AI through six imperfect yet helpful metaphors
These reveal AI’s capabilities, limitations and mysteries, helping us understand an inherently untidy and unpredictable technology.
Metaphors are by definition inexact shortcuts for understanding the world. Not only do they enrich language, they impact what gets noticed, what gets ignored and therefore what action we take. They help us see more clearly, but only in a particular way.
Life is like a journey – we talk about being at a crossroads, going in the right direction, taking a different path and reaching a destination.
Reality TV shows are like a journey – see above.
Time is like money – we spend time, save time, waste time and invest time.
Arguments are like a fight – we defend a position, attack an argument, shoot down an idea and win or lose.
The internet is no stranger to the world of metaphors either.
The web is like a place – we visit websites, go online, enter a site, leave a page and return home.
Ecommerce is like a physical shop – we browse, place items in baskets and proceed to checkout (we even used to abandon shopping trolleys, albeit that particular metaphor has gently retired).
It should be no surprise that as technologists, business leaders and society more generally wrestle with AI, metaphors have emerged to help us determine what to notice, what to ignore and what action we might take. Below, I outline six that I have found particularly useful.
The jagged frontier
This concept was first introduced by Harvard's Digital Data Design Institute (now renamed the Harvard Business School AI Institute) in 2023 in their paper Navigating the Jagged Technological Frontier. Chair and co-founder Karim Lakhani and others explored the evidence of the effects of AI on 758 Boston Consulting Group consultants.
Unsurprisingly, the researchers found that AI capability wasn’t distributed equally across tasks and challenges. For tasks inside AI’s competence, consultants using the tool completed more work, more quickly and to a higher quality. However, for challenges sitting outside core competence, consultants were more likely to reach the wrong answer.
So far so unsurprising.
The surprising part is the apparently arbitrary nature of what lies inside and outside that border. AI is poor at things you might consider straightforward and surprisingly good at things which seem more difficult.
Why does this matter? It’s good to be reminded that AI’s intelligence is artificial and so it’s good to know which side of the frontier we are located in, that the frontier is always moving, and that for the foreseeable future it will remain jagged.
Where does it fall short? The jagged edges are getting smoother all the time, so there’s every chance it just won’t be true in a few years.
Spicy autocomplete
It’s difficult to attribute this phrase to an individual author, rather it appears to have emerged from the tech geekosphere (Reddit / X / Hacker News) in early 2023 and subsequently amplified and codified by respected commentators such as Dan Shapiro and Mike Solomon on his The Cleverist blog.
It’s an evolution of an earlier idea that AI is merely autocomplete on steroids. The word ‘spicy’ suggests that AI is unpredictable, exciting (sometimes) and flavoursome in how it sees its job. Yes, it autocompletes, but with creativity and not always with obviousness.
Why does it matter? It is a reminder that large language models merely predict the completion of a conversation by outputting text, images, video or software code. It doesn’t retrieve an answer from a database, rather it generates an output one small unit at a time based on what has gone before. Flavoursome autocomplete, if you will.
Where does it fall short? It’s a particularly harsh way to describe a prediction process which is carried out across billions of parameters and enormous amounts of context. It’s a little like describing a jet engine as a device which moves air backwards.
Uncanny valley
Japanese roboticist Masahiro Mori introduced this term way back in 1970, to describe the phenomenon that human affection for robots increases as they become more human-like, but only up to a point. As the robot begins to look and act in a more human way, the small differences between the robot and the human become exacerbated.
Robot arm working on a production line? Cool.
Cartoonish speaking robot with big eyes? Cute.
Prosthetic hand not coloured to match its wearer’s skin? Yikes.
Humanoid face which makes strange shapes? Weird!
In the context of AI, a rudimentary chatbot which offers three options is clearly a machine, and we tend to judge it as an interface. However, once the chatbot speaks conversationally, remembers context and appears to display empathy, we tend to judge it as a human and so are harsher when it falls shy of full human characteristics.
Why does it matter? Humans are properly unsettled by uncanny valley, so if what you publish is entirely AI-generated, you will pay a high price when your customers (or any human, actually) identify that, and they very probably will. Much AI-generated prose is identifiable (or at least suspicious) and AI-generated images are often recognisable (with the current overuse in pub menus and promotional flyers a particular bête-noire for this author).
Where does it fall short? It doesn’t, it’s an excellent description of how humans perceive technologies differently and judge them more harshly as their form and function become increasingly human-like.
The stochastic parrot
This term was introduced by Emily Bender, Timnit Gebru, Angelina McMillan-Major and Margaret Mitchell, a group of linguists, ethicists and data specialists with a particular interest in understanding, managing and avoiding bias in large data sets. Their 2021 paper introduces the term in the title, On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?
‘Stochastic’ means involving probability or randomness and so a parrot with this characteristic is mimicking human language, assembling sequences of human sounds, based on pattern matching, without necessarily understanding their meaning. The term was nominated as the 2023 AI-related Word of the Year by the American Dialect Society.
Why does it matter? The phrase is a timely reminder that producing meaningful language and possessing meaning aren’t the same thing. It can predict without understanding, justify without reasoning, and convince without insight.
Where does it fall short? Humans aren’t perfect insight, reasoning and understanding machines either, and we use biases and shortcuts all the time to jump to conclusions we don’t fully understand. So, it’s somewhat hypocritical to look down on machines exhibiting those identical characteristics. As well as this, much of human learning involves positive mimicry, particularly in childhood (copying parents, siblings and friends) and early careers (observing and replicating the behaviours and skills of more experienced peers and bosses). The ability to mimic intelligence is still an impressive feat, certainly some distance from squawking “who’s a pretty boy then” in pet shop windows.
AI is simultaneously a prediction engine, a creative assistant, a reasoning partner, a statistical machine, a productivity tool and a system whose inner workings remain partly mysterious.
The cognitive prosthetic
This term evolved over a number of decades and through a number of lenses. It was first introduced in the late 1980s by a group of academics from the University of St Andrews. Their work introduced the concept that practical cognitive prosthesis could function to support and compensate for patient cognitive limits or impairments in the same way physical prosthetics compensate for physical impairments. This was evolved, in the context of intelligent systems, by Kenneth Ford and Patrick Hayes at the Florida Institute for Human and Machine Cognition in 1997 to explore how such systems could complement and enhance human cognition, not merely compensate for cognitive limitations.
Why does it matter? A prosthesis doesn’t pretend to be a person, rather it extends, restores or supports a particular human capability. The corollary is that the metaphor encourages us to focus less on whether the machine is intelligent and more on whether the human-machine combination performs better than the human alone. This surely, is the right way to consider the impact of AI.
Where does it fall short? It doesn’t, it accurately outlines what the everyday experience of using AI well is like. The human retains responsibility for purpose, judgement and consequences, with the machine providing support and leverage.
The shoggoth
These enormous, shapeless, mercifully fictional creatures from the shared universe of Cthulhu Mythos were created in 1936 by horror writer Howard Phillips Lovecraft. They are powerful artificial servants that eventually become difficult for their creators to control. Their modern AI manifestation was conceived by pseudonymous X user @TetraspaceWest, who posted an illustration in 2022 depicting GPT-3 as a monstrous shoggoth. The Twitterati quickly connected with it, recognised it as a meme, and spread it.
Tetraspace’s shoggoth boasts a smile, representing the polite, helpful interface presented to users. Behind the visage lies a vast, poorly understood mathematical system trained on human language.
Why does it matter? The metaphor is a reminder that we don’t fully understand what happens inside LLMs and diffusion models when we access them via their attractive and simple interfaces. The friendly front-end make helpful suggestions, apologises for mistakes and appears to exhibit empathy. Yet behind the surface, inside the beast, is an enormous set of numerical relationships that even its creators can’t interpret in a simple, complete way.
Where does it fall short? While the metaphor is probably overly dramatic, it is nonetheless a useful warning against anthropomorphising AI. The machine isn’t human, but neither is it a monster.
In conclusion
AI is simultaneously a prediction engine, a creative assistant, a reasoning partner, a statistical machine, a productivity tool and a system whose inner workings remain partly mysterious. And of course, much more besides.
