Recent developments in cognitive science and social research indicate a shift toward the utilization of artificial intelligence to challenge traditional heuristic biases and enhance the predictive accuracy of experimental outcomes.

認知科學與社會研究的最新發展表明，目前的趨勢是利用人工智慧來挑戰傳統的啟發式偏差，並提高實驗結果的預測準確度。

The prevailing adherence to the principle of parsimony, commonly identified as Occam’s razor, posits that the simplest explanation fitting the available data is preferable. However, research conducted by Marina Dubova at the Santa Fe Institute suggests that this preference for simplicity may impede the identification of complex underlying structures. Through the deployment of computational agents, Dubova demonstrated that models prioritizing complexity can achieve predictive parity or superiority over those adhering to parsimony. Furthermore, the study indicates that exploratory, novelty-driven experimentation often yields more accurate representations of ground truth than theory-guided approaches, which are frequently susceptible to confirmation bias and the rigidity of prior beliefs.

目前普遍遵循的簡約原則（通常被稱為「奧卡姆剃刀」）認為，在符合現有數據的解釋中，最簡單的一個是較佳的。然而，聖塔菲研究所的 Marina Dubova 進行的研究表明，這種對簡單的偏好可能會妨礙對複雜底層結構的識別。透過部署計算代理，Dubova 證明了優先考慮複雜度的模型，在預測能力上可以達到與遵循簡約原則的模型相當甚至更優的水平。

Parallel to these epistemological inquiries, the application of Large Language Models (LLMs) in the social sciences has demonstrated significant utility in forecasting experimental results. An analysis of 70 preregistered survey experiments in the United States revealed that GPT-4 and other open-weight models could simulate treatment effects with a correlation strength comparable to pooled human forecasts, even for data not present in their training sets. While these models exhibited a systematic tendency to overestimate effect sizes, their capacity to mirror human responses suggests a potential for the automation of pilot testing and the optimization of intervention selection. Consequently, the transition toward 'AI scientists' necessitates a critical evaluation of whether traditional scientific methods are optimal or merely reflections of human cognitive limitations.

與這些認識論探究平行的是，大型語言模型（LLM）在社會科學中的應用，在預測實驗結果方面展現了顯著的實用性。一項對美國 70 個預先登記的調查實驗分析顯示，GPT-4 與其他開源權重模型可以模擬處理效應，其相關強度與人類綜合預測相當，即使是對於訓練集中不存在的數據也是如此。雖然這些模型表現出系統性高估效應值的傾向，但它們模仿人類反應的能力，顯示出自動化初步測試與優化干預選擇的潛力。因此，向「AI 科學家」轉型的過程，需要對傳統科學方法是否最優，或者僅僅是人類認知局限的反映進行批判性評估。

The convergence of high-dimensional computational modeling and LLM-driven simulation offers a mechanism to accelerate discovery by bypassing human cognitive constraints and traditional heuristic biases.

高維計算建模與 LLM 驅動模擬的融合，提供了一種透過繞過人類認知限制與傳統啟發式偏差來加速發現的機制。