AI Changes the Person by Rehearsing a Way of Deciding

AI changes people less by supplying beliefs than by repeatedly training how they frame choices, allocate agency, and judge themselves.

The deepest personal effect of interacting with AI may not be that the machine changes what someone believes, but that it repeatedly rehearses a particular way of deciding. An AI interface can partition a problem, rank considerations, request confirmation, or present a recommendation, and each arrangement gives the user a role with different kinds of attention and responsibility . Evidence from AI-framed decision support indicates that analytic structuring increases reported psychological empowerment, which in turn predicts higher perceived decision quality . That finding complicates the familiar fear that assistance simply makes people passive: structured assistance can make users feel more agentic, even while the system determines the structure within which that agency is exercised . The person may therefore become more decisive without becoming more self-directing, because confidence in making a selection is not the same as authorship of the terms on which alternatives were constructed . AI changes the person, on this account, through repeated role practice: every interaction teaches not only an answer, but what kind of decider the user is expected to be.

This distinction between empowerment and authorship matters because an interface can strengthen reflective engagement while quietly narrowing what reflection concerns. In a scenario-based experiment with 250 participants, analytic structuring was positively associated with both psychological empowerment and self-reported trust regulation, while trust regulation strengthened the relationship between empowerment and perceived decision quality . The apparent benefit did not come merely from exposing an internal model explanation; it came from how the decision itself was analytically organized for the participant . A user repeatedly exposed to such organization may learn to experience a well-decomposed choice as a well-understood choice, although decomposition can omit considerations that do not fit the supplied categories. This is an inference rather than a measured personality effect, but it follows from the study’s central mechanism: structural presentation affected agency and evaluation, not just the informational content of the advice . The consequential design question is therefore not only whether users trust AI too much, but whether they gradually adopt its preferred grammar of relevance as their own.

Workflow research shows how this grammar can become a practical discipline rather than mere deference. A system combining a local language model, a digital twin, and human review used a Propose-Verify-Decide process for task adjustment across four virtual surgical-instrument sorting lines . Only 18 of 28 workflow cases met expectations, autonomous strategy-workflow success was 3 of 10, and failures included semantic distortion, incomplete evidence, and missed invalid inputs . Yet seven of eight invalid inputs were correctly rejected, and all four cases that cleared prior checks, produced complete evidence, and reached final engineering review passed that review . The lesson is not that review makes AI reliable in general; it is that a staged interaction can train a person to treat a fluent proposal as the beginning of judgment rather than its conclusion . Repeatedly answering dilemmas on Decide could likewise be personally consequential when the act is understood as practice in noticing one’s own criteria, because the durable change would lie in how a person approaches subsequent choices rather than in any isolated response.

The harder problem is that decision habits do not merely organize information; they can reorder values. In a simulated two-objective Q-learning ultimatum game, increasing fairness pressure generally produced fairer outcomes, but moderate pressure caused responders to reverse strategy and accept low offers more readily . The authors attribute this “forgiving” pattern to competition between payoff-maximizing and fairness-oriented preferences rather than to the simple disappearance of either objective . That mechanism suggests why AI-guided reflection could change someone without overt persuasion: repeatedly adjusting the relative prominence of objectives may alter which internal conflict is resolved first. A tool that constantly foregrounds efficiency, consistency, or fairness need not convince the user that another value is false; it can make that other value less decisive at the moment of choice. Person-level change may thus appear as a new style of compromise—greater willingness to tolerate unfairness under moderate moral pressure, for example—even when the user continues to endorse fairness in the abstract .

AI interaction can also transform people socially by changing whose judgment becomes psychologically adjacent to their own. Research on distributed inference found that communication can bring agents into agreement while making their decisions worse, and systems with identical eigenvalue and singular-value spectra produced opposite effects when only message orientation changed . In the reported example, changing orientation raised accuracy from 72.6% to 91.2% in one case but lowered it to 65.9% in another . Under community-shared bias, higher average individual accuracy could coexist with lower global-vote accuracy or harm to communities not carrying that bias, while calibration constraints reduced observed harm without guaranteeing protection . Although these results concern distributed agents rather than human personality, they expose a mechanism relevant to AI companionship and advice: the direction from which a message enters deliberation can matter more than aggregate measures of connectivity or apparent consensus . A person who routinely consults an AI is not simply adding another opinion; they may be installing a stable conversational orientation that changes which doubts receive reinforcement, which alternatives feel socially supported, and which disagreements survive long enough to matter.

This is why “remaining in control” is too weak a test for whether the person remains the same kind of agent. A review of human-machine systems argues that successful interaction requires balancing autonomous functionality with user control, illustrating the balance with assisted wheelchairs that combine user direction and automated safety features . But control over the final action can coexist with dependence on the machine for framing, search, and validation, as the Propose-Verify-Decide workflow demonstrates . One axiomatic account of substantive agency separates coherent plurality, causal openness, anticipatory non-pointing, act-level singularisation, and endogenous sourcehood; it argues conditionally that exact passive pre-act selection conflicts with retaining both singularising priority and invariance under causally inert informational extension . The paper explicitly does not establish that human decision-making satisfies the premises required to apply this result, so it cannot prove that AI prediction or advice diminishes human agency . It nevertheless sharpens the concern: if the meaningful human contribution is reduced to approving an already selected path, nominal control may preserve responsibility on paper while weakening the user’s experience of being the source from which a decision emerged.

Collaboration can produce the opposite development when the division of labor makes human judgment more practiced rather than more ceremonial. One mixed-methods study reported 91% decision accuracy and mean trust of 6.3 out of 7 in its collaborative condition, outperforming human-only, AI-only, and sequential human-AI conditions . Across eight collaboration cycles, decision time fell by 52% and decision quality rose by 26 percentage points, while AI-led ownership was about 52 percentage points higher in analytical stages than in judgment and ethical stages . These results support a complementarity model in that setting, but they do not establish that ethical judgment remains robust merely because a workflow labels one stage “human” . If AI handles comparison, retrieval, and option generation, the human may gain more occasions to make evaluative judgments while losing practice in discovering which evaluations the situation calls for. A person can therefore become better at choosing among machine-legible alternatives and worse at recognizing alternatives that were never rendered legible—a possibility consistent with failures involving semantic distortion and incomplete evidence in the engineering workflow .

The right question is consequently not whether AI interaction changes people, as if change required a dramatic conversion, but which decision dispositions the interaction repeatedly exercises. The available studies directly measure workflow performance, perceived empowerment, trust regulation, simulated preference competition, and distributed accuracy; they do not demonstrate durable personality change in human users . Any claim that AI is already remaking identity must therefore remain provisional, while the mechanisms identified across these studies justify examining change at the level of framing, value arbitration, social orientation, and experienced authorship . A useful approach to human and AI decision-making would track not only whether assistance improves outcomes, but whether users retain the ability to generate criteria, challenge decompositions, detect missing evidence, and decide when optimization is the wrong activity. Those capacities distinguish a person who uses an efficient instrument from one whose practical identity has been reorganized around the instrument’s way of making matters decidable . AI’s most important formative power may therefore be neither manipulation nor replacement, but repetition: it gives people thousands of small rehearsals in what to notice, what to delegate, what to verify, and what it feels like to call a decision their own.

Decisions · Articles · The Trolley Problem · Game Theory · Human–AI Alignment · Resonant