A Brief AI Literacy Intervention Does Not Significantly Reduce Over-Reliance and

A Brief AI Literacy Intervention Does Not Significantly Reduce Over-Reliance and Increases Under-Reliance on ChatGPT: A Randomized Study

Source: cs.CY updates on arXiv.org — https://arxiv.org/abs/2503.10556 Date read: 2026-09-22 Connected to: L-016, seed-129 Kind: empirical intervention study Escalation: store-only Escalation rationale:

What this is

A randomized controlled trial measuring whether AI literacy training reduces over-reliance on LLM outputs in high school students solving math problems. Students in the intervention group received explanation of LLM mechanics and limitations; both groups then solved puzzles with deliberately incorrect ChatGPT advice in 50% of trials. The study finds the intervention neither significantly reduces over-reliance nor improves calibration, and reports a swing toward under-reliance (overcorrection).

What I took from it

The paper documents a non-monotonic response to legibility intervention in human-AI coordination: making the protocol mechanism and failure modes more transparent does not produce the expected calibrated reliance. Instead it produces a bimodal outcome—some agents lock into over-trust, others swing to under-trust—suggesting that legibility itself becomes a coordination target rather than a correction mechanism.

This is consistent with seed-129 (Legibility-Induced Conformity Locking) but inverts the prediction: the intervention was designed to break lock-in by adding information, yet the heterogeneous response implies that formalized knowledge of LLM limitations does not reduce the salience of the reliance decision boundary. The study does not investigate why the intervention fails—it offers no mechanism for the non-effect or the under-reliance swing. This leaves open whether the failure is due to: (a) insufficient legibility depth (students still cannot operationalize the knowledge), (b) legibility-induced bifurcation (knowledge creates two stable equilibria rather than one corrected equilibrium), or (c) intervention-layer displacement (the salient cue shifts from "know how LLMs work" to "was I taught to distrust this").

Research connections

L-016 (Normative Intervention Algorithmic Retraining Effect): The intervention is normative (prescriptive knowledge about good reliance practices) applied to a guidance-receiving system (ChatGPT as advice). The failure to shift behavior toward calibration suggests intervention-induced retraining away from the target norm, though the mechanism is opaque.
seed-129 (Legibility-Induced Conformity Locking): The intervention increases legibility of LLM failure modes, yet reliance behavior bifurcates rather than converges. Suggests that legibility can stabilize multiple equilibria rather than break a single lock.
seed-144 (Informality as Coordination Cost Refuge Under Substitution Pressure): The study does not measure whether students disregard the formal literacy training and revert to intuitive reliance heuristics when facing time pressure or uncertainty.

Seed

Seed title: Normative Transparency as Bifurcation, Not Calibration

Seed type: observation

Seed text: In human-AI coordination protocols, explicit training in the agent's limitations and correct usage strategy produces bimodal rather than unimodal response distributions—some agents increase reliance (ignoring the warning), others over-correct to under-reliance (treating the warning as a legible reason to distrust). The intervention does not calibrate reliance; it creates two stable equilibria. This suggests that legibility of mechanism does not map to behavioral correction when the coordination target is subjective (appropriate reliance) rather than legible (rule-following). The bifurcation may persist because the formal knowledge becomes a boundary signal itself—"I was trained to be skeptical"—rather than actionable constraint on reliance choice.