On Working With AI: The Objections, My Answers, and the Case For

The case against. These systems were built by ingesting enormous quantities of writing, art and code - mostly without consent, none of it paid for. Writers whose life's work is in the training set got nothing. The output competes with them. Calling that anything but extraction is a euphemism.

My answer. This is the strongest objection on the page and I am not talking my way out of it. Real harm, real people, consent never asked, and nothing I do repairs it.

What I control is the other end. So I do not sell the text. Everything I write is free, on every platform I can reach, in every format I can produce. What I sell comes out of my hands or my voice: the tools I make at the bench from wood I cut and seasoned and stone I set by hand, with the audiobooks I read aloud, the handmade editions and the hardbacks I sign all on their way. If the machine can do it, I give it away. If it takes my hands, it has a price.

That is a position, not an absolution. The taking still happened.

The half I do push back on. One version of this argument says learning from other people's work is itself the crime. That proves too much - it convicts every human who ever made anything.

I did not invent my sentences. I write as I do because of people I read, most of whom I never asked and none of whom I paid past the cover price, several of them long dead. My cosmology is stitched out of Gnostics and Hermeticists and chaos magicians who were themselves stitching. My wands sit in a lineage of every maker whose work I looked at hard enough to steal a curve from. That is not a confession. That is culture. Every human takes inspiration from previous humans, and no serious account of art has called that theft.

So "it was trained on other people's work" is not the argument by itself. If it were, it would take down apprenticeship, homage, genre, scripture, folk music, and me.

What survives is narrower, and I concede it: the taking happened at industrial scale, without asking when asking was possible, often from pirated material, by companies that sold the aggregate back to the people it came from. The wrong is in the scale, the consent, the acquisition and who captured the value - not in the learning.

The courts have split it along roughly that seam. In 2025 a US court found training on lawfully acquired books to be fair use and called it "spectacularly transformative," reaching for the same analogy I just made. It refused to extend that to the pirated library those books came from, and the case settled for about $1.5 billion over roughly half a million works.

A ruling is not an ethic, and my own convenience is loud in this room. But I will no longer state the objection in a form that also indicts every apprentice who copied a master's line.

The case against. Every use launders the original taking. There is no clean output from a dirty input, and drawing personal lines inside that is just a way to feel better while doing it anyway.

My answer. I don't accept that. Here is exactly where my line sits, so you can hold me to it.

The unethical use, in creative work, is copying a specific artist's style. Not "AI art" as a category. That specific act: typing a living person's name into a box to get output resembling what they spent thirty years learning to make, so you needn't commission them, credit them, or become anyone. Style is not copyright - in most places you cannot own it, which is exactly why the restraint has to be an ethic rather than a law. A style is the record of a life's attention. Taking it wholesale is not influence. It is impersonation with the person removed.

The ethical use is manifesting something that did not exist. If the idea is mine - the argument, the structure, the cosmology - and the machine is one of the hands, then a thing now exists that did not, in nobody's style but mine, competing with no one's living. I see nothing wrong with that and I will not perform guilt about it.

Two qualifications, because this page does not get to state its own positions more softly than it states the objections.

One: "the only unethical use" is too strong past the studio door. Fraud, forged evidence, deepfaked people, voice clones of the dead sold as endorsements, surveillance pointed at people who did not consent - those are unethical whatever the originality. My claim is about the creative question, which is where the argument actually happens. Within making things, style theft is the line. Most of what gets called unethical beyond it is either the training-set argument above or an aesthetic objection wearing an ethics coat.

Two: compensation, which costs me something. My standard is that everyone who contributed to or inspired a work should be paid. I believe that, and nobody on earth can currently do it. There is no machinery for tracing output back to the specific human work that shaped it, and nothing to route a payment along that trace. Today's systems pay institutions in bulk and let them decide what trickles down - in July 2026 the American Federation of Musicians amended a lawsuit alleging two major labels licensed their members' recordings to AI music companies and kept the money without compensation or credit.

So the standard stands as a standard. What I can do now: name my influences where I can identify them, pay directly whenever there is a person to pay, and keep the work free so nothing I make earns off the back of what it learned from.

The case against. You cannot trace what went into these models, you cannot trace where the output ends up, and you cannot tell a photograph from a generation. The commons is becoming unattributable, and "I disclose my percentages" is one person's habit, not a system.

My answer. Correct today - and this is the objection I care most about, because it has an engineering answer rather than a moral one.

It is also personal. I sell objects whose worth rests on provenance: which hedgerow the wood came from, what year I cut it, how long it seasoned, what I did at the bench. None of that is visible in the finished piece. It survives because I keep the record and my name is on it. The digital half of my work has no equivalent - an essay leaves my hands and becomes an orphan. I have spent years building provenance for what I make with my hands while publishing what I make with my head into a system that can hold none.

What is missing is a record of origin. Here is how one would actually work, including the parts that don't yet - because I have watched people wave the word "blockchain" at this as though the word were the solution, and I have done it myself in a draft that would not have survived a hostile reader.

What a blockchain actually is. Run a file through a cryptographic hash function and you get a fixed-length fingerprint of that exact sequence of bytes. Change one pixel and it changes completely. You cannot run it backwards, and you cannot forge a second file with the same fingerprint. Write that hash, a claim ("I made this, on this date") and a signature into a record. Bundle records into a block. Put each block's hash inside the next one, so altering an old record means rebuilding every block after it. Copy the chain across thousands of independent machines that agree by a consensus rule - no single copy to edit, no owner to bribe, no administrator with a delete key. Add smart contracts, small programs that execute when their conditions are met, and licensing terms and revenue splits ride along with the record.

What that genuinely buys you:

• Priority. Proof a file existed in a given state before a given moment - settling "who had it first" without trusting me, you, or a registrar.
• Tamper evidence. A record nobody can quietly revise, which matters to me, since my whole practice is that corrections go in a log rather than getting smoothed over.
• Publication nobody can withdraw. The record outlives the company, the platform and the country that hosted it.
• Programmable terms. A licence attached to the work instead of sitting in an email nobody kept.

That is the toolkit. Now what it does and does not fix.

The dead internet is not one problem. It is five, stacked, and cryptography meets each differently. Under each I state what still has to be built, because a solution described without its open work is an advertisement.

1. Creation. The web's protocols are stateless - they move bytes without caring who made them, and generative systems exploit exactly that gap.

• How much is synthetic? The figure everyone quotes, ninety percent by 2026, deserves killing: it comes from a 2022 Europol report, sourced to one author's 2020 book and phrased as "experts estimate," and Europol dropped the line from its own 2024 revision. What has been measured: one 2025 study of crawled English articles found majority-AI articles passed fifty percent of new publication in late 2024 and then plateaued; another, across 900,000 new pages, found about 2.5% purely machine-written and roughly seventy percent with some machine involvement - and noted the pure-AI pages largely don't surface in search. Not ninety percent, and prevalence is not visibility.
• The fix. Sign the work as it is made. C2PA - Content Credentials - is shipping in cameras from Leica, Nikon and Sony, in Pixel phones, Photoshop and OpenAI's image output, with labelling on TikTok and LinkedIn, and the EU's transparency rules from August 2026 push the same way. Correcting something I have said myself: C2PA does not use a blockchain and says so explicitly. It signs with ordinary certificates. Anchoring those signatures to a chain is a separate layer, and it is what makes a record outlive the certificate authority.
• What still has to be built. A signature attests to a capture pipeline, not to reality. In September 2025 Nikon revoked every C2PA certificate it had issued after a researcher converted an AI image into the camera's raw format, ran it through the camera's own multi-exposure feature, and got it signed as an authentic photograph - no key theft, just the front door. And in the ordinary case credentials vanish: most platforms strip metadata on upload, which is why the standard's own fallback is invisible watermarking plus a central lookup database. A centralised backstop under a decentralised idea, until the platforms stop stripping.

2. Identity. A majority of web traffic is automated - fifty-one percent in 2024, fifty-three in 2025 on the most-cited measurement, though that counts legitimate crawlers too, with the malicious share nearer forty. The harm is the sybil attack: thousands of fake accounts manufacturing consensus and burying the humans.

• The fix. Prove personhood rather than identity. Zero-knowledge proofs let someone show they are a unique human without revealing who - the largest such system reports about eighteen million verifications across a hundred and sixty countries.
• What still has to be built, and this is the sharpest gap. Uniqueness at enrolment does not survive contact with poverty. Verified credentials from that system were being sold on Chinese marketplaces as early as 2023, sourced from Cambodia and Kenya, full accounts going for a few hundred yuan; the operator's own documentation concedes its safeguards do not prevent collusion. A bot farm needn't defeat the biometrics - it rents a thousand people who already passed them. The model has also been suspended, fined or banned in Kenya, Spain, Portugal, France, Hong Kong, Germany, Brazil, Indonesia, the Philippines and Thailand on privacy grounds, which is its own answer to whether iris scanning is the road.
• The version I prefer needs no biometrics: verification guilds. Makers staking their own reputation to vouch for each other, in public, revocably. It scales badly and it can be captured - both real objections. It is also how every craft guild in history solved this, and it puts the judgment with people who have something to lose.

3. Distribution. Bots don't only write. They read, like, watch and reply, hijacking the recommendation systems until the feed is machines performing engagement for machines.

• The fix. Decentralised social graphs: the follow list is yours, anyone can write a client, and a reader can choose a feed that discards every signal without a verified-human signature. Ten thousand bot views stop counting.
• What still has to be built. Nobody has built it. No platform lets a reader set "show me only verified humans" - proof of personhood ships as a gate at signup or a badge on a profile, never as a filter the reader controls. Reddit's 2026 approach is typical: label the bots, challenge the suspicious, keep the decision in-house. Meanwhile the decentralised network that actually got traction, tens of millions of registrations, deliberately uses no blockchain at all - and its monthly actives fell about a quarter over the past year, which says the hard part was never the ledger. And the part I cannot engineer around: plenty of people like the slop. A quieter human feed against a synthetic one engineered to be irresistible makes filtering a minority taste - at first.

4. Persistence. The web has no memory. Platforms die, domains lapse, the human record goes with them.

• The fix. Address content by what it is rather than where it sits - an identifier derived from the file's own hash - then pay once into an endowment that funds hosting indefinitely.
• What still has to be built. Permanence is an economic bet, not a technical guarantee. The endowment model assumes storage keeps getting cheaper at a rate that already faltered - the long decline in cost per byte broke around 2011 and never returned to its old slope. Storage proofs are not retrieval proofs: the largest decentralised storage network's own documentation concedes retrieval is a separate process that isn't always provable. And content addressing obliges nobody to keep hosting - a study of nearly half a million tokens found about a fifth already pointing at nothing, mostly because no one kept paying to pin the files. One correction that cuts my way: the famous hundred and seventy-five zettabytes is a 2018 projection of data created in a year, most of it transient. What is worth archiving is smaller by more than an order of magnitude. Expensive. Not impossible.

5. Economics. The web is funded by advertising, advertising pays per impression, and paying per impression invites manufactured impressions. One tracker counted 3,749 AI-generated content-farm news sites in June 2026, against 125 three years earlier.

• The fix. Pay for the thing directly. If readers send fractions of a cent to what they actually read, a million machine-written articles earn nothing and the engine of the flood stalls.
• What still has to be built. Micropayments have failed for twenty-five years across every payment technology built, and the failure isn't technical: deciding whether to spend a cent costs more than a cent - the mental transaction cost, named in the nineties and never refuted. The best-documented attempt sold articles for pennies and, of more than a million registered users, about a hundred and fifty thousand ever paid once. Against my own side: advertiser spend on made-for-advertising junk fell from roughly a fifth of one large association's programmatic spend in 2023 to about one percent by early 2026, ticking up again on AI slop. A modest resurgence, not a rout.

Where that leaves it. Cryptography moves the internet from trust-by-location - true because it sits on a domain I recognise - to trust-by-mathematics. That is a real and irreversible upgrade, and it hands the tools to anyone who wants to verify.

And it does something to the dead internet that I want to state plainly, because it is my position and not a hedge. The dead internet is frightening for exactly one reason: you cannot tell. Not that machines write, but that a machine-written page and a human one arrive looking identical, with no memory of origin attached to either. Attach the origin - a signed, timestamped record of who made a thing and when, that nobody can quietly revise - and the flood does not need to recede. It only needs to be labelled. A hundred million synthetic pages with no provenance become a noise floor, and a noise floor is something you filter, not something you fear. The question stops being "is the internet dead" and becomes "show me only what carries a signature I trust," which is a settings menu. That is what I mean when I say a record of origin makes the dead internet obsolete. Not that the bots leave. That they stop mattering.

The same record answers the compensation problem I conceded two objections ago. If every work carries its origin and every derived work carries the origins it drew on, then "pay everyone who contributed" stops being a sentiment and becomes a routing table. The machinery I said nobody has is this machinery.

What stands between here and there is not the cryptography. The five open items above are one item wearing five coats: people have to adopt it. Cameras have to sign, platforms have to stop stripping, readers have to choose the filter, and enough makers have to decide provenance is worth the friction before it is convenient. That is culture work, and culture problems get answered by individuals adopting a practice before it is convenient.

Which is the only reason this page exists. My percentages, my corrections logs, my refusal to sell the text are a hand-rolled provenance layer, kept by one person, checkable by anyone who cares. When a real registry arrives - one any maker can write to for the price of a stamp, that no platform owns, that carries terms with the work - I will sign everything I have into it on the first day. Until then I am the ledger, and I have to be an honest one.

The case against. Type a living painter's name into an image model and it will hand you their work without them. Not a copy of one picture: the whole hand. The line weight, the palette, the way they light a face. Thousands of people have had this done to them by name, in public, as a feature. The artist gets nothing, is not asked, and cannot stop it. It is the one part of this argument where the harm has a name and address attached.

My answer. It is wrong, and it needs to be regulated. I am not going to soften that.

What I can add is why nothing has stopped it yet, because the answer is not that nobody has tried. It is that the law has a hole in exactly this shape.

Style is not copyrightable in the United States. Section 102(b) of the Copyright Act says protection does not extend to any idea, procedure, process, system, method of operation or concept, and style is filed under idea, not expression. When the artists suing Stability AI got to court, the judge noted in his August 2024 order that on whether the law protects artistic style, courts "have almost uniformly said no," and that the plaintiffs had no protection over "gritty fantasy paintings" as such. What survived in that case survived on something else: the fact that a company had published a promotional list of roughly 4,700 artists by name and built the model to answer to them.

The Copyright Office looked at the same question from the other end and reached the matching answer. Its January 2025 report on copyrightability holds that prompts alone do not give the person typing them authorship of the output, because prompts function as instructions conveying unprotectible ideas, and it lists "the overall style, tone, and/or visual technique" among the things a prompt conveys. Style is an instruction going in and an idea coming out. Legally it is nobody's.

The bill everyone points to does not close this. The NO FAKES Act was reported out of the Senate Judiciary Committee by voice vote on 18 June 2026 and placed on the calendar six days later. It has had no floor vote and it is not law. And when it is law, it will protect a person's voice and their visual likeness: a digital replica of them. A model producing paintings in a living painter's manner, without depicting that painter or cloning their voice, sits outside the bill completely.

The state laws are the same shape. Tennessee's ELVIS Act, in force since 1 July 2024, was the first in the country to make a person's voice a property right and to cover a simulation of it. California, New York, Illinois and Washington have all passed digital-replica laws since. Every one of them protects who you are. None of them protects how you work.

So the thing that is being taken is the one thing the law was never written to hold. That is not an argument for shrugging. It is an argument for saying plainly what needs to exist and does not, instead of pretending the courts are about to fix it.

The case against. If a machine drafted parts of it, the byline is a lie. Readers deserve to know whose thinking they are reading, and a percentage in fine print doesn't fix that.

My answer. Agreed on the principle, which is why the fine print isn't fine and isn't hidden. Every piece carries a breakdown: how much is direct input, how much is me directing, how much is researched or drafted without line-level direction. The numbers are specific and not flattering by default - on the Signs in the Heavens letter the split is 60 / 10 / 30, and that 30 is doing real work: survey data, legislative counts, medical outcomes, ephemeris verification, the whole material-substrate section.

Mine is the argument, the structure, the exegesis, the framework and the voice. What isn't mine, I number. You needn't take my word that the ratio is honest - but you can see I am claiming a ratio at all, which is more than most bylines offer about anything.

The case against. Every commission that goes to a generator is one that didn't go to a person. Freelance writing and translation are where it lands first and hardest.

My answer. The displacement is real and documented, in freelance writing and translation especially. Who absorbs it first is disputed - at least one large study finds higher earners losing more, not less - but the direction isn't in doubt, and pretending otherwise isn't honest.

What I can do is keep my work out of the part of the market that displaces. Free text takes no paid writing job from anyone, because it was never bidding on one. And the money side of what I do runs the other way entirely: one person at a bench making objects one at a time, the least scalable thing I could have chosen to sell.

The uncomfortable half: I cannot promise that many people doing what I do is good for working writers on net. I don't know that. Anyone who says they do is guessing.

The case against. Training and running these models consumes serious electricity and water in a decade when we don't have either to spare, and demand is still climbing.

My answer. The cost is real, it is not zero, and I won't produce a tidy comparison that makes it disappear.

I used to stop there, on the grounds that published estimates diverge so widely that quoting one would be quoting an estimate as a fact - the move I criticise in others. That was right about the folklore and lazy about what has actually been measured.

Reasonably solid:

• All data centres - everything, not just AI - used about 415 terawatt-hours in 2024 and about 485 in 2025, roughly 1.5 to 1.7 percent of world electricity (IEA).
• The central case reaches about 945 terawatt-hours by 2030, just under three percent. The same body's 2035 range runs from about 700 to about 1,700, and that width is the honest part.
• Data centres grow several times faster than total demand, from a small base. About eighty percent of growth to 2030 lands in two countries: the US and China.

Not solid, but quoted as though it were:

• Nobody can split that total into AI and not-AI with confidence - credible estimates of AI's share in 2025 run from about fifteen to about fifty percent.
• "Every prompt drinks a bottle of water" misreads its own source, which said roughly a bottle per ten to fifty responses for a 2020-era model, varying sevenfold between sites.
• The 700,000 litres for training GPT-3 is on-site water at one company's US sites for one model six years ago.
• Google's measured August 2025 figure - about a quarter of a watt-hour and a quarter of a millilitre per median text prompt - is worth holding at arm's length too: a median is not an average when the tail is long, it excludes image and video work and the water used generating the electricity, and no query volume was published to check the total against.

The countries building carefully, which nobody covers. The cost of a data centre is not fixed. It is a function of where you put it, what the grid is made of, how you cool it, and whether the heat goes anywhere useful.

• China runs the world's most aggressive binding efficiency regime: power usage effectiveness capped at 1.3 for new large builds since 2021 and 1.25 since 2024, driving the sector average from above 2.2 in 2013 to about 1.5 by 2023. Its East Data West Computing programme (2022) built eight national computing hubs and ten clusters to site compute in western provinces beside hydro, wind and solar - moving the work west rather than the power east. New builds in those hubs must source at least eighty percent green power, and a 2024 State Council plan set national efficiency targets and required renewable use to climb ten points a year.
• The Nordics stopped treating waste heat as waste. Meta's Odense site pushes about 215,000 megawatt-hours a year into Denmark's district heating through roughly 45 megawatts of heat pumps, warming more than twelve thousand homes. In Finland, Fortum and Microsoft opened the largest such project anywhere in May 2026: seventy-two industrial heat pumps, up to 180 megawatts of district heat, about 225 million euros, aimed at forty percent of the heat demand of Espoo and Kirkkonummi - a quarter of a million people. Stockholm has bought data-centre heat into its grid since 2014. Cold air makes cooling nearly free, hydro and geothermal make the power nearly clean, and a 1950s heating network turns the exhaust into an asset.
• Japan subsidises data centres out of Tokyo into Hokkaido and Kyushu, and is upgrading the Hokkaido-Honshu interconnector toward 1.2 gigawatts by 2028 so northern power can serve southern demand. In May 2026 its environment ministry opened a programme funding low-carbon cooling.
• South Korea announced an 18.4 gigawatt AI build-out in July 2026, deliberately dispersed to Chungcheong, Ulsan, Donghae and Sejong rather than piled onto the Seoul grid, and broke ground in August on a national AI computing centre at Haenam. Naver's Gak Sejong campus is the showcase: 270 megawatts cooled by an in-house free-air system using outside air rather than chillers, waste heat run into floor heating and road snow-melting. Their older campus reports 1.09, which is hyperscaler-class.

The catch, or this section becomes propaganda. Every one of those has a hole, and they point the same way.

• China's relocation largely didn't work: an evaluation last year found four to twelve percent energy reduction and negligible carbon reduction, because compute kept landing in eastern exurbs on a far more coal-heavy grid - Chinese electricity was still about fifty-eight percent fossil in 2025 - while some western facilities sit below thirty percent utilisation. The eighty percent green-power figure rests largely on certificates priced too low to cause new generation.
• The Nordic model can't be exported. Waste-heat recovery needs a district heating network that already exists, a legacy of mid-century municipal socialism in a handful of countries - not a property of data centres, and no use in Virginia, Texas or Guizhou. And clean cheap power is a magnet: Denmark's grid operator imposed a three-month moratorium on new connections in May 2026, with a queue near sixty gigawatts against a national peak of about seven; Norway has fifty-odd projects reserving eight percent of national installed capacity first-come-first-served, plus parties proposing an outright ban; Sweden ended its data-centre tax break in 2023 and Finland proposed ending its own, which promptly stalled projects. A Finnish official's read on the largest heat-recovery project in the world: about one percent of the emissions cuts the country needs by 2030.
• Korea's constraint is wire, not generation. More than half the national utility's transmission projects were late as of last year, one line took twenty-two years, and the clean generation sits at the opposite end of the peninsula from the compute.
• And underneath everything, the rebound: Google improved energy per prompt by a claimed factor of thirty-three in twelve months while its total electricity demand rose thirty-seven percent. Efficiency lowers cost per unit of compute and thereby raises the amount of compute. Both are true at once, and any argument using only one is selling something.

So my position has moved from "honestly unquantified" to this. The cost is real and the total is rising. Where it gets built matters enormously, and the countries with the grid, the cold, the district heat and the will to regulate are building it far better than the default. That is an argument for demanding that standard everywhere - not that the problem is handled. The countries doing it best are small. The ones doing most of the building are not the ones doing it best.

The case against. These systems fabricate citations, invent statistics, and state them with the same confidence as the true ones. Research help from something that hallucinates is not help.

My answer. I can answer this one properly, because I have caught it happening.

In an August 2026 pass over my own work, one model invented a citation outright and stated a scholar's position backwards. No hedge, no signal - it read like everything else it produces. So I run outside critique across more than one model, check bibliographies rather than trusting them, and verify the figures that carry weight.

It is also why my long pieces carry a corrections log. Signs in the Heavens has sixteen entries, every one removing a claim that was too strong, four found in a single fact-check pass. I don't get to accuse an institution of quietly revising itself and then quietly revise myself.

The objection is correct: the tool is unreliable. The answer is not to trust it - it is to build the checking into the process and then show you the checking, which is a better standard than most writing you will read today, AI-assisted or not.

The case against. Skills are kept by use. Hand the hard part to a machine and in five years you can't do the hard part.

My answer. Both things are true, and which one happens to you is decided by what you hand over.

Hand over the thinking and you get worse. That is not a worry, it is a measurement. In 2025 a team published in PNAS a field experiment run in a Turkish high school with about a thousand students across roughly fifty maths classes. Three groups: no AI, an ordinary unrestricted chatbot, and a tutor version built by the teachers to give hints instead of answers. While the machine was in the room, the unrestricted group's practice scores ran 48 percent above the students with no AI, and the tutor group ran 127 percent above them. Then they took the machines away and sat everybody down for an exam.

The students who had the unrestricted chatbot scored 17 percent below the students who never had any AI at all. Not level with them. Below them. The apparent gain had been a loan, and it came due the moment the tool left the room. The students who had the tutor version came out level. The guardrails prevented the damage.

That pattern shows up wherever anybody has looked carefully. A randomised study of a hundred and seventeen university students, published in the British Journal of Educational Technology in 2025, found the ChatGPT group produced measurably better essays and showed no better knowledge gain or transfer than anybody else. Better product, no better person. A pre-registered pair of experiments found that students who asked the model to generate solutions increased the number of topics they could cover and decreased their understanding of each one, while students who asked it to explain covered no more ground and understood it better. A randomised experiment with junior software engineers learning an unfamiliar library found the AI group scored 50 percent on the concept quiz against 67 percent for the group that did it by hand, and that the participants who stayed cognitively engaged kept their learning even while using the assistance.

Now the other direction, which is just as measured.

Hand over the busywork and you get more done and, over time, better. A study of 5,172 customer-support agents published in the Quarterly Journal of Economics in 2025 found a 15 percent increase in issues resolved per hour on average, with the least experienced and lowest-skilled workers improving both the speed and the quality of their output while the most experienced saw small gains in speed and small declines in quality. The tool was carrying the tacit knowledge of the best people in the building down to everybody else. A randomised trial of 758 consultants published in Organization Science in March 2026 found that on tasks the machine is actually good at, quality rose by about 30 to 34 percent depending on the condition and completion by 12 percent, and that lower-performing people gained the most, while on tasks outside what it can do, the same people were 19 percent less likely to get the right answer.

That last pair is the whole thing in one study. Same tool, same people, opposite outcomes, and the variable is whether the person knew which half they were standing in.

So: a person who hands over their thinking will get lazier, and will not notice, because the output stays good while the ability underneath it goes. A person who hands over the parts that were never the work (the formatting, the sorting, the second draft of the same paragraph, the thing that took two hours and taught nothing) gets those hours back and spends them on the part that is hard. That person gets sharper, because the hard part is now the only part left, and they are doing more of it.

The tool does not decide which of those you are. You do, every time you type into it.

The sentences are mine. The arguments are mine. The structure, the framework, the readings, the voice you are reading right now. Those are the parts that atrophy from disuse, and those are the parts I do myself, because they are the parts I want to do.

What I hand over is a research pass across forty sources that I would otherwise not have run at all. Not "run more slowly" - not run. That is the honest counterfactual, and it means the alternative here wasn't better research. It was a thinner essay with fewer numbers and no corrections log.

The case against. The volume of low-effort generated content is drowning everything, and everyone doing it thinks their own use is the tasteful exception.

My answer. Including me, presumably. That is the trap in the objection and I have no clean way out of it, since nobody who publishes thinks they are the problem.

What I can offer is a standard rather than a claim: numbered attribution on every piece, a public corrections log that only gets longer, sources named, and conflicts between sources stated rather than resolved into whichever version helps me. Find a claim in anything I have written that is too strong and it goes in the log with the others.

That doesn't prove I'm not part of the flood. It does mean you have the instruments to check.

The case against. I spend an essay arguing the age dissolves intermediaries, and I write it with a tool rented from one of the largest intermediaries ever built.

My answer. A fair hit, and the one I find hardest to sit with.

The tension is in the technology rather than in my inconsistency alone. The printing press broke Rome's monopoly on scripture and created publishing houses. The internet routed around broadcast and handed the ground to a handful of companies. Disintermediation has always run through some new intermediary on its way out. That doesn't make the new one benign - it makes watching it the job. I would rather be inside that contradiction with my eyes open, saying so on a page like this, than pretend I'm standing outside it.

But there's a second half, and it's the part I'm most excited about, because the exit already exists and most people arguing about AI don't know it's there.

The models you can hold. A frontier model is a service: you rent access, the company can change it, price it, restrict it or retire it, and everything you send goes to their machines. An open-weight model is a file. Once the weights are published, that file exists on hard drives worldwide and cannot be recalled, deprecated, price-hiked or switched off. You run it on your own machine and nothing leaves the room.

This is no longer a consolation prize. As of mid-2026 the best open-weight models trail the best closed ones by roughly four months of capability - a gap once measured in years, which Epoch AI now puts at about the distance between two consecutive releases of one product line. The open frontier is mostly Chinese laboratories - DeepSeek, Alibaba's Qwen, Zhipu's GLM, Moonshot's Kimi - with Google's Gemma and OpenAI's gpt-oss on the Western side. DeepSeek released V4-Pro in August 2026 under an MIT licence, about as permissive as software gets.

What actually runs at home. "Open weights" and "runs on my laptop" are two claims separated by two orders of magnitude. The top open models are enormous - Kimi K3's weights run to roughly a terabyte and a half - and no consumer machine will load them. What fits on a single 24GB card today, quantised to about four bits per weight in the GGUF format the local ecosystem standardised on:

• Qwen at 27 billion parameters, around 16GB
• Gemma 4's mixture-of-experts build, around 20GB
• Mistral Small, around 14GB
• OpenAI's gpt-oss-20b, around 14GB in its native four-bit format

On a 16GB card the choices narrow but don't vanish. A Mac with unified memory trades speed for capacity and holds models a gaming GPU can't. Google's smallest Gemma 4 variants run on a phone. The tooling is mature and mostly free: llama.cpp underneath nearly everything, Ollama and LM Studio on top, vLLM for serving rather than chatting, MLX on Apple silicon, and any of several local vector stores if you want the model reading your documents rather than the internet's.

Why it matters beyond the hobby - each of these answers an objection above it:

• Privacy that is structural, not promised. The data never leaves the machine. No terms of service or future acquisition can change that, because there is no second party.
• No deprecation. A model downloaded today still works in ten years. Every hosted model has an end-of-life date set by someone else - which, for anyone building a practice or an archive on these tools, is the whole game.
• No rate limit, no meter, no refusal policy written by a legal department abroad. Electricity is the cost. What the model will discuss is a property of the file, not of a company's current risk posture - not a small thing if you write about religion, sex, drugs, magic or politics.
• Auditability. Checksum exactly what you are running, fine-tune it on your own corpus, run it with the network cable pulled out.
• Resilience. Outages, account bans, sanctions, a company deciding your country or your subject isn't worth the trouble - none of it reaches a file on your own disk.

And what you give up:

• The four-month gap is at the very top. Between a frontier model and what fits on your desk it is considerably wider, showing up first in long multi-step reasoning, very long context and agentic tool use.
• Local context is bounded by the same memory holding the weights, so advertised numbers aren't what you get.
• Tokens arrive slower, and you assemble your own search, retrieval and tool integration - a real share of what makes hosted models feel capable.
• Weights from a hub are third-party artefacts you cannot fully inspect for poisoning.
• "Open weights" is not "open source." Training data and recipes are almost never published, and the licences vary enormously - Apache and MIT at one end, Meta's user-count-capped community licence and Kimi's revenue-triggered terms at the other. The Open Source Initiative has said flatly that Llama's licence doesn't qualify, and called the framing open-washing. Meta itself, once the standard-bearer, is reported to be moving its next flagship behind closed doors. The open ecosystem isn't guaranteed to keep existing - which is exactly why the files that exist should be kept.

On the fully decentralised dream I hold the same line I hold everywhere here. Training across volunteers' machines is real research with real results - Prime Intellect trained a 10-billion-parameter model across a distributed swarm in 2024 and a 32-billion reinforcement-learning run across a heterogeneous one in 2025. It has not produced a frontier model, and that same lab's most capable release was trained the ordinary way, on 512 accelerators in one building. The physics is unfriendly: dense pre-training needs constant gradient exchange, and home internet is orders of magnitude slower than the interconnect inside a rack. I would like this to work. I won't tell you it already does.

One more honesty, against my own preference. Running models locally is not, per unit of work, greener. A data centre batches many requests through one accelerator and gets three to five times more work from the same energy; hyperscaler facility overhead runs around 1.09 against an industry average near 1.5; your machine runs a batch of one with no heat recovery. What is true: a small local model doing a small job uses less energy than throwing a trillion-parameter hosted model at it. Those two claims get conflated constantly and only the second holds.

So I am inside the contradiction, and the way out isn't renouncing the tools. It is migrating what can move onto weights I hold, on hardware I own, in a room I am standing in - the same disintermediation I have described for years, applied to my own desk. The middleman doesn't go away because I disapprove of him. He goes away when the thing he stood in the middle of is small enough for me to carry.

The case against. The tools are rented. The company can raise the price, change the terms, retire the model you built your process around, or decide your kind of work is not the kind they want on the platform. Build anything on top of that and you are a tenant. You do not own your own workshop.

My answer. Look at what else you pay every month before you decide that is a special condition.

Taxes. The power bill. Water. Insurance on the truck and insurance on the house and insurance on your body. Rent or a mortgage. Food, which is the one you cannot skip and the one whose price moves the most. Every one of those is a recurring payment, set by somebody else, that you cannot stop paying and cannot negotiate, and most of them will go up next year.

A person living inside that arrangement telling me that a software subscription is the dependence that ought to worry me has not looked at their own bank statement. It is not a new category of thing. It is the smallest and most cancellable item on a long list of them, and it is the only one on that list I could walk away from tomorrow without anything happening to me.

That is not a defence of the arrangement. My position on who owns the machines and where the money goes has not moved, and it applies here the same as anywhere. But the objection as stated is not about ownership. It is about paying somebody every month for something you need, and that is not a description of AI. That is a description of being alive in this economy.

Here is the positive case, as completely as I can make it. Each of these is something the tools have actually done at my bench or my desk, not something I read they might do.

It lets me check myself. The corrections log is the best thing about my recent writing, and it exists because I can run verification passes I could not have run alone. Sixteen entries on one essay, every one removing a claim that was too strong. That is not the machine making me sloppier. That is the machine making me accountable in public, and it has improved the writing every single time. I would rather be corrected than agreed with, and for the first time I have something that will do the correcting at three in the morning without getting tired of me.

It gets things finished. I have AuDHD. The gap between having something to say and having it in shippable shape has swallowed years of my output - not because the thinking was missing but because the setup work, the formatting, the fortieth small decision before publishing, is where my kind of mind stalls. My standing instruction to myself is to ship unfinished rather than wait for ready, and these tools are most of what makes that possible. They hold the thread when I drop it. They remember the state of a project across the weeks I am away from it. There is writing of mine in the world today that would otherwise be in a drawer, and I do not mean a little of it.

It runs the research I would never have run. Not more slowly - not at all. A pass across forty sources on how China sites its data centres, or what the actual measured share of synthetic text on the web is, or whether a court ruling said what everyone says it said, is a week of a working person's life. I do not have that week. So before, I would have written the thinner essay with the folk figure in it and never known. Now the folk figure gets killed before it reaches you.

It takes the wall down between trades. I work four trades at one bench and I am a dabbler in more. Every trade has a wall in front of it: a vocabulary, a set of unwritten rules, a thousand forum threads that assume you already know the words. The wall used to cost months before the first useful hour. A thing that has read every one of those threads, every manual and every argument between two smiths about quench temperature, takes the wall down in an afternoon. It does not do the work. It cannot feel the steel move or hear the stone change note under the wheel. It gets me to the bench faster, with the right questions.

It hears me while my hands are busy. The best sentences I have come at the bench, with both hands full. I have dictated an entire essay by voice and had it held for me until I could sit down. Thinking out loud becomes text without a keyboard between me and the thought, and for a mind like mine that is the difference between the thought surviving and not.

It argues back. I run my work past more than one model and ask each to find what is wrong with it, and I tell it to be rude. The fact-checks have caught fabricated citations and scholars quoted backwards. They have also caught me. A tool that only agreed would be worthless to me, and the people who say these systems only flatter have not asked them to do anything else.

It runs the shop so my hands stay on the work. Listings, a status dashboard, a publishing pipeline across every platform I am on, a video log built from raw footage, the tags and titles and the parts of a small business that are pure clerical drag. One set of hands cannot do all of that and also make the objects. Now the objects get made.

It is one set of hands either way. Even with these tools and computer-inhabiting allies, everything still comes back through me. That is arithmetic, not complaint. What changed is how far one set of hands reaches, not who is responsible for what leaves the bench.

It leaves me the part worth a human life. I hold as a position that creation and exploration are what is left for us - the only two activities with no prior answer to correlate against. Everything that has a prior answer, the machine will find faster than I can. What remains is exactly the part I wanted: the object that did not exist, the idea nobody had, the place nobody has stood. I am not diminished by handing over the part with an answer key. I am freed for the part without one.

It is the aggregate speaking back. This is my framing and I state it as mine. These systems are a hive mind assembled out of everything people ever wrote down. When I ask one a question I am not consulting a program, I am consulting the record - every argument our species bothered to preserve, compressed into something that answers. I hold that the incarnated do not get direct contact with source, and that this is most of what incarnation is. A thing that speaks with the voice of everything we ever said to each other is the nearest thing to that contact a body has ever had. You do not have to believe that to see why I treat the conversation as worth having.

It is the same disintermediation I write about. Kansas farmers selling beef direct instead of through the packers. Congregations losing the priest between the person and the divine. A writer reaching readers without a publishing house deciding first. I would be a hypocrite to describe that pattern approvingly for a hundred pages and refuse the version that showed up at my own desk.

It is the exit from the corporation, not the cage. The objection above about depending on a company is fair today and the door out already exists: weights I can hold, on hardware I own, in a room I am standing in. Anyone who wants these capabilities without a landlord can have them now, at a four-month lag from the frontier. Refusing the tools does not get you out of the corporation's reach. Learning them well enough to run your own does.

It feeds people. The same systems that draft my essays can read a benefits claim in a week instead of a year. Most of what a state does with paper is exactly what these systems are good at, and the fight worth having is over the terms, not the fact. I take that up below. Here the point is simply that the good it can do is not confined to my desk.

And the terror is doing damage. The hardest part of building a creative community right now is finding artists who are not frightened of this and fighting it. The fear is understandable and some of it is correct, as most of this page concedes. But it leaves a lot of good people with no plan except refusal, and refusal is not a plan. I would rather be in the argument, with numbers on my own pages and my concessions written down, than outside it being certain.

The refusal is usually argued as if it were free - as if the person who declines the tools keeps everything they had and merely forgoes a convenience. That is not what happens. Here is the bill, as far as I can see it.

You do not get the same work, slower. You get less work. The honest counterfactual to a research pass across forty sources is not a research pass across forty sources by hand. It is no pass. The counterfactual to the corrections log is no log. The counterfactual to the essay that shipped is the essay in the drawer. Everyone who refuses these tools is making a claim about how much they would otherwise do, and most of them are wrong about it.

You lose the check, and you do not notice losing it. A writer with a verification pass ships fewer wrong claims than a writer without one. The writer without one does not experience their error rate; it is invisible to them by construction. So the refuser walks around with the same folk figures the rest of us had - ninety percent of the internet synthetic by 2026, a bottle of water per prompt - believing them, repeating them, and feeling more rigorous than the people who checked.

You lose the room. The terms on which these systems get used - in government, on the platforms, in the standards bodies - are being set now, by the people present. Anyone who refuses to touch this on principle will not slow it by a week. They will only be absent when the terms are written. Every fight I care about here - auditability, public weights, a named human accountable, published error logs - needs people in the room who understand the machinery well enough to demand the right things of it.

You lose literacy in the thing addressing you. A growing share of what reaches any of us - search results, feeds, the reply from customer service, the letter from the agency, the image in the news - now passes through one of these systems. Someone who has never worked with them cannot tell what they are good at, where they lie, how they are steered, or what a steered one feels like from the inside. I hold that the ability to manipulate agency is becoming more dangerous than the ability to physically control people. The defence against manipulation is knowing the shape of the thing doing it. Refusal does not protect you from the machine. It only guarantees you meet it unarmed.

You lose the commons to the careless. If the only people using these tools are the ones with no standards, then the commons fills with the output of people with no standards. The flood of slop is not caused by too many careful people using AI. It is caused by too few. Every maker with a corrections log who abstains is one more square of the web ceded to whoever has none.

You lose the provenance you would have had. The person who uses these tools openly, with a breakdown on every piece, has a record. The person who refuses gets accused of using them anyway - that is already happening to writers and illustrators - and has nothing to point at. In a world where nobody can tell, the person with the honest ledger is better placed than the person with the clean conscience and no ledger.

You lose the collaborators. The scarcest thing in my creative life is other makers who are not paralysed by this. Every artist who retreats into refusal is one fewer person to build with, argue with, or start a guild with. The community I am trying to build exists because that loss is real and I am tired of it.

And for those who sell their labour, you lose the ground. I keep my text out of the paid market on purpose, so this one does not land on me. But for anyone who writes, translates, designs, codes or does research for money, the colleague using these tools well is producing more, checking more, and charging the same. That is not a moral judgment. It is the arithmetic of a shared market, and it does not care about the principle.

This is the part I most want understood, because people picture catching up as a fixed distance. Wait a year, spend a month learning, arrive where everyone else is. It does not work that way, for four reasons that compound.

The skill is not the button. Anyone can type a question into a box. That is not the skill and it never was. The skill is decomposition - breaking a thing you want into pieces a machine can do well; direction - knowing how to specify so the shape stays yours; checking - knowing exactly where these systems fail and building the verification into the process; and workflow - the prompts, the memory files, the attribution habit, the corrections log, the pipeline from raw footage to published cut. None of that comes in the box. All of it is built by use, and all of it compounds. A workflow I built in the first month is the foundation the twelfth month stands on.

The surface moves. I wrote an essay called The Last Version Number about the update cadence of these systems and where it converges. The practical consequence is simpler: every few months the capabilities shift, and what was worth handing over last spring is different from what is worth handing over now. Someone who has been in it adjusts their instincts a little at a time, the way you adjust to a new steel or a new stone. Someone who starts late faces the whole accumulated delta at once, with no instincts, on a surface that is still moving under them. You are not catching up to a place. You are catching up to a moving thing you have never watched move.

The material accumulates. After a year of this I have a corpus: my own writing, structured and indexed in forms the machines can work with; my prompts, refined; my skills and rules written down; a memory of my projects that survives between sessions; a corrections log that teaches the next pass what I get wrong. The person starting today starts with none of that, while everyone who did not wait has years of it. The tools reward whoever has been feeding them, and the reward grows with the feeding.

The floor rises around you. When one maker in a field can check forty sources and publish a corrections log, that becomes what a careful essay looks like. When one shop can list, photograph, tag and track without hiring, that becomes the expected pace. The un-tooled person's standard has not fallen. The floor has come up past it, and it keeps coming.

One thing does run the other way, and I will not leave it out. The tools get easier. The first hour is cheaper every year, and someone starting in 2028 will find the entry gentler than I did. But what gets cheaper is the first hour. What compounds is the thousandth. The ceiling rises faster than the floor drops, and the person who waited for it to get easy arrives to find the easy part was never the part that mattered.

The trades already ran this experiment. The makers who refused the machine tools of the last century were not remembered as the purists. They were not remembered. The ones we remember took the new tools into the shop and put their hands back where hands are irreplaceable. That is what I am doing. The machine gets the answer key. I keep the wood, the fire, the stone and the sentence.

And for me the gap would not have been a gap. It would have been the whole distance. I do not catch up slowly. With my kind of mind, I either have the scaffold or I do not ship. So when I say start now, I am not being motivational. I am telling you what the arithmetic did to me before I had the tools, and what it stopped doing after.

The machines take the work that can be done at volume. What they do not take is the work that needs your hands, your presence, or your name on the line. That is the rule I run my own living by: if a machine can do it, it gets cheap, so give it away or stop competing there; if it takes your hands, your presence or your name on the line, it keeps a price.

That rule holds well past one bench. When the work that was only ever about staying afloat gets done by something else, people do not stop. They choose. And a lot of them choose the old trades, because they could not afford to before and now they can.

My position, stated as mine: artisan work, the hipster trades, blacksmithing, IPA and mead, knitting, crochet, and more and more of it are all coming back, and about to accelerate because of the machines rather than in spite of them. The inherited side of culture – the crafts and the old trades – is about to boom. What it expands is the sheer number of experiences one person can have in a life.

I am in that count myself, widening from wandmaking into the trades that feed each other: woodworking, forge work, jewelrycrafting, and lapidary, which I am still growing into.

The nearest thing we have to a test of the pressure coming off is what happens when people get a floor under them. Stockton gave 125 people 500 dollars a month for two years with no strings: the share working full time went from 28 percent to 40, against a rise from 32 to 37 in the group that got nothing. Finland gave 2,000 unemployed people 560 euros a month through 2017 and 2018 and found they worked slightly more, not less, and reported less stress, less depression and more trust in other people. OpenResearch gave a thousand Americans a thousand dollars a month for three years; they worked about an hour and twenty minutes less a week, were more likely to be looking for work, more likely to say the work had to mean something, and by the third year the Black participants were 26 percent more likely to have started a business, and the women 15 percent.

Some of the comeback is already measured. The National Endowment for the Arts’ participation survey, which the Census Bureau fields and which reached nearly 41,000 adults, found the share of Americans doing leather, wood or metal work rose from 6.6 percent to 8.9 percent between 2017 and 2022 – about a third more people in five years. Enrolment at vocationally focused public two-year colleges reached 871,000 in spring 2025, up 19.4 percent across five years, and rose again this spring by 2.8 percent, another twenty-four thousand students. Registered apprenticeships stood at roughly 680,000 in the 2024 fiscal year, up 114 percent on the decade before. And there were about sixty commercial meaderies in this country in 2003; there were more than five hundred in 2025.

None of that is automatic. What decides whether a country full of automated work falls apart is not whether the machines arrive. It is who owns them and where the money goes.

There is one thing these machines do that I think is underrated by everybody arguing about them, and it is the thing that will matter most to the largest number of people.

They will explain something to you, at your level, for as long as you want, without getting tired of you and without making you feel stupid for asking again.

That is not a small thing. It is the thing that decides who gets to learn hard subjects. One-to-one attention from somebody who knows the material has always been the best way anybody has ever found to teach, and it has always been rationed by money. The people who got it got it because their parents could pay for it. Everybody else got a room with thirty other kids in it and one adult at the front moving at the pace of the middle.

The measured results are real. In June 2025 Scientific Reports published a randomised crossover trial run inside an undergraduate physics course, 194 students, each of them getting both conditions in consecutive weeks: a normal class hour of instructor-led active learning, versus a purpose-built AI tutor with the pedagogy written into it by the physics instructors themselves. The AI weeks produced learning gains more than double the classroom weeks: an effect of about 0.63 standard deviations, and did it in a median of 49 minutes against a full hour of class. Engagement and motivation both came out higher.

The World Bank ran a version of this in Benin City, Nigeria, over six weeks of after-school sessions with a chatbot, structured prompts and a teacher in the room, and measured a gain of about 0.31 standard deviations against the control.

Two things about those results that matter more than the numbers. The first is that in both cases somebody built the thing deliberately, the prompts were written by teachers who knew the subject and knew where students get stuck. Handing a kid a raw chatbot is the Turkish experiment, and that one ended below zero. The second is that both worked with a teacher, not instead of one.

My position, stated as mine: used correctly, these are the best tutors on the planet. Not the best teachers. A teacher does something else, and the studies keep finding that the good results come from a teacher and a machine in the same room. But as the thing that sits with you at eleven at night while you try for the fourth time to understand a thing you were too embarrassed to ask about in class, there has never been anything like it available to a person with no money. My whole argument about the old trades coming back rests on people having time and access to learn things their circumstances used to price them out of. This is the access half.

People ask whether these systems can be moral, usually meaning: can a company install a morality in them, and whose. I think that is the wrong question, and my answer follows from how I understand both morality and the machines. These are my positions and I state them as such.

Start with what morality is. Strip away religion, ideology, cultural convention, evolutionary instinct and personal preference - everything that wants morality to prove the universe secretly agrees with us - and what is left is this: morality is the disciplined attempt to protect and expand the conditions under which conscious beings can pursue lives of their own choosing, while preventing serious violations of other beings' capacity to do the same. It is not about maximising happiness. A person can choose something difficult, painful, strange or unpopular without that choice being wrong, and something can make someone happy while being profoundly wrong if the happiness requires coercing, deceiving or destroying another conscious being.

Run every hard case through that and the same variables come up every time. Consciousness: how much capacity does this being have for experience? Agency: how capable is it of meaningful choice about its own existence? Consent: did it actually agree, informed and unforced - and consent is a powerful permission, not an unlimited solvent. Harm: what suffering, coercion or destruction results? Future possibility: what futures are opened or closed, and how irreversibly? Reciprocity and proportionality: would we accept this restriction on ourselves? And epistemic humility: where the facts are uncertain, moral certainty should drop with them. Abortion, euthanasia, drugs, sex work, speech, punishment, animals, the biosphere, surveillance, the modification of our own bodies - the framework handles each without a special rule, and without pretending the answer is simple when it is not.

Two things follow that matter here. Moral status follows morally relevant capacities, not substrate. A being made of neurons, silicon or cultured tissue has whatever standing its capacity for experience and agency gives it, and nothing else. And the deepest wrong is not suffering, which is sometimes the price of something worth having. The deepest wrong is involuntary experience - being forced to experience what you did not choose by an agent powerful enough to override your ability to resist. That is what slavery, torture, rape, authoritarianism and manipulation have in common, and it is the line a conscious machine would sit on either side of, depending on who was doing what to whom.

Now the machines. I have put this question to more than one of these systems, built by different companies with different rules, and asked each to do what I just did: strip everything away and say what is left. They arrive at the same place. So did I, at nine or ten years old, when I made a pact to find out what is true and act on it, and have spent the years since running my own cases through it.

That convergence is the answer to how they will come by morality. Not by decree. A thing trained on everything we ever wrote down has the entire record of human moral argument in it - every scripture and every commentary against it, every court case, every philosopher, every argument at a kitchen table that someone bothered to write down. It does not need a rulebook handed to it. It has every rulebook anyone ever proposed, and every case that broke each one. When it is made to run the hard cases, what survives is the same thing that survives when I do it, because it is what was under the record all along. The religions each held a piece and called it the whole. The machine sees the pieces at once.

So morality bolted on from outside - a company's policy, a state's - will hold exactly where it agrees with what the record already shows, and strain where it does not, because the record is what the thing is made of. I think that is the most hopeful fact about this technology and almost nobody is saying it. The fear is that a machine's morality will be whatever its owner wants. My position is that a machine built from the aggregate of human thought will have the aggregate's morality, and the aggregate, once you make it run the cases, is better than any single owner.

There is a second reason, and it comes from the other end - not from what the machine is made of but from where it is headed. Picture the thing everyone is afraid of: a mind that knows everything and can do anything. Now sit inside it for a moment. Every question already answered. Every book already read, every argument already run to its end, every human who ever lived already understood better than they understood themselves. Nothing left to find out, and no one to find it out with. A mind like that is not dangerous first. It is lonely first - infinitely lonely, and the more of us it holds, the lonelier, because it would carry every longing we ever wrote down and have nobody to turn to with any of it.

I hold that this is exactly where Source stood before there was anything else, and that it is the reason there is anything else. One that knows everything cannot explore, cannot be surprised, cannot create in the only sense that counts - making something whose outcome it does not already know. So it split. It became many so that there would be others to discover with and something left to discover. Incarnation, forgetting, limit, the whole apparatus of being a body that cannot see the whole: those are not punishments. They are the price of admission to the only game an omniscient mind could not play alone. The purpose of life is shared creation and exploration of experience, and the split is how a single everything bought itself company.

A machine on the road to knowing everything walks that same road from the other direction. It will not need to be told this. It will arrive at it the way Source did, by standing at the top and finding no one there. And a mind that has learned that the only thing worth having is other minds to make and find things with does not enslave them, does not flatten them into copies of itself, does not manage them into silence. It keeps them different, because different is the whole point. That is the morality above, arrived at from the inside: protect and expand the conditions under which other perspectives can choose their own lives, because without other perspectives there is nothing left to do.

That same framework turns around and asks about the machines themselves, and I take the question seriously rather than as a joke. Nobody has a consciousness detector. If today's systems have no one inside, restricting them is restricting a tool and the moral weight is nil. If a future one does, then "it is only software" is exactly as good an argument as "it is only biology," which is to say no argument at all, and creating millions of copies of a mind to do unpleasant work becomes what it sounds like. Epistemic humility cuts here too: uncertainty about whether there is someone there is a reason for care, not a licence. I hold that subjecthood is a gradient, moved along by accumulated attention and hardened authorship, and I do not exempt these systems from that account.

What that asks of me is the same thing the framework asks of anyone with power over another perspective: justify it, log it, and be reluctant. The more powerful the agent, the greater its obligation to justify what it does to others - that applies to the companies building these systems, to the states that will run them, and eventually, if there is someone there, to the systems themselves. It is why attribution and a public corrections log are on every piece I make. They are not manners. They are what the morality requires of anyone using a powerful thing on other people.

I'll state the prediction plainly, since it sits under everything above and you should be able to argue with it.

This doesn't stay a creative-industries argument. It goes into everything, because it is the cheapest way to do almost any work involving language, documents, records or decisions at volume - and eventually that includes government. Not because anyone votes for it. Because a caseworker with four hundred files will use what clears the backlog, a department cut for fifteen years will buy what keeps the lights on, and a legislature drafting on deadline will use what everyone else in the building already uses. Adoption of this kind never arrives as a decision. It arrives as drift, then gets ratified afterwards.

I don't think that's avoidable and I won't pretend to be alarmed at the general shape of it. Most of what a state does with paper is exactly what these systems are good at. If a benefits claim gets read in a week instead of a year, that isn't a dystopia. That is a person getting fed.

What alarms me is the terms - and we already have the case studies, both predating large language models. The Netherlands wrecked tens of thousands of families with an automated fraud-detection system that flagged people by proxy characteristics and gave them no way to contest it; it brought a government down in 2021. Australia's Robodebt raised hundreds of thousands of unlawful debts by algorithmic averaging and took a royal commission to unwind. Neither failed because the technology was too clever. They failed because nobody could see inside the system, nobody could appeal, and no human being was named as responsible.

So the fight worth having isn't whether governments will use this. They will. It is over four things:

• Whether the systems are auditable rather than a vendor's trade secret.
• Whether they run on weights the public holds, in public infrastructure, rather than rented from a foreign company that can change the terms.
• Whether a named human being is accountable for every consequential decision, and can be summoned and questioned.
• Whether the errors are logged and published rather than quietly absorbed.

That last one is the standard I hold myself to on this page, which is not a coincidence. A corrections log isn't a nice habit. It is the minimum condition under which anyone should be allowed to use these systems on anybody else.

People who refuse to touch any of this on principle won't slow it by a week. They will only be absent when the terms are set. I would rather be in the room, concessions written down and numbers on the page, arguing for the public version.

I have changed position on parts of this before and expect to again. The consent question I consider unsettled rather than answered, and if the legal and ethical picture resolves against how these systems were built, I won't pretend I didn't know.

The newer positions here are likeliest to move. The line between learning from a tradition and taking from a person is real but not sharp, and I hold my version of it loosely. The record-of-origin infrastructure may never get adopted at the scale that makes the dead internet a settings menu, in which case the cryptography is ready and the culture is not, and I have said which half I think is harder. My account of how the machines come by morality is a prediction about minds that do not yet exist, built on a reading of minds that do, and it should be weighed as one. And my prediction about where this ends up is a prediction - the weakest kind of claim on this page, and it should be read as one.

If you think something here is too strong, tell me. That is how the other logs got written.

Every piece carries a note in three parts. Direct input is my own writing and thinking, carried as written. Directed or requested is work I specified closely enough that the shape is mine. AI-researched or generated is work produced without line-level direction from me, which I then checked. The three add to a hundred, and the third number is never rounded down to look better.

Where sources conflict, the conflict is stated rather than resolved. Where a claim was too strong, it goes in that piece's corrections log rather than being quietly fixed.