A manager is preparing for a difficult performance conversation. She opens an AI assistant and asks it to sharpen her notes. A few weeks later, she uses the same system to identify patterns in the employee’s performance. Then to compare that employee with others. Then to suggest who looks ready for promotion. During a restructuring, the tool begins surfacing positions that might be eliminated.
Each step is easy to justify. The system is fast, useful, and apparently helpful. Yet somewhere along the way, a tool that began by improving the language of a decision has started to participate in making it.
This is how normalization works. Large shifts often arrive as a series of small conveniences. Repetition removes friction; familiarity lowers scrutiny. Generative AI is moving rapidly into knowledge work because it drafts, summarizes, compares, predicts, and recommends. The same pattern is entering private life. People ask AI whether to confront a colleague, leave a job, trust a partner, discipline a child, forgive a friend, or end a relationship.
What happens when AI is present at so many stages of deliberation that we stop asking whether it belongs there?
When Help Starts to Shape Judgment
The central risk is the slide from assistance to substitution. Asking AI to help articulate a view is different from asking it what view to hold. Asking for additional options expands a decision; allowing the system to define the options begins to structure it. Fluent answers can make that shift hard to see because fluency feels like reasoning, and confidence can feel like authority.
A 2025 study among knowledge workers found that greater confidence in generative AI was associated with less critical-thinking effort. The workers did not simply stop thinking. Their cognitive work moved toward checking, integrating, and managing what the system had already produced. Verification still requires thought. When you verify, you ask a narrower question: “Does this look plausible?” The harder work comes earlier: “What do I believe and on what grounds, and what am I willing to be responsible for?”
This is where agency decay becomes useful as an analytical lens. It can unfold through four stages: experimentation, integration, reliance, and addiction. A tool begins as an option. It becomes part of the workflow. The workflow adapts around it. Eventually, working without it feels slower, incomplete, or strangely difficult. The psychological shift is subtle: “Could AI help me think this through?” becomes “Why would I think this through without AI?”
A 2026 study on cognitive agency transfer draws a related distinction between autonomous offloading, where AI supports a person’s own thinking, and dependent offloading, where the system increasingly structures the problem, weighs the evidence, and supplies the conclusion. Dependent use was associated with weaker perceived independent judgment, deep processing, and autonomous capability. The study does not establish long-term cognitive decline. Its more immediate warning is enough: Autonomous and dependent users can experience similar short-term benefits. The email is finished. The analysis is faster. The recommendation arrives on time. Productivity looks fine while agency may be shifting underneath.
The psychological shift is subtle: “Could AI help me think this through?” becomes “Why would I think this through without AI?”
The Moral Problem Is Attention
The stakes rise when AI enters decisions with a moral dimension. Ethical judgment begins with noticing that a moral question exists.
Imagine an HR leader reviewing dozens of compensation recommendations. One employee receives a smaller increase after returning from parental leave. On its own, the discrepancy may trigger concern. Inside a long sequence of AI-generated recommendations, each accompanied by a plausible rationale, it can pass as another data point.
A 2025 study, “Accepting the Unacceptable in the AI Era,” examined a similar mechanism. Across experiments involving managers, AI recommendations could reduce moral awareness when people were handling multiple organizational decisions, making unethical recommendations easier to accept.
This matters because questionable decisions rarely arrive labeled as ethical problems. They arrive as staffing choices, rankings, procurement trade-offs, insurance assessments, pricing decisions, productivity targets, or cost reductions. A recommendation can be analytically coherent while affecting fairness, dignity, or opportunity. Those consequences enter the decision only when somebody notices them. In a fast optimization process, the unacceptable does not need enthusiastic approval. It only needs to stop attracting attention.
The Same Subtle Drift Enters Private Life
Private decisions make the problem even more intimate. A person pastes messages into a chatbot and asks whether a partner is manipulative. A parent asks how to respond to an estranged child. A professional asks whether leaving a career is the “right” choice. Someone asks AI to write a breakup message, then asks whether sending it is morally justified.
A 2026 experiment found that ChatGPT advice influenced people’s judgments in a moral dilemma even when the advice contained no supporting moral argument. The influence, in other words, did not depend entirely on persuasion through reason. Sometimes an answer has power because it reduces the burden of choosing.
That relief deserves scrutiny. Doubt can force us to examine our motives. Hesitation creates time for us to notice who else is affected. Unease may signal that something is genuinely in tension with our values. Human judgment is often slow because reality is morally untidy.
There is a second effect. Research on human-AI feedback loops found that repeated interaction with AI could amplify biases in later human judgments, while participants often underestimated the influence. Once a system repeatedly frames what matters, which variables deserve attention, and what counts as evidence, its logic can begin to shape our own. AI does not need the final word to influence the architecture of a decision.
When what once felt unusual becomes routine, convenient, semi-automated, and no longer in need of explanation, it is accepted.
The Management Risk Hiding in Plain Sight
Business leaders already worry about privacy, cybersecurity, hallucinations, intellectual property, discrimination, and compliance. Another worry is situated beside them: What is repeated AI use doing to the quality of human judgment inside the organization?
A person can remain formally “in the loop” while moving substantively downstream of the machine. The analyst still reports. The manager still signs. The executive remains accountable. Yet the initial framing, shortlist of options, and first recommendation may all have been generated elsewhere. Human responsibility stays in place while human authorship thins out.
That creates an awkward governance arrangement: Organizations retain the liability while gradually outsourcing parts of the capacity that make responsibility meaningful.
Create Friction Before the Decision
Generative AI can enlarge human capability when it expands reflection, comparison, and imagination. The practical discipline is to preserve human authorship at the moments that matter. Before a consequential use, four questions can create enough friction to make that visible:
- What part of my judgment am I about to delegate?
- What did I think before the AI answered?
- Who will live with the consequences?
- Can I defend the final choice on my own account?
These questions take less than a minute. Their purpose is to interrupt the progression from convenience to habit and from habit to dependence.
The capacity to choose has always been so basic to human life that we never perceived or treated it as something requiring active protection. AI changes the conditions around that capacity. It can now enter the space between perception and decision, carrying part of our cognitive and moral load.
The normalization of the unacceptable happens one small prompt after the other. Familiarity makes the transition from independent judgement to assisted reasoning to delegated judgement smoothly, until it feels like a given of life. It is a slippery slope: When what once felt unusual becomes routine, convenient, semi-automated, and no longer in need of explanation, it is accepted.
Vigilance means noticing that transition while the boundary is still visible. Otherwise, outsourcing judgment will eventually come to feel entirely ordinary — and we may eventually fail to remember why keeping hold of it once seemed important.
