2 AI Is Your Arm, Not Your Brain
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. Decide what must happen between an AI output and evidence you are willing to use. Name the check, record the exchange, and keep responsibility for the result.
2.1 Why this decision matters
The decision on the table: what has to happen to an AI’s output before you are willing to call it evidence.
You arrive with a curiosity commitment and one logged line for every AI activity that shaped it. Now the stakes change: a brainstorm can produce a possibility, but it cannot make a factual statement reliable.
Picture an audit review. A certified public accountant (CPA) is an accountant licensed to sign off on a company’s financial statements, and that signature makes them personally answerable for the numbers. Example: the CPA who signs a firm’s annual statements is who investors turn to if the figures prove wrong.
When a CPA signs a statement, they are not saying the accounting software ran without errors. They are saying, “I checked this, and I will answer for it.” That is the posture this book asks you to adopt from page one. An AI tool can fetch a statistic, draft a paragraph, or write working code in seconds, but it cannot sign your work. You can. So the question that matters is not “what can the tool produce?” It is “what did I hand it, and how did I check what came back?”
2.2 The concept
The rule that names the chapter is AI as your arm, not your brain: the tool can reach, fetch, draft, and compute for you, but it does not get to decide what is true or what you claim. Example: you let it draft a summary of a government statistics release, then confirm every number in that summary yourself.
Two moves put the rule into practice. Delegation is handing a well-specified task to the tool, such as “find three published sources for last year’s county unemployment rate.” Verification is an independent check, run by a method outside the model, that the result is actually true. Example: you open the statistical agency’s real data table and read the number yourself instead of trusting what the tool typed.
Verification cannot be skipped because a wrong AI answer looks exactly like a right one (Ji et al. 2023). A hallucination is confident, polished output that is simply made up, such as a citation to a report that does not exist. Example: a tool quotes a median starting salary and attributes it to a “National Graduate Earnings Survey” that was never published. Tone is identical for real and invented, so tone cannot be your filter. Retrievability is. A source is retrievable when you can open the actual document and confirm it, such as finding the real report through a library and reading the line.
Some judgments never get delegated. A never-delegate decision is a call that stays yours no matter how good the tool becomes: which question you pursue, what your evidence licenses, the ethics of the work, and how you state your uncertainty. Example: a tool may list several ways to adjust old prices for inflation, but deciding which one your analysis requires is yours to own. The habit that ties this together is short: Ask, then Verify, then Document. Use the tool ambitiously, confirm every fact and number yourself, then log the tool, the task, and the check in a record that travels with your work (Autio et al. 2024).
Why draw the line exactly there? The curiosity studio opened with the discovery cycle Zahavy presents through Einstein: experience, an explanatory jump, predictions, and a return to evidence (Zahavy 2026). There you worked the upstream side of inquiry, where a possible explanation begins. This lesson starts downstream. Once an idea exists, AI can help locate evidence, draft text, transform data, or inspect code, and those abilities are valuable precisely when their products are verified. That is why the ledger opens now: every line from the curiosity studio becomes a row of its own, marked retrospective, because the date a ledger opens does not erase an earlier delegation.
2.3 A worked example
You want to know whether meals at restaurants near you got more expensive after a local minimum-wage increase in 2019. You cannot compare a 2019 menu price to today’s price as they stand, because a dollar buys less now than it did then. To compare them you need the Consumer Price Index (CPI), a published number that tracks what a typical household’s basket of goods costs over time, where a higher value means the same basket costs more.
Sort the work. You can delegate locating candidate published price series and listing the inputs an inflation adjustment needs. You keep three things: deciding that a consumer price index (not a producer or wholesale index) is the right measure for what households pay, choosing the base year you convert every price into, and judging whether a source is authoritative. You ask an AI assistant for the series and it returns “consumer prices rose roughly 23 percent between 2019 and 2025,” citing a statistical agency.
Now verify. You open the agency’s actual data table, read the two index values yourself, and recompute the percentage change by hand. If it confirms the number, you log a verified input. If it shows a different value, or cannot be found at all, the number the tool handed you was never evidence. Either way, your comparison rests on a number you retrieved, not one a model typed.
The habit this example drills, opening the source rather than trusting the sentence, is the documented defense against fluent invented detail (Ji et al. 2023).
The block below builds this example’s data and runs the adjustment. Notice what it cannot do for you: the two index values are inputs you retrieve and verify, and the code takes them on faith exactly as the tool did.
import numpy as np, pandas as pd
SEED = 464
rng = np.random.default_rng(SEED)
# Thirty 2019 menu prices from restaurants near campus.
prices_2019 = np.round(rng.normal(12.50, 2.10, size=30), 2)
# These two index values are the input you must RETRIEVE yourself, from the
# agency's own table. They stand in until you do; replace them and rerun.
cpi_2019, cpi_2025 = 100.0, 123.0
adjusted = prices_2019 * cpi_2025 / cpi_2019
print(f"index 2019 {cpi_2019} -> index 2025 {cpi_2025}")
print(f"price change implied : {(cpi_2025 / cpi_2019 - 1) * 100:.0f}%")
print(f"mean 2019 menu price : ${prices_2019.mean():.2f}")
print(f"same meal in 2025 dollars : ${adjusted.mean():.2f}")2.4 An AI failure case
You ask for the median starting salary of graduates in your field and a source. The tool returns a crisp value and attributes it to “the 2024 National Graduate Earnings Survey, Table 3.” It reads like every real citation you have seen: an official-sounding survey, a year, a table. It is fabricated, or the number is not on that page.
Here is exactly how you catch it. You do not trust the citation as delivered. You search for the survey in a library catalog and, if it opens, turn to the named table and read the line. If the survey does not exist, or has no such value, the confident citation collapses. Fabricated sources sound just as authoritative as real ones, so the only reliable filter is opening the document.
2.5 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 arrive from Studio 1 with a chosen research problem. There are no data yet, only the problem and the two things everything else will rest on: the record you keep and the habit of checking. This practice opens the record and runs your first verified delegation, on your own problem.
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.
Commit your own answer first, then delegate. And expect a loop rather than a single shot: you ask, you read what comes back, you push on the part that looks thin, you ask again. That back-and-forth is how real AI work goes, and it is fine. What matters is where the loop stops. It stops when you have run the check yourself, not when the tool finally sounds sure.
Each prompt below hands out a task you can check, with a verify note naming the failure it defends against.
These stay yours, no matter how fluent the tool sounds. Deciding which measure fits your question and which base year the comparison requires. Judging whether a source is authoritative enough to build on. Setting the claim you will defend and stating its uncertainty. The tool gathers options and drafts prose; choosing among them, and answering for the choice, is the researcher’s job.
Open a blank spreadsheet or document and title it AI Research Ledger. Give it eight columns: task delegated, tool used, prompt, output summary, decision, verification method, remaining concern, and responsible researcher. Put your own name in that last column now, and leave it there for the rest of the book.
Open your ledger’s first rows backward, one for each Studio 1 activity you logged: the brainstorm, the source search, the candidate list, and the red-team. Task, tool, what you kept and rejected, and how you checked it. Where a field was never recorded, write “not recorded” rather than invent it. A delegation that happened before the ledger existed still belongs to the record; that is what the record is for.
Under your chosen problem, write the single fact you would need to know before you could take it further.
Optional depth. The milestone needs only this lesson’s core cycle; run this second angle when you want more practice.
Turn a decision into a checklist (you audit the list).
For converting a 2019 restaurant price into 2025 dollars, list every input my calculation needs. For each, say why it belongs and which single choice most governs the result.After running, verify (counters illusion of completeness): a long, tidy list can still omit the choice that governs your answer. Compare it against a checklist you wrote first, and confirm the governing choice is really the one the comparison rests on.
Take the fact you just named, and ask an AI tool for three published sources that speak to it. Before you read the reply, write down what you expect it to say.
Locate candidate sources (you verify each one).
Act as an economic-research assistant. I need annual consumer price index values for 2019 and 2025. List 3 candidate published sources, each with the exact table or series name and the agency that publishes it. Only include documents you are confident exist; mark anything uncertain.After running, verify (counters confident fabrication): open each document yourself and read the number. A table you cannot retrieve is invented until you find it.
Log the exchange in your AI Research Ledger, and verify at least one output with a named method from the Verification Guide. For a citation the method is primary-source reading: open the document and read the line. An AI reviewer may run the check with you; the decision to accept or reject stays yours.
Optional depth. The milestone needs only this lesson’s core cycle; run this second angle when you want more practice.
Red-team your own choice (you keep the decision).
Here is my price index and base year for comparing 2019 and 2025 menu prices: [paste it]. Act as a hostile referee. Name every place the choice reaches past my evidence. Do not rewrite it for me; list the weaknesses.After running, verify (counters sycophantic agreement): if every objection is mild or it calls your choice “well-balanced,” push back and ask for the single worst flaw.
The ledger tells you what happened. The next decision is what should be handed to a tool in the first place.