The choice that arrives pre-shaped
A person opens an app to decide what to eat.
The system already knows the address, the time, the recent orders, the price range, the weather, the restaurants likely to convert, and which options can arrive fastest.
A few taps later, dinner is decided.
Nothing about the scene is alarming.
The person wanted food. The system reduced search. The result may be exactly right.
But a small philosophical question is hiding inside an ordinary convenience.
Where did the choice begin?
The same question now appears in more consequential forms.
Which article is worth reading?
Which candidate looks promising?
Which route is safest?
Which customer is risky?
Which sentence should be sent?
Which diagnosis should be considered first?
Which problem deserves attention?
Which future is presented as likely enough to organize around?
Technological systems increasingly participate before the moment we experience as decision.
Agency under technological acceleration begins there.
Definition
Agency is the human capacity to originate, regulate, revise, and enact action in relation to goals, reasons, values, consequences, and social conditions.
Agency Under Technological Acceleration is the Institute’s synthesis frame for studying what happens to that capacity when technologies increasingly predict, recommend, generate, rank, automate, personalize, and speed the environments in which decisions are made.
The phrase does not assume that technology destroys agency.
Tools can increase agency by expanding access, reducing burdens, creating alternatives, supporting communication, and making action possible for people whose environments previously constrained them.
The concern is more precise.
Agency can be expanded at one level while narrowed at another.
A person may gain more options while exercising less authorship over how the options were framed.
They may act faster while having less time to notice what they actually want.
They may retain formal approval while losing practical authority to question the process.
They may experience convenience as freedom while the architecture of choice becomes less visible.
Agency is more than choosing
Albert Bandura’s social cognitive account of agency is useful because it is richer than a simple decision point.
Bandura describes core features including intentionality, forethought, self-regulation, and self-reflection.
An agent does not merely select among presented options.
An agent can form intentions, imagine consequences, regulate behavior, and examine their own functioning.
Self-determination theory adds another important dimension. Ryan and Deci distinguish more autonomous forms of motivation from more controlled ones and place autonomy, competence, and relatedness at the center of psychological functioning.
These traditions differ in theory and emphasis.
Together they make one thing clear.
Human agency cannot be reduced to whether a person performed the final action themselves.
A choice is not fully agentic merely because the finger that confirms it is human.
Acceleration changes the interval of agency
Agency needs time, though not always much.
There is often an interval between impulse and action in which a person can notice, interpret, compare, remember, imagine consequence, and revise.
Digital systems can compress that interval.
Recommendations arrive before search.
Autocomplete arrives before formulation.
Notifications arrive before intention.
Generated drafts arrive before language has fully become the person’s own.
Decision support arrives before the group has defined the problem.
The system does not have to coerce anyone.
It only has to make one path easier, earlier, smoother, or more authoritative than the alternatives.
Technological acceleration therefore changes agency partly through tempo.
The faster the environment moves, the more deliberate agency must sometimes be protected rather than assumed.
The automation lesson
Human factors research has studied related problems for decades.
Parasuraman, Sheridan, and Wickens described automation across different functions, including information acquisition, analysis, decision selection, and action implementation. Their model emphasized that automation changes the human role rather than simply removing it.
Parasuraman and Manzey’s review of automation complacency and automation bias showed that people can miss automation failures or follow incorrect automated advice, including in situations involving experienced users.
This is directly relevant to agency.
A person can remain physically present and behaviorally active while the system has already captured the frame.
They can approve a recommendation they did not meaningfully evaluate.
They can become responsible for an action whose logic they did not originate and cannot reconstruct.
They can be left with nominal agency and practical dependence.
The lesson is not that automated support should be avoided.
It is that agency has conditions.
Technological mediation makes the frame visible
Philosophy of technology gives a second vocabulary.
Peter-Paul Verbeek’s work on technological mediation examines how technologies help shape human perception and action. Don Ihde similarly described different human-technology relations through which technologies become embodied, interpreted, encountered, or backgrounded.
These approaches resist the idea that a tool is neutral until a human decides what to do with it.
Tools participate in what becomes salient and actionable.
A map highlights roads and hides other paths.
A dashboard foregrounds measurable indicators.
A recommender foregrounds items that fit its model of relevance.
A generative model foregrounds plausible continuations.
Agency therefore includes frame awareness.
Can the person notice that the options were selected?
Can they ask which variables were excluded?
Can they generate a possibility the interface did not offer?
Can they recognize that the metric is not the goal?
Can they leave the system’s frame entirely?
Convenience is not the opposite of agency
It would be easy to turn this into a moral defense of friction.
That would be mistaken.
Many forms of friction reduce agency.
Inaccessible bureaucracy reduces agency.
Language barriers reduce agency.
Unnecessary cognitive load reduces agency.
Physical barriers reduce agency.
Administrative repetition can consume the time people need for meaningful decisions.
AI and other technologies can remove these burdens.
A person who can use speech-to-text may gain expressive agency.
A person who receives translation may gain access to institutions.
A worker whose repetitive administrative tasks are automated may gain time for judgment and relationship.
The relevant question is not whether the system makes action easier.
It is whether ease expands the person’s substantive ability to choose and act, or whether it narrows action around choices the system has already made convenient.
Agency can be outsourced gradually
Most losses of agency are not experienced as dramatic surrender.
They arrive as relief.
The calendar decides when to leave.
The recommender decides what to watch.
The feed decides what is salient.
The assistant drafts what to say.
The ranking system decides who deserves attention first.
The predictive model decides which case appears risky.
Each delegation can be reasonable.
The cumulative pattern deserves study.
When people repeatedly delegate memory, framing, comparison, language, and recommendation, they may also delegate opportunities to practice preference formation, uncertainty tolerance, question formation, and judgment.
The evidence does not yet justify sweeping claims about long-term agency decline from generative AI.
But cognitive offloading research gives a framework for asking the question carefully. People already use external resources to reduce internal demands, and offloading choices are influenced by task demands and metacognitive beliefs.
Agency research now needs to ask not only what people offload, but which forms of self-direction depend on the processes being offloaded.
Institutional agency
Individuals do not exercise agency in a vacuum.
Organizations decide which tools are available, which recommendations appear first, how quickly responses are expected, who may override the system, and whether dissent is rewarded or punished.
An employee may be formally empowered to reject an AI recommendation while knowing that doing so will slow the team and require an explanation to a senior manager.
A public servant may retain discretion in principle while caseloads make manual review impossible.
A clinician may be told that an AI output is only advisory while the interface places it at the center of attention.
A student may be told to think independently while every platform is designed to answer before the struggle of thinking becomes necessary.
Agency is therefore partly institutional architecture.
The question is not only “Can this person choose?”
It is “What choices are realistically available, visible, supported, and survivable here?”
This connects the Institute’s agency framing to the capability tradition associated with Amartya Sen. Substantive freedom depends not only on internal ability but on the real opportunities a person has to act.
Human oversight as an agency test
The AI governance debate often uses the phrase human oversight.
That phrase can hide weak agency.
A reviewer who has no time to inspect the evidence is not exercising the same agency as a reviewer who can slow the process.
A reviewer who cannot see how a recommendation was produced has a different role from one who can interrogate its basis.
A reviewer who can technically override but is punished for doing so does not have the same practical freedom as one whose dissent is institutionally protected.
Meaningful human oversight therefore requires more than presence.
It requires competence, information, time, authority, and an escalation path.
European AI regulation already recognizes the human side of system use through Article 4’s AI-literacy obligation. In high-risk contexts, broader AI Act provisions also place importance on human oversight as part of risk governance.
The Institute’s contribution is to ask whether the humans named in governance structures possess the agency those structures presuppose.
Agency includes the power to change the objective
Perhaps the most important form of agency appears before optimization.
A system can help pursue a goal.
An agent can question the goal.
This is where automation and agency diverge most sharply.
A system can rank applicants according to a defined objective.
A human can ask whether the objective reproduces an unjust history.
A recommender can maximize engagement.
A human can ask whether engagement is a worthy measure of value.
An AI assistant can make a workflow more efficient.
A team can ask whether the workflow should exist in its current form at all.
Agency includes the power to revise the problem.
If humans retain only the role of choosing among machine-generated solutions, the deepest layer of agency has already narrowed.
How should agency be measured?
Not by counting choices.
More options do not necessarily mean more agency.
Measurement should examine at least four levels.
Intentional agency: can the person form and articulate goals that are not simply inherited from the system?
Reflective agency: can they examine assumptions, confidence, incentives, and influence?
Practical agency: can they act differently, override, refuse, or leave when necessary?
Institutional agency: does the environment provide the authority, time, information, and protection required for those actions to be real?
Behavioral experiments can test deference to automation under varying reliability and workload. Self-report can assess perceived autonomy or control, though perception alone is insufficient. Qualitative work can reveal how people experience recommendation, pressure, and authorship. Organizational measures can assess override rates, escalation, dissent, and whether formal authority corresponds to practical power.
No single metric captures agency.
The construct is too context-sensitive and normatively loaded for a vanity score.
What the evidence does not show
The evidence does not show that recommendation systems or generative AI inevitably reduce agency.
It does not show that frictionless design is inherently manipulative.
It does not show that autonomous motivation can be inferred from outward choice.
It does not show that automation bias makes people generally incapable of supervising systems.
It does not show that human decision-making is naturally unbiased or superior.
It does not show that a person must originate every option independently for a choice to count as their own.
It does not show that agency is purely internal. Real options, power, rights, material conditions, and institutional design matter.
And it does not show that “Agency Under Technological Acceleration” is one validated psychological construct.
The phrase is an Institute synthesis for organizing a research agenda.
The option that was not offered
The person is still looking at the screen.
Several options are waiting.
One is marked recommended.
They are all reasonable.
Then the person notices that none of them answer the question they actually care about.
They close the menu.
They write a different question.
This is not heroic autonomy.
It is a small act of authorship.
The system can still help.
But the goal has returned to the person.
Agency under technological acceleration will not be preserved by keeping humans nominally present at the end of automated processes.
It will be preserved when people and institutions remain able to form intentions, question frames, resist momentum, revise objectives, and bear responsibility for what they choose to make real.
Research references
Bandura, A. (2001). Social Cognitive Theory: An Agentic Perspective. Annual Review of Psychology, 52, 1–26.
DOI: https://doi.org/10.1146/annurev.psych.52.1.1
Ryan, R. M., & Deci, E. L. (2000). Self-Determination Theory and the Facilitation of Intrinsic Motivation, Social Development, and Well-Being. American Psychologist, 55(1), 68–78.
DOI: https://doi.org/10.1037/0003-066X.55.1.68
Deci, E. L., & Ryan, R. M. (2000). The “What” and “Why” of Goal Pursuits: Human Needs and the Self-Determination of Behavior. Psychological Inquiry, 11(4), 227–268.
DOI: https://doi.org/10.1207/S15327965PLI1104_01
Parasuraman, R., Sheridan, T. B., & Wickens, C. D. (2000). A Model for Types and Levels of Human Interaction with Automation. IEEE Transactions on Systems, Man, and Cybernetics Part A, 30(3), 286–297.
DOI: https://doi.org/10.1109/3468.844354
Parasuraman, R., & Manzey, D. H. (2010). Complacency and Bias in Human Use of Automation: An Attentional Integration. Human Factors, 52(3), 381–410.
DOI: https://doi.org/10.1177/0018720810376055
Verbeek, P.-P. (2006). Materializing Morality: Design Ethics and Technological Mediation. Science, Technology, & Human Values, 31(3), 361–380.
DOI: https://doi.org/10.1177/0162243905285847
Risko, E. F., & Gilbert, S. J. (2016). Cognitive Offloading. Trends in Cognitive Sciences, 20(9), 676–688.
DOI: https://doi.org/10.1016/j.tics.2016.07.002
Sen, A. (1993). Capability and Well-Being. In M. Nussbaum & A. Sen (Eds.), The Quality of Life, pp. 30–53. Oxford University Press.
DOI: https://doi.org/10.1093/0198287976.003.0003
European Union. Regulation (EU) 2024/1689, consolidated 27 July 2026.
URL: https://eur-lex.europa.eu/eli/reg/2024/1689/2026-07-27/eng