Too much AI transparency can harm decision-making

AI explanations, the study shows, trigger metacognitive processing. Users do not simply absorb the explanation. They evaluate it, compare it with their own reasoning, and reassess their confidence. This recursive self-monitoring consumes working memory. When cognitive resources are sufficient, this process enhances decision quality. When resources are strained, the same process depletes autonomy, making users feel less in control even if the AI is correct.

Artificial intelligence (AI) systems are increasingly required to explain themselves. With explainable AI becoming standard rather than optional, a key question has emerged: when does transparency help, and when does it quietly undermine human control?

A new research paper titled “The Transparency Paradox in Explainable AI: A Theory of Autonomy Depletion Through Cognitive Load,” published as an extended multi-phase study on arXiv, offers a unified theoretical explanation for why identical AI explanations can improve decision-making in some contexts while impairing it in others. Rather than treating transparency as a static design feature, the research reframes it as a dynamic cognitive intervention whose effects depend on timing, workload, and individual capacity.

Why more explanation can make decisions worse

The research confronts what it calls the transparency paradox. Transparency is widely promoted as a universal good in human-AI interaction. Explanations are expected to improve understanding, calibrate trust, and enable users to detect errors or bias. However, empirical studies across healthcare, automation, and decision support show mixed results. In some cases, detailed explanations improve accuracy. In others, they lower performance, increase mental effort, or cause users to disengage entirely.

Existing theories struggle to explain this contradiction. Trust-based accounts assume that more information leads to better calibration between humans and machines. Cognitive load theory explains why too much information can overwhelm users, but treats overload as a static threshold rather than a process that evolves over time. Self-determination theory emphasizes the importance of perceived control, but does not specify how information itself can undermine that control during decision-making.

The authors argue that these gaps stem from a failure to model autonomy as a dynamic psychological state. In their framework, autonomy is the subjective sense of being in control of one’s decisions while interacting with AI. This sense of control fluctuates continuously as tasks unfold, shaped by workload, attention, fatigue, and information demands.

AI explanations, the study shows, trigger metacognitive processing. Users do not simply absorb the explanation. They evaluate it, compare it with their own reasoning, and reassess their confidence. This recursive self-monitoring consumes working memory. When cognitive resources are sufficient, this process enhances decision quality. When resources are strained, the same process depletes autonomy, making users feel less in control even if the AI is correct.

The study formalizes this process using stochastic control theory, modeling autonomy as a continuously evolving variable influenced by information-induced cognitive load. Under low to moderate transparency, autonomy can be maintained or even strengthened. Under high transparency, especially when explanations accumulate, autonomy declines predictably until users disengage.

This mechanism explains why transparency effects are context-dependent. Early in a task, when cognitive resources are available, explanations help. Later, when mental load has accumulated, the same explanations overwhelm remaining capacity. The paradox is not a contradiction but a consequence of dynamic cognitive depletion.

Cognitive limits, individual differences, and the timing problem

Transparency cannot be separated from timing. Most explainable AI systems treat explanations as fixed outputs delivered whenever a decision is made. The research shows that this approach ignores how cognitive state evolves during interaction.

The model predicts that the same explanation can have opposite effects depending on when it is presented. Early explanations may anchor reasoning or improve understanding. Delayed explanations, delivered after users have formed an initial judgment, can reduce bias and improve performance. Conversely, continuous or excessive explanations can accelerate cognitive fatigue, reducing both autonomy and accuracy.

Working memory capacity plays a critical moderating role. Individuals differ systematically in how much information they can process before autonomy begins to collapse. The study models this through a cognitive boundary that determines when disengagement occurs. Users with higher working memory capacity can tolerate more explanation and remain engaged longer. Users with lower capacity reach overload faster, even under moderate transparency.

This finding has significant implications for fairness and personalization. Fixed transparency policies implicitly privilege users with higher cognitive capacity, while penalizing others by overwhelming them. The study suggests that explainable AI systems designed without regard to individual differences risk becoming exclusionary, despite their ethical intent.

The research also challenges the idea that transparency should be maximized to satisfy accountability demands. The model demonstrates that both maximum transparency and zero transparency perform worse than adaptive strategies. Maximum transparency rapidly depletes autonomy, leading to disengagement. No transparency preserves autonomy but deprives users of decision-relevant information. Optimal outcomes occur at intermediate levels that shift over time based on cognitive state.

This dynamic perspective reframes transparency as a scarce resource rather than a default setting. Information has value, but it also has cost. The balance between the two changes continuously during interaction.

Toward adaptive explainability and policy implications

When autonomy is high and information load is low, more detailed explanations can be beneficial. As cognitive resources decline, transparency should be reduced to preserve control and engagement. This threshold-based approach outperforms static policies in maintaining decision quality while minimizing disengagement.

The research outlines several practical strategies consistent with this framework. Transparency should be delivered incrementally rather than all at once. Systems should track cumulative explanation complexity rather than treating each explanation in isolation. Individual differences in working memory capacity should inform personalization. Time pressure should also be considered, as urgency can justify higher cognitive costs near decision deadlines.

These insights challenge current regulatory and ethical approaches that mandate transparency without specifying how it should be delivered. The study warns that indiscriminate transparency can undermine the very goals it is meant to serve, including oversight, accountability, and human agency.

The authors note that similar autonomy depletion effects occur in education, finance, and information-rich environments more broadly. Excessive instructional scaffolding can undermine student autonomy. Over-disclosure in financial decisions can reduce confidence and lead to avoidance. The transparency-autonomy tradeoff is not unique to AI but is amplified by it.

The study also raises ethical concerns about how adaptive transparency might be implemented. Systems that adjust explanations based on inferred cognitive state may rely on behavioral or physiological data, raising privacy and consent issues. There is a risk that adaptive transparency could be used to optimize compliance rather than support user welfare. The authors argue that transparency systems themselves must be transparent, with safeguards to ensure they enhance rather than manipulate human agency.