Learning Machines: An introduction to AI and IP for small and medium-sized enterprises - What is AI?

The term “artificial intelligence” or “AI” has perhaps inevitably become something of a buzzword. It is most frequently used as an umbrella term to refer to a collection of related technological processes and tools, including machine learning, neural networks, generative AI (GenAI) and large language models.

This guide adopts the definition of AI from Getting the Innovation Ecosystem Ready for AI: An IP policy toolkit (WIPO, 2024):

“Artificial intelligence (AI) refers to the branch of computer science and engineering that focuses on creating systems capable of performing tasks that typically require human intelligence. These tasks include understanding natural language, recognizing images, making decisions and learning from data”.

It is important to distinguish between artificial general intelligence (AGI), and specific artificial intelligence (sometimes called "narrow" or "weak" AI). AGI refers to an AI system capable of matching human level intelligence across all domains, performing any task that could be performed by a human, and applying reasoning and problem-solving to entirely new situations without training. Most independent researchers see AGI as theoretical for now. By contrast, specific AI can perform a particular task or narrow group of tasks to a similar standard that a human could, and in some cases consistently outperforms human experts in certain domains. Specific AI is already here and widely used for specialized tasks in many fields.

Even though AGI is not here yet, specific AI still challenges how we think about AI as a creative actor, resulting in many of the IP tensions explored in this guide. Traditionally, AI has been understood as a tool operated by a human, who remains the author or inventor behind the output. If a human uses a hammer to build a table, we do not ask if the hammer is the creator of the table. But when an AI system can generate a novel or a drug compound or a piece of music with minimal human input or direction, the question of who, or what, created it becomes difficult to answer. This is what makes the question of AI inventorship and copyright in AI-generated works so complex.

The term “machine learning” was coined by computer scientist Arthur Samuel in 1959. He defined it as a field of study that gives computers the ability to learn new things through example and experience, as opposed to being explicitly programmed. While machine learning techniques have been around for many decades, three big developments have brought about a visible revolution in this area in recent years. These are:

the growth and widespread availability of massive digital data sets, i.e., “big data”;

unprecedented leaps in computing power; and

advancements in computer science.

Big data are the biggest contributor to recent advancements. Machine learning processes feed on data: the more data and the better its quality, the more accurate the systems and their outputs become. For example, a human, or even a team of humans, may not be able to make much sense of tens of millions of purchase records from a retail chain. But these data can be fed into a computer system and, through machine learning, used to extrapolate sales trends, create inventory reports, detect theft and automate labor. The more purchase records the system has access to, the more accurate its outputs will become.

In the past, computer programs relied heavily on code produced by a human programmer, who would pre-define the features of the sought outcome. This was a time consuming and imprecise process, requiring a lot of computer power. But now, big data, coupled with rapid increases in computing power, have resulted in a genuine shift in algorithms: computational systems can learn and improve as they are exposed to many examples and no longer rely on a priori definition by a programmer of the sought outcome.

An important caveat here is that not all big data are created “equal.” The use of biased and inaccurate data in machine learning can result in biased and erroneous outputs. If a machine learning system is trained on out-of-date, unreliable, or unverifiable data (in some cases, to minimize costs), the results they produce will also be unreliable and out-of-date, and their application in particular fields (such as medicine, transportation or finance) could lead to severe and even fatal consequences.

In 2019, the Max Planck Institute (MPI) for Innovation & Competition engaged in a comprehensive study of the technical aspects of AI from an IP law perspective. The study identifies the “trained machine learning model” (for short, the “model”) as foundational to what we understand as “AI”. These models are based on mathematical functions and generate outputs through the process of learning patterns in the data they are exposed to.

An artificial neural network is a specific type of trained model that draws inspiration from the structure and operation of the human brain to acquire patterns and representations from data. These models follow the connectionist logic of neuroscience, where neurons organized in networks establish associations according to the strength of the synapse connection. Similarly, neural network models consist of nodes, or neurons (which are mathematical functions) connected by “weights” (which represent a numeric value). “Deep” neural networks are arranged into many layers through which they can process data to produce non linear outputs.

To create an artificial neural network, the architecture must be developed by a programmer prior to the entire training process. This architecture is fixed and does not evolve during the machine learning process, and so is sometimes called a “hyperparameter.” However, the weights connecting the neurons are adaptable or “trainable” parameters, meaning they can evolve autonomously as they learn from the data, rather than being preprogrammed.

GenAI, refers to AI tools capable of generating new content based on a user’s prompt. A prompt is usually a short, written description of the desired output. The new content can take the form of text, computer code, images, audio, sound, and video.

The widespread availability of GenAI tools has significant IP implications across many industries, presenting both opportunities and risks. Crucially for IP purposes, GenAI models do not kick off on their own: they must be prompted by a human in a certain way in order to generate an output. Indeed, “prompt engineering” (the process of creating and refining input instructions to cause GenAI models to produce better outputs more quickly) has emerged as a distinct skill that requires an understanding of an AI model's capabilities and limitations.

What is a large language model?

Large language models (LMMs) are a type of neural network model. LLMs power many of the popular GenAI tools available on the open market today. They are so named because they are well suited for processing and correlating language. LLMs can perform various natural language processing tasks, including text generation, language translation, text summarization and more.

LLMs are not confined to processing human language. In machine learning terms, "language" refers to any system of symbols (such as words or code) that conveys meaning through structure and relationships. Just as human language derives meaning from both word choice and grammatical relationships, programming languages follow similar patterns of syntax and context. This is why some LLMs can generate computer code in common programming languages like Python or JavaScript. To effectively capture these complex relationships across any type of language, LLMs are designed to process sequences, handle large data sets and maintain sufficient contextual memory.

Launched in November 2022 by OpenAI, ChatGPT was the LLM chatbot that brought AI to the forefront of public consciousness, quickly transforming how people in many fields work and create. ChatGPT uses natural language processing and deep machine learning to generate written content in response to user prompts. While it was the first of its kind to achieve mainstream adoption, competitors such as Anthropic, Google and Meta released similar chatbots soon after.