ISSUE 53 Expert Witness Journal - Journal - Page 41
The image generation lawsuit
The lawsuit involving image generation AI
technology, distinct from text-based models like GPT3, illuminates the nuanced differences in AI's approach to different mediums and the implications for
copyright law.
Tremblay, Awad et Al. vs OpenAI (the fiction case)
I have chosen this case as emblematic of the cases
brought by various authors of fiction against ChatGPT.
While it obviously contains some elements that are
unique to the plaintiffs the underlying technical issues
are common across all the cases.
Image generation AI, such as DALL-E or similar
models, operates on principles similar to text-based
LLMs but with crucial technical differences. While
LLMs analyze and generate text, image generation AI
is trained on a vast dataset of images and corresponding descriptions. It learns to create new images based
on the patterns, styles, and visual elements observed in
its training data. This process involves complex algorithms like convolutional neural networks (CNNs),
which are particularly adept at processing pixel data
and identifying intricate patterns in images.
The lawsuit brought by Tremblay and Awad against
OpenAI delves into the complex interplay between
the creation of fiction by AI and the principles of copyright law. This case focuses on whether the AI's use of
narrative structures and themes, long-standing in the
literary world, constitutes infringement.
In literature, many stories draw upon common
archetypes and scenarios. The concept of a "stranger
coming to town" or a "hero's journey" exemplifies narrative archetypes that have been used and reinterpreted throughout literary history. These themes are
so ubiquitous because they resonate with fundamental human experiences and emotions. Fiction, by its
nature, often involves reimagining, reinterpreting, or
building upon these existing archetypes and themes.
One salient difference between text and image
generation lies in the interpretative nature of visual
art. While language has a more defined structure and
semantics, visual art is highly subjective and open to
interpretation. This subjectivity poses a unique challenge in determining the originality of AI-generated
art and its relation to the training data. An AI model
does not "understand" art in the human sense but instead generates images based on statistical correlations
within its training data.
In the context of AI-generated fiction, the situation
becomes more nuanced. Large Language Models like
GPT-3 are trained on extensive text corpora, including myriad literary works. These models do not 'understand' narrative in the human sense but can
generate text that follows common narrative structures observed in their training data. The AI's output,
therefore, might echo familiar storylines or themes,
but it does this through pattern recognition and statistical modeling rather than conscious creative
thought.
In the realm of human creativity, artists have long
been inspired by existing works, often reinterpreting
or building upon them to create new art. This process
involves a conscious understanding and transformation of influences, leading to original creations that reflect the artist's unique perspective and intent. In
contrast, while AI can produce novel images, it lacks
conscious understanding or intent. Its creations
are the result of algorithmic processing and pattern
recognition, not creative deliberation.
The key issue in the Tremblay and Awad case likely
revolves around the degree to which AI-generated fiction can be considered original or derivative. While all
fiction draws upon a collective pool of narrative structures and themes, human authors apply their unique
perspective, creativity, and conscious intent to these elements. AI, conversely, lacks personal experience or
intent and generates content based solely on the data
it has been trained on.
The lawsuit in question likely revolves around
whether the training process for these AI models,
which involves the use of copyrighted images, constitutes fair use. Plaintiffs may argue that the AI’s training on copyrighted works constitutes infringement,
particularly if the AI generates images that are closely
reminiscent of specific copyrighted artworks. On the
other hand, defendants might argue that the AI’s output is transformative, a key aspect of fair use, as the AI
does not replicate but rather reinterprets and recombines elements to create something new.
The question of whether merely copying the outline
scenario of a story constitutes copyright infringement
is complex. In traditional copyright law, protection is
not extended to ideas, concepts, or basic plot elements
but rather to the expression of these ideas. Thus, the
mere use of a common archetype like the "stranger
coming to town" would not typically constitute infringement. However, if an AI were to produce a work
closely mirroring the specific expression of a story –
its unique plot details, characters, and dialogue – this
could raise legitimate concerns of infringement.
A critical point in this case is the distinction between
inspiration and replication. While human artists are
inspired by existing works, they apply a conscious creative process to produce something new. AI, lacking
consciousness and intent, does not replicate this process in the same way. However, the output of AI can
still be considered original in the sense that it is not a
direct copy but a novel combination of learned elements. The legal challenge lies in defining and evaluating this "originality" in the context of AI and
copyright law.
EXPERT WITNESS JOURNAL
In the case at hand, the plaintiffs might argue that the
AI has infringed upon specific, expressive elements of
their works. The defense, however, might contend
that the AI-generated content, while inspired by various inputs, constitutes a new and transformative expression, especially given the AI's lack of conscious
mimicry or intent to replicate specific works.
39
F E B R UA RY 2 0 2 4