A Human Agency Testfor AI Adoption

Executive Summary

It used to be that an organization could adopt a technology, train staff, set a policy and move on. With AI adoption this decision is much more far reaching. It is now a decision about the future distribution of knowledge, capability, decision-making, and human development inside an organization.

AI already has a massive global, organizational, and cognitive reach. The global reach analysis estimates that more than 3 billion people can access some form of AI, frontier models operate across more than 100 languages, and AI is being integrated into nearly every major sector, including government, healthcare, education, churches, ministries, and humanitarian organizations. The next wave of compute infrastructure is designed to extend both the scale and the autonomy of these systems.

This AI progress raises important organizational questions. AI ethics commonly asks whether a human consented, whether a human approved an AI output, whether the model was transparent, and whether a decision could be appealed. Those safeguards are necessary. But they are no longer sufficient. By the time a choice occurs, an AI system may already have shaped which options were available, how they were framed, what evidence was seen as relevant, what kind of expertise is needed, at what speed, and what the institution should consider a success.

The Human Agency Test

Before approving an AI system, leaders should be able to answer five questions:

1. A Question at the Center of AI Governance

AI is shaping humans and institutions before they have thought through its role.

Technology used to be the concern of select teams in an organization such as IT or HR. But with the advent of AI, it is becoming a general layer through which many employees search, write, learn, translate, allocate resources, counsel, plan, interpret evidence, and lead. The global reach of AI spans across every continent, across major languages, and across nearly every organizational category. AI increasingly shapes what people read, write, think, learn, decide, create, pray, and communicate.

The physical infrastructure beneath AI reinforces this point. The immense expansion of chips, data centers, electricity, networks, land, water, and capital are terrestrial expressions of AI’s vast reach. Ownership of compute is highly concentrated, which means many organizations may use systems they have not built or inspected. What this means is that a ministry, nonprofit, school, hospital, or public institution may be able to access AI but without having any transparency.

2. Disempowerment Often Occurs Upstream of Human Choice

Governance and policy tend to focus on the moment of a decision: Did the user consent? Could the recommendation be rejected? Was a human involved? Was the model transparent? Could the outcome be appealed? Those questions matter because they can preserve procedural control. But this focus could hide the loss of human and organizational agency.

Before the user acts, the AI system may already have influenced:

• which options are visible and which alternatives are no longer an option;
• how the problem is framed and which evidence is treated as relevant;
• what the person should expect, prefer, fear, or consider normal;
• how much time is available for reflection or dissent;
• which human skills should still be practiced, funded, and rewarded;
• which forms of expertise should exist inside the organization;
• which goals the system is optimized to serve.

A person can retain the right to click, reject, customize, or approve while the organizational environment has already narrowed the field from which she chooses. The resulting preference may feel sincere. The resulting decision may be voluntary. Yet the person or organization may possess little authority over the conditions that produced it.

Seven potential organizational risks

• The illusion of control: Leaders sign final decisions but cannot independently evaluate the analysis, alternatives, or assumptions within the AI systems.
• Adaptive preferences: People gradually prefer what the system repeatedly makes visible, convenient, or normative.
• Capability asymmetry: The system continually improves while humans stop learning and growing, making further delegation appear increasingly rational.
• Environmental manipulation: Interfaces, incentives, and persuasion shape behavior in ways humans cannot completely understand.
• Symbolic human participation: Human review of system output becomes a mere ritual rather than an exercise of wisdom and judgment.
• Value compression: Complex human goals and purposes are reduced to measurable units; what cannot be measured receives less weight.
• Unequal disempowerment: People with fewer resources, lower literacy, trauma histories, disabilities, linguistic barriers, precarious employment, or limited alternatives bear greater risk of being persuaded by AI and possess less power to refuse.

3. Organizations Must Protect Human Influence At All Levels

Human influence is a necessity for human benefit

Influence is easier to maintain than to rebuild. Once people no longer know how to perform the work, organizations no longer employ people who understand it, and institutions no longer fund the infrastructure supporting it, human oversight becomes mostly symbolic. The question is not whether a human can intervene. It is whether a human will still exist, remain available, possess sufficient authority, understand the system, have enough time, and be able to exercise an alternative course of action.

Protect and cultivate goal-setting and not only task execution

An organization can become more efficient while losing the ability to determine why it exists. AI systems should not be assessed only by whether they accomplish a task faster or more accurately. Leaders must also ask whether humans still define the higher-order purpose, decide what success means, identify values, revise goals when circumstances change, and recognize when a measurement causes mission drift. When the system shapes the goals and humans merely optimize within them, the organization could be susceptible to missional disempowerment.

Protect and cultivate meaningful friction

Speed and frictionlessness are not always in an organization’s best interest. Some friction preserves human judgment, responsibility, memory, humility, and forges wisdom. Productive friction includes independent problem formulation before AI consultation, explanation of reasoning, dissenting reviews, documentation of recommendations that are rejected, non-automated pathways, and forums for consequential decisions. This is an agency-preserving exercise.

4. The Wisdom of Measuring Human Influence in Institutions

Most governance policies and programs measure accuracy, bias, privacy, security, incidents, compliance, and return on investment. Very few measure whether human influence is withering in an organization over time. In the Age of AI, that could be a major blind spot. A system may meet every technical metric while at the same time hollowing out the human influence needed for meaningful oversight and mission success.

Early warning indicators for the five human capacities

5. A Research Agenda For Organizations

• Longitudinal human-capability studies. Track changes in critical thinking, memory, problem framing, writing and reasoning, professional judgment, moral deliberation, metacognitive calibration, confidence without competence, and performance after AI withdrawal.
• Controlled comparisons of assistive and substitutive designs. Compare human-only work, AI-generated answers, AI critique of human work, human critique of AI work, Socratic AI, and explanation requirements.
• Goal-setting and mission-drift research. Examine whether AI changes which goals people and institutions form, how ambitious or narrow those goals become, and whether the mission gradually shifts toward what systems can readily optimize.
• Institutional-capacity studies. Measure skill decay, workforce restructuring, concentration of knowledge, loss of non-AI workflows, vendor dependence, reduced dissent, automated escalation, and disappearance of resilience processes.
• Vulnerable-population and cross-cultural research. Study the effect of generative AI on people with trauma, disability, cognitive impairment, loneliness, low literacy, economic precarity, adolescence, older adulthood, linguistic marginalization, and relational or communal conceptions of autonomy.
• Recovery and re-empowerment research. Develop methods for restoring agency and autonomy, confidence, institutional memory, alternatives, and meaningful authority after dependence has already formed.

6. The Attention Inc. Contribution to the Study of Human Agency in the AI Age

Attention Inc. helps individuals and institutions preserve the human capacity to form goals, exercise judgment, contest AI influence, maintain capabilities, and act faithfully within AI Land.