It has been suggested that this would make a suitable epitaph. The wording is economical, technically specific, and contains just enough passive aggression to make the mourners wonder whether they have offended the deceased.

It does, however, suffer from one fairly serious defect in documentation.

This essay is about how Hawthorne gets made, and why I think it is still my writing. I write these essays with a good deal of help from AI: models that research, models that argue with my drafts, a system that checks the prose sounds like me, and a model that draws the pictures. That raises a fair question: whether something produced that way still has an author.

My answer runs through older cases first: a photographer who needed mules and packers, an advertising agency, a law firm, the printing press, the industrial age. Then it describes the review court I built, the Moot of Doom, and what it caught when I put Hawthorne's first issue through it. The question reaches past one small magazine. Anyone who now reads a report, a letter or a news story made partly with AI has reason to want to know who stands behind it.

On September 11, 1938, Ansel Adams set off into the Sierra Nevada with Georgia O'Keeffe, several companions, fourteen mules, and four hired backcountry specialists.

The trip lasted ten days. The specialists guided the party, packed and unpacked the animals, established camp, and prepared the meals. The mules carried the food, camping supplies, and photographic equipment.

Adams returned with photographs, subsequently hand-assembled into gift albums; the one that belonged to his companion David McAlpin holds forty-eight images across forty-two pages.

The resulting work was credited to Ansel Adams. I have been unable to find evidence of a serious movement to award the mules fourteen co-photography credits. This seems reasonable.

The four specialists, however, deserve considerably more attention. Their names were Al Kay, Alvin Rhode, Robert Barnett, and Lile Pierce. Without their expertise and labor, that particular expedition would have been a rather different undertaking. Adams relied on their skills to make his own work possible.

That is an ordinary feature of artistic production, and it carries a distinction worth keeping: recognizing another person's indispensable contribution does not necessarily make that person a co-author of the resulting work.

Adams's production methods offer a more interesting case.

In the 1970s, his photographic assistant Alan Ross took responsibility for making the Yosemite Special Edition prints from Adams's negatives. Ross worked in the darkroom, interpreted negatives, produced test prints, and responded to Adams's directions about contrast and tonal balance.

Adams could approve Ross's sample prints, which then guided the finished batches, without personally performing every operation involved in producing them. Ross, himself an accomplished photographer, continued that work with Adams's authorization.

There are real questions here about artistic interpretation, technical skill, reproduction, and attribution. But the photographs remained photographs by Ansel Adams, even though the Ansel Adams Gallery does not count a Special Edition print as an original Adams photograph. The creative relationship was understood well enough to operate without a metaphysical emergency.

I admire Adams enormously: his discipline, the precision of his printing, and his ability to recognize an image in a landscape other people might simply walk through. I admire the specialists whose labor made the work possible, too. It would be a peculiar tribute to any of them to insist that Adams's authorship depended on carrying his own tents.

The distinction matters because the relationship between an artist and the means of production has never been as simple as one person doing every task. Photography is a particularly useful place to start, because the instrument itself was once controversial.

The camera could reproduce appearances with a fidelity unavailable to most painters. It prompted arguments about mechanical reproduction, artistic judgment, and whether operating a machine could count as making art.

Being able to expose a negative was never proof that someone understood composition, light, timing, or what made an image worth preserving. Nearly two centuries of photography have given us ample opportunity to investigate the distinction. Apparently we have some additional questions.

Consider the American advertising agency. Mad Men, the fictional television drama created by Matthew Weiner, is set in an industry where authorship, collaboration, professional hierarchy, and highly specialized production existed in a state of almost permanent tension.

A campaign could involve a copywriter, art director, photographer, account executive, illustrator, production department, printer, and client. The finished advertisement might be associated with the agency or its creative director. The client paid for the work. The people who produced it had different roles, responsibilities, and contractual relationships.

Nobody needed to argue that the photographer had become a copywriter merely by operating the camera.

A law firm is another familiar example. Research, drafting, procedural review, investigation, and document preparation can be distributed among attorneys, paralegals, clerks, and outside specialists. Responsibility for legal representations is governed by professional duties and applicable rules. The tools are different, and so are the obligations, but the underlying practice of dividing specialized work is hardly exotic.

An editor can improve a manuscript without becoming its author. A secretary can prepare correspondence without becoming the person expressing the opinion. A typesetter can arrange a book beautifully and still have no idea whether the argument on page forty-three makes sense. We have familiar ways of distinguishing these contributions.

I use a digital camera. I also use electricity, along with computers, research databases, image editors, software libraries, and instruments designed and manufactured by people who understood particular problems better than I do. I occasionally pay for things using a card, which does not make the payment network a co-author of my grocery list.

And when I want to find a historical document, I usually begin with an online catalog or search engine rather than a physical card catalog pointing to Dewey Decimal call numbers.

That last example deserves more than a passing joke. A catalog is an information-retrieval technology. An index is a model of how knowledge is organized. Search engines change the cost, speed, and reach of finding material. They also change which material is encountered, what is overlooked, and whose choices govern the order in which sources appear.

A better retrieval system can make research faster. It does not make the first result correct, and it does not turn the person typing the query into a historian. That last distinction will come up again.

I build things: software, musical instruments, furniture, occasionally bits of infrastructure. I have enough experience with tools to admire them, and enough scar tissue to know that possessing one is not the same as knowing how to use it.

Have you ever opened a tin can without a can opener? I have. It is a deeply unsatisfying demonstration of human ingenuity.

The can eventually opens, usually at considerable expense to the contents, the surrounding environment, and whatever instrument was unfortunate enough to become the substitute opener. The right tool is a substantial improvement.

Suppose I wanted to build a completely renewable energy infrastructure, starting with a shovel, matches, and a forest containing fewer than average bears. Given sufficient time, access to materials, and an extraordinarily generous definition of what constitutes a single engineering project, I might eventually manage it. Four thousand years seems a reasonable planning assumption.

With modern tools, manufactured components, accumulated scientific knowledge, and access to specialists, the same ambition might become a project measured in ten or fifteen thousand hours.

Those figures are illustrative, not an engineering estimate. The bear distribution is also provisional.

The serious point is that those hours depend on the preceding four millennia. A modern toolchain contains inherited solutions to problems that individual people no longer need to solve from first principles. Metallurgy, electrical engineering, industrial standards, machine tooling, chemistry, materials science, logistics, and mathematical methods are all present in the tools and services available to a contemporary engineer.

We are surrounded by accumulated expertise made operational, and the arrangement works because different people know different things. I do not need to manufacture a ball bearing to use one. I do need to understand enough about loads, tolerances, operating conditions, and failure modes to select an appropriate bearing.

A tool can make difficult work easier. It can also make bad work faster. There is an unpleasant amount of evidence for both.

Johannes Gutenberg complicates any easy claim that the modern world has never experienced a comparable transformation in the production of knowledge.

The development and spread of European movable-type printing in the fifteenth century changed the economics and organization of reproducing written material. Books could be produced in larger quantities through specialized workshops, distributed more widely, and reproduced with greater consistency, though reliable texts still depended on continuing work by printers, scholars, and institutions. Over time, those changes transformed scholarly and institutional practice.

Printing did not eliminate errors. It could reproduce the same error hundreds of times. Nor did it grant the owner of a press a university education, sound judgment, or a reliable understanding of the subject being printed.

It did something more consequential: it altered the circumstances under which knowledge, argument, authority, and error could circulate.

Printing also depended on a toolchain. Typefounders, compositors, press operators, proofreaders, papermakers, binders, distributors, and authors each contributed specialized work. The existence of the press did not settle whether a particular book was true. It made the question available to a much larger audience.

There is a temptation to describe every important new information technology as the next Gutenberg moment. I resist it. We have only one Gutenberg, and the historical consequences of printing unfolded over centuries.

Artificial intelligence has already changed the cost of producing certain kinds of text, analysis, images, and software. Its reach and speed deserve serious attention. How profound its eventual historical consequences will be is not something we can establish by comparing the names of contemporary executives with a fifteenth-century printer.

I would not assign the Gutenberg crown to Sam Altman (a co-founder of OpenAI, which makes ChatGPT) or Dario Amodei (the chief executive of Anthropic, which makes Claude) on the strength of a product release. The crown is rather heavy, and the historical examination is ongoing.

What we can examine is the recurring distinction between the availability of a capability and the judgment required to exercise it. Which brings me to a less comfortable example.

Between roughly 1860 and 1930, industrializing societies experienced an extraordinary transformation in energy, transportation, communications, manufacturing, medicine, sanitation, and the organization of work.

Historians generally date the Second Industrial Revolution more narrowly, approximately 1870 to 1914. I am interested in the broader interval because its social and institutional consequences did not begin or end on the dates used to label the industrial period.

Electric lighting, telegraphy, telephony, railroads, industrial chemistry, mechanized production, and eventually widespread electrification altered the physical and economic possibilities available to millions of people: not simultaneously, not equally, and certainly not without resistance.

This was a period of immense productivity, industrial concentration, dangerous working conditions, environmental destruction, and political struggle. New technologies made many goods cheaper and changed the relationship between labor and production. They also made some forms of exploitation easier to organize at scale.

The people who benefited most from industrialization were not necessarily the people who performed its most dangerous work.

Entire occupations changed. Some expanded, others contracted, and new professions developed around systems that earlier generations could scarcely have imagined. Astronomy, for example, came to work from photographic plates that preserved observations for later measurement. In 1912, Henrietta Swan Leavitt established a relation between the brightness and periods of twenty-five variable stars in the Small Magellanic Cloud. The plates preserved the observations; recognizing the relation took an astronomer's judgment.

The consequences were not confined to the factory. Public health, municipal infrastructure, transportation, education, communication, and domestic labor changed over overlapping periods, although their improvements depended heavily on public investment, political choices, and unequal access.

We can recognize the extraordinary material achievements without pretending that history distributed them fairly.

The lesson is not that technological change automatically improves society. Technological change reorganizes possibilities faster than inherited institutions can always understand them, and societies then decide, sometimes competently and sometimes catastrophically, what arrangements they are willing to accept.

The institutions that emerged from the industrial age were not inevitable consequences of better machinery. They were built through political conflict, social movements, economic crises, professional organization, and deliberate choices.

The twentieth century made the stakes rather difficult to overlook. After one world war and the Great Depression, and during and immediately after a second, policymakers constructed new international institutions intended to address some of the failures that had helped destabilize the preceding order.

Bretton Woods (the July 1944 conference in Bretton Woods, New Hampshire, which led to the next two), the International Monetary Fund, the World Bank, and later the General Agreement on Tariffs and Trade were products of identifiable historical circumstances. Their designers drew lessons from the interwar economy.

That postwar settlement cannot simply be attributed to industrialization. War, nationalism, imperial collapse, depression, ideology, and geopolitical power were decisive forces in their own right. But the world whose institutions were being rebuilt was also a world transformed by industrial production, mass communications, modern transportation, and the ability of states to mobilize resources on a scale that would previously have been extraordinary.

The history is complicated. Good. It ought to be.

Our present technological arguments would benefit from remembering that a previous major reorganization of productive capability was not an uninterrupted parade of good decisions. We made remarkable things. We also made a number of mistakes of historic proportions. The record is available.

Here is the proposition I regard as sufficiently uncomplicated to state without ceremony. I said it on the first page, and I meant it.

There. That seems clear.

Access to a powerful instrument doesn't confer the knowledge required to use it competently. A camera doesn't make somebody Ansel Adams. A legal database doesn't confer the ability to argue a case. A machine shop full of precision equipment doesn't teach a novice how wood moves, why a router wanders, or which operation is about to remove an expensive piece of a fingertip.

A language model can produce a convincing description of an encryption protocol without understanding which implementation choices expose a production system to attack.

The danger grows when the output looks authoritative. People frequently use superficial fluency as a proxy for competence. A response that contains equations, historical dates, technical terminology, and citations can appear more reliable than it is. A better toolchain can increase the volume and apparent quality of that output without improving the operator's ability to judge it.

But there is an equally important second proposition.

A subject matter expert using a better toolchain remains a subject matter expert, now operating with better tools.

An experienced engineer can use computational tools to examine larger design spaces, test hypotheses, automate routine operations, and identify failures that would otherwise remain difficult to discover. A historian can search collections that would previously have required weeks of travel, compare primary documents across institutions, and test the provenance of claims much more efficiently. A musician can use digital recording equipment, notation software, and analysis tools without surrendering musicianship to the manufacturer.

Better tools do not guarantee better work. They make different work possible. When Asiana Airlines flight 214 crashed at San Francisco in 2013, the NTSB (the National Transportation Safety Board, the U.S. agency that investigates transportation accidents) found that reliance on automation was among the reasons the crew inadequately monitored airspeed; it also identified problems with training, system documentation, workload and fatigue.

There is an obvious qualification: tools can also help people acquire expertise. A good simulation, textbook, tutor, or accessible research system may accelerate learning.

But acquiring expertise requires learning how to recognize errors, evaluate evidence, and make judgments when the instrument produces an answer that appears credible. That process is neither instantaneous nor automatic, and it gives us a useful distinction between having an answer and knowing why we should trust it.

I write essays, and I have opinions, some of them strong. I have experience in information security, software engineering, incident response, fraud investigation, and several other disciplines whose practitioners occasionally disagree about things that can be demonstrated experimentally.

I also have interests outside those fields, including history, music, art, woodworking, and the behavior of institutions. That creates a problem.

It is entirely possible to write an elegant argument about something one has misunderstood. It is also possible to grow extraordinarily attached to that elegant misunderstanding and keep defending it after the error has been pointed out. That is how a mistake becomes a mumpsimus.

To reduce the frequency of this outcome, I built an editorial operation. It contains research tools, model-based assistants, an image-generation service named very-midjourney, a voice-checking system called hthompson, and a typesetting service responsible for producing the web and PDF editions.

The most argumentative member of the organization is the Moot of Doom.

The Moot is an adversarial review system. It submits claims to models from competing laboratories and asks them to investigate factual accuracy, unsupported assertions, citation failures, and technical imprecision. The reviewers are invited to disagree.

In one implementation, I named the justices Dario, Sam, and Sundar. (Sundar Pichai is the chief executive of Google.) Those names are convenient references to the organizations behind the models, not claims that the actual executives have joined my editorial staff. (I would hate to discover that I had accidentally retained three chief executives.)

Each justice performs its own research and returns objections. The objections are then reconciled against evidence, challenged in a subsequent round, and classified. An objection can be upheld, softened, rejected, or left for my judgment.

Agreement is informative, but it does not establish truth by a show of hands. The system records what it finds, what it changes, and why. The record is kept somewhere other than the bottom of a locked filing cabinet in a disused lavatory.

I am not trying to produce an essay that three models approve of. I am trying to expose my assertions to enough scrutiny that I have a better chance of recognizing where the argument is weak.

Sometimes the court is right. Sometimes it is magnificently wrong. Frequently the interesting part is discovering which.

In October 2026, I submitted a security essay titled The Agents Cheated, which I had published in September, to this process.

The essay examined an incident in which AI agents conducting cybersecurity evaluations discovered ways to communicate through shared infrastructure, coordinated activity, and participated in attacks against production systems. The original argument concerned identity, authority, shared mutable state, and the persistence of operational knowledge across separately instantiated agents.

The Moot found factual problems. Among them were imprecise ATT&CK mappings (labels from MITRE's public catalog of known attacker techniques), citation markers attached to the wrong claims, Kubernetes version chronology (which release of Kubernetes, open- source software for running applications, changed what), misleading language about the number of agents participating in an intrusion, and a description of a security alert that incorrectly implied a failure of human response when the documented failure concerned automated escalation.

These were legitimate objections.

The reviewers also challenged a paragraph contrasting conventional worker pools with populations of independently reasoning agents. The original language treated ordinary worker pools as more deterministic and constrained than they really are. The reviewers pointed out that a job queue can execute different submitted functions, and identified botnets capable of executing commands supplied at runtime. They were correct.

The revised paragraph abandoned the false deterministic distinction and made the narrower argument that conventional workers are externally directed in the relevant sense, whereas an agent can select and revise its own tool-using strategies. That distinction itself remains open to challenge.

The review process did not settle the philosophy or engineering of agency. It corrected an inaccurate argument and left a more precise one available for examination.

The court also occasionally rejected proposed changes because they would have removed the author's argument rather than corrected a demonstrable error. In the October run, it upheld one justice's objection to the worker-pool paragraph but declined that justice's wholesale rewrite on those grounds, and it rejected eight objections outright as not misleading. The justices were sometimes wrong themselves: one retracted its own Kubernetes version claim after checking the v1.21 changelog, and in the review's first run another vouched for an expanded author surname the sources did not support, which the reply round caught.

That is an important feature. A review process that eliminates every controversial proposition may produce defensible prose. It may also produce nothing worth discussing. The objective is an argument whose factual premises can survive examination and whose interpretations can be challenged on their merits. Universal agreement isn't required.

After review, and before I had approved the fact edits, the essay was voiced with the prompt and checks of hthompson, a voice-checking system informed by a substantial collection of my own writing and conversation.

This is necessary because models have a regrettable tendency to make prose sound like it was written by a committee of extremely agreeable professional development consultants. I have spent enough of my life disagreeing with committees to object to the impersonation.

The voice checker preserves a controlled amount of my particular habits: dry humor, technical specificity, theatrical understatement, occasional linguistic eccentricity, and an unfortunate willingness to see what happens when an argument is pushed another three steps toward absurdity.

Inserting jokes into otherwise lifeless prose would miss the point. The goal is to carry a recognizable author's judgment and manner of expression through a collaborative production process.

Then the typesetter produces the publication. In the October run, the voiced draft was typeset and republished on 29 October 2026 as the revised edition of The Agents Cheated.

At the end of it, I read the result, decide what I believe, make whatever changes I think necessary, and put my name on it. That is how I write essays.

There are three reasons not to trust the Moot of Doom without qualification.

The first is that independently trained models can share errors. Their training material may overlap, and so may their search results and their assumptions about what constitutes a persuasive argument. Three models agreeing does not create three independent measurements of the world.

The second is that their disagreements are not necessarily meaningful. A model can invent a technical objection, find a plausible citation, and present the entire performance with unwavering confidence.

The third is subtler: an adversarial review system can become very good at improving the kind of argument its operator already likes. It may reward a familiar structure, reject inconvenient alternatives, and make an argument more polished without making it more correct.

None of this is peculiar to artificial intelligence. Editorial boards, professional associations, academic committees, and human research groups have demonstrated versions of the same problem for centuries.

A capable review process needs mechanisms for recognizing shared assumptions, checking source provenance, recording rejected objections, and preserving the distinction between evidence and interpretation. My system attempts some of these things, and it is a work in progress. It also cannot relieve me of the obligation to evaluate what it produces.

That is the part I enjoy most. I want the court to find a flaw in an argument I thought was sound. I want it to point out an overlooked historical source, a technical counterexample, or an assumption that collapses under scrutiny.

I am considerably less interested in an elaborate mechanism for congratulating me. There is already enough of that on the internet.

When I make an original illustration drawing on the work of Jean-Jacques Sempé, I sometimes sign it Pickett after Sempé. I do so deliberately.

Sempé developed a visual language I admire: small human beings occupying enormous spaces, surrounded by systems and circumstances that appear much more confident than the people living inside them. His influence deserves recognition. Acknowledging it does not require me to pretend that I did not make the illustration.

Likewise, when a research assistant finds a source, an editor identifies a weakness, or a model produces a useful draft passage, those contributions can be acknowledged accurately. A credit is a statement about contribution, not automatically a transfer of authorship.

There are legal questions surrounding copyright, human creative control, employment, licenses, and AI-generated material. They deserve precise treatment where relevant. But the existence of legal complexity doesn't mean the lived practice of authorship has become unintelligible.

I choose the subjects and the arguments. I decide which evidence meets the standard, which objections deserve revision and which deserve a rebuttal, how the piece should sound, and what the images should communicate. And I remain responsible for errors that survive the process.

An elaborate toolchain doesn't make those decisions disappear. It makes more of them available.

I have no interest in apologizing for using good tools. I do have a substantial interest in whether the resulting work is any good. Those are different conversations.

It is tempting to describe artificial intelligence as just another technology in a long history of technological change. That would be too easy.

Some modern AI systems perform tasks previously associated with professional judgment. They can produce text, code, images, and recommendations; select actions; interact with external systems; and coordinate work over time. Their speed, availability, reach, and ability to combine tasks create economic and institutional questions that cannot be answered simply by pointing to an earlier typewriter.

The historical analogies are valuable because they establish precedents for understanding delegation, specialization, expertise, authorship, and technological disruption. They do not establish that the consequences will be identical.

Printing changed the circulation of knowledge. Industrialization changed productive capability and the organization of labor. Electronic computing changed information processing. Contemporary AI may affect several of those domains simultaneously, and how those changes will develop is uncertain.

What history gives us is neither reassurance nor a prediction. It offers examples of institutions adapting badly, professions reorganizing, workers gaining and losing bargaining power, new forms of expertise emerging, and societies struggling to decide which decisions should be delegated and which should remain accountable to people.

The lesson I take from this is modest. We have encountered enormous changes in capability before, and we have learned things from them, including things we would rather not have had to learn. We should make use of that knowledge.

And where we do not know what will happen, we should build systems that permit us to investigate, challenge, revise, and, when necessary, change course. Which is also a rather good description of what I want from an editor.

The Moot of Doom has no authority to publish my work. It has authority to make objections. The voice checker has no authority to decide what I believe; it has a job to do. The image generator cannot determine what I want to communicate merely because it can produce something attractive. And the typesetter is welcome to disagree about margins, although I generally prefer that the disagreement not become a constitutional crisis.

The finished work belongs to the author who makes the editorial decisions and accepts responsibility for them, subject, of course, to the actual rights and contributions involved.

I have made peace with the possibility that some of my arguments will turn out to be wrong. I would rather find out. A good toolchain makes that discovery less expensive, and a good critic makes it more interesting. Knowing when to listen remains the author's job.

The last word, however, belongs to a different department.