生成式 AI 資訊呈現模態對人類決策之影響

Abstract

隨著生成式人工智慧（Generative AI）技術的快速成長，大型語言模型（LLM）的智慧決策輔助系統（Copilot）已被廣泛導入高精密製造、半導體產線監控等高錯誤成本之關鍵任務領域。當前消費級 AI 介面普遍採用模擬人類思維步調之「動態逐字輸出（Dynamic Streaming）」呈現模態；然而，此種視覺律動型態是否會系統性地影響人類決策者的信任校準，進而造成自動化偏誤（Automation Bias），在現有人因工程（Ergonomics）與人機互動（HCI）領域中仍缺乏系統性的量化實證。有鑑於此，本研究旨在探討生成式 AI 不同資訊呈現模態（動態逐字輸出 vs. 靜態完整呈現）與外部環境壓力（生理壓力、心理警戒）對人類決策品質、主觀認知負荷及隱性注意力分配之重塑機制。

本研究採用二因子受試者內設計（2呈現模態 x 2壓力情境），透過兩階段解耦之網頁實驗平台，招募32位有效理工背景受試者之完整行為與心理數據矩陣。實驗中植入高流暢度之 AI 錯誤建議（AI 幻覺陷阱題）作為訊號，並導入高頻滑鼠軌跡（AOI 切換頻率）進行隱性注意力監測，最終結合訊號偵測理論進行深層認知決策建模。

推論統計與模型校準結果顯示：呈現模態對決策績效具備決定性影響，在正常試次中雙模態正確率皆為100.0%，但在 AI 幻覺之陷阱題中，靜態模態下高達 91.67% 的錯誤偵測率，在動態輸出模態下降至43.75%，證實動態呈現是誘發自動化偏誤之關鍵原因。NASA-TLX 主觀負荷分析表明，動態模態因暫時資訊作用顯著使用了大腦工作記憶，使其心理需求（83.26分）與時間需求（87.63分）上升；然而受試者之自我表現滿意度卻保持不變，呈現出「不自知的盲目信任」。SDT 模型精確還原，證實動態呈現會導致個體區辨幻覺之客觀敏感度指標降低（自2.09降至1.01），而內在判斷準則對數值則非理性向上偏移（ln β自0.15上升至0.40），大幅提高懷疑 AI 建議之心理門檻。滑鼠軌跡分析證實，雙區切換頻率與陷阱題正確率具備正相關（r = .676），而動態模態之視覺牽引導致核對行為下降（自 9.77 次降至 4.91 次），成為促使過早決策（Premature Closure）的外顯行為。

With the explosive growth of Generative Artificial Intelligence (AI) technologies, intelligent Copilot systems based on Large Language Models (LLMs) have been widely integrated into mission-critical domains characterized by high error costs, such as semiconductor manufacturing and precision chemical process monitoring. While consumer-grade AI interfaces heavily favor "Dynamic Streaming" (word-by-word token generation) to mimic human cognitive pacing, whether this visual rhythmic pattern systematically degrades human trust calibration and catalyzes automation bias remains critically under-explored in contemporary cognitive ergonomics and Human-Computer Interaction (HCI). This thesis systematically investigates how generative AI presentation modalities (Dynamic Streaming vs. Static Display) and environmental stressors (physiological stress and mental alertness) reshape human decision quality, subjective work-load, and implicit attention allocation.

A 2 (Presentation Modality) x 2 (Stress Scenario) within-subject factorial design was employed. Utilizing a decoupled web-based experimental platform, complete behavioral and psychological data matrices were captured from 32 validated participants with engineering backgrounds. Fluent yet erroneous AI recommendations (AI hallucination traps) were embedded as target signals. High-frequency mouse-tracking via Areas of Interest (AOI) switching frequency was utilized for implicit attention monitoring, and Signal Detection Theory (SDT) was applied for cognitive decision modeling.

Inferential statistics and model calibration yielded several critical findings: (1) Presentation modality exerts a decisive main effect on decision quality. While performance achieved a ceiling of 100.0% accuracy across both modalities in normal trials, the error detection rate on trap trials suffered a catastrophic collapse from 91.67% under the Static modality to a mere 43.75% under the Dynamic Streaming modality, identifying dynamic rendering as a key interface catalyst for automation bias. (2) NASA-TLX workload profiles indicated that the dynamic modality severely drained working memory via the transient information effect, causing subjective mental demand (83.26) and temporal demand (87.63) to skyrocket. Crucially, users' self-reported performance satisfaction remained statistically invariant, revealing a deceptive "unconscious overreliance." (3) SDT modeling decrypted the cognitive black box, demonstrating that dynamic streaming halved objective sensitivity (d', from 2.09 to 1.01) while causing an irrational upward shift in the log-response criterion (ln β, from 0.15 to 0.40), significantly elevating the psychological threshold required to reject AI advice. (4) Mouse-tracking telemetry verified a robust positive correlation between AOI switching frequency and trap detection accuracy (r = 0.676). However, the dynamic visual pull halved verification behaviors (from 9.77 to 4.91 switches), serving as the overt behavioral signature of "Premature Closure" and cognitive tunneling.