August 30, 2024, by Brigitte Nerlich
AI, LLMs and an explosion of metaphors
I would love to write a blog post entitled “What is it like to be an LLM”, but I have to leave that to others who are really into large language models or to the LLMs themselves. Instead, I have to resign myself to thinking about the question: “What is an LLM like?” This means collecting and analysing the many metaphors people have begun to create, use and circulate to think about, talk about and interact with LLMs and other AI systems.
This explosion of metaphors is indicative of a lively discourse about LLMs since they became a matter of public consumption at the end of 2022 with the advent of ChatGPT.
Collecting metaphors for AI systems
I have started to collect some metaphors for LLMs and AI system myself, but when I began to play with ChatGPT (OpenAI), I also asked the chatbot to collect some metaphors for me and I have now done the same with Claude (Anthropic). In both cases I asked the chatbots about the metaphors they would use for themselves and also for other chatbots. This is not really collecting metaphors at the coalface though. It’s more like chasing metaphors projected onto the walls of Plato’s caves. But it’s a start.
Other people have collected and thought about metaphors for LLMs in more systematic ways and, at some point one would have to merge all these collections and analyse them (see Furze, Trott, Figar, Lukes etc.).
People might ask why I don’t just do my usual media thing. Go to Nexis, bung in a few keywords, collect a sample, read it and analyse it, as I have done so many times before for Covid, climate change, epigenetics and much more. I tried, but for various reasons it just doesn’t work. Most importantly, there is too much stuff out there – at least for me and my deeply hands-on qualitative approach.
So, let’s get back to how put chatbots to work. I combined all the metaphors for AI systems previously proposed to me by ChatGPT and Claude into one list, which became quite long! I won’t inflict that list on you. If you are interested, I can provide it. I then asked Claude to sort the list into categories, which it did very well. I then asked it to sort the metaphors into those highlighting benefits and those highlighting risks. It did that too, even adding a few instances of metaphors that can flip both ways.
I will inflict these shorter lists of categories on you, as they are quite interesting and you might want to add to them or to criticise them or use them in some other way. Let’s start with a list of categories for various perspectives and roles attributed to AI systems.
Categories of metaphors for AI systems
Claude proposed the following list of metaphor types based on the list of metaphors I had fed it (which itself was based on metaphors generated by ChatGPT and Claude:
- Tool-based Metaphors:
- Digital Assistant/Virtual Assistant [This one reduces assistants to tools – perhaps not so great]
- Toolbox/Swiss Army Knife
- Force Multiplier
- Digital Compass
- Knowledge and Information Metaphors:
- Knowledge Engine/Information Hub
- Infinite Encyclopedia
- Knowledge Ocean
- Boundless Library
- Cognitive and Mental Metaphors:
- Cognitive Extension/Cognitive Prosthesis
- Language Model as a Brain
- Sentiment Analyst
- Collaborative and Social Metaphors:
- Digital Colleague/Collaborator
- Creative Collaborator
- Conversational Companion
- Robotic Companion
- [seems to miss the now ubiquitous ‘Co-pilot’ and the above ‘assistant’]
- Predictive and Insight Metaphors:
- Crystal Ball
- Oracle
- Time Traveler
- Reflective Metaphors:
- Mirror
- Cultural Mirror
- Adaptive and Versatile Metaphors:
- Digital Chameleon
- Linguistic Acrobat
- Creative and Generative Metaphors:
- Virtual Artist
- Virtual Wordsmith
- Storytelling Djinn
- Specialized Role Metaphors:
- Data Doctor
- Simulated Teacher
- Algorithmic Chef
- Financial Advisor
- Environmental Steward
- [here again one could add Co-pilot]
- Language and Communication Metaphors:
- Language Wizard
- Multilingual Polyglot
- Parrot or Mimic
- Transformative and Impactful Metaphors:
- Game Changer
- Disruptive Force
- Nuclear Explosion/Bomb [transformative!]
- Cautionary Metaphors:
- Black Box
- Double-Edged Sword
- Pandora’s Box
- Analytical and Processing Metaphors:
- Knowledge Distillery
- Cosmic Observatory
- Augmentation Metaphors:
- Augmented Reality
- Protective Metaphors:
- Privacy Guardian
As I said in a previous post, most of the metaphors that ChatGPT and Claude found in their databases had a rather magical aura. So, what did Claude do when I asked it to sort the metaphors into those highlighting risks and those highlighting benefits? It did that and, whereas before it ‘just’ provided category names, this times it lingered more on what the metaphors do in discourse.
Metaphor highlighting risks and benefits
Metaphors Highlighting Risks:
- Black Box: Emphasizes the lack of transparency and interpretability in AI systems.
- Double-Edged Sword: Suggests AI can have both positive and negative impacts.
- Pandora’s Box: Implies AI could unleash unforeseen consequences.
- Nuclear Bomb/Explosion: Highlights potential for rapid, destructive impact.
- Disruptive Force: Emphasizes potential for upheaval in various sectors.
- Mirror: When used to reflect societal biases, this highlights risks of perpetuating or amplifying existing problems.
- Parrot or Mimic: Suggests potential for mindless repetition without true understanding.
Metaphors Highlighting Benefits:
- Digital Assistant/Virtual Assistant: Emphasizes AI’s helpful, supportive role.
- Knowledge Engine/Information Hub: Highlights AI’s vast information processing capabilities.
- Cognitive Extension/Prosthesis: Suggests AI can enhance human cognitive abilities.
- Digital Colleague/Collaborator: Emphasizes AI’s potential to work alongside humans productively.
- Crystal Ball/Oracle: Highlights predictive capabilities.
- Swiss Army Knife/Toolbox: Emphasizes versatility and usefulness for various tasks.
- Force Multiplier: Suggests AI can significantly enhance human productivity and capabilities.
- Game Changer: Emphasizes potential for positive transformation.
- Navigator: Highlights AI’s ability to guide through complex information or decisions.
- Time Traveler: Emphasizes predictive analytics capabilities.
- Augmented Reality: Suggests AI can enhance our perception and understanding of the world.
- Virtual Artist/Wordsmith: Highlights creative potential.
- Data Doctor/Financial Advisor: Emphasizes specialized problem-solving capabilities.
- Language Wizard/Multilingual Polyglot: Highlights advanced language processing abilities.
- Environmental Steward: Emphasizes potential for positive environmental impact.
Metaphors highlighting both risks and benefits, depending on context:
- Digital Ecosystem: Could emphasize interconnectedness (benefit) or vulnerability (risk). [here one could mention the ‘pollution’ metaphor and the ‘collapse’ metaphor
- Cultural Mirror: Could highlight increased self-awareness (benefit) or reinforcement of biases (risk).
- Digital Chameleon: Could emphasize adaptability (benefit) or potential for deception (risk).
After accomplishing this categorisation task, Claude pointed out the dual use of metaphor: “This categorization helps to illustrate the dual nature of AI technologies, emphasizing both their potential advantages and the concerns surrounding their development and use.”
Conventional and creative metaphors
Many of the metaphors unearthed by the chatbots have already become conventionalised, such as the metaphor of the digital assistant, the library, the mirror and of course the (stochastic) parrot.
The chatbots did not find the metaphors of ‘knowledge pollution’ or ‘knowledge collapse’ though, which was strange – don’t they read my blogs? They also don’t seem to have found the metaphor of ‘bullshit engine’ – yep, who wants to be called that….
There were some metaphors, however, that looked a bit fresher, such as the ‘linguistic acrobat’, the ‘knowledge observatory’ or the ‘knowledge distillery’.
I wondered how I could get at a few fresh and creative metaphors in a quick and easy way myself. In the end, I put “LLM AND metaphor” into Twitter/X and also Bluesky and collected what came up. I stopped when the search sort of petered out, which was after about ten to 20 hits. I’ll cite a few interesting finds, not all.
Critiquing and commenting on conventional metaphors
One tweeter reflected on the mirror metaphor and said: “Looking into a LLM is like looking in a distorting mirror: you are looking at a distorted version of yourself and the longer you look the weirder it gets.” Here the negative aspects of mirrors are highlighted, while somebody else reconfigured the mirror metaphor in a positive light and said: “a domain-specific app that uses an LLM is like an array of mirrors that concentrates light”.
Another person criticised the parrot metaphor and said that it “overemphasizes the statistical part, but underemphasizes the ‘deep’ hierarchical and structured part of these models. Both are relevant”.
This shows that conventional metaphors have not yet become dead metaphors. They still trigger thought and reflection. But what about novel and creative metaphors?
Creating novel metaphors
Somebody called LLMs a mega vac, sucking up all information and another called them Ouija boards. These are perhaps not overly creative metaphors, but fairly novel ones.
There were quite a few more creative metaphors drawing inspiration from food as the metaphor’s source domain.
On Twitter/X Anne Beaulieu, a former colleague, used ab extended food-based metaphor and said: “An LLM is like a massive fridge full of items. Just because everything is available doesn’t mean you should use it. Some items are rotten (biases) while others are too bland or overly spicy. A real chef understands these nuances and their impact. Read more.https://tinyurl.com/guidingAI” – in the document you’ll find a longer extension of that metaphor.
Another metaphor cited in a tweet is also inspired by food but also features a bear: “LLMs are like a trained circus bear that can make you porridge in your kitchen” – put it can also ransack your kitchen.
I loved a creative food-related metaphor cited on Bluesky which manages to combine Plato, Kant and Diet Coke: “LLM articles are like diet coke, they are the appearance of the thing but not the thing itself.”
I also found an extended metaphor based on the comparison of LLMs with eatable or poisonous mushroom soup, or a reference to chalk that you put in milk and much more….
There are probably many many more metaphors out there based on people’s everyday experiences, hobbies, philosophies and scientific background knowledge. Somebody said, for example, that “prompts are like biochemistry, operating on the substrate of the LLM architecture”…. Such metaphors won’t travel as far as the food-based ones, I would guess.
A metaphor observatory
But why should one make chatbots collect metaphors for chatbots and why should one go out collecting metaphors for chatbots in the wild? What is the point? Ok, one point is: it’s fun. Like collecting seashells on the seashore and saying: “Oh, look at this one. Isn’t it pretty?”
Another point is a bit deeper and possibly totally misguided. By collecting and analysing AI metaphors from a variety of sources and at different points in time, researchers may gain deeper insights into the individual and collective linguistic activity and creativity that structures the fast-changing landscape of artificial intelligence and LLM discourse.
One could perhaps pinpoint hotspots of creativity and critique, of comment and contention. But to do this properly, to establish an AI metaphor observatory if you like, a more systematic approach is needed. I haven’t found one yet.
Image: Robert Delaunay, Rythmes sans fin – Joie de vivre, 1930, Wikimedia Commons
No comments yet, fill out a comment to be the first
Leave a Reply