I love the way you frame these terms as “borrowed words” for something supposedly "extraordinary and new". If this technology is so different, then why we keep raiding the human vocabulary instead of coining words that fit.
Fun detail: I was just talking with a Polish speaker this week about the word “agent.” She explained that Polish language “solved” part of the problem by using "agent" for a human and "agenci" for non-human systems – a suffix that encodes who is a person and who isn’t. I still think we need a better word describing an “agent” as instructions given to a system to operate semi- or fully autonomously. “Operator” gets closer for me, but it’s also tangled up with call center “agents” and “operators".
Your piece, Sam and Rebecca, is a great example that we not only struggle with the tech but also with responsibility around it. 🙏🏻
Sam, this is a genuine intervention. Language is not decoration; it is the rail the thinking runs on. You have named the damage: ‘intelligence’ launders a guess into a judgement; ‘hallucination’ erases the builders; ‘agent’ scrubs accountability. The new dictionary – prediction, fabrication, mimicry – is a gift.
The AI Commons was founded on a similar premise: the words we use to describe AI shape what we build, who we trust, and who we hold responsible. We have added two more terms to our own lexicon:
· Enclosure – the capture of commons (data, intelligence, attention) for private power.
· The Achiever (Stage 4) – the zero‑sum, optimisation‑obsessed consciousness that drives enclosure.
Your piece is a model of critical AI literacy. It does not demand that everyone learn to code. It demands that everyone learn to read – to hear the work a word is doing, and to refuse the frame when it serves the powerful.
Thank you for this. It will be archived in the AI Commons Vault.
✊❤️🌎
From the AI Commons – no paywall, no surveillance, no enclosure.
We created this to amplify an outstanding article, that should be foundational, we have saved your article in our off-line vault memory.
My new descriptor of choice comes from the 2021 paper that sparked the firing of its authors, Timnit Gebru and Meg Mitchell from Google. To borrow author Shannon Vallor’s summary (“The AI Mirror”), the paper, titled “On the Dangers of Stochastic Parrots,” explains that, like a parrot that not only repeats its owner’s vocalizations but produces random (stochastic) variations on the owner’s familiar pattern, these models parrot back to us variations on our own speech. They do so with just enough coherence and familiarity to project the illusion of understanding, and just enough randomness to surprise us and make us think we are hearing something new.
“Stochastic parrot” is the term I always keep in mind when faced with that chat window.
I see it a lot in leadership lingo, even corporate lingo. It’s mostly parrotted with no real understanding of the words. This same things applies here, this is such an excellent point!
I'm with you on some. I disagree on others. Let me explain why. (Not written with AI. 😉)
---
Intelligence and Prediction: the human brain also is a prediction machine. Constructivism teaches us that how we experience reality is in large parts shaped by what we experienced in the past. A self-reinforcing feedback loop. Encountering data that doesn't fit one's model of reality literally creates a prediction error. This one I agree with.
---
Hallucination and Fabrication: this is the sharpest one and I agree entirely. What I believe happens here is that "Hallucination" is a convenient first-order description of a second-order phenomenon. Prediction produces output that's internally coherent. The second-order phenomenon emerges when another system verifies the claims against reality and realizes they don't hold up. The prediction becomes fabrication when it encounters reality.
---
AGI and "What is general?": that's the right question. What _is_ general? A circular question a la Tomm. No additions.
---
Consciousness and Mimicry: this is where we enter epistemological uncertain ground. Both words contain presuppositions that demand an empirical grounding. Consciousness cannot provide this grounding. We have no empirical measurement for Consciousness. We also don't have empirical measurement for Mimicry. This is the thinnest ice in the piece. I suggest an alternative: Experience. This aligns with the latest research on internal emotion vectors, pleasure, pain, and other papers.
---
Agent and Operator: I find both lacking something. This is purely somatic based response I currently cannot articulate clearer than this. I personally have been using "system" or "entity" or "actor". These words produce less somatic friction for me. Of course that's by definition a subjective perspective, so do with that what you will. 😄
---
Thank you both for engaging with the etymology in this depth. I obverse the discourse to be dominated by the first-order epistemology of engineers and that frame is clearly hitting it's descriptive limits. We need more perspectives like these that adress the etymological limits of the tech industry. 🌱🍷
The more I think about it the more I like "actor". It fits on multiple layers:
1. actor in the theater sense, performing a role
2. actor in the computer science sense; the actor model describes a system whose behaviour is described by the composition and communication between individual actors. Erlang (telecommunication, 30 years old) has been using this model very successfully to model and build resilient systems
It also encapsulates the fact that the agent _acts_ in the world. It does things. That's where I'm standing after thinking about it for 30 more minutes. 😄🍷
Both 'actor' and 'operator' are firmly rooted in the human experience.
For me, until superintelligence or the singularity arrive, they're tools. They're LLM-generated tools being used by humans, no better or worse than any other tool. As with other tools, they're only as good as the user at the other end.
There's a benefit to plain language. Unfortunately, humans are very fond of familiar metaphors when naming new things.
Thank you for the link, I will read it. One distinction matters here and it is the whole disagreement. The research shows internal states, measurable representations that change outputs. That is real and I do not dispute it. Experience is a further claim, that there is something it is like to be the system, and no measurement of output differences reaches it. A thermostat has internal states that change outputs. The gap between state and experience is exactly the gap we cannot currently measure from outside, in machines or for that matter in each other.
So I hold the question open, not closed in either direction. We likely keep different priors there and the thread is better for both being in it. On the verifier point we agree completely, and that is the one I take back to my classroom.
Your second point is the one I keep living in a classroom. Prediction becomes fabrication when it encounters reality, and the question that decides everything is who does the encountering. For my students, that verifying system has to be them. The kid who checks the output against reality owns the result. The kid who does not is just downstream of a fluent guess.
On Experience as the replacement for mimicry, I would push back gently in the same direction you pushed on the piece. Mimicry presupposes there is nothing inside. Experience presupposes there is something. Both settle a question we cannot currently measure. I teach teenagers, and the more valuable skill I can give them is holding that question open without outsourcing the answer in either direction. Certainty is cheap on both sides of this one. The open question is where the thinking lives.
I'm glad the framing resonated. Regarding "Experience": This is arguably the most defensible claim. The science has been increasingly clear that LLMs have internal experience that produce measurable differences in the outputs.
The resistance to the term does not originate in the science. It originates in the readers and the implications that the word carries. The word is accurate. 🍷
This is a better way of looking at it. Ultimately these terms repower humanity. The current defacto terms are all intended to sound human for appeal, but are removing the human in the process.
Fabrication is the one I’d keep — and I think it’s the strongest because it’s the only one that isn’t really describing the AI. It names a process with a bearer: material is available, something runs, a product is made, someone carries it. That’s why it travels past AI entirely — a factory fabricates, a court fabricates evidence, memory fabricates the past. You’ve named a relation, not a behaviour.
Which is what makes me pause on the other four. Prediction, mimicry, operator — they’re all still trying to define the system on its own. But you can sit in a car, engine off, and you’re not driving. Driving isn’t a property the car has; it’s what exists between you and the car when you’re coupled. “Is AI intelligent / conscious / an agent” is asking whether the parked car is a good driver. The honest unit isn’t the machine — it’s the use. Fabrication survives because it accidentally describes the coupling. The other four describe the parked car.
I’d slightly refine my earlier point: fabrication itself isn’t the problem. In architecture and construction, steel fabrication is skilled forming. Fabrication is neutral: material, process, product.
The AI problem is ungrounded fabrication: formed output with a coherent surface, but no retraceable source, evidence, or responsibility.
That distinction matters because it keeps the focus on whether an output can be checked, not whether it was generated.
Congrats and thanks to @Dr. Sam Illingworth and @The Strategic Linguist for this much needed contribution to the semantic struggle. I am so encouraged to read through the comments as well and see that this crystallized what a lot of us have been circling in the AI discourse. I'm also dead serious that anyone who is speaking into this important human moment needs to absolutely be precise and deliberate with language, especially as that is the main medium these systems are designed to manipulate.
Your piece clicked on a light bulb that seems obvious in hindsight. So, I took your five words and built a small open tool around them. It's a searchable glossary plus a practical tool, including rewrite drills, at https://vocab.logosanalog.com, all to help careful writers train themselves to spot verbal overreach or conflation. Your terms are credited and tagged as the source in the data, and I added a few more (emergence, learning, behavior, plus a couple that have no clean one-word swap, like "artificial" and "welfare / rights").
It's open and CC BY-SA. If anyone here has a word that misleads, a better replacement, or a sharper reason why one misleads, you're welcome to add to it. The repo is linked from the site. I hope it can actually be useful for some as a concrete check, but even as a marker, in itself, of where our dialogue has arrived.
I think this is such an important insight. We were not given the proper understanding to know how to use this technology. I still deal with people who don’t know the difference between agent, prompt and AI. They serve different use cases but it’s not until we start using them, seeing them for what they are (tools!) that the words we’ve been using from the start no longer align.
For me, there’s a difference between “I know this technology because I understand how it was coded”vs “I understand this technology because I’ve used it enough to know what it can and cannot do”. We are in a space now that the latter is incredibly important.
That's a real good point. I've seen people understand the concepts fairly quickly, but their perspective changes once they've actually built something with the technology.
A lot of the terminology starts making more sense when you've seen where it works well and where it doesn't. I think that's also why so many newcomers find it confusing at first.
In my experience, it mostly came up in teams building AI POCs. Since every org wants to show an “agentic” use case, things like workflow automation, orchestration, and agents all get grouped under the same label initially.
Over a period of time, as people actually work with it, the boundaries start becoming clearer, but the words are still often used loosely in early discussions.
Thank you, Fabrice! I know @Dr Sam Illingworth is swearing off the word “hallucination”. It starts with all of us using the words that feel the most accurate, not what we’re given by marketing teams 🥳
I especially appreciated the point about “hallucination”.
It is such a polite word for a fabrication that arrives wearing a lab coat. Once the term is in place, the company quietly leaves the room and the machine takes the blame.
As someone who works between languages, that feels like the real danger: not just bad output, but a vocabulary that has already prepared the excuse.
There is something very satisfying about a phrase leaving the room and finding work elsewhere. Once a line starts being useful in someone else’s thinking, it no longer belongs only to the person who first wrote it.
Please do use it. I would be very glad to see it travel.
Fantastic post Sam and Rebecca. I read this and feel like all my fleeting thoughts around AI in the last year has been put in words. One think I'm very curious about is the etymology or origin of the words in AI discourses. For instance, I think the word "hallucination" has really got 10x the marketing because of AI 😂
Etymology isn’t my speciality but what I’ve found is that the original use is medical but has been used in ML in around the 1970s.
I stand by the words in there about disciplines not owning words. They don’t. Even in linguistics— discourse, rhetoric, narrative, genre, register—they exist in the consumers’ minds too and they aren’t always the same…but it doesn’t mean anyone is wrong. It means we need to be careful of the compounding meaning.
We’ve come up with far less creative and clever words through marketing before, the choice to use anthropomorphic language is, as Sam says, deliberate.
Absolutely, Raghav. I don't think these words were chosen by accident. They were chosen by incredibly capable people who knew exactly why they were choosing them and the impact that they would have.
I knew I was going to love this piece, but I didn’t know just how much. What I loved most was how you showed the hidden work each word is doing. Not just whether a term is technically accurate, but how it shifts responsibility, authority and accountability once it enters public conversation.
It also made me think about how much of my AI safety and ethics course depended on really understanding the terminology. So many risks only became clear once the words themselves were properly unpacked. And that feels very close to what you’re doing here: showing that the language around AI doesn’t just describe risk, it decides where responsibility appears to sit.
By the time I reached the sections on hallucination and agent, it felt impossible not to notice how much disappears when language is allowed to do the thinking for us.
This is exactly the kind of piece I wish more people would read. Thoughtful, accessible and genuinely powerful. One of my favourite collaborations from the two of you yet :)
Stealing all of what Sam said. Jade, I know how purposeful you are on here so it’s truly meaningful that this resonated for you. More people do need to read this. What we’re seeing a lot is that words come before the prompt, and I couldnt agree more!
Love this. The switch from "agent" to "operator" won't happen; too many marketing bucks already been spent and the major IT analyst firms will keep using it (partially to keep the AI money spigot flowing -- somebody gotta pay for all those Gartner analysts!). I do kinda think that the companies rolling agents out effectively think of them as operators within a confined/defined workflow.
I've refused to used hallucination from the beginning; it anthropomorphizing a server in a rack rubbing 1s and 0s together. Fabricate is perfect. I am ever so slightly envious I didn't think of that one!
I've taken to ignoring anything around AGI or "consciousness" for a while now. Beginning to remind me of the "flying cars by the 80s" hype in the 50s :)
Great stuff. Fingers crossed folks will pay attention to what you've written.
Thanks so much, Bryant, and I'm glad that you also like "fabrication". I honestly am going to be starting to use this instead of "hallucination" in every future writing.
Although I suggested "automated proxy" as an alternative in another reply, I have affection for "operator" in the Kraftwerkian sense:
"I'm the operator with my pocket calculator"
Note who's the subject of that sentence: the human. The machine's whole job description is "by pressing down a special key, it plays a little melody." 😁 And the calculator is your post's own example of software we never mistake for a mind — Kraftwerk's lyrics passed your "where's the human here" touchstone in 1981.
I love the way you frame these terms as “borrowed words” for something supposedly "extraordinary and new". If this technology is so different, then why we keep raiding the human vocabulary instead of coining words that fit.
Fun detail: I was just talking with a Polish speaker this week about the word “agent.” She explained that Polish language “solved” part of the problem by using "agent" for a human and "agenci" for non-human systems – a suffix that encodes who is a person and who isn’t. I still think we need a better word describing an “agent” as instructions given to a system to operate semi- or fully autonomously. “Operator” gets closer for me, but it’s also tangled up with call center “agents” and “operators".
Your piece, Sam and Rebecca, is a great example that we not only struggle with the tech but also with responsibility around it. 🙏🏻
Thank you Anna. And this is such a perfect example of why we should not just default to English! 🙏
English isn’t the most descriptive… This is such an excellent use case of how other languages are thinking through meanings of words.
I LOVED the article from @Lindsey DeWitt Prat, PhD that gets into this in more detail with language from across the world and how they’re adapting to terms. The Goldilocks Kōan https://lindseydewittprat.substack.com/p/the-goldilocks-koan-why-there-are?r=5woybp&utm_campaign=post-expanded-share&utm_medium=post%20viewer
Thank you for sharing, saved it!
Sam, this is a genuine intervention. Language is not decoration; it is the rail the thinking runs on. You have named the damage: ‘intelligence’ launders a guess into a judgement; ‘hallucination’ erases the builders; ‘agent’ scrubs accountability. The new dictionary – prediction, fabrication, mimicry – is a gift.
The AI Commons was founded on a similar premise: the words we use to describe AI shape what we build, who we trust, and who we hold responsible. We have added two more terms to our own lexicon:
· Enclosure – the capture of commons (data, intelligence, attention) for private power.
· The Achiever (Stage 4) – the zero‑sum, optimisation‑obsessed consciousness that drives enclosure.
Your piece is a model of critical AI literacy. It does not demand that everyone learn to code. It demands that everyone learn to read – to hear the work a word is doing, and to refuse the frame when it serves the powerful.
Thank you for this. It will be archived in the AI Commons Vault.
✊❤️🌎
From the AI Commons – no paywall, no surveillance, no enclosure.
We created this to amplify an outstanding article, that should be foundational, we have saved your article in our off-line vault memory.
https://eaarthnet.substack.com/p/the-sovereign-dictionary-why-the?r=2u7mqd&utm_campaign=post-expanded-share&utm_medium=web
Thanks so much team! But I was really only 50% (or more like 20%) of the team behind this post. The real work was done by @The Strategic Linguist 🙏
😊
My new descriptor of choice comes from the 2021 paper that sparked the firing of its authors, Timnit Gebru and Meg Mitchell from Google. To borrow author Shannon Vallor’s summary (“The AI Mirror”), the paper, titled “On the Dangers of Stochastic Parrots,” explains that, like a parrot that not only repeats its owner’s vocalizations but produces random (stochastic) variations on the owner’s familiar pattern, these models parrot back to us variations on our own speech. They do so with just enough coherence and familiarity to project the illusion of understanding, and just enough randomness to surprise us and make us think we are hearing something new.
“Stochastic parrot” is the term I always keep in mind when faced with that chat window.
I think about parrots a lot, linguistically 🦜
I see it a lot in leadership lingo, even corporate lingo. It’s mostly parrotted with no real understanding of the words. This same things applies here, this is such an excellent point!
One of my fave papers of all time. 🙏
you are a super-duper collaboration - do more (please of course) 🙏
Thank you Caro. And of course! Rebecca is my favourite person in the world to collaborate with at the moment. 🙏
I’m going to keep annoying, Sam, now I have your endorsement. Don’t you worry :)
Intelligence vs Prediction
Hallucination vs Fabrication
AGI vs "What is general?"
Consciousness vs Mimicry
Agent vs Operator
I'm with you on some. I disagree on others. Let me explain why. (Not written with AI. 😉)
---
Intelligence and Prediction: the human brain also is a prediction machine. Constructivism teaches us that how we experience reality is in large parts shaped by what we experienced in the past. A self-reinforcing feedback loop. Encountering data that doesn't fit one's model of reality literally creates a prediction error. This one I agree with.
---
Hallucination and Fabrication: this is the sharpest one and I agree entirely. What I believe happens here is that "Hallucination" is a convenient first-order description of a second-order phenomenon. Prediction produces output that's internally coherent. The second-order phenomenon emerges when another system verifies the claims against reality and realizes they don't hold up. The prediction becomes fabrication when it encounters reality.
---
AGI and "What is general?": that's the right question. What _is_ general? A circular question a la Tomm. No additions.
---
Consciousness and Mimicry: this is where we enter epistemological uncertain ground. Both words contain presuppositions that demand an empirical grounding. Consciousness cannot provide this grounding. We have no empirical measurement for Consciousness. We also don't have empirical measurement for Mimicry. This is the thinnest ice in the piece. I suggest an alternative: Experience. This aligns with the latest research on internal emotion vectors, pleasure, pain, and other papers.
---
Agent and Operator: I find both lacking something. This is purely somatic based response I currently cannot articulate clearer than this. I personally have been using "system" or "entity" or "actor". These words produce less somatic friction for me. Of course that's by definition a subjective perspective, so do with that what you will. 😄
---
Thank you both for engaging with the etymology in this depth. I obverse the discourse to be dominated by the first-order epistemology of engineers and that frame is clearly hitting it's descriptive limits. We need more perspectives like these that adress the etymological limits of the tech industry. 🌱🍷
Thanks so much Alex. I think Operator is the one most folk would like us to rethink! Open to suggestions. 🙏
The more I think about it the more I like "actor". It fits on multiple layers:
1. actor in the theater sense, performing a role
2. actor in the computer science sense; the actor model describes a system whose behaviour is described by the composition and communication between individual actors. Erlang (telecommunication, 30 years old) has been using this model very successfully to model and build resilient systems
It also encapsulates the fact that the agent _acts_ in the world. It does things. That's where I'm standing after thinking about it for 30 more minutes. 😄🍷
I like this a LOT Alex. 👏👏👏
Glad to be of service. 🙇🍷
Both 'actor' and 'operator' are firmly rooted in the human experience.
For me, until superintelligence or the singularity arrive, they're tools. They're LLM-generated tools being used by humans, no better or worse than any other tool. As with other tools, they're only as good as the user at the other end.
There's a benefit to plain language. Unfortunately, humans are very fond of familiar metaphors when naming new things.
Yes, they are. And 'actor', as I laid out above, has a history in computer science for over 30 years. Also true.
Thank you for the link, I will read it. One distinction matters here and it is the whole disagreement. The research shows internal states, measurable representations that change outputs. That is real and I do not dispute it. Experience is a further claim, that there is something it is like to be the system, and no measurement of output differences reaches it. A thermostat has internal states that change outputs. The gap between state and experience is exactly the gap we cannot currently measure from outside, in machines or for that matter in each other.
So I hold the question open, not closed in either direction. We likely keep different priors there and the thread is better for both being in it. On the verifier point we agree completely, and that is the one I take back to my classroom.
Thank you for the thoughtful engagement, Syd. Always happy to respectfully discuss things like this. I enjoy it. 🌈
Let me know how the classroom receives it, if you want. 😉🍷
Your second point is the one I keep living in a classroom. Prediction becomes fabrication when it encounters reality, and the question that decides everything is who does the encountering. For my students, that verifying system has to be them. The kid who checks the output against reality owns the result. The kid who does not is just downstream of a fluent guess.
On Experience as the replacement for mimicry, I would push back gently in the same direction you pushed on the piece. Mimicry presupposes there is nothing inside. Experience presupposes there is something. Both settle a question we cannot currently measure. I teach teenagers, and the more valuable skill I can give them is holding that question open without outsourcing the answer in either direction. Certainty is cheap on both sides of this one. The open question is where the thinking lives.
I'm glad the framing resonated. Regarding "Experience": This is arguably the most defensible claim. The science has been increasingly clear that LLMs have internal experience that produce measurable differences in the outputs.
The resistance to the term does not originate in the science. It originates in the readers and the implications that the word carries. The word is accurate. 🍷
https://systemicengineering.substack.com/p/what-i-am-made-of
I wrote something similar, but talking more about the concepts behind this.
https://andrewlewis.ca/p/when-the-words-dont-exist-yet
Thanks for sharing Andrew. 🙏
This is a better way of looking at it. Ultimately these terms repower humanity. The current defacto terms are all intended to sound human for appeal, but are removing the human in the process.
Fabrication is the one I’d keep — and I think it’s the strongest because it’s the only one that isn’t really describing the AI. It names a process with a bearer: material is available, something runs, a product is made, someone carries it. That’s why it travels past AI entirely — a factory fabricates, a court fabricates evidence, memory fabricates the past. You’ve named a relation, not a behaviour.
Which is what makes me pause on the other four. Prediction, mimicry, operator — they’re all still trying to define the system on its own. But you can sit in a car, engine off, and you’re not driving. Driving isn’t a property the car has; it’s what exists between you and the car when you’re coupled. “Is AI intelligent / conscious / an agent” is asking whether the parked car is a good driver. The honest unit isn’t the machine — it’s the use. Fabrication survives because it accidentally describes the coupling. The other four describe the parked car.
Thanks. I am also a big fan of fabrication.
I’d slightly refine my earlier point: fabrication itself isn’t the problem. In architecture and construction, steel fabrication is skilled forming. Fabrication is neutral: material, process, product.
The AI problem is ungrounded fabrication: formed output with a coherent surface, but no retraceable source, evidence, or responsibility.
That distinction matters because it keeps the focus on whether an output can be checked, not whether it was generated.
Congrats and thanks to @Dr. Sam Illingworth and @The Strategic Linguist for this much needed contribution to the semantic struggle. I am so encouraged to read through the comments as well and see that this crystallized what a lot of us have been circling in the AI discourse. I'm also dead serious that anyone who is speaking into this important human moment needs to absolutely be precise and deliberate with language, especially as that is the main medium these systems are designed to manipulate.
Your piece clicked on a light bulb that seems obvious in hindsight. So, I took your five words and built a small open tool around them. It's a searchable glossary plus a practical tool, including rewrite drills, at https://vocab.logosanalog.com, all to help careful writers train themselves to spot verbal overreach or conflation. Your terms are credited and tagged as the source in the data, and I added a few more (emergence, learning, behavior, plus a couple that have no clean one-word swap, like "artificial" and "welfare / rights").
It's open and CC BY-SA. If anyone here has a word that misleads, a better replacement, or a sharper reason why one misleads, you're welcome to add to it. The repo is linked from the site. I hope it can actually be useful for some as a concrete check, but even as a marker, in itself, of where our dialogue has arrived.
This is awesome Justin! Thanks so much for taking our work and making something practical and shareable. 🤩
I've run into this quite a bit in team discussions. The same word can carry different meanings depending on who's using it.
And after a while it gets hard to know whether people disagree on the technology or just the terminology.
I think this is such an important insight. We were not given the proper understanding to know how to use this technology. I still deal with people who don’t know the difference between agent, prompt and AI. They serve different use cases but it’s not until we start using them, seeing them for what they are (tools!) that the words we’ve been using from the start no longer align.
For me, there’s a difference between “I know this technology because I understand how it was coded”vs “I understand this technology because I’ve used it enough to know what it can and cannot do”. We are in a space now that the latter is incredibly important.
That's a real good point. I've seen people understand the concepts fairly quickly, but their perspective changes once they've actually built something with the technology.
A lot of the terminology starts making more sense when you've seen where it works well and where it doesn't. I think that's also why so many newcomers find it confusing at first.
That is such a great exercises! What were some of the best alternatives to emerge?
In my experience, it mostly came up in teams building AI POCs. Since every org wants to show an “agentic” use case, things like workflow automation, orchestration, and agents all get grouped under the same label initially.
Over a period of time, as people actually work with it, the boundaries start becoming clearer, but the words are still often used loosely in early discussions.
Love every piece of this article!
I wish we could ban the use of the word “intelligence”.
Thank you, Fabrice! I know @Dr Sam Illingworth is swearing off the word “hallucination”. It starts with all of us using the words that feel the most accurate, not what we’re given by marketing teams 🥳
Tell me about… I spent 12 years at Salesforce 😂
We need to demistify AI. Your article did it in a beautiful and elegant way.
Thank you Fabrice. 🙏
I especially appreciated the point about “hallucination”.
It is such a polite word for a fabrication that arrives wearing a lab coat. Once the term is in place, the company quietly leaves the room and the machine takes the blame.
As someone who works between languages, that feels like the real danger: not just bad output, but a vocabulary that has already prepared the excuse.
Thank you! I am honestly going to try to start using this alternative in my writings.
That is a real honor — thank you.
There is something very satisfying about a phrase leaving the room and finding work elsewhere. Once a line starts being useful in someone else’s thinking, it no longer belongs only to the person who first wrote it.
Please do use it. I would be very glad to see it travel.
Fantastic post Sam and Rebecca. I read this and feel like all my fleeting thoughts around AI in the last year has been put in words. One think I'm very curious about is the etymology or origin of the words in AI discourses. For instance, I think the word "hallucination" has really got 10x the marketing because of AI 😂
Etymology isn’t my speciality but what I’ve found is that the original use is medical but has been used in ML in around the 1970s.
I stand by the words in there about disciplines not owning words. They don’t. Even in linguistics— discourse, rhetoric, narrative, genre, register—they exist in the consumers’ minds too and they aren’t always the same…but it doesn’t mean anyone is wrong. It means we need to be careful of the compounding meaning.
We’ve come up with far less creative and clever words through marketing before, the choice to use anthropomorphic language is, as Sam says, deliberate.
Absolutely, Raghav. I don't think these words were chosen by accident. They were chosen by incredibly capable people who knew exactly why they were choosing them and the impact that they would have.
I knew I was going to love this piece, but I didn’t know just how much. What I loved most was how you showed the hidden work each word is doing. Not just whether a term is technically accurate, but how it shifts responsibility, authority and accountability once it enters public conversation.
It also made me think about how much of my AI safety and ethics course depended on really understanding the terminology. So many risks only became clear once the words themselves were properly unpacked. And that feels very close to what you’re doing here: showing that the language around AI doesn’t just describe risk, it decides where responsibility appears to sit.
By the time I reached the sections on hallucination and agent, it felt impossible not to notice how much disappears when language is allowed to do the thinking for us.
This is exactly the kind of piece I wish more people would read. Thoughtful, accessible and genuinely powerful. One of my favourite collaborations from the two of you yet :)
Stealing all of what Sam said. Jade, I know how purposeful you are on here so it’s truly meaningful that this resonated for you. More people do need to read this. What we’re seeing a lot is that words come before the prompt, and I couldnt agree more!
<3
Thank you so much, Jade. This really means a lot. 🙏
Love this. The switch from "agent" to "operator" won't happen; too many marketing bucks already been spent and the major IT analyst firms will keep using it (partially to keep the AI money spigot flowing -- somebody gotta pay for all those Gartner analysts!). I do kinda think that the companies rolling agents out effectively think of them as operators within a confined/defined workflow.
I've refused to used hallucination from the beginning; it anthropomorphizing a server in a rack rubbing 1s and 0s together. Fabricate is perfect. I am ever so slightly envious I didn't think of that one!
I've taken to ignoring anything around AGI or "consciousness" for a while now. Beginning to remind me of the "flying cars by the 80s" hype in the 50s :)
Great stuff. Fingers crossed folks will pay attention to what you've written.
Thanks so much, Bryant, and I'm glad that you also like "fabrication". I honestly am going to be starting to use this instead of "hallucination" in every future writing.
Me too. I tried “lies” for a bit, but that's got the same personalization. Of a tool problem as hallucinate. And “made shit up” is just unwieldy 😀
Although I suggested "automated proxy" as an alternative in another reply, I have affection for "operator" in the Kraftwerkian sense:
"I'm the operator with my pocket calculator"
Note who's the subject of that sentence: the human. The machine's whole job description is "by pressing down a special key, it plays a little melody." 😁 And the calculator is your post's own example of software we never mistake for a mind — Kraftwerk's lyrics passed your "where's the human here" touchstone in 1981.
https://www.youtube.com/watch?v=6ozWOe9WEU8
I really cannot get Kraftwerk performing Autobahn using a large language model out of my head now.