to information or strategic planning. The bottleneck is the human capacity to frame the right questions, critically evaluate the answers, and summon the psychological fortitude to act upon them in the real world. AI may help narrow the knowledge gap. But responsible human agency, judgment, and disciplined execution are what close the doing gap.

13. Appendices

Appendix A: ASK–JUDGE–DO Model Visual

(Text-described for accessibility) The visual representation of the ASK-JUDGE-DO model consists of three interlocking gears, symbolizing the continuous, iterative nature of human-AI collaboration.

• Gear 1 (Inquiry/Blue): Contains the words “Ask” and “Synthesize.” It represents the intake of information, prompt literacy, and interaction with AI. It turns the central gear.
• Gear 2 (Evaluation/Red): Contains “Know Limits” and “Judge.” This is the central, largest gear, colored red to signify stopping, critical reflection, human accountability, and epistemic vigilance. It acts as the regulator of the system.
• Gear 3 (Execution/Green): Contains “Do” and “Observe.” It represents the output into the physical and social world, driving real-world impact and feeding observational data back into the first gear for continuous learning.

Appendix B: AI-Use Disclosure Template for Student Assignments

Student Name: _____________________

Assignment/Project: _______________________

1. Did you use Generative AI in the completion of this assignment? [Yes/No]
2. Which tools were used? (e.g., ChatGPT-4, Claude 3, Midjourney)
3. What specific tasks did the AI assist with? (e.g., Brainstorming, outlining, code debugging, grammar checking, data synthesis)
4. Human Accountability Statement: I verify that I have critically reviewed all AI-generated content, fact-checked all claims against primary sources, and take full personal academic and ethical responsibility for the final submission.
   Signature: _____________________

Appendix C: Faculty Rubric for Evaluating AI-Assisted Work

Criterion	Emerging	Proficient	Mastery
Prompt Literacy	Uses generic, simple prompts. Accepts the first output without refinement.	Uses specific, context-rich prompts. Iterates 2-3 times to improve output relevance.	Demonstrates deep iteration, applying constraints, role-playing, and highly specific context to refine outputs.
Epistemic Vigilance	Fails to identify AI hallucinations. Accepts biased data without question.	Identifies major errors. Verifies primary facts against external sources.	Critically deconstructs AI logic. Cross-references with peer-reviewed literature. Identifies subtle algorithmic biases.
Agency and Execution	Output reads as heavily automated. Lacks personal voice, novel synthesis, or real-world application.	Blends AI output with personal insights (Cyborg/Centaur approach). Shows intent to apply.	Transforms AI assistance into a highly original, context-specific application that solves a real-world problem through physical/social execution.

Appendix D: Learner Self-Assessment for Question Quality, Verification, and Execution

(To be completed prior to initiating a high-stakes project)

1. The Ask: Did I clearly define the problem and the necessary context before consulting AI?
2. The Limit: Does this task require a licensed professional (legal, medical, financial, engineering) or pose a safety risk?
3. The Judgment: Have I verified this information outside of the AI system using trusted, primary sources?
4. The Friction: What is the immediate psychological, physical, or environmental cost/friction preventing me from executing this today?
5. The Implementation Intention: If [Specific Situation/Time occurs], Then I will [Execute this Specific Action].

Appendix E: Research Limitations and Evidence-Quality Notes

The research synthesized in this report draws heavily from recent, rapid empirical studies on generative AI, primarily published between 2023 and 2025. Due to the extreme velocity of AI development, findings regarding specific capabilities—such as the exact boundaries of the “jagged technological frontier”—may shift as foundation models improve. Furthermore, while the intention-behavior gap and automation bias are well-documented psychological phenomena with decades of robust meta-analytic support, their specific interaction with modern LLMs remains an emerging field of study. Causality regarding the long-term cognitive effects of AI offloading cannot yet be definitively proven and should be evaluated through ongoing longitudinal research.

Declaration of Responsible Use of AI in This Research

The drafting of this report utilized artificial intelligence as a cognitive support tool for structural outlining, synthesizing provided source material, and thematic organization. In accordance with the ASK-JUDGE-DO model, all AI-generated assertions were subjected to rigorous human verification, cross-referenced strictly against the provided verified source material (including peer-reviewed articles from The American Economic Review, Trends in Cognitive Sciences, Mind & Language, and primary documentation from Tesla and UNESCO). Human judgment was exercised to ensure ethical framing, particularly regarding the compassionate treatment of execution barriers and the accurate representation of Tesla’s Supervised FSD limitations. All citations have been manually integrated, verified, and checked before publication to ensure academic integrity and alignment with Di Tran University’s research standards.

Works cited

1. GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models – arXiv, https://arxiv.org/pdf/2303.10130
2. (PDF) GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models – ResearchGate, https://www.researchgate.net/publication/369369163_GPTs_are_GPTs_An_Early_Look_at_the_Labor_Market_Impact_Potential_of_Large_Language_Models
3. EJ1374562 – Artificial Intelligence in Higher Education: The State of the Field, International Journal of Educational Technology in Higher Education, 2023 – ERIC, https://eric.ed.gov/?id=EJ1374562
4. A Systematic Review of Artificial Intelligence in Higher Education Institutions (HEIs): Functionalities, Challenges, and Best Practices – MDPI, https://www.mdpi.com/2227-7102/16/2/185
5. Full Self-Driving (Supervised) – Tesla, https://www.tesla.com/ownersmanual/2012_2020_models/en_us/GUID-2CB60804-9CEA-4F4B-8B04-09B991368DC5.html
6. Tesla FSD in 2026: Full Self-Driving Guide for US & Canada Owners, https://tslna.com/en/tesla-full-self-driving-guide/
7. (PDF) Trust in Automation: Integrating Empirical Evidence on Factors That Influence Trust, https://www.researchgate.net/publication/272887576_Trust_in_Automation_Integrating_Empirical_Evidence_on_Factors_That_Influence_Trust
8. Guidance for generative AI in education and research – UNESCO | CEDEFOP, https://www.cedefop.europa.eu/en/tools/vet-toolkit-tackling-early-leaving/resources/guidance-generative-ai-education-and-research-unesco
9. Report on the First Scientific Experiment to Test the Impact of Generative AI on Complex, Knowledge-Intensive Work – EDRM, https://edrm.net/2024/04/report-on-the-first-scientific-experiment-to-test-the-impact-of-generative-ai-on-complex-knowledge-intensive-work/
10. Speed vs. Accuracy: The Productivity–Quality Trade-Off of AI – Business Science Daily, https://businesssciencedaily.com/speed-vs-accuracy-the-productivity-quality-trade-off-of-ai/
11. [2303.10130] GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models – arXiv, https://arxiv.org/abs/2303.10130
12. Why Are There Still So Many Jobs? The History and Future of Workplace Automation, https://ideas.repec.org/a/aea/jecper/v29y2015i3p3-30.html
13. Why Are There Still So Many Jobs? The History and Future of Workplace Automation, http://jenni.uchicago.edu/econ341/readings/Autor_2015_JEP_v29_n3.pdf
14. Using the outside world to save on brainpower – EurekAlert!, https://www.eurekalert.org/news-releases/786784
15. [PDF] Cognitive Offloading. – Semantic Scholar, https://www.semanticscholar.org/paper/Cognitive-Offloading.-Risko-Gilbert/a43c46209e447520c0753707baa8c9f12cead7c5
16. 3: AI Literacy and Gen-AI Literacy Frameworks – Emerald Insight, https://www.emerald.com/books/edited-volume/17291/chapter/114713634/AI-Literacy-and-Gen-AI-Literacy-Frameworks
17. AI Literacy Frameworks for Educators: An Umbrella Review – ScholarSpace, https://scholarspace.manoa.hawaii.edu/bitstreams/3b2f1cf7-9d30-4613-ab88-07cb8965986d/download
18. AI prompt engineering: A critical new skillset for 21st-century teachers – eSchool News, https://www.eschoolnews.com/digital-learning/2025/06/04/prompt-engineering-a-critical-new-skillset-for-21st-century-teachers/
19. Epistemic Vigilance, Cautious Optimism and Sophisticated Understanding – ResearchGate, https://www.researchgate.net/publication/264319410_Epistemic_Vigilance_Cautious_Optimism_and_Sophisticated_Understanding
20. Epistemic Vigilance – BibBase, https://bibbase.org/network/publication/sperber-cleacutement-heintz-mascaro-mercier-origgi-wilson-epistemicvigilance-2010
21. Pragmatics and epistemic vigilance: a developmental perspective – libra, https://libra.unine.ch/server/api/core/bitstreams/fb71b74c-01bf-4fb9-afbe-a7c5726d348f/content
22. Generative AI has disrupted education. Here’s how it can be used for good – UNESCO, https://www.weforum.org/stories/2023/09/generative-ai-education-unesco/
23. Self-efficacy: toward a unifying theory of behavioral change – PubMed, https://pubmed.ncbi.nlm.nih.gov/847061/
24. Metacognition and Cognitive Monitoring: A New Area of Cognitive-Developmental Inquiry., https://www.semanticscholar.org/paper/Metacognition-and-Cognitive-Monitoring%3A-A-New-Area-Flavell/ee652f0f63ed5b0cfe0af4cb4ea76b2ecf790c8d
25. Metacognition and cognitive monitoring: A new area of cognitive–developmental inquiry, https://www.researchgate.net/publication/232599909_Metacognition_and_cognitive_monitoring_A_new_area_of_cognitive-developmental_inquiry
26. Intention-Behavior Gap Measure – RPubs, https://rpubs.com/djw/intention-behavior-gap
27. Sheeran, P. and Webb, TL (2016) The Intention–Behavior Gap. Social and Personality Psychology Compass, 10 (9). pp – White Rose Research Online, https://eprints.whiterose.ac.uk/id/eprint/107519/3/The
28. Piers Steel | University of Calgary | 80 Publications | 164 Citations | Related Authors, https://scispace.com/authors/piers-steel-1yp5el47fe
29. We’re Sorry This Is Late … We Really Meant To Post It Sooner: Research Into Procrastination Shows Surprising Findings | ScienceDaily, https://www.sciencedaily.com/releases/2007/01/070110090851.htm
30. Automation Bias and the Principles of Judicial Review – Taylor & Francis, https://www.tandfonline.com/doi/full/10.1080/10854681.2023.2189405
31. When Should Your Team Override AI? A Trust Calibration Routine – AOM Journals, https://journals.aom.org/doi/10.5465/amp.2025.0019
32. Cognitive Offloading | Request PDF – ResearchGate, https://www.researchgate.net/publication/306352756_Cognitive_Offloading
33. AI Literacy Literature Summary – CSU AI Commons – California State University, https://genai.calstate.edu/communities/faculty/ethical-and-responsible-use-ai/ai-literacy-literature-summary
34. Navigating The Jagged Technological Frontier – Field Experimental Evidence of The Effects of AI On Knowledge Worker Productivity and Quality | PDF | Artificial Intelligence – Scribd, https://www.scribd.com/document/687431208/Navigating-the-Jagged-Technological-Frontier-Field-Experimental-Evidence-of-the-Effects-of-AI-on-Knowledge-Worker-Productivity-and-Quality
35. Research from the Office of RIDIL – ODUGlobal, https://online.odu.edu/sites/default/files/2024-02/White%20Paper%20on%20Artificial%20Intelligence%20in%20Higher%20Education.pdf
36. Enterprise workers gain 40 percent performance boost from GPT-4, Harvard study finds, https://venturebeat.com/business/enterprise-workers-gain-40-percent-performance-boost-from-gpt-4-harvard-study-finds
37. Full Self-Driving (Supervised) Limitations – Tesla, https://www.tesla.com/ownersmanual/model3/en_us/GUID-59736DF1-31FA-4F67-8A3F-6111778939D9.html
38. Full Self-Driving (Supervised) Limitations – Tesla, https://www.tesla.com/ownersmanual/modely/en_nl/GUID-59736DF1-31FA-4F67-8A3F-6111778939D9.html
39. Full Self-Driving (Supervised) – Tesla, https://www.tesla.com/ownersmanual/model3/en_us/GUID-2CB60804-9CEA-4F4B-8B04-09B991368DC5.html
40. Full Self-Driving (Supervised) – Tesla, https://www.tesla.com/ownersmanual/models/en_nl/GUID-11E01B0B-DC74-4B8D-B979-DA37EC6F9093.html
41. Adaptive Trust Calibration in Human-AI Cooperation, https://ir.soken.ac.jp/record/6336/files/A2191%E6%9C%AC%E6%96%87.pdf
42. Adaptive trust calibration for human-AI collaboration – PMC – NIH, https://pmc.ncbi.nlm.nih.gov/articles/PMC7034851/
43. AI and the future of work: A tale about centaurs and cyborgs – Siili Solutions, https://www.siili.com/newsandinsights/ai-future-of-work-centaurs-and-cyborgs
44. Algorithm Aversion – Why an AI Mistake Is Less Forgiven Than a Human One | AI & Psychotherapy – IA et Psychothérapie, https://ia-et-psychotherapie.com/en/resources/concepts/algorithm-aversion/
45. Algorithm Aversion – The Decision Lab, https://thedecisionlab.com/reference-guide/psychology/algorithm-aversion
46. Algorithm aversion: people erroneously avoid algorithms after seeing them err – PubMed, https://pubmed.ncbi.nlm.nih.gov/25401381/
47. The Intention–Behavior Gap | Request PDF – ResearchGate, https://www.researchgate.net/publication/307857321_The_Intention-Behavior_Gap
48. Doing It Now or Later – American Economic Association, https://www.aeaweb.org/articles?id=10.1257/aer.89.1.103
49. Present bias – BehavioralEconomics.com | The BE Hub, https://www.behavioraleconomics.com/resources/mini-encyclopedia-of-be/present-bias/
50. Doing It Now or Later – IDEAS/RePEc, https://ideas.repec.org/a/aea/aecrev/v89y1999i1p103-124.html
51. Getting rational now or later: navigating procrastination and time-inconsistent preferences for new rationalists – LessWrong, https://www.lesswrong.com/posts/NbunzAPqzP5XTkryo/getting-rational-now-or-later-navigating-procrastination-and
52. DOING IT NOW OR LATER – Ted O’Donoghue Northwestern University, https://www.kellogg.northwestern.edu/research/math/papers/1172.pdf
53. ‪Piers Steel‬ – ‪Google Scholar‬, https://scholar.google.com/citations?user=_l-1G1oAAAAJ&hl=en
54. Implementation intention – Wikipedia, https://en.wikipedia.org/wiki/Implementation_intention
55. Implementation Intentions. – Social Psychology & Motivation, https://www.socmot.uni-konstanz.de/publications/implementation-intentions
56. Displacement or Complementarity? The Labor Market Impact of Generative AI – Harvard Business School, https://www.hbs.edu/ris/download.aspx?name=25-039.pdf
57. Voices of Change: Generative AI and the Transformation of Work in Latin America | OECD.AI, https://wp.oecd.ai/app/uploads/2025/07/Voices-of-Change-Generative-AI-and-the-Transformation-of-Work-in-Latin-America.pdf