Not All Explanations Are Sought: Information-Seeking Psychology for Human-Centered XAI

TEXT START: This position paper argues that human-centered explainable AI (HCXAI) should incorporate insights from the psychology of information seeking.

The Dissection

The paper relocates the problem of increasingly autonomous AI from political economy to user psychology. Its central move is to ask how systems can make humans seek explanations at the right moments, using instrumental, hedonic, and cognitive utility as design variables.

That is a legitimate interface problem. It is also a narrow one. The paper is designing the control panel for a machine whose control is migrating elsewhere. It treats human understanding, attention, and intervention as the scarce resources to optimize while leaving ownership, authority, and economic necessity largely untouched.

The Core Fallacy

The core fallacy is assuming that better explanation-seeking preserves meaningful human agency after cognitive automation has made human labor structurally inferior.

Under the Discontinuity Thesis, P1 makes AI cheaper and more capable across cognitive work. P2 prevents institutions from preserving stable human-only domains at scale. P3 removes the majority from economically necessary production. In that environment, explanations do not restore the mass employment-to-wage-to-consumption circuit. They only improve how a shrinking class of authorized humans supervises, verifies, or services AI capital.

The paper also risks confusing epistemic access with operational control. A user may understand why an agent acted without possessing the authority, time, resources, or bargaining power to stop it. “Making explanations sought” can therefore produce better-informed servitors without producing sovereigns.

Hidden Assumptions

• Humans will remain legitimate decision-makers over agentic systems rather than ceremonial approvers.
• Users will have enough attention and cognitive capacity to evaluate cascading actions as systems become more complex.
• Explanations will be timely, accurate, and sufficiently complete to support intervention rather than post hoc rationalization.
• Cognitive biases can be managed through interface design instead of being reinforced by dependence on superior automation.
• Those who seek explanations will have the authority to act on them.
• Human-centered design remains the governing objective under competitive pressure to remove costly human oversight.
• Information-seeking is the decisive bottleneck, rather than ownership of AI capital, access to infrastructure, and control of deployment.
• Institutions can coordinate stable human oversight domains, despite the thesis’s coordination-impossibility constraint.

The most dangerous assumption is that explanation creates control. Often it creates compliance with a more legible machine.

Social Function

This is partial truth functioning as transition management and ideological anesthetic. It correctly identifies that people will seek too much or too little information, and that agentic systems create new risks around intervention. But it frames the transition as a problem of calibrated transparency instead of displacement and power transfer.

For sovereigns, these techniques may provide auditability and risk control. For indispensable servitors, they may become survival equipment and a form of verification arbitrage. For the displaced majority, they offer little more than transparency theater: knowing why the system excluded them does not make them economically necessary again.

The Verdict

Technically useful, systemically inadequate. The paper may improve human supervision of agentic AI, but it does not challenge the mechanism that makes most human supervision replaceable. It makes the transition less confusing for those allowed into the control room; it does not prevent the control room from becoming unnecessary, automated, or privately owned.