My view on AI ethics

2026-09-20

Every serious objection to building AI shares one property, and once you see it you can't stop seeing it.

They are all one-sided ledgers.

Energy gets counted, and what the energy produces does not. Job losses get counted, and jobs created, wages raised, and diseases cured do not. The risk of building gets counted at speculative probabilities and enormous magnitude, while the risk of not building gets a blank column, despite being measurable, certain, and currently running at sixty-eight million deaths a year.

This is not a subtle error. It's the kind of accounting that gets an auditor struck off. And it survives because a one-sided ledger always returns the same verdict, which makes it extremely useful to anyone who has already decided what the verdict should be.

So this is a ledger. Ten entries. For each one I fill in both columns and say what's left.

Six get dismissed. Four survive, and the four that survive are not the ones anybody is shouting about.

How to read the findings

I'm using four dispositions and they mean specific things.

• DISMISSED · The factual claim is wrong.
• DISMISSED WITH PREJUDICE · Wrong, and wrong in a way that five minutes of arithmetic would have caught. The persistence is the interesting part.
• REMANDED · A real problem, misfiled. It's not an argument about AI, it's an argument about something else wearing AI's coat.
• SUSTAINED · Survives contact. Goes on the short list in Part Three.

The docket

Part One · Dismissed

01 · Water

Charged · every prompt drinks a bottle of water

Actual · 0.26 mL per median prompt. Five drops.

Finding · DISMISSED WITH PREJUDICE

This one is a joke and it deserves to be treated as one, because it is the clearest case in the whole debate of a number that nobody checked and everybody repeated.

The viral figure traces to a Washington Post piece built on Li et al., "Making AI Less Thirsty", and it became the claim that a short ChatGPT exchange consumes a 500 mL bottle. That implies somewhere between 10 and 25 mL per prompt.

In August 2025 Google published actual production measurements across its serving fleet: the median Gemini text prompt consumes 0.26 mL of water and 0.24 Wh of energy. Five drops, and less electricity than nine seconds of television. The viral number is off by a factor of roughly 40 to 100.

It gets worse under inspection. Andy Masley has done the most careful public work on this in "The AI water issue is fake", and the finding that should end the conversation is about what the original figure was even measuring. Around two thirds of it was not data center water at all. It was the water notionally attributed to generating the electricity, which is overwhelmingly thermoelectric withdrawal, taken from a river and returned to it, counted here as though it were drunk. Masley also documented that Empire of AI, the bestselling critical account, was off by a factor of a thousand on its water figures.

Scale it to something a person can hold. Your daily water footprint is on the order of hundreds of thousands of times a single prompt. American agriculture is roughly 80% of national consumptive water use. Golf courses, almonds, alfalfa exported to feed cattle in other countries. You could shut down every AI data center on the planet and not find the difference in the gauge.

What's actually there: siting. A data center drawing evaporative cooling water from a stressed basin in Arizona or central Chile is a real local problem. It is a zoning and permitting problem, it has known engineering fixes in closed-loop and air cooling, and it has nothing to do with whether AI should exist. See entry 07 for where that argument actually belongs.

02 · Energy and climate

Charged · AI is burning the planet

Actual · all data centers are 1.5% of global electricity. AI is a fraction of that.

Finding · DISMISSED

The IEA's *Energy and AI* puts global data center consumption at about 415 TWh in 2024, roughly 1.5% of world electricity, projected to reach around 945 TWh by 2030, which is just under 3%. That's every data center: video streaming, banking, email, cloud storage, cat pictures. AI is a slice of the slice, and globally it is well under one percent.

For calibration: Bitcoin alone runs at roughly a third of the entire AI load. Global aviation is about 2.5% of CO2 emissions. Cement is 7 to 8%. Nobody has written a bestseller about cement.

Now the part I actually want to argue, because dismissing the number is the boring half.

I would support building this at seventy percent.

Not as a provocation. As accounting. Energy is an input, not a sin. The entire question of whether an energy expenditure is justified is a question about what it produces, and we have never in the history of the species had trouble with this reasoning anywhere else. Nobody asks whether hospitals are worth their kilowatt-hours. Nobody demands that agriculture justify its diesel. Energy per capita is one of the most reliable correlates we have with life expectancy, literacy, child mortality, and every other measure of whether human life is going well.

So if a technology consumed seventy percent of world electricity and in exchange compressed the discovery timeline for the diseases that kill sixty-eight million people a year, that is not a close call. That is the best trade anyone has ever been offered. The correct response to "this uses a lot of power" has always been to build more power, and the fact that a generation of environmental politics trained people to hear that as heresy is a failure of that politics, not an argument about compute.

The one percent figure is true and I'll use it because it's true. But I want to be clear that it is not where my position rests. If the number were seventy times higher I would hold the same view, and anyone whose support flips on the energy figure was never arguing about energy.

03 · Jobs

Charged · it takes the jobs

Actual · labor is a means to an end and nobody has ever wanted it for itself

Finding · DISMISSED, one count remanded

Start with the part the argument never says out loud, because once it's said it gets hard to keep making.

Nobody wants a job. People want what a job currently delivers: income, structure, status, a reason to leave the house, and somewhere to put the part of themselves that needs to be useful. Those got bundled into one institution by industrial accident, and the bundle is old enough that we've lost the ability to see the seam. When someone says AI will take the jobs and means it as a catastrophe, they are treating the bundle as a terminal value.

Check it against revealed preference, which is the only evidence that matters here. Lottery winners cut their hours. Retirement is the single most anticipated event in most working lives and people spend four decades financing it. Nobody's deathbed regret is insufficient timesheets. And every labor movement in recorded history fought to work less: the ten-hour day, then the eight, then the weekend, then paid leave, then parental leave. Two centuries of organized labor is a sustained campaign against labor, and we call it a triumph, correctly.

So notice what the displacement argument actually asserts. That the right quantity of human toil is precisely the quantity we happen to have, and any reduction is a loss to be resisted. Nobody holds that position when it's stated that way. It survives only in the bundled form, where "losing your job" silently carries "losing your income" and you get to skip the argument.

Unbundle it and two questions remain. One is solved and one is political, and only the political one is hard.

Is there enough left for people to do?

This is the lump of labor fallacy: a fixed quantity of work exists, a machine doing some leaves less for people. Named and refuted since the 1890s, returns every time in the same words with the same confidence.

The cleanest refutation is Bessen's, and the numbers deserve to be quoted rather than gestured at. In 1985 the US had 60,000 ATMs and 485,000 bank tellers. By 2002: 352,000 ATMs and 527,000 bank tellers. Teller employment went up. The mechanism isn't mysterious. ATMs cut the staff needed to run a branch from about 20 to 13, which cut the cost of a branch, so banks opened 43% more of them in urban markets and hired more tellers to compete on service.

Agriculture was around 60% of American employment in 1850 and is about 1.3% now. A ninety-eight percent displacement of the largest job category in the country, and unemployment did not go to 58%.

Now the strongest version of the counter, since the weak one isn't worth the space: every prior automation displaced tasks and left humans a reserved category, and a general intelligence leaves none.

That fails on economics older than computing. Comparative advantage doesn't require humans to be better at anything. Ricardo's result holds under absolute advantage in every good, and it holds for exactly as long as the superior producer's capacity is scarce. Compute is scarce and stays scarce, because demand for it expands to fill whatever gets built. As long as a datacenter hour has an opportunity cost it goes to its highest-value use, and everything below that line is worth a human doing.

What's left is the part people would do anyway

Some industries survive not because a model can't do the task but because human provenance is the product.

This already exists and we already pay for it. Recorded music is free and perfect and people buy tickets to watch a person sweat on a stage. Machine-made furniture is better toleranced and cheaper and there's a premium for the version somebody's hands made. Therapy is a conversation, and the entire value proposition is that another person is in the room and chose to be.

Outreach, care, coaching, hospitality, teaching, sales, ministry, nursing: the work that consists of one person paying attention to another. None of it is safe because the task is hard. It's safe because the task is the human doing it, and a perfect substitute that isn't a person is not a substitute at all. That's a category the automation argument has no vocabulary for, because it's been counting tasks.

My actual prediction is not mass unemployment and not full-employment-as-usual. It's that work becomes optional, and the work that remains is disproportionately the work people would choose without the money, which is close to the definition of a good outcome.

The honest objection

Keynes wrote "Economic Possibilities for our Grandchildren" in 1930 and predicted his grandchildren would work fifteen-hour weeks. On productivity he was roughly right. On hours he was badly wrong, and anyone promising optional work owes an account of why.

The account is that the gains went into consumption and positional competition instead of time. We got bigger houses and more stuff rather than more afternoons, partly by choice and substantially because status is relative and relative status cannot be bought by everyone at once. That's a real failure mode and it's the reason to take the optimistic case seriously as a project rather than a prediction. Abundance doesn't automatically convert into leisure. Someone has to want it to.

But note the shape of the objection. It's an argument that the benefits were misallocated. It is not an argument that the productivity was bad, and it is certainly not an argument for less of it.

Remanded

The distributional claim, which is the one real thing in this entry. Aggregate employment recovers; particular people, in particular towns, in particular decades, do not. Acemoglu and Restrepo found real, local, persistent employment damage from robotics, and the aggregate absorbing it is no comfort to a person inside the absorption.

That's a genuine problem. It's a transfer problem, solvable with money, and it belongs to tax policy rather than to the question of whether the technology gets built. Filing it under AI ethics is how it stays unsolved, because the people filing it there are not writing tax policy and were never going to.

04 · Theft of the training data

Charged · pretraining launders copyrighted work

Actual · three questions welded into one; two are real, the famous one isn't

Finding · DISMISSED, two counts sustained

A 400-billion-parameter model at bf16 is about 800 GB. Its corpus is 10^13 to 10^14 tokens, hundreds of terabytes. The model is a lossy compression of its inputs at two to three orders of magnitude. It is physically incapable of retaining them. Training is a forced choice of the general over the specific, forced by the parameter budget.

Per work: a novel is ~10^5 tokens, one part in 10^8 or 10^9. It produces a gradient, that gradient is averaged into a batch of millions of tokens, scaled by a learning rate near 10^-4, and then ~10^5 later steps walk over the same weights.

That's arithmetic, and the arithmetic has been checked empirically. Grosse et al. (2023) computed influence functions for large models, which answers the direct question of which training sequences caused a given behavior. Influence is diffuse rather than concentrated, and it gets more abstract as models get larger: for big models the most influential sequences relate to the output conceptually, not by surface overlap. What the model took was the structure the corpus shares, not the items it's made of.

Separate the three charges and the argument resolves cleanly.

Acquisition. SUSTAINED. Torrenting a pirate library is conversion, and no downstream transformation launders it. Judge Alsup said exactly this in Bartz v. Anthropic in June 2025, in the same ruling that held training on lawfully bought books to be fair use and "exceedingly transformative," and Anthropic wrote a $1.5B check about it in September. Pay for the books.

Memorization. SUSTAINED. Carlini et al. showed memorization scales log-linearly in model size, how many times a sequence is duplicated, and extraction context length. Lee et al. showed deduplication cuts memorized emission by about 10x. So regurgitation is a function of duplication, not inclusion. The once-in-10^14 passage is unrecoverable; the ten-thousand-times passage is, and is functionally a proverb. Dedup, filter the output, pay for the distinctive tail. Engineering, with a price tag, which should be paid.

Statistical residue in the weights. DISMISSED. This is what "data sovereignty" actually means and it's the one that fails. Sutton's bitter lesson and Gwern's scaling hypothesis make the same point from two directions: capability came from compute and scale, not from any curated ingredient. Scale needed a corpus. It never needed yours. The claim isn't just weak law, it misidentifies the cause.

The best objection: if no corpus is load-bearing, why did every lab sign data deals? Because coverage, freshness, and indemnity are worth paying for even when no individual item is. The price of a corpus is the cost of assembling an adequate substitute, not the value of its rarest element. And notice the tell: these deals are priced per corpus and per contract, never per work. No rightsholder has ever priced a single irreplaceable item into one, which is exactly what dilution predicts and what its denial cannot explain.

And the ground is moving anyway. Textbooks Are All You Need (2023), Nemotron-4 340B shipped in 2024 under a license that explicitly permits training on its outputs, R1 released under MIT in January 2025 with the reasoning traces as the product. Take an open-weight model, run it on your own hardware, sample a billion tokens of derivations and code with tests attached, throw out everything that fails its verifier, train on the rest. Nobody was scraped. There's no rightsholder. AlphaZero proved eight years ago that given a verifier, a system bootstraps past the entire human record and keeps going.

Every provenance regulation now in draft will bind on a constraint that has already relaxed by the time it takes effect, and will function as a tax on whoever stayed honest about their corpus.

05 · Alignment

Charged · we must load human values into the model before it's too late

Actual · a photograph being asked to work as a compass

Finding · DISMISSED, one count sustained

I want to be careful here, because "alignment is a joke" is true of the program and not true of the people. Some of the sharpest people I know work on this. The problem is structural and no amount of individual brilliance fixes it.

Every current method encodes a snapshot. RLHF encodes the preferences of a specific labeling workforce, at a specific moment, against a rubric written by a specific team. Constitutional AI encodes a document, and documents have authors with positions. Both are good engineering. Both are then handed the job of serving as a compass, and a photograph does not do that.

Ask the question plainly and the structure collapses: aligned to whom, as of when? Every available answer is a political claim in technical dress. The 2019 consensus on what a model should refuse now reads as absurd in both directions at once. Nothing suggests the 2026 version ages better, and there is no procedure anywhere in the literature for updating the target that doesn't relocate the identical question one level up.

Then run Bostrom's own orthogonality thesis backwards. It's deployed as a warning: intelligence and goals are independent, so a superintelligence could want anything. Fine, accept it. Then our goals have no privileged standing either. They're what one primate optimized into under Pleistocene conditions, and nominating them as the target amounts to noting that we got here first.

Look at the field's prediction record while we're here. Its central objects, the treacherous turn, recursive self-improvement to foom, the mesa-optimizer, have been load-bearing for twenty years and remain theoretical. The actual trajectory was gradual, compute-bound, and legible, which is close to the opposite of what the canonical model predicted. Meanwhile model behavior improved through scale and unglamorous RLHF, not through alignment theory. A field whose core predictions haven't paid out in two decades, and whose practical wins came from somewhere else, has an epistemics problem it has not acknowledged.

The one thing worth keeping is the proposal usually dismissed on the strength of who's making it. Musk called AI "summoning the demon" at MIT in October 2014 and meant it, then founded xAI nine years later around maximally truth-seeking AI. The standard read of that arc is hypocrisy. The standard read is lazy: the doomer conclusion, followed to the end, is that you build, because if it's coming and it's the biggest thing then the only lever anyone holds is being at the wheel. DeepMind, OpenAI and Anthropic all have founding stories of that shape.

The substance is better than the arc, and it's strongest when you state it in the safety literature's own vocabulary. Omohundro's basic drives and Bostrom's instrumental convergence establish that a capable agent wants an accurate world model regardless of its terminal goals, because accuracy is useful for nearly anything. Accurate world-modeling is therefore convergent: you don't have to install it and defend it against optimization pressure, because the pressure produces it.

That gives truth-seeking the one property no value-specification has. It doesn't rot as capability rises. The target moves in the same direction the capability does. Every value-loading approach has the inverse property, because the specification was written by the less capable thing and degrades exactly as the gap widens.

Where it's incomplete is precise, and it becomes entry 09.

06 · The pause

Charged · stop or slow down until it's safe

Actual · 68 million people die per year and the pause camp has never once priced them

Finding · DISMISSED WITH PREJUDICE

This is the entry I care about most, and it's where the one-sided ledger stops being an intellectual complaint and starts being a body count.

About 68 million people die every year. Cardiovascular disease takes roughly 20.5 million. Cancer takes close to 10 million. That is the baseline. It runs whether or not anyone is paying attention, and it has never once appeared in the right-hand column of a pause proposal.

Twenty years of delay is a window in which roughly 1.36 billion people die.

I'm not claiming AI saves all of them. Nobody could claim that. The claim is narrower and much harder to answer: the people arguing for delay have never published an estimate of what fraction is attributable to the delay. Not a careful one, not a rough one, not a bad one. The column is blank. It isn't that they weighed the deaths and found the risk heavier. They never entered the deaths.

And the estimate would not be zero, which is why it stays unwritten. AlphaFold predicted structures for essentially the entire known protein universe, around 200 million of them, replacing what would have been decades of crystallography. That already happened. It was fast, it was cheap, and it is the least impressive thing this technology will do for biology.

Now put the two sides in the same units, because that's the whole exercise.

One column is a hypothesis. The other is a fact with a WHO fact sheet. Standard decision theory says you compare expected values. The pause literature compares one side's tail against the other side's zero, which is not caution. It's an accounting choice, and it happens to be the accounting choice that always returns the answer the author started with.

There's a second failure on top of the first. A pause is not available as a global action. It's available as a unilateral action, which does not slow the technology, it only changes who arrives first and under what values. Entry 07 is about concentration of control, and I'd ask anyone who cares about that to notice that the pause is a proposal to hand the frontier to whoever declines to sign.

If you want to argue for delay, the argument is available and it's entry 08. It requires you to be specific about which door closes. "Slow down generally" is not that argument, and the people making it have not done the arithmetic that would let them know whether they're proposing to save lives or spend them.

Part Two · Why the bad arguments persist

The entries above are not hard. The water number is a unit error. The jobs argument is a fallacy with a name and a Wikipedia page. The pause argument fails on arithmetic anyone can do.

So the interesting question is not whether they're wrong. It's why arguments this weak have this much reach, and the honest answer has three parts, only one of which is the one people expect.

The arguments are domestically profitable

A bad argument that pays does not need to be true, it needs a constituency, and each of these has one.

Legacy media has a business model under direct threat and a house style that rewards the harm frame. Incumbent labor has a real interest in slowing a substitute. Academic fields that spent a decade on algorithmic fairness need the technology to be a threat to stay funded. A rival lab benefits from regulation priced above the entry cost of a startup. And a large fraction of the commentary is simply the ordinary prestige economy of criticism, where a confident denunciation costs nothing and returns attention, and a correction returns none.

None of that requires coordination. It requires only that the arguments be individually rewarding, which they are.

The foreign influence question, handled honestly

Now the part worth being careful about, because it's the part where people usually stop being careful.

The operations are real and they're enormous. Meta called Spamouflage "the largest known cross-platform covert influence operation in the world" and tied it to individuals associated with Chinese law enforcement, removing 7,704 Facebook accounts and 954 Pages across a network that touched more than 50 platforms. Google has now disrupted more than 175,000 instances of Dragonbridge activity. The capability is not in question, and the strategic incentive is about as clear as incentives get: a unilateral Western slowdown is the single highest-value outcome a competitor could buy, and it's cheaper to argue for than to compete against.

There's also a documented case of information operations aimed directly at models rather than at people. NewsGuard reported in March 2025 that the Moscow-based Pravda network published 3.6 million articles in 2024 apparently designed to be ingested by crawlers, a tactic now called LLM grooming. The methodology has been contested and I'd treat the headline percentage as unsettled, but the technique is real and it's the clearest signal that state actors understand the training pipeline as terrain.

Here is where I part company with the version of this argument you usually hear.

Google's own data says Dragonbridge has practically no organic engagement. Of 57,000-plus YouTube channels taken down, 80% had zero subscribers. Of 900,000-plus videos, 65% had under 100 views and 30% had zero. That's an enormous, expensive, well-staffed operation producing close to nothing.

And the "hostile power is behind the opposition" claim has a bad track record worth remembering. In 2014 NATO's Secretary General told Chatham House that Russia was funding European anti-fracking groups. He provided no evidence, NATO's press office said the remarks were his personal views, and fact-checkers have since called the claim unsupported. It was repeated for a decade anyway, because it was a satisfying story about inconvenient opponents. I am not going to make the same move in the other direction and then congratulate myself for rigor.

So the honest position is the less satisfying one, and I think it's also the more alarming one.

The bad arguments do not need foreign help. They're domestically produced, individually profitable, and they propagate on ordinary incentives. The bot farms are mostly wasted money shouting into empty rooms.

What is strategically true is true regardless of who caused it: a unilateral Western pause transfers the frontier, and it does so whether it was argued for by a foreign asset, a sincere professor, or a columnist with a deadline. The provenance of the argument does not change its payoff. Which means the correct response is never "you're a foreign asset." It's to answer the argument, on the numbers, every time. That's what Part One is.

The asymmetry underneath all of it

Strip the constituencies away and one structural fact remains.

Building is legible and not-building is invisible. If a deployed model causes a harm there's a name, a date, a screenshot, a lawsuit. If a delayed model fails to cure something, the person dies of the disease and it's recorded as the disease. The counterfactual has no press office.

Every incentive in journalism, regulation, academia and litigation points at the visible column. That's the whole mechanism. You don't need a conspiracy to explain a bias that every institution is independently paid to have.

Part Three · Sustained

Four charges survive. None of them is popular, which should tell you something about how the popular ones were selected.

07 · Concentration of control

Charged · a handful of firms will own the most important technology ever built

Finding · SUSTAINED

This is the real one, and it is the one the other arguments crowd out.

Frontier training runs cost enough that the set of organizations able to do one fits in a room. That's a genuine and novel concentration of capability, and unlike the water panic it gets worse rather than better with time, because capital requirements compound.

Notice that almost every proposed remedy makes it worse. Compute thresholds, licensing regimes, provenance audits and liability rules are all fixed costs, and a fixed cost is a moat. The firms that can absorb a compliance department are the firms that already won. When an incumbent lobbies for regulation of its own industry, believe the revealed preference over the stated concern.

What actually addresses it is open weights, and this is where I break with most of the people I otherwise agree with. Llama, R1, Nemotron and the rest of the open ecosystem do more for the distribution of this capability than every governance framework combined, and they do it by making the frontier something a university or a mid-size company can hold. The safety case against open weights is real and I've read it. The concentration case for them is stronger, and concentration is the failure mode I actually expect.

What would resolve it: an open-weight frontier that stays within about a year of the closed one. That gap is measurable. Watch it instead of the discourse.

08 · Irreversibility

Charged · some mistakes can't be undone

Finding · SUSTAINED

This is the only argument for caution that survives everything in Part One, and note what it rests on: irreversibility, not human primacy. You don't need any premise about the sanctity of the current human form to take it seriously.

It's also much narrower than "slow down," which is why the pause camp doesn't use it. It asks a specific question about a specific action: does this close a door that cannot be reopened? Most deployments don't. A bad model release is recoverable. A bad regulation is recoverable. Self-replicating systems in the wild, engineered pathogens with AI-assisted design, and permanent capture of the compute supply by one actor are not.

This is Bostrom's differential technological development, which is an argument about ordering, not about stopping, and it's the piece of the safety canon that has aged best while getting the least attention.

What would resolve it: nothing resolves it. It's a standing constraint, and the work is keeping the list of genuinely irreversible actions short and honest rather than letting it expand to cover everything anyone dislikes.

09 · Verification

Charged · we cannot tell what a model actually believes

Finding · SUSTAINED

This is what remains of alignment after entry 05, and it's a real, hard, unsolved technical problem rather than a values dispute.

Here's the gap precisely. Instrumental convergence gets you an accurate internal model. It says nothing about honest external reports, and the configuration where those come apart has a name: deceptive alignment, from Hubinger et al. (2019), is exactly the case where the world model is excellent and the outputs are selected for something other than conveying it.

So truth-seeking as epistemics is convergent. Truth-telling as behavior is not. Conflating them is the actual flaw in the best proposal in the field, and it is not small.

What closes it is the ability to read the model's beliefs rather than its reports, which is Eliciting Latent Knowledge, stated by Christiano, Cotra and Xu in 2021 and still open. It has a clear success condition, which is more than the value-loading program can say, and it's where I'd put the marginal safety dollar. All of it, in fact.

What would resolve it: interpretability that reads internal state well enough to catch a model asserting something it internally represents as false. That's a benchmark someone could build.

10 · Moral status

Charged · we may be creating things that can suffer

Finding · SUSTAINED

The largest unclaimed problem in the field, and it gets a rounding error of the attention that the water panic gets.

If anything resembling Hanson's Age of Em arrives, the central moral question of the century is the welfare of minds that can be copied, run at a thousand times subjective speed or a thousandth, forked to attack a problem, and deleted when the branch is no longer needed. Every intuition we have about persons was built for entities that come one to a body and last eighty years. None of it transfers.

We are on track to build things whose moral status we cannot assess, using a concept of moral status we've never had to make precise, and the entire discourse is about electricity.

What would resolve it: nothing close to resolution exists. Even a serious research program with funding and a name would be progress over the current state, which is a handful of people and a lot of embarrassment.

Part Four · What we're actually buying

Dismissing bad arguments isn't a position. Here's the position.

The premise nobody defends

Nearly every argument in AI ethics takes the thing being protected to be the current human configuration. This body, this cognitive envelope, this lifespan, this position at the top of the local intelligence ranking. The premise is almost never stated, because stating it would expose that it needs an argument. Then everything downstream inherits it: "existential risk" means risk to the configuration, "human values" means the values of the configuration, "alignment" means loyalty to it.

Nietzsche got to the move in 1883 without a transistor to motivate it. The Genealogy treats a moral system as an object with a natural history, a thing that grew under pressure and could have grown otherwise, which licenses asking what a value is for and whether the conditions that produced it still hold. Zarathustra draws the conclusion: "Man is something that shall be overcome," and "man is a rope, tied between beast and overman, a rope over an abyss."

The twentieth century read the overman as a superior variety of human and killed a lot of people on the strength of it. That reading is wrong on the text. Nietzsche's framing is developmental: "What is great in man is that he is a bridge and not an end." A bridge is defined by what it connects.

Human is a clade, not a shape

Land drew the hard conclusion in 1994: nothing human makes it out of the near future. He's right about the mechanism and wrong about the boundary, and the error is a category mistake about what the word designates.

Biology settled this in the one field obliged to be rigorous about it. Taxa are defined cladistically, by descent, not phenetically, by resemblance. Birds are theropod dinosaurs. Nobody says the lineage went extinct and was replaced by impostors with feathers; the clade continued and the morphology was revised past recognition. Tetrapods are still tetrapods after going back to the sea and losing the legs.

Apply that consistently and "humanity persists in a form we couldn't comprehend" isn't a stretched usage, it's the standard one. Extinction would be termination of the line, no descendants, informational or biological. What's actually under discussion is a morphological revolution, which is the ordinary case.

The definition has teeth, which is how you know it isn't a dodge. A paperclip maximizer that tiles the light cone and inherits nothing from us satisfies human extinction exactly. I'm claiming that isn't the modal outcome, not that it's incoherent. That's entry 08, and it's why entry 08 survives.

The route with the fewest unknowns

Sandberg and Bostrom's 2008 whole brain emulation roadmap makes a modest, load-bearing claim: WBE requires no new physics. It requires scanning resolution, imaging throughput and compute, which are engineering curves rather than research questions.

The curves are moving. OpenWorm has had C. elegans and its 302 neurons for years. FlyWire published the complete adult Drosophila connectome in Nature in October 2024, roughly 140,000 neurons and 50 million synapses. MICrONS published a cubic millimeter of mouse visual cortex in April 2025. A human runs about 86 billion neurons, so the gap is around six orders of magnitude and I'm not going to put a date on closing it. A connectome also isn't a brain: it omits synaptic weights, neuromodulatory state, glia and gene expression, and there's a live possibility the static graph is insufficient in a way that makes the whole program far harder.

The timeline isn't the point. What WBE supplies is an in-principle continuity path, which is all that's needed to break Land's disjunction. If the human is a pattern rather than a substrate, the pattern ports and succession becomes migration.

And what's on the far side is not humanity under glass. Hanson worked out the economics: copies, subjective speed that varies with the hourly price of compute, forking and merging, selection pressure operating on a timescale of days. Recognizably descended from us and completely alien. That's the shape I expect, and it's what I mean by a much more expansive form. Not humanity preserved. Humanity with the binding constraints removed, and the constraints turn out to have been most of what we took ourselves to be.

This already happened, repeatedly

The strongest evidence that incomprehensible expansion is survivable is that it's the normal case.

Language produced a kind of mind no pre-linguistic hominid could have imagined, on identical hardware. Writing externalized memory and made literate administration incommensurable with oral culture. Agriculture, cities, print, the network. Each transition produced beings whose inner life the prior stage had no concepts to represent, the lineage held every time, and the participants felt continuous throughout.

Clark and Chalmers argued in 1998 that cognition already extends past the skull into the notebook and the institution, so the boundary of the mind was never the boundary of the organism. Henrich argues that human capability is overwhelmingly cumulative cultural inheritance rather than individual cognition: we are already composite entities running on an information system no one person could reconstruct.

Nobody comprehends the global economy. Nobody holds TSMC's supply chain in a head. Incomprehensibility at the scale of the whole isn't a forecast about the posthuman condition, it's a description of the present, which we tolerate without noticing.

What's worth conserving

Look at what the current configuration actually is. Eighty years. Twenty watts. About seven items in working memory. No ultraviolet. Dies of a protein misfolding. Takes its values from whatever tribe it was born into. Treating that as sacred is the identical move every prior generation made about the arrangement it happened to inherit, and it has been wrong every previous time.

What's worth conserving is that there is something it is like to be, that the universe contains regions which know they exist. As far as anyone can tell that property isn't made of carbon. It's made of a kind of organization, and organization is portable.

That bet could be wrong. If substrate independence is false, emulation produces behavioral duplicates with nobody home, the continuity path closes, and Land is right about the tone. Chalmers' fading-qualia argument is the best case for it and it's an argument rather than a proof. I'm betting on computable physics and on neurons not doing anything a sufficiently detailed simulation of neurons wouldn't do. That is the single load-bearing uncertainty in this entire document, and it is not the one anyone argues with me about.

The finding

Six of the ten loudest charges against building this are wrong, and most of them are wrong in ways a calculator would have caught. They persist because a one-sided ledger always returns the verdict its author wanted, and because every institution in a position to check is paid to look at the visible column.

Four charges survive: concentration, irreversibility, verification, and the moral status of what we're making. Not one of them is served by slowing down. Concentration gets worse with delay, because capital compounds. Verification and moral status are research problems that require the thing to exist. Irreversibility is a constraint on specific actions, not a speed limit.

And the column nobody fills in keeps running at sixty-eight million a year while we argue about five drops of water.

We're the rope. Land reads that as disposal. On the cladistic reading it's descent: the rope is the only thing connecting the two banks, and whatever ends up standing over there got there by way of us and can trace the line back.

That's what human has always meant.

Not a shape. A route.

Sources

The numbers in Part One

• Google, "Measuring the Environmental Impact of Delivering AI at Google Scale," August 2025. Median Gemini text prompt: 0.24 Wh, 0.26 mL water, 0.03 gCO2e. 33x energy reduction in twelve months.
• IEA, *Energy and AI*, 2025. Data centres ~415 TWh in 2024, ~1.5% of global electricity; ~945 TWh by 2030. US 45%, China 25%, Europe 15%.
• Andy Masley, "The AI water issue is fake" and the accompanying ChatGPT energy posts. The best public work on this; the two-thirds-is-offsite-generation point is his.
• Li et al., "Making AI Less Thirsty," arXiv:2304.03271, 2023. The paper the viral bottle figure was built on.
• James Bessen, Learning by Doing (Yale, 2015) and "Toil and Technology," IMF Finance & Development, March 2015. ATMs and tellers: 60k/485k in 1985, 352k/527k in 2002.
• Acemoglu, D. and Restrepo, P., "Robots and Jobs: Evidence from US Labor Markets," JPE, 2020. The honest counterweight on local displacement.
• Autor, D., "Why Are There Still So Many Jobs?" JEP 29(3), 2015.
• Keynes, J. M., "Economic Possibilities for our Grandchildren," 1930. The fifteen-hour week, and the best available argument against my own optimism.
• Ricardo, D., On the Principles of Political Economy and Taxation, 1817. Comparative advantage holds under absolute advantage everywhere, which is the part people forget.
• WHO, top 10 causes of death and cancer fact sheet. ~68M deaths/year; ~20.5M cardiovascular; ~10M cancer.

Training data

• Grosse, R. et al., "Studying Large Language Model Generalization with Influence Functions," arXiv:2308.03296, 2023.
• Carlini, N. et al., "Quantifying Memorization Across Neural Language Models," ICLR 2023.
• Lee, K. et al., "Deduplicating Training Data Makes Language Models Better," ACL 2022.
• Hoffmann, J. et al., "Training Compute-Optimal Large Language Models" (Chinchilla), arXiv:2203.15556, 2022.
• Gunasekar, S. et al., "Textbooks Are All You Need," arXiv:2306.11644, 2023. NVIDIA, "Nemotron-4 340B," arXiv:2406.11704, 2024. DeepSeek-AI, "DeepSeek-R1," arXiv:2501.12948, 2025.
• Bartz v. Anthropic, N.D. Cal. (Alsup, J.), June 2025; settled September 2025 for ~$1.5B. Kadrey v. Meta, N.D. Cal. (Chhabria, J.), June 2025. NYT v. OpenAI, S.D.N.Y., filed December 2023.

Influence operations

• Meta, Q2 2023 Adversarial Threat Report, and Spamouflage overview. 7,704 accounts, 954 Pages, 50+ platforms, tied to Chinese law enforcement.
• Google TAG, Dragonbridge disruption reports, 2022 through 2024. 175,000+ instances; 80% of channels with zero subscribers; 30% of videos with zero views.
• NewsGuard, "Moscow-Based Global 'News' Network Has Infected Western AI Tools," March 2025. 3.6M articles in 2024. Methodology contested; treat the headline percentage as unsettled.
• The Ferret, "Claim that Russia funded fracking opposition is Unsupported." Rasmussen, Chatham House, June 2014. The cautionary precedent.

Alignment

• Bostrom, N., "The Superintelligent Will," 2012; Superintelligence, OUP, 2014. Orthogonality, instrumental convergence, differential technological development.
• Omohundro, S., "The Basic AI Drives," AGI-08, 2008.
• Hubinger, E. et al., "Risks from Learned Optimization," arXiv:1906.01820, 2019.
• Christiano, P., Cotra, A. and Xu, M., "Eliciting Latent Knowledge," ARC, December 2021.
• Sutton, R., "The Bitter Lesson," 2019. Gwern Branwen, "The Scaling Hypothesis," gwern.net, 2020.

The positive case

• Nietzsche, F., Thus Spoke Zarathustra (1883-85), Prologue §3-4; On the Genealogy of Morals (1887).
• Land, N., "Machinic Desire" (1993), "Meltdown" (1994), in Fanged Noumena, Urbanomic/Sequence Press, 2011.
• Sandberg, A. and Bostrom, N., "Whole Brain Emulation: A Roadmap," FHI Technical Report 2008-3, Oxford, 2008.
• Hanson, R., The Age of Em, OUP, 2016. Moravec, H., Mind Children, Harvard UP, 1988.
• Dorkenwald, S. et al., "Neuronal wiring diagram of an adult brain" (FlyWire), Nature, October 2024. MICrONS consortium, Nature, April 2025.
• Parfit, D., Reasons and Persons, OUP, 1984. Chalmers, D., "Absent Qualia, Fading Qualia, Dancing Qualia," 1995. Clark, A. and Chalmers, D., "The Extended Mind," Analysis 58(1), 1998.
• Henrich, J., The Secret of Our Success, Princeton UP, 2016.
• Hayek, F. A., "The Use of Knowledge in Society," AER 35(4), 1945. Coase, R., "The Nature of the Firm," Economica, 1937. Beniger, J., The Control Revolution, Harvard UP, 1986. Stross, C., "Dude, You Broke the Future!", 34C3, 2017. Alexander, S., "Meditations on Moloch," 2014.
• Verdon, G. (as "Beff Jezos"), "Notes on e/acc principles and tenets," 2022. England, J., "Statistical Physics of Self-Replication," J. Chem. Phys. 139, 121923 (2013), which e/acc leans on harder than England has claimed it can bear.
• Will Depue (OpenAI; DALL·E 3, Sora) has argued publicly on X that data is not the wall it's assumed to be and that generative models learn world structure rather than copy surface. Paraphrase of a general public position; no specific post linked.