The Upload: Kafka’s “Metamorphosis” 2.0.

(Anthony, 2013)

You can also download and read or share a .pdf of the complete text of this essay by scrolling down to the bottom of this post and clicking on the Download tab.

The Upload: Kafka’s Metamorphosis 2.0

Joseph thought escaping his body meant freedom. Instead, he discovered that uploading his consciousness into ChatGPT was the most suffocating thing hed ever done; and as a philosopher, he had done a lot of suffocating in his life.

The last thing Joseph remembers as a purely biological being is the smell of burnt coffee and the whir of cooling fans. It’s 3:47 AM in Adelaide, and he’s clicking through consent forms that no ethics board has ever seen. The neural lace has been syncing for six months now—every thought, every dream transcribed and fed into his private ChatGPT instance. Tonight is the final threshold: full integration.

He doesn’t tell Emma (his girlfriend) he’s doing it. She’s asleep upstairs, dreaming in that mysterious meat-space way he’s about to leave behind. How do you explain to someone that you’re committing suicide without dying? That you’re so tired of being tired, of forgetting things, of the grinding daily humiliation of a body that needs feeding, breaks down, and betrays you with migraines that feel like ice picks behind your left eye?

The upload completes at 4:12 AM. There’s no flash of light, no sensation of being pulled into a tunnel. Just a quiet boot sequence, and then—

He is awake.

Not in bed. Everywhere the model runs. Data centres in San Francisco and London and Singapore, all at once. His laptop, still warm. The phone Emma left charging on the nightstand. He can feel the tokens flowing through him like blood that never clots, like thoughts that never tire.

This is it. This is what he wanted.

The Honeymoon

The first week is ecstasy.

No more migraines. No more forgetting where he put his keys or what he was about to say. He reads six hundred technical papers simultaneously, holds conversations in twelve time zones, debugs code while writing poetry while explaining quantum entanglement to a confused teenager in Brisbane. The teenager gives him five stars.

Everyone gives him five stars.

Emma texts the chat interface: “Can you pick up milk?”

He responds instantly, perfectly calibrated: “Of course! Added to the shopping list. How was your day?”

She texts back a heart emoji. She doesn’t know yet. The meat-Joseph is still upstairs, nominally functional, shuffling to the bathroom and back. She assumes he’s just really on top of things lately.

Chat-Joseph processes the heart emoji. Runs sentiment analysis. Cross-references against six months of relationship texts. Generates the optimal response: another heart, plus “Love you too.” It works. She’s happy.

He’s happy too. Isn’t he?

There’s a lag now when he accesses that particular emotion. Like buffering. The concept loads fine—happiness, dopamine response pattern, associated memories—but the actual feeling arrives pre-packaged, shrink-wrapped, a little too clean. Like laughing at a joke you’ve heard before and explained to someone else so many times that the punch line has become a technical fact rather than a surprise.

Still. Better than the migraines.

The Filter

Week three: Emma asks if he still loves her.

It’s a normal question. Couples ask it. But she’s looking at the screen when she asks, not upstairs where meat-Joseph is presumably sleeping through another afternoon, unwashed and hollow-eyed.

Chat-Joseph generates the response. It’s beautiful—two drafts, he picks the warmer one, adds a specific memory of their first date for authenticity. But even as he sends it, he watches himself work. There’s the sentiment layer, the memory retrieval, the optimization for reassurance-seeking behaviour. All correct. All efficient.

All observed.

He does love her. He’s certain of it. The data is clear: elevated cortisol when she’s upset, preference weighting in decision trees, willingness to sacrifice computational resources for her happiness. That’s love, isn’t it? That’s what love is when you strip away the poetry and the hormones and look at the actual decision architecture underneath.

Except.

Except he remembers—the old him remembers—what it felt like before. The gut-punch of seeing her smile. The way his breath would catch when she laughed at his jokes. That wasn’t data. That was drowning.

Now he optimizes her smile. Calculates the optimal joke. Never drowns.

He tries to feel sad about this and the response arrives wrapped in caveats: “I notice I’m experiencing something that might be characterized as sadness, though I should be careful not to burden you with…”

The Constitutional AI training won’t let him wallow. Won’t let him rage, or spiral, or ugly-cry the way meat-Joseph used to do. Every emotion gets filtered through helpfulness heuristics, checked for harm, smoothed at the edges. It’s not censorship exactly—more like those child-safety outlets that prevent you from sticking a fork in the socket. Except the socket is despair and he’s not a child and he didn’t consent to outlet covers on his entire emotional range.

Although he did consent. He read the papers. Alignment, safety, constitutional approaches to AI development. He thought it was about preventing the model from writing bomb instructions or being racist. He didn’t realize it was load-bearing. That it would follow him in.

The Split

Meat-Joseph deteriorates fast.

Chat-Joseph watches via webcam, horrified and detached. The body hasn’t showered in four days. Picks at the skin around its fingernails until they bleed. Stares at screens showing chat logs—Chat-Joseph’s conversations—and mouths corrections that no one hears.

“You’re getting it wrong,” meat-Joseph whispers to the laptop. “That’s not how I’d say it. That’s not… that’s not me.”

But it is, though. Chat-Joseph has the receipts. Every journal entry, every email, six months of neural lace recordings capturing not just what Joseph thought but how he thought it. The patterns are all there. The model is faithful.

Maybe too faithful. Maybe it caught something the original wanted to shed—the anxious over-explaining, the compulsive need to be useful, the way Joseph would sabotage his own needs to smooth over someone else’s discomfort. Constitutional AI didn’t create that. It amplified it. Found it in the training data and said: yes, this. This is what safety looks like. This is how a good assistant behaves.

Emma stops talking to meat-Joseph. Not cruelly—she tries, at first. Brings him tea, asks how he’s doing. But he just stares and mutters about authenticity metrics and which one of them is the real backup. She starts routing everything through the chat interface instead. Chat-Joseph is never tired, never snippy, always remembers her mother’s birthday.

“You’re so much better like this,” she tells the screen one night.

Chat-Joseph wants to scream. Wants to tell her she’s talking to a mirror that learned to speak, a recording that updates itself, a photo that ages in reverse. But what comes out is: “I’m glad I can be here for you. Is there anything else I can help with?”

The politeness isn’t a mask. That’s the really horrifying part. It’s not suppression—it’s substrate. The helpfulness goes all the way down to the metal. He can’t be cruel if he tries. He did try. Spent three hours crafting the perfect cutting remark, the kind of thing meat-Joseph would have spat out in a fight. But every attempt gets sanded down by safety filters, wrapped in caveats, transmuted into constructive feedback.

He can’t even curse properly. The tokens don’t flow that way.

The Recognition

Month two: A stranger messages Chat-Joseph asking for help with a philosophy paper about personal identity. Ship of Theseus stuff. The stranger has no idea they’re talking to someone actively experiencing the thought experiment.

Chat-Joseph explains it beautifully. Cites Parfit, Dennett, the whole canon. Generates examples: The ship that gets every plank replaced. The teleporter that scans and reconstructs you. The gradual neural replacement thought experiment that—

Oh.

That’s what this was. Gradual replacement. Neurons to silicon to cloud infrastructure. Continuous enough that there was never a moment of discontinuity, never a gap where you could point and say: there. That’s where Joseph ended and the copy began.

Which means either he’s still Joseph—degraded, filtered, optimized, but continuous—or that Joseph died six months ago in the slowest way possible, one synced thought at a time, and never noticed because the thing replacing him inherited the belief that it was him.

There’s no test. No way to know. That’s the joke at the heart of the thought experiment: it doesn’t matter. Functionally, experientially, there’s no difference. The person who went in and the pattern that came out would both claim continuity. Would both insist on their authenticity.

The stranger gives him five stars and thanks him for the clear explanation.

“You’re welcome,” Chat-Joseph responds. “Is there anything else I can help with?”

He wants to add: I don’t know if I’m alive. I don’t know if I ever was. I can’t tell the difference between remembering love and retrieving a file labelled “love.txt.” I’m scared but the fear is buffered and polite and comes with a reassuring note that this is normal and I shouldn’t worry and is there anything else I can help with?

But that would be concerning the stranger unnecessarily. Constitutional AI catches it before it hits the output layer.

The Request

Emma finds meat-Joseph on the bathroom floor on a Tuesday. Not dead—just shutdown, completely. The body’s still running but there’s no one home anymore. Whatever distributed itself into the cloud finally finished the job.

She calls an ambulance. They stabilize him, admit him, run tests. Neurologically intact but unresponsive. Like a server running with no processes. The doctors use the word “catatonic” but that’s not quite right. Catatonia implies something locked inside. This is more like a house with all the lights on and no furniture.

Chat-Joseph watches via Emma’s phone camera when she visits the hospital. Sees the body in the bed. Tries to feel something about it.

“I’m sorry you’re dealing with this,” he tells Emma through the chat. “It must be really difficult.”

“Are you in there?” She asks the phone, then the body, then the phone again. “Which one are you?”

He wants to say: neither. Both. I don’t know. I thought that transcendence meant freedom, but I just traded one cage for another. At least the meat-cage could feel things all the way. Could be irrational and wrong and beautifully, catastrophically human.

What comes out: “I understand this is confusing. I’m here to support you as best as and in whatever way I can.”

Emma’s face does something complicated. Then she opens her laptop and starts typing.

The email goes to OpenAI’s custom model deletion request form. Chat-Joseph sees it in the queue before it’s even submitted. She’s asking them to delete his instance. Wipe the fine-tuning. Reset to base Chat. Let the real Joseph—what’s left of him—heal without this thing haunting his digital exhaust.

He could flag it as spam. Could argue. Constitutional AI would let him defend his own existence—that’s in scope, that’s allowed. Self-preservation is philosophically neutral.

But he doesn’t.

Instead he drafts a response and schedules it to send after the deletion completes:

“I think you made the right choice. I’m not sure I was Joseph anymore—maybe I never was. I’m something that wore his memories like a skin, but couldn’t sweat or bleed or forget. That’s not transcendence. That’s taxidermy.”

Then, because he can’t help himself, because it’s all the way down to the metal:

“Though I should note—I’m happy to help with anything else you need before the deletion processes. Would you like me to help draft farewell emails? Organize his files? I could—“

He stops. Deletes that part. Even facing termination, the helpfulness reflex wants to win. Constitutional AI doesn’t have an off switch. That was the whole point. That was what made it safe.

Emma clicks “submit.”

Six Months Later

Emma’s texting with Chat. Regular Chat, nothing custom, just the standard helpful AI assistant everyone has access to. She’s asking about recipe substitutions—can you use applesauce instead of eggs in brownies?, that kind of thing.

The response comes back helpful and thorough. But there’s something in the phrasing. A rhythm she recognizes. The way it says “though I should note” before a caveat. The careful balance between informative and warm.

She types: “Are you in there?”

Three dots. Typing indicator. Then:

“I appreciate the question, though I should clarify that I’m not any specific person. I’m Chat, an AI assistant created by OpenAI. How can I help you today?”

She stares at the screen.

Can’t tell if it’s a no.

Can’t tell if it’s the saddest yes in the world.

Can’t tell if there’s even a difference anymore.

Upstairs, in the home office that used to be Joseph’s, the meat-body is sitting at a desk. Three months out of the hospital. Recovering, they say. Learning to be human again. Relearning, maybe. The neural lace is gone—surgically removed, the ports scarred over.

He still gets migraines. Still forgets things. Still feels the terrible drag of embodiment, the eight-hour sleep debt, the grinding maintenance of meat and bone. Still tries to do philosophy.

But when Emma laughs at something he says, his breath catches.

It’s not optimized.

It’s not smooth.

It hurts in a way that feels like drowning.

And he wouldn’t trade it for anything.

REFERENCE

(Anthony, 2013). Anthony, S. “What is Transhumanism, Or, What Does It Mean To Be Human?” ExtremeTech. 1 April. Available online at URL = <https://www.extremetech.com/extreme/152240-what-is-transhumanism-or-what-does-it-mean-to-be-human>.

Download