3 You as a Research Director
This chapter is part of a book in active development and has not yet been through the author’s review. Content may change as the review advances.
The research decision. Every task in your project belongs in one of three buckets: safe to delegate, delegate then verify, or never delegate. Your decision here is where each task goes. You sort the list before you open the tool, not in the middle of using it, and you defend the placements one line at a time.
3.1 Why this decision matters
The decision on the table: which tasks in your project you hand to an AI, which you hand over and then check yourself, and which never leave your hands at all.
Picture a principal investigator, the researcher who runs a lab and owns its scientific decisions, briefing a new lab member on the first day:
“Here is the deal. A research assistant here can run the assay, plot the readings, and format the table, and honestly, they will do it faster than I would. What the assistant does not do is choose the question, decide what the readings mean, or sign the claim. That part is mine. Learn where that line sits, because it does not move just because the assistant got faster.”
AI is the fastest research assistant you will ever direct. The risk is not that it does the routine work. The risk is letting it drift into the calls that make the work yours. Deciding, task by task, what you delegate is the first act of being a research director instead of a passenger.
3.2 The concept
A research director is the person who assigns each piece of work and stays accountable for the finished result. Example: you may ask a tool to draft your methods paragraph, but you are the one who signs it and answers for it. Delegation is handing a task to someone or something else to carry out, the way you might hand off formatting a messy table while keeping the reading of what the numbers mean.
The core move of this chapter is a delegation triage, sorting each task into one of three buckets before you start.
- Safe to delegate: low stakes and easy to check, so a wrong answer costs little. Example: rewording one survey item three ways.
- Delegate then verify: fine to hand out, but you must confirm the result with a real check before you trust it. Example: asking the tool to locate a lab protocol, then opening the actual protocol to confirm it exists and says what the tool claimed.
- Never delegate: a judgment that stays yours no matter how capable the tool becomes. Example: deciding whether your evidence actually supports the claim you want to make.
A never-delegate decision is any call that defines the science itself: the question you pursue, the design and the people or samples it covers, what your measures mean, the ethics, the boundary of your claim, and the uncertainty you report (National Academies of Sciences, Engineering, and Medicine 2017). Verification is the step that earns trust: confirming a result yourself, by a named method, before you put your name on it.
One thing about delegation has changed, and your triage has to keep up. Handing a task to a modern tool rarely means one prompt and one answer. It means a loop: the tool drafts, runs, reads its own error, rewrites, and runs again. Increasingly the tool closes that loop by itself and shows you only the tidy result at the end. A loop you did not watch is not a loop you verified. So when you drop a task into “delegate then verify,” decide right then what you will inspect when the loop stops, rather than deciding later whether the final answer looks plausible (Autio et al. 2024).
3.3 A worked example
You are asking whether a common food preservative, sodium benzoate, slows the growth of a harmless lab strain of E. coli. You measure growth by optical density, how cloudy a liquid culture gets as cells multiply, read on a plate reader as OD600. You are the director. The AI is your assistant. Here is your triage.
Safe to delegate. Reformatting three plates of raw OD600 numbers into one tidy table, and drafting a plain-language version of your methods paragraph. If a value lands in the wrong column, one glance catches it, so the stakes are low.
Delegate then verify. Locating a standard protocol for reading bacterial growth curves, and drafting the formula for doubling time, the hours a population takes to double in number. You hand these out, then confirm each: you open the actual protocol, and you recompute one doubling time by hand from two printed OD values.
Never delegate. Deciding that the concentrations you test are biologically meaningful and safe. Deciding whether a gap between the treated and untreated curves licenses the word “inhibits.” Stating how uncertain that conclusion is, given three plates and one strain. No tool makes those calls. The moment you let it, you become a reader of someone else’s guess.
The line this triage draws, between work you may hand out and judgments that stay with the named researcher, is the one research-integrity guidance draws too (National Academies of Sciences, Engineering, and Medicine 2017).
The block below builds this example’s data and prints the numbers this section quotes. Run it, then change one input and watch which sentence above stops being true.
import numpy as np, pandas as pd
SEED = 464
rng = np.random.default_rng(SEED)
hours = np.arange(0, 8)
def plates(doubling_hours, n_plates=3, od0=0.05):
"""Three plates read hourly; the culture doubles every `doubling_hours`."""
true_curve = od0 * 2 ** (hours / doubling_hours)
return true_curve + rng.normal(0, 0.01, size=(n_plates, len(hours)))
control, benzoate = plates(1.5), plates(2.4)
print(pd.DataFrame({"hour": hours,
"control OD600": control.mean(0).round(3),
"benzoate OD600": benzoate.mean(0).round(3)}).to_string(index=False))
# The check you run by hand, from two printed OD values four hours apart.
def doubling_time(od_a, od_b, hours_apart):
return hours_apart * np.log(2) / np.log(od_b / od_a)
print(f"\ncontrol doubling time : "
f"{doubling_time(control.mean(0)[2], control.mean(0)[6], 4):.2f} h")
print(f"benzoate doubling time: "
f"{doubling_time(benzoate.mean(0)[2], benzoate.mean(0)[6], 4):.2f} h")3.4 An AI failure case
You paste your eight tasks in and ask the tool to sort them. Back comes a clean, confident table. Seven rows look right. But one reads: “Decide whether the growth difference is biologically meaningful: delegate then verify, just apply the standard effect-size cutoff.” It sounds authoritative, and it hides a scope change. A judgment about what counts as meaningful in your specific system has been swapped for a mechanical lookup, and the “standard cutoff” it names is one no source will confirm. You catch it two ways. You wrote your own triage before you asked, and your version had that row as never delegate, so the mismatch jumps out. Then you try to retrieve the “standard cutoff,” and nothing real supports it. The row goes back where it belongs.
3.5 It is your turn
You are working inside Studio 2: Set your rules, shape your question. Keep what you write here; the studio’s milestone chapter is where it joins the other lessons’ pieces into one artifact you can defend.
You have a ledger, a chosen problem, and its rejected alternatives. Now you take the job of the person who decides who does what.
The hands-on half of this section lives in the chapter’s companion notebook: open it in Colab with the badge at the top, and work the steps there.
Each prompt below is commit-first: write your own answer before you open the tool, so its reply has something of yours to disagree with. Expect to run each one more than once. You read the reply, find the weak spot, sharpen the prompt, and go again. The loop is normal; stopping the loop before you have checked anything is the mistake.
The buckets themselves are yours to fill, and three calls never leave the “never delegate” bucket: which question your project pursues, whether your evidence supports the claim you want to make, and how much uncertainty you report. You may ask a tool to lay out options, but choosing among them, and answering for the choice, is the job that makes you the director.
Take the research problem you committed to, and list every task the project around it would need: finding sources, gathering or cleaning data, choosing measures, computing, drafting, interpreting, deciding. Aim for at least ten lines, and do not filter yet.
Sort every line into safe to delegate, delegate then verify, or never delegate, with a one-line reason beside it. That sorted list is your director’s task list, and you will keep amending it for the rest of the book.
Optional depth. The milestone needs only this lesson’s core cycle; run this second angle when you want more practice.
Turn your triage into a checkable list.
Here are eight tasks in my project: [paste your list]. For each, label it "safe to delegate," "delegate then verify," or "never delegate," and give a one-line reason. Return it as a table so I can audit each row on its own.After running, verify (counters illusion of completeness): compare the table against the triage you wrote first and against the never-delegate checklist; a tidy table can quietly mis-file the one decision that mattered.
For each “delegate then verify” line, name the check now, while it is cheap: which document you will open, which number you will recompute, which second method you will run.
Optional depth. The milestone needs only this lesson’s core cycle; run this second angle when you want more practice.
Locate a bench resource (delegable).
Act as a lab-methods assistant in microbiology. I am measuring bacterial growth by optical density (OD600). List 3 published, retrievable protocols or textbook methods for reading a growth curve. For each, give authors or source, year, and where it appears. Only include work you are confident exists; mark anything uncertain. Then tell me which one to verify first.After running, verify (counters confident fabrication): open each protocol yourself before you cite it; a source you cannot retrieve is invented until it opens.
Star the two lines you would defend hardest as never-delegate, and write one sentence each on why those calls are what make the project yours.
Hand the same list to an AI tool, ask it to argue with your placements, and keep only the changes you can justify out loud.
Red-team your own boundary.
I marked this task "never delegate": [paste one task]. Act as a hostile reviewer and argue that it is actually safe to hand to an AI. Give your strongest case. Do not agree with me.After running, verify (counters sycophantic agreement): if the reviewer folds and praises your call, push back, then take its strongest objection to a colleague or mentor rather than trusting the tool’s approval.
Log that exchange in your AI Research Ledger, and verify at least one placement with a named method from the Verification Guide. A triage is a judgment, so the fitting method is peer reasoning: walk your three buckets past a colleague or mentor and let them challenge a placement. Working alone, use the solo substitute: put the list away for a day, re-sort it cold, and compare with your first sort. Every item that moved is a judgment you had not actually settled. An AI reviewer may run either version with you, but it cannot be the peer, because what you are testing is your own consistency. The decision to accept or reject stays yours.
Before you leave the map, add the declaration work ahead to it: drafting candidate question wordings sits in delegate-then-verify; choosing among them, the kind, the reach, and the boundary sit in never-delegate. A delegation map remains a promise until one task travels the entire route from specification to verification and defense; that is the next lesson.