17th January 2024
In 2024, there seem to be as many definitions for artificial intelligence as there are strong opinions about the technology. This appears to be because AI just won’t stop changing. As AI continues to advance, the scope of what this type of technology is capable of grows. At the same time, as AI becomes more of an everyday technology for people globally, what AI is to any individual person is constantly changing. A good definition of AI must therefore encompass the full capabilities of the technology now, AI’s likely future capabilities, and be relevant to how the average user experiences AI on a daily basis. Starting to see why there are so many different definitions out there?
Picking a single definition from the existing pile is difficult. While the obvious solution appears to be to just come up with our own definition, we also don’t want to add to the existing pile if we can avoid it. As such, the goal of this briefing is to review a handful of existing definitions that we like. For each definition, we’ll highlight language that we feel captures some aspect of AI particularly well and expand on why. This also allows us to highlight commonalities between different ways of defining AI. By sharing the process as much as the solution, our goal is to furnish you the reader with the tools to define AI for yourself.
“Systems that think or act like humans”
Sometimes, less is more. We like how succinct IBM’s definition of AI is, especially given that it is one of the top Google search results for “define AI”. As a place to start, this definition is excellent.
It’s worth noting that IBM’s “What is artificial intelligence (AI)?” article discusses two approaches to defining AI. These are the Human Approach found above, and the Ideal Approach, which defines AI as systems that think or act rationally. We find the Human Approach to be the better of the two for a few reasons.
First, it avoids the messy discussion needed to define rationality. If you need to define part of your definition for AI, then your definition is likely too complicated. The second reason we prefer the human approach is that it implies fallibility. Humans make mistakes, and often make decisions based on incorrect information. AI suffers from the same issues, seen in AI hallucinations, or in AI models developing racist biases. The perception that AI is somehow infallible or perfectly objective contributes to several AI-related issues and is something we feel is critical to discuss. By defining AI by human thought and action, we bake that fallibility into the definition, opening the door to these critical conversations.
While IBM’s human approach definition is an excellent starting point, it lacks precision. While such a broad definition is perfect for anyone who only needs a basic understanding of artificial intelligence, those of us grappling with AI issues and policy need a more nuanced definition.
National Institute of Standards and Technology (NIST)
“Engineered or machine-based system that can, for a given set of objectives, generate outputs such as predictions, recommendations, or decisions influencing real or virtual environments. AI systems are designed to operate with varying levels of autonomy”
We like this definition of AI from NIST, as we feel that the inclusion of specific language makes the definition precise. However, compared to IBM’s short-and-sweet definition, this one is certainly less approachable.
Like IBM, the NIST definition frames artificial intelligence as a type of system. Defining AI as a type of system accurately represents how the technology functions, as collection of interacting programs, models, and physical pieces of technology. Conceptualizing AI as a system is more accurate than thinking of it as a singular object, as is common in everyday conversation (“an AI”).
This definition also specifies that AI systems generate outputs, and that these outputs can influence both real and digital environments. Real and virtual environments are occupied by people. If AI influences these spaces, then AI can influence the people occupying them. Understanding this concept is critical to discussing the impact AI technologies have on society, so we like the inclusion of this language in the definition.
While other definitions implythat AI systems operate autonomously to some degree, we like that NIST specifies that the level of autonomy varies. We feel that discussing levels of autonomy can help newcomers to AI understand just how many interactions they have with AI systems daily. From NPCs in video games (low autonomy) all the way up to self-driving cars (high autonomy), AI systems are everywhere. Mentioning how AI autonomy varies opens the door to discussions where these everyday interactions can be used as illustrative examples.
Urban Libraries Council (ULC)
“The use of data, machines, and algorithmic processes to simulate or extend human capacity to perform tasks or make informed decisions.”
While not as clearly stated as it is in the previous two definitions, ULC’s definition again frames AI as a system (“data, machines, and algorithmic processes”). In this definition, we also again see AI defined by what humans are capable of. As noted above in the discussion of the IBM definition, we tend to favor this approach for a few reasons. Beyond the point about baking fallibility into the definition of AI, defining AI by human capability makes understanding AI more accessible. Human capacity is a reference point that everyone has. Including this common reference point when attempting to define AI gives even someone with no experience with the technology a starting point.
What we appreciate most about ULCs definition is that it goes beyond defining AI by human capacity, further defining the technology by how people can interact with it. The specific language “…to simulate or extend human capacity” is where we find this. AI technology can extend/enhance human capacity to do certain tasks or make certain decisions, but the technology can also simulate these capabilities. This definition recognizes that AI technology is both a tool to be used by people, as well as a technology capable of replacing people in certain roles. Some of the most important discussions about AI are centered around how the technology has the capacity to improve human productivity, while at the same time posing a threat to millions of jobs worldwide. Grappling with these issues is critical if we want AI to improve the quality of life of its human users. As such, we appreciate the recognition of AI’s dual capacity to either simulate or extend human capabilities in ULC’s definition of AI.
Defining by Commonalities
Having reviewed three different ways of defining artificial intelligence, we can pick out some commonalities. These commonalities are a collection of approaches to understanding AI, as well as specific language used to describe it. While not a definition in the traditional sense, the list below is intended to help both beginners and experts better understand what AI technology is, what it is capable of, and how people can interact with it.
- Artificial intelligence technology (AI) is best understood as a system
- AI systems try to replicate the capabilities of humans in one or more ways
- AI systems, like humans, are fallible
- The level of autonomy an AI system operates with is variable
- AI systems generate outputs that can influence the real and virtual worlds
- An AI system’s outputs are determined in part by its inputs