Data Science, Econometrics, and Research in the Age of AI

Producing with AI, verifying with discipline

LearnNotebookLM

→

ExploreGoogle Colab

→

ResearchGitHub

Carlos Mendez

Nagoya University

2026

The New Trade-off in the Age of AI

The problem

Producing content has never been cheaper

AI writes the essay, the code, and the chart in seconds — for almost nothing.

But can we trust what it produced?

Open on the shift everyone already feels: production is nearly free. The provocation — cheap to make is not the same as safe to use. That gap is the whole talk.

Cheap to produce, expensive to verify — the new danger zone

Cost of verification

highlow

NEW DANGER ZONEcheap to make,
expensive to confirm

Costly all aroundslow to make and
hard to check

Trivialeasy to make,
easy to check

OLD WORLDmanual effort
embedded the checking

lowhigh

Cost of production

AI slashes production cost — but not verification cost — so work migrates into the danger zone.

Adapted from Clifford Russel. Old world (bottom-right): making something by hand was costly, and that manual effort embedded the checking. AI moves us to the top-left: production collapses toward zero, but verification cost stays high — now it is the binding constraint. The rest of the talk is about tools that pull verification back down.

The real risk is black-box work — in learning and in research

When we can’t check the output, we’re trusting a black box.

• In learning — plausible answers that quietly mislead.
• In research — results no reviewer, including you, can audit.

The fix isn’t to avoid AI. It’s to design a workflow that separates production from verification.

Name the stakes for both halves of the talk. Black-box learning erodes understanding; black-box research erodes credibility. Bridge to the response: intentional workflow design.

One principle, three tools, one workflow

• One principle — separate what AI produces from what you verify.
• Three tools — NotebookLM (learn), Google Colab (explore), GitHub (research).
• One workflow — plan, produce with AI, verify at every step.

The roadmap. Each tool maps to a stage of the research life cycle and to a place where verification has to live. Keep it a promise, not a summary.

NotebookLM — Learn

Tool 1

NotebookLM grounds AI in your sources — a protected study environment

You give it your notes, PDFs, slides, and data. It answers only from those sources.

Grounded in your materials → far less hallucination, and every answer is traceable.

The key idea for a general audience: unlike an open chatbot, NotebookLM is walled to the documents you upload. That protected environment is exactly what makes it safe for study — answers stay tied to trusted sources.

One set of notes → podcasts, videos, summaries, quizzes, and a grounded chatbot

Your sourcesnotes · PDFs · slides · data

→

🎙 Podcast 🎬 Video overview 📝 Summary ❓ Quiz 💬 Grounded chat

One upload becomes many ways to learn — the AI podcasts and AI tutors behind metricsAI work exactly this way.

NotebookLM is the content-transformation layer: a single source set fans out into audio, video, text, self-test, and conversation.

In practice: NotebookLM “Deep Dives” you can listen to

Every chapter of the metricsAI project becomes an AI-generated audio discussion — published as a podcast on Spotify.

Listen on Spotify

Concrete example: the metricsAI podcast series is NotebookLM “Deep Dives” — two AI voices discussing each chapter. Students listen on their commute; the source is always the chapter, so it stays grounded.

Google Colab — Explore

Tool 2

Colab: a data-science lab in the browser, with AI inside

• zero installation — nothing to set up
• free CPUs and GPUs
• AI inside — it writes code and explains errors

Start on your laptop at a café; the lab is the same everywhere.

Colab removes the single biggest barrier for beginners — installation. For a general audience, stress: it is just a browser tab, but it is a real computational notebook with AI help built in.

metricsAI: a full econometrics course in Colab notebooks

Every chapter ships as a Google Colab notebook — code, output plots, and an AI assistant beside it.

Course: quarcs-lab.github.io/metricsai  ·  Example: Chapter 1 in Colab

This is what “a lab in the browser, with AI inside” looks like: a Colab notebook running the chapter’s code, with plots inline and an AI Summary/assistant panel. No install — just open the link.

With the right credentials, Colab reaches planetary-scale data

Connect Google Earth Engine and query satellite data straight from a notebook:

• pollution — aerosols, air quality
• energy — nighttime lights
• social — population and development indicators

Big geospatial data, no supercomputer — just a notebook and credentials.

This is the big-data turn the workshop emphasizes. Earth Engine hosts petabytes of satellite imagery; from Colab you query only what you need. Pollution, energy, and social indicators all become variables you can analyze.

Exploration in Colab becomes a research project — managed in GitHub

A notebook is where discovery happens. Now make the work reproducible and reviewable.

A promising exploration → a project anyone can re-run and audit → GitHub.

Bridge to Tool 3. A Colab notebook is where discovery happens; GitHub is where it becomes credible research. This is the natural hand-off.

GitHub — Research

Tool 3

GitHub makes AI-assisted research transparent and auditable

Collaborate with AI agents the way you would with a careful coauthor — in the open.

Claude Code + GitHub + Overleaf: code, results, and manuscript, all version-controlled.

The research counterpart to skills-and-planning. When an AI agent proposes changes through GitHub, every decision is recorded and reviewable.

Everything lives in one repository

The responsible-ai-assisted-research-101 template: analysis code, data, the manuscript, and an agent operating manual — all versioned in one place.

github.com/quarcs-lab/responsible-ai-assisted-research-101

A concrete template (github.com/quarcs-lab/responsible-ai-assisted-research-101): numbered analysis scripts, immutable raw data, code-emitted tables/figures, a DECISIONS.md and LOG.md, and a CLAUDE.md operating manual for the agent. One place, fully reproducible.

Change the unit of verification: review a story step-by-step

Not a monolithic block of text — a sequence of small, legible changes.

• each commit is one decision
• each diff shows exactly what changed
• the history is the audit trail

Goldsmith-Pinkham’s framing. Reviewing a 40-page output is hopeless; reviewing 40 small commits is tractable. Version control changes what you verify — from a result to the chain of steps that produced it.

800 lines to trust, or 12 to check?

Left: an 800-line dump you’d have to trust wholesale. Right: a bounded 12-line diff and the commit story you actually review.

The unit of verification, made visual: a wall of 800 lines is cognitive overload; a 12-line bounded diff plus a readable commit history (validate measures, add CAR analysis, add stages…) is something a human can actually check, one decision at a time.

A work environment built for verification — the antidote to black-box research

GitHub manages the discussion of the code, the results, and the story of the manuscript.

Every claim traces back to the commit that produced it.

Tie back to the opening trade-off. Black-box research is unverifiable by construction; a version-controlled environment makes verification the default. And in the age of AI the barrier is low — your AI agent can teach you Git as you go.

But doesn’t AI just do the work for me?

Objection. If the AI produces everything, what is left for the researcher?

Response. AI produces; you design, verify and decide. The judgment — what question, which check, whether to trust the result — stays human. The tools make that judgment easier to exercise, not optional.

The ethos slide. Steelman the worry honestly, then answer: the division of labor is production (AI) versus verification and judgment (human). That is the whole thesis in one exchange.

One Integrated Workflow

Putting it together

Learn → Explore → Research: NotebookLM, Colab, GitHub in one loop

LearnNotebookLM

→

ExploreGoogle Colab

→

ResearchGitHub

↻ iterate — and verify at every step

The payoff picture. Understand a topic with NotebookLM, explore and analyze data in Colab, then formalize it as reproducible research in GitHub — looping back as you learn more. Verification runs through all three.

Three takeaways

1. Understand the trade-off — AI cuts production cost, so verification becomes the job.
2. AI assists both learning and research — but only with verification built in.
3. Integrate the tools — NotebookLM, Colab, and GitHub together serve both production and verification.

Keep it to three; each maps to one act of the talk.

In the age of AI, the scarce skill is not production — it’s verification.

The single sentence to remember. It resolves the Act-I hook: production got cheap, so the value — and the discipline — moves to verification and judgment. That is what these three tools are for.