1 From Curiosity to a Research Problem

Under development

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. Begin with an experience that will not leave you alone. Decide which curiosity you are willing to own, let AI widen the possibilities without choosing for you, and commit to one research problem with your reasons attached. The formal research question comes later, after your working rules and ledger exist.

1.1 Why this decision matters

The decision on the table: which one research problem, out of everything your curiosity could become, gets months of your work.

Imagine finding a restaurant menu from 2019. A meal that once cost eleven dollars now costs sixteen. Your first reaction may be immediate: prices near campus have exploded.

That reaction is useful. It is also not yet a finding. Perhaps restaurant prices rose faster than prices generally. Perhaps the old menu was unusually cheap. Perhaps portions changed. Perhaps fees moved out of the listed price. Perhaps the example is memorable precisely because it is unusual. Research begins when the first explanation stops being the only explanation you can see.

You do not need a polished question yet. You need a curiosity strong enough to survive alternatives. AI can help here: it can propose explanations, comparisons, affected groups, or implications you did not think to name. Use that help, mark what the tool introduced, and preserve the exchange. The final commitment remains yours, because you will be the one who must pursue, revise, and defend it.

1.2 The concept

In the account of scientific discovery that Zahavy develops from Einstein, inquiry moves through a cycle (Zahavy 2026):

flowchart LR
  E["Experience<br/>something in the world catches"] -- "the jump: a possible explanation<br/>(abduction, yours to make)" --> J["Ideas and assumptions<br/>what could explain it"]
  J -- "derivation<br/>(delegable, then checked)" --> P["Predictions<br/>what else should be true"]
  P -- "verification against evidence<br/>(delegable, then checked)" --> E

The cycle is redrawn from Zahavy’s reconstruction of Einstein’s account of discovery; the notes on the arrows, marking what you may delegate and what you may not, are this book’s (Zahavy 2026).

Experience does not logically dictate one explanation; the researcher makes a jump toward an idea that could make sense of it. The jump has a name. Abduction is proposing what might explain a pattern in the first place, where deduction asks what must follow if an explanation is true and induction looks for patterns across observations. Abduction creates a candidate, not proof. Example: restaurant prices seem to change every third week, and “payday timing moves menu updates” is an abductive possibility. It gives you something to check; it does not become evidence by sounding plausible.

Why does a book about AI-driven research open here? Zahavy’s position paper, written at Google DeepMind, uses Einstein’s route to general relativity to argue that important scientific jumps do not always begin with a loud error signal: Newtonian physics already fit the data well, and the reorganizing idea came from thought experiments about falling observers, not from compressing observations. On that basis the paper argues that current language models are strong at recombining learned patterns and increasingly useful for derivation, while lacking the grounding in experience from which this kind of jump begins (Zahavy 2026). Treat that as an argued position, not a settled boundary; position papers exist to be tested, and tools keep improving. The practical point holds either way: you answer for the claim, so the jump is yours to make and to defend, and this book begins with yours.

The jump is not a solo performance after its first line is written. Once your curiosity is on the page, AI is welcome at the table as a divergence partner: ask it for alternative explanations, overlooked comparisons, or questions that would challenge your starting belief. Mark every person, population, outcome, mechanism, or comparison it introduces as NEW, accept no novelty claim or citation without checking, and keep or reject each direction with a reason. What the tool cannot do is care; you choose what you would chase even if every tool disappeared tomorrow.

With the jump committed, the rest of the lesson reasons downward, in four moves (Booth et al. 2024):

Experience. What happened, or what did you notice?
Curiosity. What do you want to understand about it?
Topic. What area of inquiry contains that curiosity?
Research problem. What is not yet understood well enough for the decision, explanation, or audience you care about?

Stop there. Units, outcomes, conditions, question kind, and claim reach belong to the formal declaration that closes the next studio.

Then test the problem three ways (Hulley et al. 2013). Importance asks who could use the answer, and for what. Feasibility asks whether you can reach suitable evidence with the time, skills, and permissions you actually have. Possible contribution asks what the project might add — and stays provisional, because you cannot know a gap is real until Studio 3 maps the existing evidence.

One more lens sharpens all three. A 2026 essay in Science, “Advancing science by designing for surprise,” starts from a blunt fact: an answer that confirms what everyone already believes teaches almost nothing (Evans et al. 2026). The essay makes the idea usable: elicit informed predictions first, then compare them with what the evidence actually shows. The gap it names sits at the level of an audience, meaning the informed readers whose expectations your result would move, not at the level of any one researcher. Before you look at any evidence, then, write the answer you currently expect and name the person for whom a different answer would be genuinely surprising. That written sentence is your expected answer. Treat it as a planning proxy for that audience gap rather than a measure of what anyone learned, which would take evidence about the audience you do not have yet. The essay’s warning travels with the tool: the more surprising the claim, the more evidence it takes, so surprise is a place to aim your question, never a license to inflate your answer.

1.3 A worked example

Start with a health curiosity: “does moving more actually help you sleep?” It is a real feeling, but you cannot study it, because it names no one and measures nothing. Give it edges and it becomes a topic: physical activity and sleep quality in undergraduates. Push further to a research problem: we do not know whether more daily walking goes with better-rated sleep among first-years living in residence halls. That points at one missing piece, not a whole subject.

Now run the three tests. It clears importance (anyone advising sleep-deprived first-years wants the answer), feasibility (step counts and a nightly rating are both collectable in a few weeks), and contribution (the activity-and-sleep link is studied broadly in adults but far less in this dorm-living, first-year group). The curiosity is now a problem you could defend. Sharpening it into the formal question — the sentence that fixes units, outcome, and reach — is the close of the next studio, after your rules and ledger exist.

The funnel from curiosity to a research problem, and the tests it has to survive, follow the standard treatment of research problems (Booth et al. 2024).

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)

# Sixty residence-hall first-years: average daily steps and a 1-10 sleep rating.
steps = rng.normal(6200, 2300, size=60).clip(1000, 15000)
sleep = np.clip(4.6 + 0.00035 * steps + rng.normal(0, 1.1, size=60), 1, 10)

group = np.where(steps > 8000, "more than 8,000 steps",
                 np.where(steps < 4000, "fewer than 4,000 steps", "in between"))
print(pd.DataFrame({"group": group, "sleep rating": sleep})
      .groupby("group")["sleep rating"].agg(["size", "mean"]).round(2).to_string())
print("\nthe question asks about the two edge groups; the middle is neither")

1.4 An AI failure case

You ask a tool to “turn my curiosity about sleep and health into a research question.” It replies, with total confidence: “Does blue-light exposure before bed reduce REM sleep in college undergraduates? Studies show blue light suppresses melatonin.” It sounds authoritative, and it is wrong for you.

It fails in two ways. First, it silently changed your scope. You never chose blue light or REM sleep, yet the question now hangs on both. That is a silent scope change, and you catch it by reading the tool’s units, outcome, and population against the ones you committed. Yours were steps and a sleep rating, not screens and REM. Second, the melatonin line is stated as settled fact with no source. Before repeating it, you find and open a real study yourself, and if you cannot, you treat it as confident fabrication and drop it.

1.5 It is your turn

You are working inside Studio 1: Begin with your curiosity. Keep what you write here; the studio’s milestone chapter is where it joins the other lessons’ pieces into one artifact you can defend.

The studio page’s opening move put your curiosity on paper. This practice stretches it, with AI welcome at the table, and lands on the research problem your project will chase.

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. Only then open the tool, so its output has something of yours to disagree with. And plan on going around more than once. The first list of candidate problems is almost never the useful one; you narrow, re-ask, and narrow again. The one thing that must not drift across those cycles is whose curiosity is being scoped.

Do not delegate

Three calls stay yours alone. Which problem you commit to. How you word the question, including its units, outcome, and population. And the judgment on whether it is important and genuinely new. A tool can widen your options and attack your choice. It cannot make the choice, because only you carry the reasons you will have to give when someone asks why this problem and not another.

Copy your opening move’s four lines here, date them, and label them version zero: the first entry of your project’s record. Then stretch before you narrow. Write two candidate directions yourself. Then brainstorm with your AI:

Stretch your curiosity (a structured brainstorm; you choose).
```
Here is my curiosity: [your one sentence]. Act as a divergence partner,
not as the author of my research question. Ask me five questions that
stretch it: nearby populations I have not named, outcomes I could watch
instead, a sharper version of the same wonder, and what answer would
genuinely surprise me. Label anything you introduce that was absent
from my sentence as NEW. Do not rank, do not choose for me, do not
claim anything is novel, and do not cite sources in this round.
```
After running, verify: answer the five questions in your own words, circle every NEW element, and reject any stretch that silently changes who, what, or why. The curiosity you carry forward must be one you would chase without the tool; this counters sycophantic agreement, where the tool inflates whatever you brought. Keep one line for each AI activity in this section: task, tool, what you kept, how you checked. One line per activity, not per chat message. The next studio opens your AI Research Ledger, and these lines become its first rows.
Take the strongest curiosity you now hold and walk it down the funnel in writing: experience, curiosity, topic, research problem. One line per stop, so you can see where the vagueness leaves. Stop at the problem; the formal question belongs to the next studio.
Run the ownership test on the problem you are converging on: would you still chase it if every AI tool disappeared tomorrow, and is the scope one you chose rather than one the tool drifted you into? Write the one-line answer to each.
Score it on importance, feasibility, and contribution, one honest sentence each. Feasibility means the access, data, and time you actually have, not the ones you wish you had.

Locate what is already known (before you claim your problem is new):
```
Act as a research-scoping assistant. My topic is [your topic]. List what is
already reasonably well established about it and where a genuine open question
might remain. For every established fact, give one real, retrievable source I
can open myself. Mark anything you are unsure exists.
```
After running, verify: open each source before you trust it. This counters confident fabrication, where a fluent answer cites a real-sounding study that does not exist.
Write a second candidate problem from the same curiosity and scope it the same way. Then choose between them, and write one sentence on why the other one lost.

Widen a curiosity into candidates you check:
```
Here is my curiosity: [one sentence]. Produce five candidate research problems
inside it. For each, name the unknown it targets, who would care about the
answer, and roughly what evidence could speak to it, in a table. Label
anything you introduce that my sentence did not contain as NEW. Do not add
any topic I did not mention.
```
After running, verify: check each row against the curiosity you actually wrote, and mark every element the table introduced that your curiosity did not contain. Keep or reject each marked element with a reason; that trail is part of your declaration. This counters silent scope change, where the tool quietly answers a nearby question and hands it back as yours.
Under the problem you kept, write your expected answer in one sentence, and name one real person or role for whom a different answer would be genuinely surprising. Keep this sentence; late in the book you will hold your evidence up against it.
Ask an AI to attack your chosen problem on all three tests, and keep only the objections that survive your own reasoning.

Red-team the problem you chose:
```
Act as a hostile grant reviewer. Here is the research problem I want to commit
months of work to: [paste it]. Argue the strongest case that it is not worth
doing. Attack it on importance, on feasibility, and on whether it adds anything
new.
```
After running, verify: keep only the objections that survive your own reasoning. This counters sycophantic agreement, where a tool praises your idea because you proposed it.
Finish by handing your problem sentence to someone who knows nothing about your project. Ask them, from that sentence alone, what is unknown, why it matters, and what evidence could answer it. You cannot run this test on yourself, because your head fills in what the sentence leaves out. Wherever they hesitate is the gap your next revision has to close. That check is called peer reasoning, and you will use it again. An AI can stand in when no one is free, but a real reader is the test, and the decision to accept or reject stays yours.

You have chosen a problem worth pursuing. Before AI performs consequential work on it, give every exchange a record and every output a standard of verification; that is the next studio’s job.

Milestone next. This was the last lesson of Studio 1. Milestone 1: Your curiosity, committed is where the lessons’ pieces become the studio’s versioned artifact. Produce it before you move on.

References

Booth, Wayne C., Gregory G. Colomb, Joseph M. Williams, Joseph Bizup, and William T. FitzGerald. 2024. The Craft of Research. 5th ed. University of Chicago Press. https://press.uchicago.edu/ucp/books/book/chicago/C/bo215874008.html.

Evans, James, Casey Petroff, and Gary King. 2026. “Advancing Science by Designing for Surprise.” Science 393 (6808). https://doi.org/10.1126/science.aej4257.

Hulley, Stephen B., Steven R. Cummings, Warren S. Browner, Deborah G. Grady, and Thomas B. Newman. 2013. Designing Clinical Research. 4th ed. Lippincott Williams & Wilkins. https://books.google.com/books?id=_b62TBnoppYC.

Zahavy, Tom. 2026. LLMs Can’t Jump. Position paper. Google DeepMind.