生成式人工智慧與大型語言模型在心理學研究中的方法論意涵：推論條件與推論邊界的再界定

Methodological Implications of Generative AI and Large Language Models for Psychological Research: Redefining Inferential Conditions and Boundaries

生成式人工智慧與大型語言模型（large language models, LLMs）正快速進入心理學研究流程，廣泛應用於研究寫作與程式輔助、刺激材料生成、自然語言資料的標註與分析，及以模型模擬人類反應乃至生成合成受試者資料（synthetic respondent data）的準實驗用途。無可否認，LLM大幅提升了研究效率，並能克服傳統人工編碼容易疲勞、標準漂移所造成的不穩定性，在設定得當的情況下提供高度一致的分類結果。然而，本文主張，LLM並非僅是單純的「效率工具」，而是實質改寫了心理學研究中資料處理、測量建構與統計推論的「推論條件」（inferential conditions）。本文首先探討LLM的機率本質如何系統性地引發偏誤，接著提出LLM在心理學研究中的四種方法論角色，並以臨床、認知與社會心理學的實徵研究為例，分析LLM扮演測量裝置或代理標註者時如何對效度論證與誤差結構提出挑戰。為避免推論層級錯置，本文提出五個階段檢核以及具備實務可操作性的推論門檻與操作指引，並呼籲學術生態系建立相應的審查機制。面對LLM引入的新變異來源，心理學界必須引進新的統計校正技術，但其終極目的，仍是為了堅守心理計量學長期發展的效度論證與誤差管控精神，亦即堅守「手段創新、標準延續」之原則，以維持心理學知識的可審查性與累積性。

Generative artificial intelligence and large language models (LLMs) are rapidly entering psychological research workflows, with broad applications in research writing and programming assistance, stimulus generation, annotation and analysis of natural language data, and quasi-experimental uses that simulate human responses and generate synthetic respondent data. Undeniably, LLMs substantially enhance research efficiency and overcome the instability resulting from fatigue and standard drift common in traditional manual coding, delivering highly consistent classification outputs when appropriately configured. Yet this article argues that LLMs are not merely efficiency tools but substantively reshape the inferential conditions of data processing, measurement construction, and statistical inference in psychological research. The paper first examines how the probabilistic nature of LLMs systematically produces bias, and subsequently outlines four methodological roles that LLMs play in psychological research. Drawing on empirical studies from clinical, cognitive, and social psychology, it analyzes how LLMs acting as measurement instruments or proxy annotators pose challenges to validity argumentation and error structures. To prevent misplacement of inferential levels, a five-stage checking procedure is proposed, together with practically operable inferential thresholds and operational guidelines, and a call is issued for the scholarly ecosystem to establish corresponding review mechanisms. While psychology must adopt new statistical correction techniques to confront the new sources of variance that LLMs introduce, the ultimate purpose remains to uphold the validity argumentation and error control traditions long developed within psychometrics. This amounts to a principle of“innovation in methods, continuity of standards,”which maintains the auditability and cumulativeness of psychological knowledge.