Try to describe AI in one sentence.
How long did you stop to think?
Most people I ask that question start to respond, but pause when they realize their initial explanation feels insufficient.
Packaging new technology into an accessible definition has always been challenging, but what what strikes me about AI is that it defies definition by both casual consumers, like my 70-year-old dad using the free tier of ChatGPT, and AI natives like me, who work for companies that invent AI tools.
AI has transcended every metaphor we’ve applied, which is a signal that it could be a truly new technology that reshapes the world around it, not just a stepwise advance in the way we execute existing work.
5 years of AI metaphors
The progression of metaphors over the last 5 years is instructive. Notice how metaphors change as the models advance.
2021: the stochastic parrot
GPT-1 was released in 2018 and ChatGPT was released in 2022. In 2021, Emily Bender described LLMs as stochastic parrots: machines that could regurgitate language, but who didn’t understand it1.
2023: the blurry JPEG, the intern, the calculator for words, the autonomous agent
By 2023, the models had advanced enough to become far more interesting, impressive and useful. This year in particular is interesting because of the proliferation of metaphors and how much they differ from each other.
The blurry JPEG
In February 2023, Ted Chiang called ChatGPT a “blurry JPEG of the web” in a popular New Yorker article2:
[ChatGPT] retains much of the information on the Web, in the same way that a JPEG retains much of the information of a higher-resolution image, but, if you're looking for an exact sequence of bits, you won't find it; all you will ever get is an approximation.
The intern
In May 2023, Ethan Mollick, a professor at the Wharton School of Management, argued in favor of humans as a metaphor for AI3:
...several billion people just got free interns. They are weird, somewhat alien interns that work infinitely fast and sometimes lie to make you happy, but interns nonetheless.
The calculator for words
In April 2023, Simon Willison, a software engineer who writes prolifically about AI, called LLMs calculators for words (citing Ted Chiang’s essay)4:
I like to think of language models like ChatGPT as a calculator for words. This is reflected in their name: a “language model” implies that they are tools for working with language.
The autonomous agent
2023 was also the year terminology around autonomous agents began to enter the mainstream lexicon as both a descriptive metaphor and a practical implementation of AI.
AutoGPT5 is an early example of running AI on loop, which is the basic architecture of an agent.
2025: the digital employee
In 2024, individuals and business started to use agents to perform tasks. As the underlying models improved, the metaphor circled back at humans, but at much higher stakes than an intern.
In a January 2025 post, Sam Altman predicted that AI agents would “join the workforce and materially change the output of companies” that same year6.
2026 (to date)
In 2026 the term agent refers to actual products and specific tools that people can build and run, or that are operated on their behalf.
The digital employee narrative is still dominant as agent adoption grows, but there is also fragmentation as we continue to grapple with how capable AI has become.
The two most interesting themes I see are on opposite ends of the spectrum.
First, there is pushback against personification. A good example is the exoskeleton metaphor7, which acknowledges the immense utility of AI, but through the lens of human augmentation (not replacement).
Second, AI is being spiritualized. AGI has been part of the AI conversation since the beginning (and conceptualized for a very long time), but I’ve heard more than one person describe some modern, AI-focused work environments as increasingly cultish. A recent post from roon8 on X captured this metaphor in undiluted form:
it is a literal and useful description of anthropic that it is an organization that loves and worships claude, is run in significant part by claude, and studies and builds claude.
AI transcends the metaphor
New technologies have impact across a set of matricies:
- Breadth vs depth
- Rate of improvement vs rate of adoption
- Novelty vs historical precedent
Most innovations index high on one or two of the plots, making them easier to describe; AI is in the top right of all three, defying easy definition. When technology transcends metaphor and even the best thinkers make wrong predictions, take note, because the impact won't be localized, it will reshape the world around it.
Join almost 3,000 subscribers via Substack to get notified about new posts.
Footnotes
-
Emily Bender and others wrote a paper titled On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜. The stochastic parrot article on Wikipedia provides a great summary of the paper as well as the broader debate it sparked. ↩
-
You can read Ted Chiang’s article on ChatGPT as a blurry JPEG on the New Yorker’s website (it’s gated to subscribers). ↩
-
You can read Ethan Mollick’s article, Onboarding your AI intern, on Substack. ↩
-
Simon Willison stays on the bleeding edge of AI and has coined many popular terms, including “agentic engineering” and “prompt injection.” You can read his 2023 article on his blog. ↩
-
AutoGPT was one of the first popular implementations of an AI system that would attempt to accomplish a goal across multiple attempts and iterations, without human intervention. ↩
-
You can read Reflections by Sam Altman on his blog. It was posted in early 2025 and is fascinating to read 18 months later. ↩
-
I found the exoskeleton metaphor on Hacker News, where a post from a company called Kasava hit the first page. The article includes rich examples from the world of robotics, but also reads as a bit of a self-serving thesis that justifies their product. ↩
-
roon’s anonymity is officially unknown, but it is widely claimed that he works for OpenAI. On X, he’s a leading voice on AI topics, providing a cocktail of razor sharp, philosophical, meme-infused esoteric perspective, and his recent post about the spiritualization of Claude is exhibit A. ↩