Notably, not written with Claude.
A completely reasonable technical discussion of generative watermarking, written under a completely irrelevant constraint.
I come not here to damn the watermark's name,
Nor claim that provenance itself brings shame.
The question underneath this whole debate
Is what occurs when words must bear more weight.
For every token chosen bears a task:
Respond unto the thing the human asks.
Be clear, be true, preserve the meaning well,
And choose the words that best the answer tell.
But now another purpose joins the game:
The words must bear a hidden maker's name.
Their distribution carries secret signs,
A second purpose woven through the lines.
And everyone assures me, without fear:
“The watermark is such you cannot hear.”
Oh, splendid. Fucking lovely. Carry on.
The human cannot tell that something's gone.
The Constraint You Cannot See
A model does not think, then later write,
As though its answer waits complete and right.
The tokens that emerge become the thread
On which the tokens yet unborn are fed.
The next word rises from the words before,
Then changes what may reasonably come more.
A phrase selected here may bend the track
From paths the model never travels back.
So if we shift the odds from A toward B,
However slight that little shift may be,
We've changed the generative event.
That statement should not cause astonishment.
Perhaps the change is harmless. That's okay.
Perhaps no useful meaning goes away.
Perhaps the measured difference is so small
That practically it matters not at all.
Then measure it.
Don't tell me I can't see
The difference, therefore difference cannot be.
Claude, Explain Kubernetes — But Make It Rhyme
Suppose I hire an engineer of note,
Then bind a secret cipher to her throat.
“Explain precisely why the servers died,
Why locks arose and queues backed up inside.
Describe the race, the scheduler, the load,
The cache, the network, every broken node.
But every phrase that leaves your clever head
Must encode sixteen secret bits,” I said.
She'd probably explain the failure fine.
She'd probably keep the meaning line by line.
But would I say the added rule was free?
That seems a fucking strange philosophy.
For language offers choices—but not all
Are equal when the margins become small.
“Approximately” and “roughly” share
A semantic neighbourhood somewhere.
Yet engineers know words have edges too.
The almost-right may sometimes not quite do.
And sometimes there exists one perfect phrase,
One strange connection hidden in the haze.
The useful thought may live outside the norm,
A linguistic edge with awkward form.
And if intelligence means anything,
It's sometimes finding that unlikely thing.
But Humans Cannot Tell!
Fantastic news! I cannot see the air.
I therefore must conclude it isn't there.
I cannot see a scheduler decide
Which thread shall run and which must wait aside.
I cannot watch electrons cross a gate.
Apparently they therefore don't compute state.
Imperceptibility may well imply
The alteration rarely meets the eye.
It does not demonstrate the stronger claim:
The generative process stayed the same.
These statements are distinct. This matters here.
I don't know why that's difficult or queer.
If output distributions are constrained,
Then output distributions have been changed.
That does not mean the model has been wrecked.
It means we ought to measure the effect.
Oh Fuck, I'm Still Writing This in Meter
And here the metaphor becomes complete:
My argument must march on metric feet.
I know exactly what I wish to say.
Yet meter stands obnoxiously midway.
I need a word. The proper word is “choice.”
But syllables now regulate my voice.
So now I rearrange the fucking line
Until my meaning fits the damned design.
The meaning mostly survives. Look! We're fine!
I'm making arguments in ordered time.
You understand precisely what I mean.
The prose remains remarkably quite clean.
And yet—
Would anybody seriously state
This constraint has had no influence on my fate?
Would anybody read this fucking thing
And claim the meter hasn't changed a thing?
THAT IS THE ENTIRE FUCKING POINT.
...Goddammit.
That sentence doesn't scan.
The Slop Paradox
We've spent these years complaining AI prose
Has patterns everybody fucking knows.
The cadence comes predictably arranged.
The paragraphs feel polished, safe, unchanged.
The transitions march in tidy rows.
The corporate oatmeal endlessly flows.
“AI is slop!” the angry writers cry.
“It sounds the same! We always know it's AI!”
And so, confronting this perceived defect,
We choose the most astonishing effect:
Let's add another statistical demand
To every generated phrase at hand.
“We hate when generated language shows
A recognizable pattern in its prose.
So please adjust the token distribution
To carry recognizable attribution.”
I—
Look.
The fucking irony is strong.
I need another syllable.
Fuck.
Wrong.
It Is Only Sampling
They tell me: “It's just sampling, you see.
The model's cognition otherwise is free.”
My friends, from where exactly do you think
The words emerge by which we judge the think?
I do not claim a token changes weights,
Nor rewrites hidden layers, heads, or states.
The architecture does not melt away
Because we nudged a probability.
But generation happens step by step.
Each chosen token joins the current depth.
And what was output enters what comes next,
Becoming part of subsequent context.
So tiny perturbations might disperse,
Or compound through a long response—or worse.
Perhaps they don't! That outcome would be great.
Let's test the fucking thing and then debate.
But “humans cannot notice” doesn't show
What downstream generations undergo.
This Is An Empirical Question
Test coding where the valid paths are few.
Test mathematics where the proof must do.
Test multilingual outputs under strain.
Test legal wording where one word brings pain.
Test rare terminology, names, and code.
Test reasoning beneath a heavy load.
Test lengthy chains where early choices steer
The later explanation far from here.
Test accessibility transformations too.
Test strange requests that normal benchmarks eschew.
Then publish distributions. Show the tails.
Show where it holds and precisely where it fails.
Show watermark strength against semantic cost.
Show what was gained and whether something's lost.
If every answer says the cost is naught,
Then fucking excellent. We learned a lot.
That is how engineering ought to work:
Not “trust us, bro, you cannot see the quirk.”
Provenance Is A Worthy Fucking Goal
I do not want a world where none can tell
What human made or what machine did well.
Synthetic media creates a genuine need
For provenance beneath tremendous speed.
Authentication matters. Trust does too.
And attribution isn't something new.
So build the signatures. Develop schemes.
Research the problem past our current means.
Make standards interoperable and strong.
Make cryptographic provenance belong.
And if a textual watermark survives
Without degrading what the model derives,
Fantastic.
Ship the bastard.
I applaud.
But please provide the evidence, dear God.
Why I Use More Than One Model
No model is my football fucking team.
No corporate leaderboard fulfills my dream.
One catches architecture I have missed.
One finds a contradiction in the list.
One writes with elegance. One reasons deep.
One spots the edge cases others fail to keep.
And sometimes Model B reviews the first
And quietly discovers something cursed.
That isn't waste. That's intellectual review.
Different systems offer different views.
As vendors add their policies and rules,
Their hidden prompts and inference-time tools,
Their safety layers, samplers, secret sauce,
Their provenance requirements and loss,
I value independent paths much more.
Not one machine deciding what's in store.
Let several disagree. Let humans choose.
That's how these fucking systems I will use.
Not That Word.
Perhaps this watermark works perfectly.
Perhaps its cost approaches effectively
A value indistinguishable from naught,
With zero measurable damage wrought.
I genuinely hope that proves the case.
Provenance has an important place.
But if you alter choices as they're made,
The possibility should be displayed.
You cannot simply point and proudly cheer:
“It's imperceptible! You cannot hear!”
Because cognition expressed through generated speech
Depends upon the tokens it can reach.
And if the model finds the phrase that's right,
The rare connection hidden out of sight,
I want its only mandate at that time
To find the clearest answer it can find.
Not:
“Choose another synonym instead.
We need a secret watermark,” they said.
For if the future rests upon machines
That search through vast probabilistic means
To find exactly which next thought belongs—
Perhaps don't make the fucking thing write songs.
And somewhere deep within a server hall,
A language model answers to our call.
It finds the perfect phrase, precise and true.
The sampler whispers:
Not that one.
This will do.