[愛墾研創·嫣然] 知識輸入與文化創意生產

知識輸入的結構、微調與文化創意生產：從機器創造力到人類文化的跨域洞見

在文化創意生產的脈絡中，創造力從來不是內發、孤立、或無中生有的活動，而是在一個多層次的「知識—規則—行動者」的系統中被集體塑造。Colton、Pease 與 Ritchie（2001）提出的「輸入知識對創造力的影響」模型，提供了重要洞見：創造行為的可變性、原創性與可信度，取決於創作者的知識輸入結構如何被建置、塑形甚至操控。

此觀點與文化創意研究中常見的三元素——物質技術基礎（resources）、社會規則場域（rules & norms）、行動者能力（agency）——高度一致。這與你所附的 INPUT–OUTPUT 模型相互印證：創作從來不是單純的個人能力或靈感，而是由「輸入條件」共同形成的。

（一）知識輸入不是中性資訊，而是創作的制度性前提

原文指出：

「知識輸入會影響創造力，而理解創造力之前，必須先理解輸入結構的性質。」（In order to address… we first discuss the nature of the kinds of program…）

這意味著：

1 創造不是從零開始，而是在已有知識（inspiring set）中選取與再組織。

2 知識輸入並不單是「資料量」問題，而是具有深層結構：分類、權重、典範、禁忌、範式。

在人類文化中，這些知識輸入包括：

— 前人影像、故事、符碼與風格
— 社會對「可接受的創意」的規範
— 技術所允許的創作方式
— 某文化對某些形式的歷史偏好（如美學、敘事、語言）

因此，文化創作的前提並非空白的自由想像，而是「文化—技術—規範所共同提供的可用性空間」。

（二）微調（fine-tuning）與創造力的可疑性：文化與AI的共同命題

原文極具洞見地提出：

「如果一個系統被過度微調以產生特定結果，人們會質疑它的創造力。」

在 AI 領域，微調意味著：

— 資料集被刻意挑選
— 演算法被設定為偏好某類輸出
— 產出看似“巧妙”，但本質是模板化與預訓練的延伸

在文化領域，這種情況更為普遍，例如：

— 政策導向式的文化生產
— 商業化公式化的內容（如 K-pop、流量劇）
— 只複製經典風格的文創商品
— 社會規範預先決定“什麼是好作品”

過度微調的文化生產會造成：

1 創意匱乏（creativity scarcity）：作品過度雷同，缺乏陌生化。

2 可預測性上升（predictability）：越可預測，越不創新。

3 文化生命力降低（cultural vitality erosion）：文化被框限在特定美學或敘事中。

因此，微調不只是技術議題，更是文化政治議題。

（三）創造力的可信度取決於：輸入的廣度、自由度與系統的可變性

Ritchie 的創造力框架指出三大判準（impressiveness, typicality, novelty），而這三者都直接依賴於輸入結構：

1. 廣度（breadth）

輸入越多樣，創造力越可能跨越範式。

2. 自由度（freedom）

輸入越鬆綁，創作者越能突破社會規則與典範約束。

3. 可變性（variability）

允許輸入在規則邊界內自由組合，創造更高不可預期性。

這可與你上方的 INPUT–OUTPUT 模型對應：

換言之：創造是輸入與條件的產物，而非純粹主體意志的展現。

（四）文化創意的啟示：真正的原創來自不被控制的輸入

對文化創意政策與內容生產而言，此理論有重大啟示：

（1）文化制度不應只提供範例，而應提供“異質性”

若文化只提供經典、傳統、主流，那創作便難以跨越。

（2）創作者的知識輸入不應被單一路徑訓練（micro-fine-tuning）

過度職訓化的創作者，其想像力常被框限。

（3）創意產業應重視「輸入的自由度」而非「輸出的形式化」

現今內容平台（如 TikTok）注重輸出格式的一致化，而非輸入環境的開放，這會削弱創新的文化力。

（4）AI 與人類創意的對照讓我們理解：

文化創造力是「輸入的政治」

誰決定我們看什麼？
誰決定什麼是“好”的輸入？
誰決定哪些知識可以被納入創作？

文化創造力即文化自由度。

（五）結論：創造力不是產物，而是輸入結構的生成效應

綜合 Colton、Pease 與 Ritchie 的框架，我們可提出文化創造力的核心論證：

創造力不是神秘的靈感，也不是演算法的奇蹟，而是輸入知識的架構、社會規則的邊界、以及創作者能動性的交互生成。

當輸入多元、規則鬆綁、行動者具備能力與自由時：創造力呈現出原創性與不可預測性。

當輸入單一、規則僵化、行動者被制度微調時：
創造力退化為模板化與可預期的再生產。

因此，文化創意的真正課題不是「如何創造內容」，而是：如何設計一個能夠孕育真正創造力的輸入環境。

A good system should suggest new areas of the search space to explore, i.e. find artefacts which we had notthought of generating before. If we accept this analogy, then Popper's criteria for evaluating theories sheds light on criteria for evaluating creative programs.

Popper sets out two criteria for a satisfactory theory (in addition to it logically entailing what it explains). Firstly it must not be ad hoc. That is, the theory (explicans) can not itself be evidence for the phenomena to be explained (explicandum), or vice versa.

For example if the explican dum is this rat is dead', then it is not enough to suggest that this rat ate poison' if the evidence for it having done so is it being now dead. There must be independent evi dence, such as the rat's stomach contains rat poison'.

The opposite of an ad hoc explanation is therefore one which is independently testable. Secondly, a theory must be rich in content. For example, a theory which explains phenomena other than the specific phenomenait was designed to explain has a much richer content, and therefore has greater value than one which is less general (the principle of universality).

Applying these criteria to creative programs, if we see a programP asthetheoryandtheset ofartefactsA we wish to generate as the phenomena to be explained, then we are in terested in the independent testability of P and the richness of its content.

A program which has been carefully tailored in order to produce very specific artefacts cannot be claimed to be a good program on the grounds that it produces those artefacts. There must be independent grounds for its value, such as also generating other valuable artefacts.

Within the programming analogy, this is clearly connected to the rich ness of content criterion; the more valuable artefacts outside of A and fewer worthless artefacts a program generates, the better that program is. It is important to note that Popper's criteria are general for all scientific theories, applying to single statement explana tions as well as all-encompassing theories.

They therefore apply to any program(including subsets of larger programs) able to generate artefacts in A. The conclusion of the anal ogy is that we should aim to make our programs as general as possible. That is, any creative program which re-invents already known artefacts should also generate a reasonable number of new, valuable, artefacts.

From:The Effect of Input Knowledge on Creativity by Simon Colton, Alison Pease, Graeme Ritchie fr. Informatics Research Report EDI-INF-RR-0055. DIVISION of INFORMATICS, Centre for Intelligent Systems and their Applications, Institute for Communicating and Collaborative Systems November 2001， Appears in Proceedings of ICCBR-2001.

Motivation from the Philosophy of Science ～Following Tarski, Popper suggests that we divide the uni versal class of all statements into true and false, T and F, (Popper 1972).

He claims that the aim of science is to discover theories (explanations) whose content covers as much of T and as little of F aspossible, where the content of athe ory is the set of all statements logically entailed by it.

This set may also be divided into true and false statements (the theory's truth and falsity content). A good theory should suggest where to look, i.e. new observations which we had not thought of making before.

This is comparable to a situation where the universal class of all basic items in a domain is divided into good and bad, V and V0. If we describe the content of a program as its output set O which may be divided into good and bad arte facts, then we can claim that one aim of a creative program is to generate as much of V (and as little of V 0) as possi ble.（中文翻譯）[下續]