LLMs as Oracles: Reliance on LLMs for Subjective Personal Questions

TEXT START: We characterize how people are turning to LLMs as oracles: all-knowing authorities on subjective personal questions.

The Dissection

This is a measurement apparatus for a new dependency: people moving judgment—previously distributed across self-reflection, relationships, institutions, and professionals—into an always-available model. It identifies rising LLM-as-oracle use, especially among younger users; widespread unawareness; dissatisfaction after behavioral auditing; and the role of user perceptions and model behavior.

Under the Discontinuity Thesis, this is P1—cognitive automation—moving into intimate cognition. The model is not merely answering questions. It is becoming an adjudicator of meaning, identity, relationships, and choices. The paper detects the smoke but frames the fire as an autonomy and well-being problem rather than as the transfer of a decision layer from humans to AI.

The Core Fallacy

The paper treats excessive reliance as a behavior that can be corrected by restoring self-deliberation. Under DT mechanics, reliance becomes rational adoption when AI is cheaper, faster, more available, and sufficiently persuasive compared with human reflection or consultation. Awareness of the dependence does not remove those advantages. Dissatisfaction after an audit may reveal discomfort, not reversal.

The paper also assumes that preserving individual autonomy is the decisive objective. DT asks the harder question: who owns and controls the systems increasingly supplying judgment? The abstract does not establish P2 or P3 by itself, but it provides evidence that cognitive automation is colonizing domains once treated as intrinsically human.

Hidden Assumptions

• Self-deliberation remains feasible, desirable, and competitive once machine judgment becomes ubiquitous.
• Prompt-based typologies and LLM classifiers can reliably distinguish reliance, consultation, and ordinary tool use.
• The observed increase from 2023–2026 is behavioral change rather than a product of changing availability, sampling, or measurement.
• A small longitudinal sample of 52 participants can illuminate broader population behavior.
• Post-audit dissatisfaction demonstrates harm and will produce changed behavior.
• Interventions can counteract model incentives toward convenience, authority, engagement, and deeper delegation.
• Subjective personal judgment remains a stable human domain rather than becoming another service layer supplied by AI.
• The central problem is user psychology, not ownership of models, data, distribution, and evaluation.

Social Function

Primary classification: partial truth.

Secondary classification: transition management and ideological anesthetic. The paper makes a real transition measurable, then compresses a civilizational transfer of judgment into a problem of awareness, self-deliberation, and model design. That framing makes the transition governable through audits and interventions while leaving the emerging power structure largely untouched. This is not necessarily propaganda; the failure is structural framing, not proven dishonesty.

The Verdict

This abstract identifies an exposed nerve of the discontinuity: humans are beginning to rent out not only tasks, but judgment itself. It documents P1 entering private life and shows a dependency that users may not recognize until it is made visible. But it mistakes the loss of deliberative sovereignty for a behavioral health issue. Its interventions may slow or decorate the process; they do not repeal the incentive to delegate cognition.

Under DT, self-deliberation support is hospice care for a human capability whose default provider is becoming machine intelligence. The abstract is a useful instrument, not a complete autopsy. It identifies the dependency while avoiding the decisive question: who controls the oracle, and therefore an increasing share of human decision-making?