Chapter 14 · The Researcher's New Craft
Chapter companion
This chapter's ladder. This chapter does not occupy one of the seven steps. It turns the whole ladder around and looks at it from the human side. The four levels are not only AI's climbing route, they are also your table of role changes, and section 14.6 gives the four names for what a human is called on each rung.
The skill on the first line of your resume, AI did it faster than you last week. This chapter wants to convince you of one thing only. What got eaten is only part of your process steps, and you, the person, are still here. Once you separate the person from the process steps, the panic turns into a list you can act on.
This chapter delivers. Three things. A split ledger of which process steps are depreciating and which are appreciating, a portrait of the new craft in five items (including the four-column dispatch brief), and a four-role ladder table seen from the human side.
14.1 The first line of your resume
Six on a Friday afternoon. A new hire three months in sends you an industry review. You had planned to spend part of the weekend giving him pointers. Instead you read the whole thing that night, paragraph by paragraph, and your red pen never came down. Clean structure, claims carrying their qualifiers, key numbers with sources, even a small table of "which two camps are fighting and what the stakes are." The flaws you finally picked out were one citation format and two sentences that read like translation.
You ask how long it took. "Two days. Most of it verifying AI's drafts. The writing was fast."
Five years ago your first review of that quality took three weeks, and it is what earned you your standing at this company. The first line of your resume still says "fast at mastering unfamiliar fields, produces high-quality reviews and judgments." That line looks a little old tonight. On the subway home you ask yourself the question every craftsperson has asked on some evening these past two years. Is my value being eaten by AI?
The first thing this chapter does is point out where the question goes wrong. It goes wrong on the unit. It treats "you" and "your process steps" as the same thing. Separate them and you see that what got eaten is the process steps, and only half of those. The other half was not eaten. It is still going up in price.
14.2 It is the process steps that depreciate
You have already met the accountant in Chapter 2, section 2.1. VisiCalc ate "recompute it," bookkeeping jobs shrank, and accountant jobs grew instead. Now unfold another precedent, the drafter.
Before AutoCAD, "design" in industry was really two bundles of process steps. One bundle decided what to draw, structure, function, tradeoffs. The other copied the decisions onto the sheet, line weights, lettering, three-view projection, a hand that must not shake. A drafter's ten years of skill all lived in that second bundle. AutoCAD ate almost the whole of it (row 11 of the transfer map). The result was asymmetric. The drafter's occupation shrank, the designer's did not. The cost of revising a drawing collapsed from a week to an hour, so designers dared to revise ten times. The craft did not disappear. The account was just split differently. The mechanical bundle went to the machine, the bundle that decides what to draw went to people, and the first bundle got more valuable precisely because the second got cheap.
To run this ledger on yourself, first give two words operational definitions. Transcription work, where the mapping from input to output is basically fixed, right and wrong are checkable on the spot, and the method is already written down in ten million precedents. Copying out, formatting, translating, boilerplate reviews, first-draft assembly, boilerplate code all count. Judgment work, where "what counts as right" is itself part of the job, with no ready answer to copy. What to ask, what to trust, what counts.
Why is the depreciation wholesale? Transcription work has cheap ground truth, its errors are visible on the spot, machines can close the loop and check themselves, supply can be copied without limit, so the unit price collapses toward zero. How much better AI is than you does not even enter this ledger. The appreciation is wholesale too. Once production gets cheap, output volume explodes, and every piece of output needs someone to answer "can it be trusted." Demand for judgment explodes and supply does not. The bottleneck sets the price, and the price rises on the link that was not automated. This script was rehearsed once on search engines, row 8. This time it is research's whole transcription bundle.
So that midnight question should be rewritten as what fraction of my time goes into process steps that are depreciating. The next section is that table.
14.3 The craft migration list
The table has one row per process step. The third column is the evidence anchor, the source to check back against. Reading columns one, two, and four is enough. Every row was argued in an earlier chapter, or is already booked on the transfer map.
| Process step | Depreciating / appreciating | Evidence anchor | What you should do instead |
|---|---|---|---|
| Reading every paper end to end and hand-writing the summaries | Depreciating | Chapter 4's paper card; row 8 of the transfer map | Have AI produce structured summaries, spot-check them, read in full only the few papers that bear directly on your question |
| Citation formatting, translation, copying out | Depreciating | Chapter 3's defining tasks for tool level | Dispatch all of it, keep only the spot check |
| Boilerplate reviews and the "related work" first draft | Depreciating | Chapter 4; row 1 of the transfer map | Have AI draft it, and check every claim heading into a decision against the original before you sign |
| Boilerplate code for data pipelines and analysis | Depreciating | Chapter 7; row 5 of the transfer map | Have AI write the pipeline, you write the criteria and the tests, run a pilot first |
| First-draft assembly (arranging existing material into a finished draft) | Depreciating | Chapter 9 | Generate it from the claims list, dispatch it |
| Running the mechanical checks (citation existence, number tracing) | Depreciating now | Chapter 12's batch dispatch at L1 | Dispatch it to an independent channel, read only the "undecidable" and "falsified" columns |
| Picking the question (what is worth asking, what shape to ask it in) | Appreciating | Chapter 5; the no-cheap-ground-truth side of 3.4 | Do it yourself, AI only produces candidates |
| Criteria design (locking in "what counts as losing" before the run) | Appreciating | Chapter 6; row 5 of the transfer map | Lock it in and sign it yourself, no changes before the run |
| Dispatch (writing the task as a self-contained brief) | Appreciating (a newborn process step) | 14.5 of this chapter; this book's own case | Practice the four-column brief, see 14.5 |
| Allocating the verification budget and the final review | Appreciating | Chapter 12; row 1 of the transfer map | Assign the layer yourself, never outsource the final review |
| Labeling confidence and reviewing status | Appreciating | Chapter 13; row 3 of the transfer map | Label the status yourself, AI only runs the searches |
| Signing your name (answering for the consequences of a conclusion) | Appreciating | The fourth criterion in 3.3 | Sign only conclusions you can answer for |
Two notes on reading the table. First, depreciating does not mean gone. The steps on the depreciating side still have to be done. They move from you doing them by hand to you dispatching and accepting. They drop from a skill to a cost line, and the responsibility stays in your name. Second, not one row on the appreciating side is newly invented. All of them were already part of the research craft, buried under the transcription workload. Once the transcription is pulled out, they become the main job.
14.4 Five things, one craft
The six rows on the appreciating side collapse into five core skills. The sixth row, signing your name, is not one of the five, and section 14.7 takes it on its own. The four that earlier chapters already established each get one sentence here to pick them up, not a rerun.
Taste in questions. Chapter 5 taught it as a capability, grinding a blur of curiosity into a falsifiable question. Now it rises into an identity. Once reviews, code, and first drafts are nearly free, almost the only thing separating you from anyone else is what question you asked. AI can generate a hundred candidate questions in one breath. Judging which one deserves three months of your life is you.
Criteria design. The core move of Chapter 6, signing off "what counts as losing" before the run. In an age of free production, criteria are the only thing standing in front of "the conclusion happens to support the plan that was finished first." Drafting can be outsourced. The right to sign cannot.
Dispatch craft. The only newborn process step among the five, taught on its own in the next section.
Verification discipline. Chapter 12 delivered the whole set, three-layer assignment, the independent channel, three-value output. Putting it in the portrait adds one sentence only. It is the one of the five that cuts across all seven steps.
Honest calibration. Chapter 13 delivered it, three statuses, the conditions that change a status, the review date. The reason it belongs in the portrait, it governs how you price your own output. Price it wrong and the first four, however well done, go bankrupt in someone else's hands.
Why exactly these five? Take the three variables from Chapter 3, section 3.4 and run them across, visibility of errors, cost of correction, whether cheap ground truth exists. All five land on the worst side, one by one. Taste in questions has no cheap ground truth on its side, no compiler can rule on whether a question is worth asking. Errors in criteria design are silent. Errors in dispatch show up latest of all, a botched brief is invisible until the work comes back. Verification discipline is a contest over the cost of correction, misallocate the budget and the errors that slip through flow quietly downstream. Honest calibration waits on the slowest feedback of all, mark your confidence too high and you wait for reality to settle the account.
Errors silent, correction expensive, no ground truth to lean on, this is exactly the class of work that section 3.4 said needs the knob turned low and a human present. Whatever a machine can close the loop on has collapsed in price. What is left is scarce, and it has nothing to do with being noble. This new-craft list is nobody's design. It is what the distribution line of cheap ground truth cuts out of your work on its own. That line is a strong correlation, not an iron law, and the line itself moves. The withdrawal condition is the same one written in row 11 of the Chapter 13 map, the row saying what gets eaten is transcription and not judgment. The day stable evidence of automated competence on judgment work appears, this list has to be redrawn.
One more structural observation. All five are meta-work. What they produce is constraints on content, and not one line of the content itself. The question constrains the direction, the criteria constrain winning and losing, the brief holds execution in place, verification guards the door, calibration caps how strongly you may state a claim. One hour upstream saves ten hours downstream. The researcher's new craft pressed into one sentence, you go from a person who produces content to a person who produces constraints.
Constraints have another side. Earlier chapters all presented them as a brake. Here is the other half, they are also an asset. The market is not short of content. It is short of what makes a pile of automatic output credible, a set of locked-in criteria, a harness that leaves errors nowhere to hide, meaning the scaffolding the experiments run in, an eval environment usable as a training signal, meaning the kind that scores repeatedly and feeds the results back so the model improves. Whoever banks these has the larger radius of letting go. This accumulation cannot be carried off and cannot be copied. It grows inside your understanding of your own problem. The discipline that guards against self-deception and the assets worth money are two sides of one thing.
The five also mesh as one set of gears. Taste picks the question, criteria set winning and losing for it, the brief sends the work out, verification decides whether what comes back gets in the door, calibration decides how strongly you may state what got in. Signing your name finally collects the whole chain's responsibility under one person. Break any link and the other four spin free. However hard your criteria are, a botched brief still gets you back a pile of plausible garbage. Chapter 15 will teach you to freeze these gears into habits that do not run on willpower. This chapter only makes you see the shape of the whole set.
14.5 Dispatch craft, writing a task into a brief AI can take
This craft has already shown up in pieces. Chapter 4's paper card is a miniature brief, and Chapter 12's verification brief goes further, giving only the claim and leaking no expectation. This chapter gathers the pieces into one craft, because every step you climb toward collaborator level hands a bigger block of work to an executor who does not share your brain, and the quality of that handoff sets the rework rate.
All the difficulty of dispatch sits on one thing. Whoever takes the job does not have your tacit knowledge. Dispatching to a colleague can be sloppy, a colleague will ask, will fill the gaps, the two of you share the air of one office. Dispatch to AI and nothing outside the brief exists. Dispatch craft is the ability to turn tacit context into explicit text. Its skeleton is four columns.
[TASK] What to produce. Verb first, one sentence. Medium, format, length spelled out.
[CONTEXT] Every fact and file the other side needs to start, written on the rule that "the
recipient knows only what the brief says." Check before dispatch, is the context pack current?
[BOUNDARIES] What is not allowed. No inventing facts or citations. Anything uncertain
gets marked. Whatever is missing, come back with a list, no filling in from imagination.
[ACCEPTANCE CRITERIA] What counts as delivered. Checks a third party can run,
not "write it a bit better."
This book is itself a first-hand case of "a researcher commanding an army of AI." Chapter drafts were drafted by writing agents and finalized by the editor-in-chief, checks ran through an independent channel, and Chapter 12 already showed you that channel's record of shooting down the book's own claims. Two dispatch lessons here, both of which really happened. The first is a rework specimen. After the Chapter 5 writing agent turned in its draft, three motive paragraphs were torn down and redone. On review the account went to the person dispatching, and the agent had done nothing wrong. The context pack it received was an old version of the case status doc, still holding a claim the verification had already overturned. An agent does not refresh facts on its own. The world in the brief is its entire world. The lesson hardened into a process step, the context pack gets updated before dispatch, not to be skipped even once. The second lesson is positive. Every brief carries a required-reading list and a boundary discipline forbidding invented case facts. At handover you check against the list, instead of judging by feel whether it reads well.
The test of a good brief is one sentence. An executor who has never met you and cannot ask you questions can start work from this brief alone, and knows what counts as delivered. Fall short of that and the rework goes down as your dispatch incident. A bad-brief and good-brief comparison, plus the full fillable template, are in this chapter's appendix.
14.6 The same ladder, seen from the human side
When Chapter 3 set the ladder up, the viewpoint sat on AI's side, watching how high it climbs. Now turn the ladder around and look at what your role is called on each rung. The four criterion questions (who drafts, who reviews, who decides, who answers when it's wrong) do not change. Your title does.
| Level | Your role | What your day looks like |
|---|---|---|
| Tool | Artisan | The craft is all in your hands, AI is a faster pen |
| Assistant | Lead writer and full reviewer | AI drafts, you read every paragraph; your eyes are the line of defense |
| Collaborator | Editor-in-chief and process designer | You no longer read sentence by sentence; you write criteria, design process, spot-check, and rule |
| Autonomous researcher | Principal (nobody answers for it) | You set the goal and the acceptance criteria; but "who answers when it's wrong" lands on nobody, and the human role at this level has never been filled in |
Artisan, lead writer, editor-in-chief, principal, these four names are this book's working terms, and Chapters 15 and 16 refer back to them.
Look at the gap from assistant to collaborator. Your work changes from "doing research" to "designing the process that makes research acceptable." Chapter 3 said that here your reviewing switches trades. This chapter's version is that this is the moment the five new-craft skills report for duty. The editor-in-chief is still an artisan, only in a different trade, and the work has not dropped by an ounce. The empty slot beside "principal" in the last row is the fourth column of that table in section 3.3, reproduced as is. Until the answering problem is solved, the top floor of the ladder houses AI, not people, and should not house people.
One old rule from section 3.4 carries over unchanged. Roles, like levels, live on "step × task," and they do not follow the person. In a single afternoon you can be the principal on citation formatting, the editor-in-chief on the data pipeline, fall back to lead writer when reading the results, and go back to artisan on "dare we state this conclusion unconditionally," writing every qualifier by hand. Climbing the ladder, seen from the human side, means "editor-in-chief" shows up more and more often in your role column, while "nobody" should never show up at all.
14.7 Operational definitions for the three things that "are always human"
"Some things will always be human." That line gets said so often it has nearly been emptied out. The only way to give it content is an operational definition. Three of them, each with a test.
One, judgments where someone has to answer when they are wrong. A judgment belongs to a human if and only if, when it goes wrong, someone has to clean up the wreckage, retract, compensate, correct, apologize. AI bears no consequences. So for any judgment whose consequences cannot be recalled, publishing, architecture selection, advice to a patient or a client, the last pair of eyes has to sit on a person who can bear the consequences. This follows straight from the fourth criterion in section 3.3, with not half a line of sentiment.
Two, the power to define "what counts as winning." Criteria are research's constitution. Whoever writes the criteria defines what is true inside that small world. AI can draft a criteria proposal, but hand over the signature and you drop from researcher to spectator of the research. The test is simple. Pull up the criteria file for the project on your desk and look at the name written where the final ruling goes. That name is this project's real researcher.
Three, a signature that carries lifetime responsibility for the output. A signature ties your name to the entire future of the output, the part of that future that gets overturned included, and the credit line is only what that looks like in passing. The spine case's hypothesis H and the subplot's persona verification are already filed in rows 13 and 14 of Chapter 13, one mixed, one falsified. The signature owning up on those filed rows is mine. Not one of the models that ran scans or wrote drafts for me is credited. The signature on the subplot plan was a human's from beginning to end. The test is colder. Before a piece of output goes out, ask one question. Three years from now it gets falsified, who steps up? Anything whose answer holds no person's name does not deserve to go out.
The three together are the entire content of the "judgment" in "command plus judgment." Not one of them can be written as a prompt. What a prompt holds is information processing. The substance of these three is a promise.
14.8 Swap in your project
Make yourself a "my process list," budget thirty minutes.
- Write down the 10-15 process steps you actually did last week, verb first, specific down to "formatted the citations for a report," and no writing "did research" at that grain;
- Label each one transcription or judgment, asking only two things while you sort, does this step have cheap ground truth, and how fast do errors show up (the 3.4 variables, now used as a sorter);
- Mark the transcription side depreciating and the judgment side appreciating; anything you cannot call, mark "mixed," then split it in two and relabel;
- Estimate against your calendar, not your impression, what fraction of last week went to the depreciating side?
- Circle the single most time-consuming step on the depreciating side and write its first four-column brief this week, then dispatch it;
- Circle the one you are weakest at on the appreciating side (check against the five), it is your private focus for the rest of this book.
- Beginner's version. If you have not banked any judgment yet, keep at least one step on the depreciating side that you do not dispatch, literature summaries or citation checking for instance, and do it by hand for a full month before you talk about outsourcing. Judgment is trained inside transcription work. Dispatch all of it and you dispatch the chance to practice along with it.
Most people, filling this in for the first time, get their jolt at step 4. People who think they live by judgment find their calendar showing six to eight tenths of their time going to transcription. That is nothing to be ashamed of. The previous generation of designers had calendars full of tracing too. Drawing the list and then not moving is the shameful part. The full fillable version of the process self-check sheet is in this chapter's appendix.
Want an agent to run it with you? Paste this to your AI assistant or coding agent:
Help me build the Chapter 14 "my process list", budget thirty minutes. I will dictate the 10 to 15 process steps I actually did last week, you record them
verb first, and send anything at the grain of "did research" back for me to split. Label each one transcription or judgment, asking only two things, is there
cheap ground truth, how fast do errors show up. I give the answers, I set the labels. Estimate the time shares against my calendar, you may not estimate from
my impression. For the most time-consuming step on the depreciating side, I write a four-column brief this week and dispatch it, and you build the skeleton
from the task, context, boundaries and acceptance criteria columns of Template 2 in docs/appendices/ch14-templates.md, contents filled by me. Remind me of the beginner's version, keep at least one step on the depreciating side and do it by hand for a full month. If any command errors, stop and show me the output.
14.9 Sober reminders
- The migration hurts. "It is the process steps that depreciate" holds only if people can move their time to the appreciating side. People who cannot move really exist. The shrinkage of bookkeeping jobs is about four hundred thousand specific people (the number Chapter 2 settled), and the drafter was a whole occupation. For someone who built an identity and an income on a process step that got eaten, panic is an accurate response. And the transition stories you hear are survivors' stories by construction (Chapter 2, failure condition five). This book can give you a list and a direction. It cannot give you a guarantee.
- Do not romanticize "commanding." The editor-in-chief's days are no easier than the artisan's. Once production is free, your day is acceptance, spot checks and rulings around the clock, and the bottleneck moved from production onto you. Dispatch craft saves transcription time. It saves none of the mental effort.
-
The apprenticeship gap, still exploring. Judgment has always been trained inside transcription work. Taste in the literature comes from having read a hundred papers to shreds by hand, a feel for criteria comes from having been burned by an unfair baseline in person, and once transcription is eaten, where does the next generation practice judgment? Coding already has an account you can check, conclusion first, among the jobs most easily displaced by AI, employment for people just entering is falling while people with more years are rising, and the gap is driven by "hire fewer juniors." The numbers go like this. A measurement based on real ADP payroll data (Brynjolfsson et al., 2025, Stanford Digital Economy Lab) finds that in the occupations most exposed to AI (software development included), employment of workers aged 22-25 fell about 16% in relative terms, while workers over 30 in the same occupations grew 6%-12%. Another study covering tens of millions of resumes (Hosseini Maasoum and Lichtinger, 2025, SSRN working paper) points the same way, firms adopting generative AI shrink junior roles while senior roles barely move. The door to practice is narrowing. This book is confident about how people who already have judgment migrate. On how judgment grows from zero, the honest answer is still exploring.
If you do not have judgment yet, do not start subtracting from this table. The paragraph above describes the apprenticeship gap as a field-level problem, while what you hold is a personal decision, so here is a default plan you can test, to keep "still exploring" from turning into a shield. The depreciation table is a ledger for people who already paid tuition. A beginner cannot treat it as a permit. Down to the actions. Those "hundred papers" in your own field, until you have read them by hand, AI may do exactly two things at the literature step, help you find which one to read, and quiz you after you have read it ("I understand it as X, how far does the original support that"). It may not write the summary for you. The summary is the very step where taste gets trained. Outsource it and the practice is gone, and what is left is collecting, which happens to feel a lot like learning. In the same way, at the criteria step, until you have been burned by an unfair baseline yourself, plans AI drafts get walked item by item against the Chapter 6 checklist, with no adopting a draft whole. The evidence grade of this plan has to be stated plainly. It is derived from transfer history, a suggestion, not a prescription that has been measured (the same class of reasoning labeled in row 3 of the Chapter 13 map). I give it a falsifiable shape. The day a controlled measurement appears showing "people who started out using AI throughout have judgment indistinguishable from the traditional path three years later," this plan is void and I will withdraw it. Until then, slower is better.
How the five skills get frozen into habits that do not run on willpower, and how they grow onto a team, is Chapter 15's job. This chapter is responsible only for making you see the ledger.
14.10 The unfair advantage you now hold
The next time that question finds you late at night, "is my value being eaten by AI," you do not have to choose between panic and self-comfort. Spread out your own process list and point to which half is depreciating, which half is appreciating, and which way your time has already started moving. Most people who ask that question have not yet separated "me" from "my process steps."