Topic guide
Human agency and responsibility
When an AI-assisted decision causes harm, the sentence "the system decided" can hide more than it explains. A model may contribute causally to an outcome without choosing the goal, the acceptable risk, the training data, the deployment setting or the conditions under which a person can challenge its recommendation.
This guide brings together Prompted Psyche materials on agency, delegation and responsibility. It focuses on the chain of human and institutional choices around a system. The aim is neither to deny the novelty of AI nor to assign every failure to one individual. It is to keep the authorship of goals, constraints and decisions visible.
Area definition
What agency and responsibility means here
Human agency is the capacity to form intentions, make choices and act within real constraints. Responsibility concerns who can be answerable for a decision, process or outcome. In AI systems, both may be distributed across managers, product teams, vendors, operators, reviewers and institutions. Distribution does not mean disappearance.
Causal contribution is different from moral agency. A system can alter what happens without possessing the kind of intentions, understanding or accountability normally required for moral responsibility. At the same time, responsibility should not automatically fall on the lowest-level person who touched the output. An operator can become a moral crumple zone when an institution gives them formal blame but little information, time or authority.
Distinctions that matter
- Cause is not moral agency
- A system can shape an outcome without becoming a blameworthy moral subject.
- Delegation is not disappearance
- Passing a task to a model does not erase the human choice to define and use it.
- Oversight is not a signature
- Meaningful review requires information, competence, time and power to intervene.
- Responsibility can be distributed
- Several actors can hold different duties without making the nearest operator the sole cause.
Key questions
- Who defined the objective and the acceptable trade-offs?
- Who selected the data, model, threshold and deployment context?
- What can a reviewer actually inspect, contest and change?
- Which incentives reward agreement with the automated recommendation?
- Who benefits, who bears the risk and who can provide a remedy?
A useful way to orient yourself
Trace the decision rather than staring only at the output. Start with the goal, then follow the chain through procurement, data, model configuration, interface, policy and final use. At each stage, ask what alternatives existed and who had the authority to choose among them. This makes agency laundering - presenting a human decision as a machine necessity - harder to sustain.
Then examine the reviewer. A human in the loop is not enough if the role is ceremonial. Meaningful control depends on whether the person understands the system, sees relevant evidence and uncertainty, has enough time and can change the result without being punished for disagreement.
Starting point
Where to start
The main essay develops the difference between AI as an amplifier, a moral buffer and a moral alibi, while preserving real limits of human control.
An essay on causal contribution, moral agency, distributed responsibility and the institutional uses of AI as an alibi.
Library
Explore the topic from several angles
Articles build the argument, concepts clarify the mechanisms and Practice turns them into concrete actions.
Articles
The related essays extend responsibility from organizational decisions to interpretation and epistemic reliance.
Concepts
These entries clarify the roles of agency, oversight, decision support and misplaced computational authority.
Practice
These exercises keep factual analysis, interpretation and the final decision in distinct places.
Notes
The notes show how responsibility can disappear inside summary and interpretation.
Sequence
Suggested reading path
- Are we afraid of AI, or of ourselves?
Build the responsibility-tracing framework and its moral limits.
- Human agency
Clarify what remains distinctly human in a constrained decision process.
- Human oversight
Distinguish meaningful intervention from ceremonial review.
- How to use AI without handing over the decision
Apply the distinction between model analysis and human judgment.
- AI does not read people. It helps make sense of the situation
See the same boundary in everyday attempts to infer people from partial text.
Connections
How the pieces connect
The main article provides a map of moral and institutional responsibility. Concepts identify the roles that can be confused in a workflow. Practice entries make the chain visible at the level of an individual decision, and notes show how quickly language can hide the shift from evidence to interpretation.
The result is not a formula for legal liability, which depends on law, jurisdiction and facts. It is a way to ask better questions before a system is treated as the author of a decision or a frontline operator is made responsible for choices designed elsewhere.