How AI Fits the Research Workflow in 2026

By 2026, artificial intelligence has quietly become part of almost every stage of academic work — but not in the way the hype suggests. It does not do research for you. It does not know whether a claim is true. It has never read your dataset, sat in your lab, or understood your field the way you do after years of study. What it does well is accelerate the supporting work around research: turning a blank page into a first draft, compressing a fifty-page report into a page you can scan, rephrasing a clumsy sentence, and helping you see the shape of an argument before you commit to it.

This course teaches you to use AI as a scholarly co-pilot across the research lifecycle — from finding literature to disseminating results — without ever crossing the lines that matter in academia. Those lines are not decoration. Cross them and you risk your degree, your funding, your publications, and your reputation. Respect them and AI becomes one of the most useful assistants a researcher has ever had.

The research lifecycle, stage by stage

Think of your work as a cycle. AI can support most stages, but its role is different in each.

1. Question and design. Brainstorming research questions, pressure-testing a hypothesis, mapping what a methodology involves. AI is a sparring partner here, not an authority.
2. Literature discovery. Finding relevant papers, understanding an unfamiliar subfield, mapping how ideas connect. Specialized tools help, but every result must be verified at the source.
3. Reading and synthesis. Summarizing dense papers, extracting arguments, building a synthesis across many sources, taking structured notes.
4. Analysis. Conceptual help with code for data analysis, explaining a statistical method, debugging a script. AI never invents data or results.
5. Writing. Structuring a manuscript, improving clarity, tightening prose — especially valuable for non-native English speakers.
6. Funding. Structuring grant proposals, tailoring language to a funder, clarifying aims.
7. Dissemination. Presentations, posters, plain-language summaries, and reading peer review more effectively.

Notice a pattern: AI is strongest at structure, language, and speed of a first pass, and weakest at truth, novelty, and judgment. Your scholarly value lives in the second set. Guard it.

What genuinely improves with AI

Let us be precise, because vague promises are how people get burned.

• The blank-page problem disappears. You always have a rough draft to react to, which is far easier than creating from nothing.
• Comprehension speeds up. A dense methods section becomes navigable when AI gives you a plain-language map first — which you then read the original to confirm.
• Language stops being a barrier. A brilliant result described in awkward English gets the clear, precise expression it deserves.
• Iteration is cheap. You can restructure an argument five different ways in minutes and pick the strongest.

What does not change — and must not

Here is the part most AI-hype skips.

• You are fully accountable for every word you submit. Not the model. You. If an AI-inserted citation is fake and it reaches a reviewer, that is your error, not the tool's.
• AI hallucinates confidently. It will invent references, DOIs, quotes, and page numbers that look perfect and do not exist. This is not an edge case; it is a defining property. We will return to it constantly.
• Integrity is non-negotiable. AI is a drafting and thinking aid, not a ghostwriter. Presenting AI-generated scholarship as your own original work — where that is prohibited — is misconduct.
• Policies vary and you must check them. Your institution and each target journal or funder has its own rules on AI use and disclosure. There is no universal rule; there is your rule for your venue, which you look up.

The one principle to carry through the whole course

AI helps you think and write better. It never knows the truth, and it is never the author of your scholarship.

Every prompt and workflow in this course is built on that sentence. When AI helps you turn a tangled paragraph into a clear one that still says exactly what your data shows, that is good scholarship made more legible. When AI supplies a citation you did not verify, invents a statistic, or drafts a "finding" your data does not support, that is a path to retraction and disgrace.

A first prompt you can use today

This warm-up prompt does not write anything for you. It helps you map your own project so everything downstream is grounded in your real work.

You are a thoughtful research mentor. Ask me one question at a time to
help me articulate my current research project. Start with my core
research question. After each answer, ask a natural follow-up to draw
out specifics: my discipline, my method, my data or sources, my stage,
and what I am stuck on. Do not propose findings, do not invent
citations, and do not write anything for me yet. At the end, summarize
back the real facts I gave you so I can confirm they are accurate.

Notice the shape: a clear role, questions that surface your real material, and an explicit ban on invention. That pattern — role, task, integrity guardrail — repeats in every prompt you will learn here.

A worked example: one week in a doctoral project

Abstract principles stick better with a concrete picture, so walk through a realistic week for a second-year PhD student in public health — call her Ana — who uses AI well. Nothing in this scenario involves the model producing scholarship on her behalf; every use accelerates work she then verifies and owns.

Monday — orientation. Ana is entering an adjacent subfield (health misinformation on social platforms) for one chapter. She asks a general assistant such as Claude or ChatGPT for a plain-language map of the subfield: main debates, typical methods, key terminology. She treats the answer as a tourist map, not a land registry — useful for orientation, guaranteed to contain simplifications and possibly errors. She writes down the terms it surfaced and takes them to Semantic Scholar and her library databases, where the real literature lives.

Tuesday — triage. Her saved database alerts delivered fourteen new papers. She pastes abstracts into her assistant and asks, for each: research question, method, population, main claimed finding — in four lines. Twenty minutes later she knows which three papers deserve a full read. She reads those three herself, in the original, because a summary of an abstract is a filter, never a substitute.

Wednesday — writing. She drafts a background section in her own words, quickly and badly — her honest, ugly first draft. Then she asks the model to critique it: where is the argument unclear, where do paragraphs repeat, which transitions fail? She revises herself, sentence by sentence. The ideas, citations, and claims are all hers; the model acted as a tireless reader, not a writer.

Thursday — analysis support. Her R script for a mixed-effects model throws a convergence warning. She pastes the code and the warning (no participant data — only the code) and asks for an explanation of what the warning means and which modelling choices typically cause it. The model explains three plausible causes. She checks each against the package documentation before changing anything, because model explanations of statistical software are sometimes confidently outdated.

Friday — integrity check. Before sending the chapter draft to her supervisor, she opens every citation and confirms it exists, says what she claims, and is correctly formatted in Zotero. She notes in her research log which AI tools she used and for what, so that when her university or a target journal asks about AI assistance, she can answer precisely instead of reconstructing from memory.

Total AI time: perhaps ninety minutes across the week. Time saved: plausibly a full day. Integrity risk added: essentially zero — because at every step, the model touched process, and Ana kept custody of truth.

Decision table: what to delegate, what to keep

A practical way to internalize the boundary is a standing decision table. Keep something like this pinned above your desk and adapt it to your field.

Two rows deserve emphasis. The citation row is absolute: in this course you will never ask a general chatbot for references, because fabricated citations are the single most common way researchers get publicly burned by AI. The confidentiality row is equally hard: material entrusted to you under peer-review confidentiality or containing personal data must not be pasted into consumer AI tools, full stop.

Five failure patterns to recognize from day one

Most AI-related academic disasters follow one of five patterns. Learn to name them now; the rest of the course teaches the antidotes.

1. The outsourced mind. The researcher stops reading originals and works only from summaries. Quality degrades invisibly until a reviewer asks a question the summaries never covered.
2. The confident fabrication. A generated reference, statistic, or quote looks perfect, gets pasted, and is never checked. Discovery — by a reviewer, an examiner, or a reader with a database — is humiliating and sometimes career-ending.
3. The silent ghostwriter. Substantial AI-generated text is submitted as original work where policy forbids it. Even when detection tools fail, collaborators talk, drafts get compared, and misconduct processes exist.
4. The leaked secret. Unpublished data, a colleague's manuscript under review, or identifiable participant information is pasted into a tool whose terms allow reuse. The breach happens at upload, not at discovery.
5. The atrophied skill. A doctoral student lets AI do all the summarizing and structuring during exactly the years those muscles are supposed to develop. The degree arrives; the competence does not.

None of these patterns is caused by the technology itself. Each is a human decision to skip verification, hide assistance, or bypass confidentiality. Which means each is fully preventable by the habits this course drills.

Your integrity baseline: a setup checklist

Before the next lesson, spend fifteen minutes establishing your baseline:

• Locate your institution's current policy on AI use in research and coursework. Save the link; policies are updated frequently, so you will re-check it per project.
• Identify whether your target journals or funders publish AI-use and disclosure policies, and bookmark those pages rather than trusting summaries of them.
• Start a simple AI-use log: a running note of which tool you used, for what task, on which project. Two lines per use is enough.
• Decide your personal red lines in writing. At minimum: no unverified citations, no fabricated content, no confidential material in consumer tools.
• Install or update a reference manager (Zotero or Mendeley) — from Lesson 10 onward it becomes your single source of citation truth.

This checklist looks bureaucratic and takes a quarter of an hour. It is also, quite literally, the difference between researchers who use AI for years without incident and researchers who explain themselves to an integrity committee.

How this course is organized

You will move through the research lifecycle in order: foundations and integrity, literature discovery, reading and synthesis, citations and the hallucination problem, research questions and design, academic writing, grants, data-analysis coding, and dissemination and career. Each lesson gives you ready-to-use prompts, worked examples, and workflows, and each one keeps the integrity spine visible.

Let us begin where every responsible use of AI in research must begin: with a clear-eyed understanding of the tools, and an unbreakable commitment to integrity.