The New Copyright Challenge in the AI Era: Does AI Creation Infringe Copyright?
In the digital creative field, artificial intelligence (AI) and its applications in generative AI have ushered in a new era, challenging the boundaries of traditional copyright law.
Contrary to common belief, the machine learning process of AI is not similar to the typical reproduction or imitation of infringers. Instead, AI’s machine learning is more akin to human learning, as it learns the logic and presentation order of colors, language, and pattern arrangements through exposure to works. It then uses this learning to produce new works that do not substantially resemble the learning materials.
This article aims to delve into the complexity of AI machine learning and clarify its legal implications by citing recent important cases to correct common misconceptions.
How does AI generate images and text?
Generative AI works by learning from data and information to create new content. Unlike typical infringement patterns such as “distribution,” “reproduction,” and “adaptation,” it primarily “learns” from existing works and data.
Image Generation:
1. Data Analysis and Pattern Learning:
Generative AI used for image creation begins by analyzing a large amount of image data. This includes identifying objects in images and understanding deeper elements such as brush texture, color gradients, lighting, and spatial relationships. For example, a mature AI generation, when learning landscape paintings, applies different elements such as brushstrokes, color blending techniques, and the interaction of light and shadow after recognizing various elements.
2. Feature Extraction:
Convolutional neural networks in AI algorithms extract specific features from image works, separating and identifying various features such as edges, shapes, and textures. “Feature extraction” is crucial for AI robots to understand different artistic styles, brushstrokes, and techniques.
3. Generation of New Works:
Once AI has learned specific techniques and artistic styles through feature extraction and data analysis, it can generate new images. This is usually done using Generative Adversarial Networks (GANs). GANs consist of image generators and image discriminators, and through their interaction and iterative process, generative AI ultimately produces images that are stylistically and feature-wise similar to the training data (often referred to as potentially infringing works), but do not have substantial similarity when compared directly to the training data.
Text Generation:
1. Data Acquisition and Language Model Building:
For text generation, AI models like ChatGPT absorb vast amounts of text data from diverse sources such as books, articles, website content, and even conversation records. AI constructs a language model that can understand grammar and infer context from this textual data.
2. Language Prediction:
The most common language prediction model in text generation AI is n-gram, which calculates the probability of words or phrases following specific words or phrases to achieve idiomatic expression, narrative structure, and subject-object agreement. However, n-gram language prediction models are limited to handling more complex text generation tasks.
3. Encoding and Text Understanding:
For complex tasks that involve extending context and generating entire texts, n-gram models fall short because they can only predict based on limited contextual information, without understanding the semantic meaning of the text. In contrast, Transformer models utilize self-attention mechanisms to transform the text into vectors through input embedding and incorporate positional encoding to understand the overall context of the text.
4. Text Generation:
After deep understanding of the text through the encoder, the decoder is responsible for generating text based on the learned text features. The Transformer model’s characteristics enable it to generate coherent and creative text that is logically consistent and original in content, based on a deep understanding of the original text’s semantic meaning.
How does AI generation differ from copyright infringement?
From the principles of image and text generation mentioned above, it is clear that the way AI generates content is vastly different from the infringement patterns defined by copyright law. The differences are particularly evident in the following points:
1. The Creative Nature of AI:
Generative AI does not simply “copy” or “reproduce” the data it learns (existing works). Instead, it learns the underlying logic, structural patterns, and styles of the text or image from a vast amount of data and then creates new works that possess novel qualities. For example, in image generation, although AI may learn from existing artworks, the resulting images are not replicas or reproductions of existing works. Instead, they are new creations that stem from the recombination and reinterpretation of the deep learning outcomes.
2. Legal Interpretation:
From a legal perspective, there is a significant distinction between AI-generated content and human reproduction. The fundamental concept of copyright law is to “protect the expression of ideas, not the ideas, concepts, or systems themselves.”
Based on the above, AI-generated works learn the underlying logic, structural patterns, artistic styles, brushstrokes, and concepts from training data (original works), but they do not aim to “reproduce” or “replicate” the expression of the training data (original works).
The way AI generates works clearly challenges the boundaries of traditional copyright infringement. In the representative case of “Andersen v. Stability AI Ltd,” the legal focus of the lawsuit was whether Stable Diffusion’s generated images constituted infringement, given that it used copyrighted images for AI training without constituting infringement.
3. Transformation and Fair Use:
When discussing whether AI-generated works constitute infringement, the question of whether they possess sufficient transformative qualities arises. This means that generative AI adds additional expressions or even imparts new meanings based on the original works. This is where the possibility of “fair use” comes into play.
This depends on the AI’s ability to create significantly different works from the original works. The current ban on providing AI tools for altering original works, such as DALL-E, is aimed at avoiding legal disputes of this nature.
The controversy over using AI to alter existing works reached its peak with the recent global sensation “Palworld,” which used generative AI to create modified versions of multiple Pokémon, including fusions of multiple Pokémon. Regarding text generation, the Thomson Reuters v. Ross Intelligence case delved into a comprehensive discussion on whether AI-generated legal documents constituted fair use and concluded with an affirmative result.
The impact of generative AI on copyright law:
The process of AI content generation, as demonstrated in the contexts of images and text, showcases a creative form that differs from direct copying or reproduction. This distinction is crucial in understanding why AI’s learning and generation methods differ from copyright infringement patterns.
As AI continues to develop, the existing legal framework and interpretations must be revised and adapted accordingly. The ever-evolving AI technology poses constant challenges to the traditional notions of copyright infringement.
The distinction between the “learning and generation process of AI” and the “resetting patterns of copyright” and the “learning of ideas and concepts by humans” goes beyond mere definitions. It also involves profound legislative logic and interpretations of creative ethics.
With the advancement of AI technology, the existing legal framework will undoubtedly undergo revisions. However, the direction of such legal amendments depends on how lawmakers balance the “AI’s innovative potential” and “the protection of original works.”
Therefore, the next time you encounter a debate on whether generative AI constitutes copyright infringement, remember that it involves a delicate balance between these two values and refrain from prematurely concluding that “generative AI infringes upon the copyright of original works.”
Opinion articles present diverse viewpoints and do not represent the stance of “WEB3+.”