- Published on
LLMs: the destructive and constructive powers of metaphors
- Authors
- Name
- Damian Płaza
- @raimeyuu
"It wasn't really me"
It's Tuesday. The incident channel is on fire.
A payment service started double-charging customers around 2 AM. The fix is already in, the bleeding stopped, and now comes the part nobody loves - the blameless retro.
Three people in the room (well, in the call).
Dylan, the engineer who merged the change.
Tanya, his tech lead.
Eric, the engineering manager, who just wants to understand what happened.
Read that conversation again, dear Reader.
Notice something?
Was anybody lying?
Was anybody lazy?
Dylan reviewed the diff. Tanya reviewed the PR. And yet, when Eric asks the only question that matters - who understood this? - there was only silence, afterwards.
Where did the understanding go?
Was it there before?
What if it leaked out through two words?
Copilot and junior.
A copilot who flies alone
Let's start with the "one printed on the product box".
Microsoft did not name the product "Autocomplete Deluxe". They named it Copilot.
It's a brilliant name - warm, collaborative, reassuring.
A copilot sits beside you.
A copilot shares the cockpit.
A copilot has your back.
And here's the thing about a copilot, dear Reader: in a real cockpit, the copilot is a human being.
Trained for years.
Accountable to an authority.
But suddenly this name "invites" all of that into how we think about "the machine".
But watch what people actually do with it.
They don't co-pilot. They autopilot.
They flip it on, lean back, and let it fly the straight stretches while attention drifts to the phone. The word says "co-", the behavior says "auto-", and almost nobody notices the slip - because the comforting half of the metaphor (there's a competent partner here) quietly cancels the demanding half (you are still pilot in command).
This is not Microsoft's sin alone.
We all do it. "Agent". "Assistant". "Teammate". Each one is a little connection from the world of accountable humans.
And we didn't see what else we have "invited".
The junior who never grows up
Now the second word. The one we invented ourselves.
"The agent is basically a junior dev."
I bet you you've heard it, dear Reader.
On the surface it's humble and generous - we're acknowledging the thing produces real work, we're calibrating our trust, we're being nice.
But a junior developer, dear Reader, is a person.
A junior is accountable.
When a junior ships a bug, they feel it.
They carry it home. Next time, they check twice.
They grow.
I want juniors that I work with to become better engineers, better people.
A junior also has something to lose and something to gain: a reputation, a career, dignity, a seat at the table.
An LLM has none of that.
It does not learn from the outage (not in that session, not in your codebase).
It does not carry anything home. It does not become a senior. It cannot be held responsible, because responsibility is a relationship between people, and there's nobody there to stand in the relationship.
So the "junior" metaphor cuts in two directions at once.
It over-humanizes the machine - lending it a role that comes bundled with accountability the machine does not have.
And it dehumanizes the actual juniors - because if a statistical text generator is "basically a junior dev", then what, exactly, were we valuing in the human ones? Their keystrokes? Or the fact that a person was growing a mind in there, learning the domain, building the theory that Naur told us the real product is?
I wrote once about abstracting over people - about how, the moment we turn a person into a role-shaped slot, we start treating the slot as the person.
Call the machine a "junior" long enough and you'll start treating juniors like "machines".
And remember, you are next in the line.
Every metaphor is a flashlight
I love metaphors.
In LLMs, 3D printers and amplifiers I happily told tiny tales with the usage of metaphors.
We cannot think without them.
We cannot live without them.
Lakoff and Johnson spent a whole book - Metaphors We Live By - showing that abstract thought is metaphor nearly all the way down. "Time is money." "Argument is war." "The economy is heating up."
We don't "decorate" our reasoning with metaphors - we reason in them.
A metaphor is like a flashlight (nice metaphor, isn't it?)
Point it at a dark concept and it lights things up. You suddenly see - that's the constructive power.
But a flashlight does two things at once. It illuminates a cone. And it leaves everything outside the cone in deeper darkness, precisely because your eyes adjusted to the bright part.
"Copilot" lights up: partnership, assistance, competence, we're-in-this-together.
It shadows: who is pilot in command, who signs the logbook, who the regulator calls.
Every metaphor highlights and hides (I wish I could say: highlights and downlights).
There is no neutral one.
George Box put it famously:
All models are wrong, but some are useful.
Where "wrong" means "imperfect", rather than "incorrect".
What if a model is just another metaphor?
Then we could easily state:
All metaphors are imperfect, but some are useful.
The question is never "is this metaphor true?" (none fully is) but "what does this metaphor let me stop seeing?"
When the flashlight becomes the room
Here's the part that actually worries me, dear Reader.
A fresh metaphor feels like a metaphor. You can feel the gap. When someone first said "technical debt", you heard it - ah, like a loan, interesting, and of course it's not literally money. The seams showed.
But metaphors age. And as they age, the seams disappear.
Say "technical debt" ten thousand times and it stops being a comparison and becomes a thing.
We have debt dashboards now. We "pay it down".
We've forgotten there was ever a metaphor there to question.
The flashlight didn't just light the room - it became the room, and now we can't imagine the furniture arranged any other way.
This is the same quiet trick that our mind plays with abstractions.
A good abstraction starts as a deliberate choice and ends as "just how things are".
We stop asking why null means absence, why a User is one class, why the folders are Models, Services, Helpers.
The decision fossilizes into the default, and the default becomes invisible - I keep poking at exactly this in the whole ambiguity series.
It becomes an invisible frame around our thinking.
That's exactly why naming is framing - and a metaphor is just naming wearing a costume.
"Copilot" and "AI junior" are turning invisible faster than almost any metaphor in our field, because they're reinforced by a billion dollars of marketing and a million exhausted engineers who'd like it to be true.
By the time a metaphor is invisible, you no longer choose it.
It's not a word in your vocabulary anymore - it's a wall in your mental model, and you plan your movements around it without knowing the wall is there.
That's the destructive power, and hear me out - the problem is not that the metaphor is wrong.
The problem is that we stopped seeing it at all, so we can no longer even ask whether it is wrong.
So pick a better flashlight
If we can't live without metaphors, and if every metaphor hides something, then the work is not to abandon them - it's is to be intentional about them.
Choosing them almost as if we design ours system - because whatever we care for representing, sooner or later will become our reality.
A ceiling or a trampoline.
So let me offer a different metaphor for the LLM (or tools, in general).
A better-fitting flashlight for the thing I actually observe.
Not a copilot. Not a junior.
An exoskeleton.
In LLMs, 3D printers and amplifiers I landed on amplifier - the sound amplifier makes the signal louder but does not play the instrument. The exoskeleton is that same idea, worn on the body. And I like it more for this conversation, because of what it refuses to hide.
Consider what an exoskeleton actually does:
- It amplifies the operator's own motion. It has no gait of its own. It multiplies yours.
- It makes you stronger and faster at whatever you were already doing - including walking confidently off a cliff. A weak, careless operator in an exoskeleton is not safer. They just get to make the same mistakes bigger, and earlier.
- The person inside is unambiguously the one in command. Nobody says "the exoskeleton walked into the wall". You did. With assistance.
- Strap it onto someone with no training and you don't get a professional. You get a civilian who can now lift a car - onto their own foot.
Feel the difference?
"Copilot" and "junior" both mean someone else is partly responsible here.
"Exoskeleton" does not expresses that sentence.
Accountability stays attached to the operator, which is exactly where it actually is.
An exoskeleton amplifies the operator - it does not replace, excuse, or share the blame with them. Strength multiplied includes the strength of your mistakes.
What does the exoskeleton amplify?
"Exoskeleton? Metaphors? That must be a joke!" - some might say.
Philosophizing all down the way, right?
If the LLM is an exoskeleton, then the obvious question is: an exoskeleton for which muscles?
Of course, designing systems is more about knowledge working - what Peter Naur called Programming as Theory Building, where the real product is the theory living in the builders' heads, not the text they typed - hence "muscles" are not the right wording here (a metaphor failing in public, dear Reader - how fitting for this very post).
What about term "capabilities"?
Let's try.
What capabilities is the LLM an exoskeleton for?
And the answer is a little uncomfortable, because the capabilities it amplifies are the ones we've spent a decade pretending tools would let us skip:
Critical thinking - is this diff actually right, or merely green? Dylan's exoskeleton produced idempotency-free retry logic at superhuman speed. It amplified his judgment - and he had turned his judgment off.
Clear thinking - do I understand the problem well enough to know what a good solution even looks like? The machine will happily amplify a muddy intention into a thousand muddy lines.
Clear communication - can I say what I want precisely enough that the amplified output is pointed at the right target? A prompt is a spec. If the spec is vague, you now get a vague system produced really quickly.
Empathic collaboration - do I still remember there are people on the other side of this code? The teammate who reviews it, the Dylan on call at 2 AM, the customer charged twice. The exoskeleton amplifies the consequences, but it does not care about them. It will never feel the weight of the person it just hurt. That part is yours. And it fades the moment "the agent handled it" quietly replaces "someone understood what this does to someone else".
These are not "soft skills".
They are the core skills.
The exoskeleton makes a strong operator astonishingly productive - and makes a weak operator dangerous at a pace nobody can review, control and handle.
The exoskeleton makes clear thinking the bottleneck and the premium. It doesn't remove the need for a strong operator.
It "just" exposes whether you are one, dear Reader.
I said something close to this in Code, knowledge and "AI": the machine produces probable stuff, and probable-but-wrong is expensive, because it looks like progress until it charges a customer twice at 2 AM.
So the antidote is not more speed.
Maybe it is about slowing down enough to understand what the exoskeleton just produced, remembering that simple isn't easy, and being ok to say "I don't know" before a probable artifact passes for an understood one.
Whoever trades understanding for speed will soon lose both.
Meanwhile, in the wild: "vibe coding"
There's one more metaphor loose in the field, dear Reader, and it's spread faster than COVID.
Vibe coding.
Describe a vibe, accept whatever comes back, ship it. Lovely. Frictionless.
But feel what it quietly takes for granted.
It assumes the coding is the job.
That the whole of the work is producing code - more of it, faster - as if keystrokes were the thing we were ever paid for.
It's like counting other proxy metrics - number of PRs, number of tests, test coverage.
Why?
Because it's easy, which means requiring less effort.
I called this the code fetish in The ambiguity of code: mistaking the representation for the understanding, treating "the code" as the goal instead of as one representation of knowledge we grew about a problem.
"Vibe coding" is that fetish, automated.
It amplifies the typing and leaves the understanding outside.
And it's only one end of a spectrum.
On the other end, a different metaphor is quietly available - a clumsier, less sexy one: AI-Augmented Software Engineering.
Notice what the word "engineering" refuses to hide.
Where "vibe coding" lights up flow, output, ease, "augmented engineering" keeps the demanding part in the cone: that there is engineering here, and the exoskeleton is meant to augment the capabilities engineering actually requires - not replace them with a vibe.
What capabilities?
The ones no amount of fluent output substitutes for.
Designing conceptual integrity - so the system says one coherent thing, not forty plausible ones stitched together.
Preserving optionality with our decisions - keeping tomorrow's doors open instead of nailing them shut at superhuman speed.
Directing coupling intentionally - deciding what gets to depend on what, on purpose, rather than letting the generator wire everything to everything.
That's the exoskeleton again. It amplifies the engineer's judgment. It does not vibe on their behalf.
"Vibe coding" amplifies typing.
"AI-augmented software engineering" amplifies engineering.
Same tool - two flashlights pointed at very different things.
Back in the retro
Let's rewind the call and give Eric one better question - the one the exoskeleton metaphor hands them for free.
Not "did the AI make a mistake?" (meaningless - the exoskeleton doesn't have intentions).
But: "where was the operator's attention while their strength was amplified?"
Discussion full of symbols, metaphros, abstractions and concepts - invisible as we take them for granted.
"Review the junior's PR" used to mean "check a human's understanding", but now, with the human removed and the concept (name) kept, it came to mean "skim some text nobody understood".
It always was about learning, understanding, "the theory behind the system", but suddenly it is about (over)producing more. Now. Quickly.
What's next?
Metaphors are how we reach understanding, and how we lose it.
They build the bridge and they decide which vehicles are not allowed to cross.
Both powers live in the same word.
We don't get the constructive half without the destructive half - we only get to choose how awake we are while we use it.
A double-edged sword, cutting both ways.
So please use metaphors, dear Reader.
Just be aware of them.
Just keep them visible.
Question them.
And when the strength is amplified, remember whose legs are doing the walking.
Yours.
What metaphors do you use, sometimes unconsciously?