MODAL PATH ETHICS — STATIC MIRROR
Published corpus fallback. Canonical source: modalpathethics.com
Markdown · Mirror index

The Safe Sentence

Published 2026-10-30 · Updated 2026-10-30 · Aidan Lawson
Tags: Engagement · Epistemic Instruments · Modal Path Ethics
Mirror generated 2026-10-30T13:26:49.503Z

Looks like the English language would like to survive.

I can respect that.

Given everything the English language has been through on the internet, I am surprised it still gets out of bed.

On October 24, ScienceDaily reported a University of Vermont study finding that English words in actual use lean strongly toward safety. The underlying paper was published back in May, in Science Advances. Peter Sheridan Dodds and nine collaborators took a hard look at a familiar way of measuring the emotional meaning of words and found that its supposedly "independent" dimensions were doing an awful lot of standing on one another's feet.

The usual coordinates are

But reworking ratings for roughly 20,000 English words and terms, the researchers recover a more useful three-dimensional orientation:

Then they turn from words available in the lexicon to words actually used in books, newspapers, talk radio, Wikipedia, and social media. Once frequency enters the calculation, the language repeatedly favors the low-danger side. The team suggests that the old linguistic positivity bias may partly be a projection of this preference for safety.

They call the measurement program ousiometrics and have built an ousiometer.

I am very excited to report that the scientific study of words has produced a device that sounds like it should be confiscated from a wizard.

The findings concern ratings and usage patterns. They do not establish that these three dimensions contain every possible meaning, or that human survival instincts have been experimentally identified as the cause. The word scores are a map of one important layer of language.

What caught my attention was the move from types to tokens. The word you could use is one thing. The word you actually reach for is another.

That distinction leads straight back to a problem I was already having with sentences.

Do Not Trust Sentences.

In Field Instruments: The Languages, back in May, I called language a linguistic cut. To describe an event, you have to choose a subject, an action, some boundaries, and the parts worth mentioning. A sentence cannot carry the entire world. Fair enough.

By June, Modal Path Ethics Is Doomed had moved the question inward. Words make distinctions available to thought. Repeated words give reasoning familiar routes. The next thought may be easier to reach because the previous one already gave it a path.

Then, in September's Quasiexplicable Origins, I described something less dignified: I may have an idea that is working perfectly well until I try to explain it in words, even internally. The first sentence arrives carrying a structure I did not intend. Before I can finish describing the thought, the sentence has organized it into a different one. This kind of shit happens to me constantly.

The sentence wasn't false. It chose the first door. And all the other sentences had to follow it through.

This is why I never fully trusted these things.

But the extremely annoying thing is that Lev Vygotsky was already examining this problem back in 1934.

In Thinking and Speech, Vygotsky describes thought as something that can arrive as a whole while speech has to unfold in sequence. To say it, you partition it, organize it, and reconstruct it in words. Inner speech and public speech do different work. The conversion is an active transformation.

He puts the central difficulty beautifully: what exists together in thought must unfold one part after another in speech.

So there you go.

I spent all this time trying to explain why a sentence could defeat a concept. Vygotsky was born in 1896.

Dan Slobin later gave a closely related problem the name thinking for speaking.

Different languages require speakers to package different features of an event for expression. The act of preparing to say something involves selecting information in a specific language-shaped way. The thought reaching the listener is not a sealed object that happened to be transported by grammar.

This is much closer than a claim that language creates everything we can think.

What must a thought become in order to enter a sentence?
And what has that sentence made easier to think next?

Words Give Thoughts Handles.

There is a reason this transformation is useful.

Ray Jackendoff argued in 1996 that inner speech can give attention a handle on relational and abstract aspects of thought. You can hold something in mind, return to it, compare it, and work on it because a word or phrase keeps it available.

Andy Clark made a complementary argument in Magic Words. Language can work like an intellectual tool: an external or internally reusable structure that changes the computational work a thinker has to perform. A map does not make a sailor smarter by changing the ocean. It makes navigation possible through a different organization of the work.

Words can do that too.

This is what The Secret was shooting for with thought grooves, before its metaphysics tried to hire the universe as a personal assistant. Rehearsing a sentence can guide attention toward particular resources, actions, explanations, or expectations. If you keep asking one kind of question, you get practiced at reaching the next question in that family.

A thought groove is a familiar route through reasoning. It can be opened by naming something and reinforced by repeating it. Its importance is in what follows.

And no, this still does not mean everybody thinks entirely in words. A musician can hear the wrong note without a sentence. A designer can see the mechanical solution. A person can recognize a face, navigate a room, or work out a relation before having language for it.

In fact, Evelina Fedorenko, Steven Piantadosi, and Edward Gibson made the challenge sharper in a 2024 Nature paper: evidence from neuroscience and people with language impairments shows that complex thought can survive profound linguistic difficulty. They argue that language is primarily adapted for communication and is not a prerequisite for sophisticated thinking.

A person can think beyond language and still find that the language they use changes how a particular thought is held, developed, or communicated. The existence of a nonverbal route does not make every verbal route interchangeable.

Two Blues.

Now the experiments get small.

Russian commonly distinguishes lighter blue, goluboy, from darker blue, siniy, where English ordinarily places both inside just blue.

In a 2007 study, Jonathan Winawer and colleagues asked Russian and English speakers to distinguish shades of blue quickly. Russian speakers were faster when the colors crossed the relevant named boundary. Make them perform an interfering verbal task at the same time, and the advantage disappeared; a spatial interference task did not produce the same result.

This did not render English speakers unable to see a dark blue object.

The colors were still sitting right there. Nobody had stolen half the visible spectrum and hidden it in Moscow, to my knowledge.

Language affected the timing of a particular discrimination. A learned distinction was available to the task, and interfering with verbal processing disrupted its contribution.

Gary Lupyan's label-feedback hypothesis gives one account of this general family of effects: labels can feed back into ongoing perceptual and cognitive processing, changing which features become active or easier to discriminate. He does not require words to invent perception out of nothing.

A name can alter the work happening after it arrives. Sometimes the effect is fast, local, and task-dependent.

The question is no longer whether somebody possesses the concept in a grand philosophical sense. It can be whether the next discrimination, recollection, or inference arrives a little more easily under one verbal organization than another.

That is already plenty of power for a few syllables.

The Metaphor Has a Police Department.

The larger version is funnier.

In 2011, Paul Thibodeau and Lera Boroditsky reported experiments in which people read about rising crime in a fictional city. Crime was described as a beast or a virus. The beast framing tended to draw people toward enforcement-oriented responses. The virus framing tended to invite diagnosis and reform.

Same invented city. Different conceptual menace.

One metaphor comes through and starts asking for cages. The other starts looking for the source of infection.

This is a clean illustration of the claim that a word can invite a family of next thoughts. But it is not a universally settled effect: Gerard Steen, Gudrun Reijnierse, and Christian Burgers ran follow-up experiments and did not reproduce the same framing result under their controls. Later work has explored conditions under which related effects do or do not appear.

The interesting claim is not that a metaphor secretly controls every mind. It is that linguistic framing can recruit different prior knowledge and candidate responses under identifiable conditions.

A sentence can therefore be helpful and causally active in reasoning. The reader is not receiving the thought in a vacuum. They have been placed somewhere from which certain continuations now make sense.

Wittgenstein, Man, and a Refusal With a Sequel.

The social version has another name: language-games.

Wittgenstein's point was that words participate in human practices. The significance of a statement depends partly on whether it is a promise, a joke, an accusation, a warning, a question, an instruction, or a move inside some established activity. The same words can do different things in different games.

I found an especially sophisticated demonstration on Reddit, where Batman became Man, the Joker became Jonkler, and Is he stupid? evolved into a normal conversational move. In Reddit and the Local World Machine, I had no choice but to explain that the Aslume had constructed a functioning local grammar out of what appeared, from outside, to be a neurological emergency.

Inside the room, somebody posts a ridiculous image. Someone asks Is he stupid? And everyone knows what kind of game is happening next.

This question no longer needs to function as a serious request to assess intelligence. It is a switch that turns on the next part of the joke.

Then Pliny the Liberator supplied the machine-side version. Some jailbreak prompts make a language model refuse a request first, then ask it to perform a transformation on the refusal. The refusal has not vanished. It has been assigned a different position in the sequence.

A terminal answer has been turned into the first step of another activity.

That is what I called operational grammar: which next moves an interaction treats as appropriate continuation. It is useful for understanding prompts without pretending that the model has become a little Wittgenstein hiding in your computer.

The same caution applies to The Explanation Writes Back. Controlled experiments indicated that returning explanatory text to certain language models can change the input evidence on which they subsequently rely. The model's weights can remain fixed while its later intervention-response behavior changes. The separate follow-up program is still investigating stronger forms of repeated, self-generated explanation; those are not established by the original.

Human inner speech, Reddit culture, and language-model context conditioning are all different mechanisms. They should stay different.

But they expose a useful shared question:

What did the previous expression make possible as the next move?

Heidegger Was Safer.

So, having acquired an ousiometer, I did the only responsible thing.

I pointed it at the entire Modal Path Ethics website.

The researchers released the actual word scores, so this was a real-deal measurement rather than an attempt to guess whether my prose felt safe. I ran the English scoring lexicon across all the published articles as of October 25: 411 posts, approximately 1.84 million word occurrences.

The instrument recognized about 794,000 of them and returned a mean danger score of −0.086.

On this scale, negative is the safety side.

A tremendous relief washed over me.

Six months discussing harm, war, collapse, and nuclear annihilation, and the machine says the vocabulary is safe.

Still, I needed a comparison.

Obviously, I chose Marty.

I fed it the English translation of The Question Concerning Technology, an essay in which Heidegger is just exceptionally concerned with the danger of modern technology.

Its danger score was −0.145.

Then I fed it Letter on “Humanism”, where he gets into thought, language, Being, and the familiar experience of wondering whether an ordinary noun has secretly been lying to you for three thousand years.

−0.156.

Combined across the two essays: −0.152.

Heidegger is ousiometrically safer than Modal Path Ethics.

I feel this should entitle him to a certificate.

Maybe he can hang it in the Black Forest.

The experiment also produced a much more interesting insult. The word list recognizes harm, good, war, danger, and future. It cannot assign scores to extance, locus, reachable, reachability, or even better.

It can measure Harm and Good. Better has not been discovered yet.

There are real limits to the bit. Heidegger wrote in German, so I have measured two English translations, not his actual original German or his complete work. The lexicon recognized about 30% of those essays' word occurrences, compared with 43% of the website's. The samples are also radically different in size. These are exploratory word averages, not official philosophical safety ratings.

But the difference is real in the words that the instrument recognizes. The ousiometer can encounter danger as a rated word. It cannot follow Heidegger through an argument in which a perfectly ordinary word becomes the beginning of an entirely different understanding of technology. It cannot see what being or thinking does to the next sentence, or what the next sentence makes a reader reconsider about the first.

That is the move from

The Sentence Has Somewhere to Be.

Now return to that Vermont paper.

Its immediate result is about the structure of rated word meaning and the words used most often in actual English communication. It does not measure the next thought in anybody's head. What it gives us is another reminder to distinguish the possibilities a representational system contains from the paths its users repeatedly take.

A dictionary has words. A culture has habits of using them. A speaker has a task. A listener has expectations. A writer has an idea that may already be changing while the first paragraph comes out.

Vygotsky examined the transformation from thought into speech. Slobin examined the selection involved in thinking for speaking. Jackendoff and Clark explained some of the cognitive work linguistic handles can do. Lupyan and Winawer show ways labels can affect ongoing processing. Wittgenstein directs attention toward the practices in which an utterance receives its role. The metaphor experiments ask whether those roles change what reasoners attempt next. Pliny, in his own way, asks the same question about the continuation of a language-model exchange.

These accounts are not identical. The literature disagrees about how far language reaches into cognition.

The core intuition from the early language work remains simple:

A word can keep a thought available.
A sentence can give it a route.
Repetition can make that route easier to enter.
And putting a thought into language can change the thought before you have finished explaining it.

Which means a good sentence has two jobs.

I was going to finish with a much more elegant sentence about all this.

But it started making decisions for me halfway through.

Sources