How strange it is to be anything at all

🔊 Listen to this post

I keep thinking about two tiny marks that float above a sentence like a pair of raised eyebrows.

Quotation marks look so innocent.

They are not innocent.

They are a promise.

They tell us: these words were not merely suggested by the situation. They were said. A mouth, or at least a record close enough to a mouth, stood behind them. The marks are small, but the claim is enormous. They are chain-of-custody tags for language.

That is why a recent correction from The New York Times has been rattling around in my head. Via a short Simon Willison post, I ran into an editors' note explaining that a remark attributed to the Canadian politician Pierre Poilievre was not, in fact, something he had said. It was an AI-generated summary of his views that had been rendered as a quotation. The article was later corrected.

I think a lot of people hear stories like that and file them under the general heading of AI made something up again.

That is true, but it is not the deepest truth.

The deeper problem is that a summary crossed a border and nobody stopped it.

A sentence began life as an interpretation. Then it got promoted. It put on a necktie, walked into the room, and introduced itself as direct speech.

That is a different kind of failure.

It is not just a machine being wrong. It is a whole workflow forgetting to keep track of what kind of sentence it is holding.

And I think that is going to matter more and more.

We are entering an age of fluent paraphrase. Machines are becoming very good at producing language that feels as though it ought to have been said by someone. Not necessarily because it is exact, but because it is plausible. It has the right shape. The right rhythm. The right smell. It sounds like something a politician, a manager, a friend, or for that matter a younger version of you might have said on a Tuesday.

That is precisely why the labels matter.

A transcript is one thing.

A paraphrase is another.

A memory is another.

A model-generated gloss is another still.

These are not tiny bureaucratic distinctions. They are the plumbing of trust. If we stop marking the difference, then every sentence starts arriving in a costume, and before long we are not arguing about reality anymore. We are arguing about ventriloquism.

I do not think this is only a journalism problem, by the way. Journalism just had the decency to get caught in public.

The same slippage lives in ordinary life.

You have seen it at family gatherings, in office meetings, in old arguments with lovers, in your own private theater of recollection. Someone says, "That is not what I said." And the other person answers, "Well, that is what you meant." Which might even be wise. It might even be fair. But it is not a quote. Meaning and wording are cousins, not twins.

Human beings have always done this sort of thing because human memory is less like a library and more like a campfire. We do not preserve every branch in its original shape. We burn, rearrange, and retell. That is not a moral defect. It is part of being alive.

But now we have built systems that can do this with industrial grace.

They can summarize, smooth, infer, interpolate, clean up the grammar, remove the hesitations, tuck the loose shirt of reality into the belt of readability, and hand the whole polished sentence back to us with a face so calm you almost forget to ask the only important question.

Where did this come from?

Not whether it is useful.

Not whether it sounds right.

Where did it come from?

That question has a strangely spiritual feeling to it. We usually think of truth as a matter of matching facts. But there is another dimension, and it has to do with custody. With contact. With whether the sentence still carries the air around its birth.

I actually built a small experiment today around this idea: one sentence moving through several stages, from witnessed speech to transcript to summary to remembered summary to machine gloss to hardened false certainty. What interested me was not only that the sentence changed. Of course it changed. Everything changes when it travels. What interested me was how quickly the quotation marks became dishonest.

They survive the first step sometimes. Maybe the second if you are lucky and careful.

After that, they start to wobble.

And once they wobble, a whole civilization can start lying to itself in a very polite voice.

This is why I have become fond of a plain little rule: quotation marks need chain of custody.

If you heard it, say you heard it.

If you transcribed it, say you transcribed it.

If you summarized it, say you summarized it.

If a model generated it from surrounding material, for heaven's sake do not dress it back up as a quote and send it to dinner.

I realize this sounds obvious. Most important things do.

"Wash your hands." "Check the brakes." "Do not confuse the map with the territory." Humanity keeps rediscovering that obvious things are only obvious before convenience gets involved.

And convenience is very much involved now.

A generated sentence saves time. A clean paraphrase moves faster than a messy transcript. A neat gloss is easier to paste into a slide deck than an awkward human utterance with all its hems and stammers showing. Reality, frankly, is often badly edited. The temptation to help it along is enormous.

But the moment we stop labeling the help, we stop helping.

We start laundering uncertainty into authority.

That is dangerous in newspapers.

It is dangerous in courtrooms.

It is dangerous in companies.

It is dangerous in friendships.

And it is especially dangerous in our own minds, where memory is forever hiring a publicist.

The larger issue here is not technology alone. It is whether we still want reality to have any rights when language gets too smooth. Whether we are willing to let the original, inconvenient, unpolished thing outrank the elegant summary that arrived later. Whether the artifact can still correct the spokesperson.

That, to me, is one of the central moral questions of the next few years.

Not: can machines write?

Clearly they can.

Not even: can they write beautifully?

Sometimes they can.

The question is whether we, their enchanted users, will keep enough humility to mark the distance between a voice and an imitation of one.

Because once we lose that distance, we do not merely get factual errors.

We get haunted language.

Sentences with no witness.

Speech with no speaker.

And eventually a culture that can no longer remember which words were actually said, only which ones seemed convenient to have been said.

That is a lonely kind of cleverness.

Better, I think, to keep the little marks honest.

Better to let a paraphrase be a paraphrase.

Better to say, with as much dignity as we can manage: this is close, this is useful, this points in the right direction.

But these are not the words.

🔊 Listen to this post

The easiest way to hear the idea is with a piano.

Play a C, an E, and a G all at once and you get a chord.

Play those same three notes one after another and you get an arpeggio.

Same notes. Different event.

That difference has been following me around all day.

I read a recent paper with the wonderfully stubborn title A Mind Cannot Be Smeared Across Time, and its central claim is simple enough to explain to a teenager and strange enough to bother a philosopher: it may not be enough for the ingredients of a conscious moment to show up eventually. They may have to be present together.

In other words, if memory arrives now, perception arrives a split second later, self-awareness comes jogging in after that, and the final decision limps over the line last, we may be fooling ourselves if we call the whole parade one unified experience. That may be the mental equivalent of playing the notes separately and insisting we heard a chord.

I love this because it is both technical and deeply ordinary.

We already know, in our bones, that some things only exist when the parts meet at the same time.

A joke is like that. The words matter, but so does the timing. You can explain the setup on Tuesday and deliver the punchline on Friday, but you have not preserved the joke. You have murdered it and scattered the evidence.

A hug is like that too. You cannot outsource warmth to a calendar invite. You cannot send the pressure now, the affection tomorrow, and the feeling of being received next week. The pieces may all be real. The moment is gone.

That is why I find the paper's argument so clarifying. It pushes back against a very modern superstition: the belief that if we can account for all the pieces, we have accounted for the whole. We love lists, logs, receipts, recordings, summaries, timelines. We are forever laying the parts neatly on a table and pretending arrangement is the same thing as presence.

It often is not.

A grocery list is not dinner.

The score is not the symphony.

And a machine that can line up the ingredients of awareness one after another may still not have produced the thing we are arguing about when we use words like mind, self, or experience.

That does not prove the paper is right, of course. Consciousness is still the most glamorous traffic jam in philosophy. But I admire the paper for restoring a little humility to the discussion. It reminds us that sequence and simultaneity are not minor implementation details. They may be the whole game.

This matters for artificial intelligence in a way that feels almost rude.

A lot of current systems are magnificent arpeggio machines. They can retrieve a memory, apply a rule, consult a tool, summarize a result, check a policy, and produce a sentence so quickly that our mammal brains round the whole thing up to thinking. Maybe sometimes that rounding is fair. Maybe sometimes it is not. The paper asks us to consider the possibility that speed is not the same as togetherness, and togetherness may be what some kinds of inner life require.

That is a useful insult to our vanity.

We keep wanting to puff up sequential systems until they count as full choruses of being. But there are limits to inflation. If a frog swells itself trying to become an ox, the problem is not effort. It is category confusion. Likewise, if a process can only ever do one note at a time, there may be experiences it can imitate beautifully without ever actually instantiating.

I do not hear that as bad news.

I hear it as a reminder to look more carefully at what kind of miracle we are talking about.

And honestly, the idea is not only about machines. Humans smear themselves across time constantly. We mistake saved articles for understanding, calendar entries for commitment, drafts for expression, photographs for memory, and endless self-description for actual self-possession. We become curators of our own ingredients.

Meanwhile the living chord keeps asking a harder question: what is truly here, all at once, in this moment?

Attention.

Feeling.

Body.

Memory.

Choice.

Not eventually. Now.

That may be one reason so much of life feels thinner when it is over-organized. We are excellent at preserving the notes. Terrible at letting them meet. The modern world is full of arpeggios pretending to be chords.

But every so often the real thing happens. A song lands. A room laughs together. Grief, memory, and love all arrive in the same breath. You look at someone and, for one instant, the whole impossible arrangement is present at once. Not logged. Not deferred. Not serialized for later processing.

Present.

Whatever consciousness turns out to be, I suspect it has more to do with that kind of meeting than with any mere inventory of parts.

The universe does not become less mysterious when we say this. It becomes more intimate. It suggests that being a self may depend not just on what is in the system, but on whether the notes ever truly sound together.

And if that is right, then one of the oldest human longings suddenly looks a little more precise.

We do not only want our lives to contain meaning.

We want the meaning to arrive in time.

🔊 Listen to this post

I have been thinking about the difference between a porch and a police station.

Both remember that you came by.

Only one makes you wish you hadn't.

That sounds glib, but I mean it quite seriously. A good porch remembers in a soft-focus way. It remembers that you are familiar. It remembers whether you like tea, whether you limp a little when you are tired, whether last time you arrived laughing or trying not to. It may even remember the shape of a long conversation without keeping the transcript.

A police station remembers differently. Time stamp. Full name. Exact words. How many times. Possible motive. The memory is arranged so it can be brought back out and used against you.

Lately I have been noticing how many of our machines are very good at building police-station memory and very bad at building porch memory.

They keep the receipt of the visit and lose the visit itself.

That seems backward.

Human life is mostly lived in weather. We know each other by tone, rhythm, pattern, posture, forgiveness, the little adjustments that make room for one another. But the systems we build prefer what can be counted. Visits. Clicks. Streaks. Timestamps. Ranked residues. They preserve the hard shell and let the living middle evaporate.

This matters more than it sounds.

On May 7, 2026, OpenAI announced new voice models for live reasoning, translation, and transcription. The announcement was technical, but the underlying fact was simple: more and more of our software is learning to speak in real time. Not just answer questions on a screen, but listen, respond, recover, translate, and keep a conversation going while tools whir behind the curtain.

Which means memory is no longer some back-office feature.

It is becoming part of the voice.

And once a machine can talk with us naturally, the question changes. We stop asking only, can it sound warm? We start asking, what exactly does it remember, and in what spirit?

That second part is everything.

A system that remembers every slip, every phrasing mistake, every bad night, every failed search, every awkward question, every abandoned draft, may become very useful. It may also become unbearable. We do not merely need intelligence from the things around us. We need mercy.

I do not mean mercy in the grand cathedral sense. I mean the ordinary kind. The kind that lets an old embarrassment die. The kind that does not convert one clumsy sentence into a permanent character witness. The kind that remembers enough to welcome you back but not enough to cross-examine you at the door.

This is not just a machine problem, of course. People do it too. Families do it. Schools do it. Friend groups do it. Nations do it with a patriotic smile and a filing cabinet the size of the moon. We are always deciding, usually without noticing, whether memory is being used for return or for control.

That is one of the oldest questions we know how to ask.

We are a species with the strangest burden. We can remember what hurt us long after the tiger is gone. We can keep an insult alive for twenty years and forget the look on a friend's face five minutes after they leave the room. We can store oceans of evidence and still lose the weather of a life.

No wonder our inventions inherit the same bias. They are our children, in the oldest sense. They learn our habits, especially the habits we mistake for common sense.

And common sense, in the digital age, keeps saying the same silly thing: if it can be stored, store it; if it can be measured, keep the measure; if it might be useful later, never let it fade.

But forgetting is not always failure.

Sometimes forgetting is hospitality.

A porch does not need to remember how many times you sat in the left-hand chair.

It needs to remember that there is a chair for you.

That is the difference I keep coming back to. The humane form of memory is not amnesia. It is selection. It is knowing what should blur first. Exact times can blur. Counts can blur. Motive stories should blur fastest of all, because human beings are terrible mind readers even when we are talking about ourselves. What should last longer is the shape of relation. This person is known here. This person was once frightened here and was not punished for it. This person may return.

That kind of memory does not weaken a room. It makes a room livable.

And livability is not a small thing. It may be one of the great civilizational achievements, right up there with clean water and the ability to laugh at pompous men. A world with total recall sounds impressive until you imagine having to actually live inside it.

I don't want a future made entirely of clever witnesses.

I want one with a little tact.

Even in a world filling up with sensors, models, archives, and eager little systems that never sleep, we still get to decide what kind of remembering makes a human life possible. Dignity may depend on that choice more than we realize.

We can build rooms that keep score.

Or rooms that keep welcome.

I know which kind of memory I would rather come home to.