English
Computation can be a medium for making art, a method for investigating aesthetic judgments, and a tool that changes creative work. These are different relationships. Writing a program that produces images, predicting which images people prefer, and explaining why an artwork matters require different kinds of evidence.
The connection is useful even for readers who do not intend to become artists. Creative coding makes choices about representation, constraints, interaction, and evaluation visible. It also shows why generating many outputs is not the same task as deciding what is worth making.
Background: Graphics (§6.1), human–computer interaction (§6.2), generative models (§5.3, §5.5), and intellectual property (§9.6).
The Relationship: Computation as Creative Medium and Creative Question
Computation as an Artistic Medium
Rules and chance were used in art before electronic computers. Computing made it practical to execute complicated procedures repeatedly, vary parameters, and respond to input during a work’s presentation. A program can describe a family of possible works as well as produce a particular image, sound, or performance.
Consider a drawing made from a grid of dots. A parameter controls the maximum displacement of each dot from its grid position. With zero displacement, the grid remains regular; increasing the parameter allows more irregular arrangements. The artist chooses the grid, displacement distribution, colors, and which results to retain. Randomness supplies variation within these choices, rather than making the choices disappear.
A pseudorandom seed helps reproduce a run. If the program, inputs, generator, and sequence of calls remain fixed, the same seed reproduces the same random-number sequence. This allows a creator to return to a result and change one property at a time. A seed alone does not guarantee identical output across different program versions or generators.
Harold Cohen’s AARON provides a historical example of explicit rules for drawing and composition. Successive versions encoded different knowledge about shapes, figures, and color. As Boden’s account emphasizes, the available representation affected what the program could draw: adding ways to represent objects and their relationships changed its possible compositions. This connects artistic practice to the general computational question of what a representation makes expressible.
Interactive works add an input–response relationship. A viewer’s movement or a musician’s action may change the output while the work runs. The timing and interpretation of that input become part of the experience. Testing therefore includes how people encounter the work, not only whether a saved image looks as intended.
The Formalization of Aesthetics and Creativity
A computational account of aesthetics must say what it measures. Symmetry, repetition, contrast, and statistical complexity are measurable properties. A predictor trained on preference ratings may estimate the judgments of a particular group under particular conditions. Neither kind of measurement is automatically a universal measure of artistic value.
Boden distinguishes three ways to produce creative novelty. Combinational creativity joins familiar elements in an unfamiliar arrangement. Exploratory creativity finds possibilities within an existing set of rules. Transformational creativity changes some of the rules that define what can be produced. These are ways to analyze a process, not a ranking of every artwork or a test that settles whether a machine has a mind.
The dot-grid example makes the distinction usable. Combining the grid with a photographic texture changes the combination of materials. Searching different displacement parameters explores the existing generator. Replacing the fixed grid with points placed in response to a viewer’s movement changes a defining rule. The classification depends on which rules and system boundary are being considered; it is not always uniquely determined.
Novelty alone is insufficient. A random sequence can differ from every sequence seen before without serving a creative purpose. Evaluation may also concern coherence, relevance to a brief, expressiveness, skill, or the effect on an audience. These criteria can conflict, and their importance varies with the work and context.
A model of a creative process is useful when it explains how possibilities are generated and evaluated. It need not reduce every human judgment to a single score. For a design task, separate hard requirements, such as legible text at the intended size, from preferences about style. A visually striking result can still fail a required use.
Generative AI in Creative Work
Generative AI changes how a creator can specify and explore possible outputs. A prompt, reference image, or partial composition can guide a learned model. The resulting workflow still contains decisions about purpose, constraints, selection, editing, and presentation. Assessing the final artifact alone may conceal which parts of the process were controlled by whom.
Separate three questions: what properties the artifact has, how it was made, and how it is presented or used. The same-looking image can play different roles as a fictional illustration, a claimed documentary photograph, or an exercise in studying a technique. Questions about truthfulness and provenance therefore depend on context, not only visual quality.
Training material and generated outputs also raise questions about permission, attribution, labor, and compensation. Models differ in their data and licenses, and some can reproduce training material. A claim that every output is a copy is too broad; so is a claim that learning makes provenance irrelevant. The legal distinctions in §9.6 apply to particular materials and activities. Economic effects on creative work require evidence about actual tasks and markets, rather than predictions that one tool will replace an entire profession.
For a concrete design exercise, specify a poster’s required text, dimensions, and intended audience. Compare generated candidates for correct wording, readability, and visual hierarchy; then document selection and edits. This gives a beginner observable criteria for evaluating assistance, while leaving room for judgment about what the poster should communicate.
What This Perspective Changes
Creative work exposes a distinction that appears throughout computing: defining a space of possible outputs is different from choosing a useful or meaningful result. Representation and constraints shape that space; human purposes and evaluation shape selection.
Build one small generator with a few interpretable parameters. Save its inputs and seed, compare variations, and explain why you kept one result. This exercise joins programming, observation, and judgment without requiring a claim that creativity has one complete computational definition.
Resources
Foundational and Bridging Texts
Boden’s The Creative Mind develops the three forms of creativity discussed above; Mind as Machine places systems such as AARON in their research context. The Whitney Museum’s AARON materials show examples and explain selected rules. Reas and Fry and Pearson provide routes into making generative work. Birkhoff and computational-aesthetics research should be read as attempts to formalize selected criteria, with attention to what the measures omit.
| Resource | Role | Type |
|---|---|---|
| Boden, The Creative Mind: Myths and Mechanisms (2nd ed.) | A framework for analyzing creative processes | Depth |
| Reas & Fry, Processing + creative coding literature | Programming as a medium for visual work | Practice |
| Pearson, Generative Art | Generative procedures and artistic practice | Entry |
| Birkhoff, Aesthetic Measure / computational aesthetics surveys | Attempts to measure selected aesthetic properties | Depth |
Practice and Current Sources
Choose one environment and make a small work before comparing many tools. The p5.js random-seed reference supports reproducible experiments with a generator. The Coding Train and community examples provide exercises; music systems offer a different medium for rules, time, and interaction. When using generative AI, record the human decisions and edits as well as the output. Discuss a particular work or workflow so that questions about value and authorship remain concrete.
| Resource | Role | Type |
|---|---|---|
| Processing / p5.js / openFrameworks (free) | Environments for interactive and generative work | Practice |
| p5.js reference and examples (free) | Documented examples for browser-based work | Practice |
| The Coding Train (free; optional paid support) | Guided creative-coding projects | Auxiliary |
| OpenProcessing (free; paid Plus options) | Examples to inspect, modify, and discuss | Practice |
| SuperCollider / TidalCycles / Sonic Pi (free) | Algorithmic composition and live coding | Practice |
| Generative AI tools (§5.3, §5.5) | Model-assisted generation, selection, and revision | Practice |
| Contemporary AI-art discourse (ongoing) | Arguments about creative work, authorship, and institutions | Reference |
Traps
| Trap | Why it misleads | Better response |
|---|---|---|
| Treating randomness as creativity by itself | Variation does not establish purpose, coherence, or value. | Explain what the rules generate and how results are evaluated. |
| Equating an aesthetic score with universal artistic value | A measure captures selected properties or judgments under particular conditions. | Identify whose judgments and which criteria the measure represents. |
| Inferring the entire creative process from a finished artifact | Similar outputs may have different origins, purposes, and human contributions. | Examine the process and presentation as well as the result. |
| Assuming a new tool settles authorship or eliminates a profession | Rights depend on particular contributions and rules; work consists of multiple tasks. | Analyze the actual workflow, permissions, and evidence of change. |
中文
计算可以成为创作艺术的媒介、研究审美判断的方法，也可以成为改变创作过程的工具。这些联系并不相同。编写生成图像的程序、预测人们更喜欢哪些图像，以及解释一件作品为何有意义，需要不同类型的证据。
即使不准备成为艺术家，这种联系也有学习价值。创意编程让表示、约束、交互与评价方面的选择变得可见，也说明生成许多结果与判断什么值得创作，是不同任务。
背景知识：图形学（§6.1）、人机交互（§6.2）、生成模型（§5.3、§5.5）与知识产权（§9.6）。
这种关系：计算作为创造媒介与创造问题
计算作为艺术媒介
在电子计算机出现以前，艺术已经使用规则与偶然性。计算使复杂过程的反复执行、参数变化，以及作品展示期间对输入的响应，更容易实现。程序既可以生成一幅图像、一段声音或一次表演，也可以描述一组可能作品。
设想一幅由点阵构成的画。一个参数控制每个点偏离网格位置的最大距离。位移为零时，点阵保持规则；增大参数则允许更不规则的排列。创作者选择网格、位移分布、颜色，以及保留哪些结果。随机性在这些选择之内提供变化，并没有让选择消失。
伪随机种子有助于复现一次运行。如果程序、输入、生成器与调用顺序不变，相同种子就会复现相同的随机数序列。创作者因此可以返回某个结果，每次只改变一种属性。仅有相同种子，不能保证不同程序版本或生成器产生完全相同的输出。
Harold Cohen 的 AARON 提供了以显式规则进行绘画与构图的历史例子。不同版本编码了有关形状、人物与颜色的不同知识。Boden 的记述强调，可用的表示影响程序能画什么：增加对象及其关系的表示方法，会改变可能的构图。这把艺术实践联系到一个普遍的计算问题：某种表示让哪些内容能够被表达。
交互作品还增加了输入与响应之间的关系。观众的动作或音乐家的操作，可以在作品运行时改变输出。输入的时机与解释方式，都成为体验的一部分。因此，测试还包括人们怎样接触作品，而不只是保存下来的图像是否符合预期。
美学和创造力的形式化
计算性的审美解释需要说明测量什么。对称、重复、对比与统计复杂性，都是可测量的属性。根据偏好评分训练的预测器，可能估计特定群体在特定条件下的判断。这两类测量，都不会自动成为艺术价值的普遍尺度。
Boden 区分了产生创造性新意的三种方式。组合式创造把熟悉元素放进不熟悉的组合。探索式创造在已有规则内寻找可能性。变革式创造改变定义可生成内容的部分规则。这些是分析过程的方法，不是对所有作品的价值排名，也不能用于直接判定机器是否具有心智。
点阵例子可以说明区别。把点阵与摄影纹理结合，改变了材料的组合。搜索不同位移参数，是在已有生成器中探索。把固定网格改成根据观众动作布置的点，则改变了一条定义性规则。分类取决于关注哪些规则以及怎样划定系统范围，并不总有唯一答案。
仅有新颖性还不够。一个随机序列可以不同于此前所有序列，却不服务于任何创作目的。评价还可能涉及连贯性、与任务要求的相关性、表现力、技巧或观众体验。这些标准可能冲突，其重要性随作品与情境变化。
创作过程模型的价值，在于解释怎样产生与评价可能方案，不必把所有人类判断压缩为一个分数。对于设计任务，应区分硬性要求与风格偏好，例如文本在目标尺寸下必须可读。视觉上醒目的结果，仍可能无法满足使用要求。
创作过程中的生成式 AI
生成式 AI 改变了创作者指定与探索可能输出的方式。提示词、参考图或未完成的构图，可以引导学习模型。由此形成的流程，仍然包含目的、约束、选择、编辑与展示方面的决定。只看最终作品，可能看不出过程中的哪些部分由谁控制。
应区分三个问题：作品具有什么属性、怎样制作，以及怎样呈现或使用。外观相同的图像，作为虚构插图、声称真实的纪实照片，或学习技法的练习时，作用可以不同。因此，真实性与来源问题取决于情境，不只取决于视觉质量。
训练材料与生成输出，还涉及许可、署名、劳动与报酬。不同模型的数据和许可证不同，有些模型可能复现训练材料。声称每个输出都是复制品过于笼统；声称学习使来源不再重要，同样过于笼统。§9.6 的法律区分需要应用到具体材料与活动。判断创作工作的经济变化，应依据实际任务与市场的证据，而不是直接预测一种工具会替代整个职业。
可以做一个具体设计练习：规定海报必须包含的文字、尺寸与目标受众。比较生成候选的文字正确性、可读性与视觉层次，再记录选择和修改。这让初学者拥有可观察的辅助效果评价标准，同时保留对海报应该传达什么的判断空间。
这种视角会改变什么
创作使计算中反复出现的一个区别变得清楚：定义可能输出的范围，与选择有用或有意义的结果，是不同任务。表示与约束影响可能范围，人的目的与评价影响选择。
可以制作一个参数少、含义清楚的小型生成器。保存输入与种子，比较变化，并解释为什么保留某个结果。练习把编程、观察与判断结合起来，无须先断言创造力存在一个完整的计算定义。
资源
基础与桥接文本
Boden 的《The Creative Mind》发展前文讨论的三种创造方式；《Mind as Machine》把 AARON 等系统放在研究背景中。Whitney Museum 的 AARON 材料展示作品并解释部分规则。Reas、Fry 和 Pearson 提供制作生成式作品的入门路线。阅读 Birkhoff 与计算审美研究时，应把它们视为形式化特定标准的尝试，同时关注度量省略了什么。
| 资源 | 作用 | 类型 |
|---|---|---|
| Boden，The Creative Mind: Myths and Mechanisms（第 2 版） | 分析创作过程的框架 | 深入 |
| Reas & Fry，Processing + 创意编程文献 | 把编程作为视觉创作媒介 | 实践 |
| Pearson，Generative Art | 生成过程与艺术实践 | 入门 |
| Birkhoff，Aesthetic Measure / 计算美学综述 | 测量特定审美属性的尝试 | 深入 |
实践与当前来源
先选择一个环境完成小作品，再比较多种工具。p5.js 随机种子参考有助于开展可复现的生成器实验。The Coding Train 与社区例子提供练习；音乐系统则为规则、时间与交互提供另一种媒介。使用生成式 AI 时，除了输出，也应记录人的决定和修改。围绕具体作品或流程讨论，可以使价值与作者身份问题保持具体。
| 资源 | 作用 | 类型 |
|---|---|---|
| Processing / p5.js / openFrameworks（免费） | 交互与生成式创作环境 | 实践 |
| p5.js 参考与示例（免费） | 浏览器创作的文档与例子 | 实践 |
| The Coding Train（免费；可选付费支持） | 创意编程项目指导 | 辅助 |
| OpenProcessing（免费；Plus 选项付费） | 可检查、修改与讨论的例子 | 实践 |
| SuperCollider / TidalCycles / Sonic Pi（免费） | 算法作曲与现场编程 | 实践 |
| 生成式 AI 工具（§5.3，§5.5） | 模型辅助的生成、选择与修订 | 实践 |
| 当代 AI 艺术讨论（持续发展中） | 有关创作、作者身份与制度的论证 | 参考 |
陷阱
| 误区 | 为什么会误导 | 更好的做法 |
|---|---|---|
| 把随机性本身当作创造力 | 变化不能证明目的、连贯性或价值。 | 解释规则生成什么，以及怎样评价结果。 |
| 把审美分数等同于普遍艺术价值 | 度量只捕捉特定条件下的部分属性或判断。 | 明确度量代表谁的判断与哪些标准。 |
| 从完成作品推断整个创作过程 | 相似输出可能具有不同来源、目的与人类贡献。 | 同时检查过程、呈现方式与结果。 |
| 认为新工具能直接决定作者身份或消灭一种职业 | 权利取决于具体贡献与规则，工作也包含多种任务。 | 分析实际流程、权限与变化证据。 |