Session 3: AI-Assisted Research & Work
Using these tools for research and real work without being misled: summarising and drafting where they help, verifying everything that matters, and handling citation and sources honestly.
Learning objectives
By the end of this session you will be able to do each of these without prompting.
- Use AI for summarising and drafting in ways that keep you in control of the content
- Build a verification routine that catches fabricated facts and sources
- Handle citation properly, including when a tool's source does not exist
- Integrate AI into a real work task end to end with checks at the right points
- Know which parts of research should never be delegated
The taught content
The division of labour that makes research safe
The useful split is not 'AI does research, you check it'. It is narrower and more practical: the tool handles language and structure; you handle facts and judgement. Summarising a document you have, restructuring your notes, drafting a section from an outline you supplied, rephrasing for a different audience — all language work, all verifiable by you quickly. Anything involving a specific fact, figure, date, source or claim about the world is yours.
That split sounds restrictive and it is not, because the language work is most of the time. The reason the division matters is asymmetric risk: a badly written paragraph costs you ten minutes, while a fabricated citation in a client report or a student dissertation can cost you your standing. Optimising for speed on the wrong side of that line is a poor trade.
So the working rule for this session is: let the tool work with material you have already established as true. Give it your notes, your sources, your data, and have it reorganise and express them. Do not ask it what is true and then check afterwards, because you will not check as carefully as you think you will — a plausible sentence reads as verified even when it is not.
Summarising: powerful, and dangerous in one specific way
Summarising is one of the strongest uses, and the safe form is summarising a document you provide. The model works from your text, so nothing needs to be invented, and because you know the source you can immediately see what it captured and what it dropped. This is genuinely useful and genuinely low-risk.
The dangerous form is summarising something you have not read, or asking it to summarise a topic rather than a document. Now there is no source in front of you, and the output may be a fluent synthesis of plausible-sounding material rather than a summary of anything real. The operation looks identical; the risk is completely different, and the difference is invisible unless you notice which one you are doing.
Two habits keep summarising useful. Ask for what is missing as well as what is present — 'what does this document not address?' turns a summary into a gap analysis, which is more valuable and forces you to engage with the source. And keep the source open beside the summary while you use it, so any claim can be traced back in a second rather than trusted from memory.
Drafting: the tool writes, you own the content
Drafting works well when you supply the substance. Give it your points, your data, your argument, and have it produce prose — that is transformation work, and it is fast. What fails is asking it to write about a subject and then adopting its framing, because its framing is generic and its specifics are unverified.
The practical method is to draft in the order that keeps you in control. Outline first, yourself — the argument, the sequence, the evidence. Then draft section by section, not all at once, because a section at a time is checkable and a whole document at once is not. Then edit for accuracy and voice, which is where you add the specificity that makes it yours and removes the padding that makes it recognisably generic.
There is an honest point here about ownership. Text you have outlined, supplied the substance for, verified and rewritten is your work — you used a tool the way you would use a spellchecker or a thesaurus. Text you accepted without reading carefully is not your work and will not survive scrutiny, in a classroom or a client meeting. The boundary is not which tool you used; it is whether you understand and stand behind what was produced.
Verification: the routine, not the intention
Everybody intends to check. The routine is what makes checking actually happen, and it is short. Every specific claim gets verified against a source you can name. Names, dates, figures, statistics, legal references, product versions, quotations — anything that could be wrong and would matter if it were. A claim you cannot verify does not go in, however plausible it reads.
Then the citation rule, which is absolute: search for every source before you cite it. A tool can produce a reference with a real-sounding author, a plausible journal and a specific year that does not correspond to any actual publication. It is not lying; it is producing the shape a citation has. Searching takes thirty seconds and it is the difference between a reference and a fabrication that ends up in a document with your name on it.
Two supporting habits. Check quotations verbatim — a paraphrase presented in quotation marks with a source attached is a misquotation, and it is a serious one in academic work. And verify the most load-bearing claim first, because if the central fact is wrong the rest of the piece is wasted effort. Verification is cheaper in proportion to how early you do it, which is an argument for checking as you draft rather than at the end.
What not to delegate, and why the list is short
The list of things that should stay with you is shorter than people fear, and it is consistent. Establishing what is true — because the tool cannot and will not tell you it is guessing. Judgement about your specific situation — your client, your contract, your local context, where generic advice misses the constraint that decides the question. Anything you cannot verify quickly, because unverifiable output cannot be safely used at all.
Then the categories where the cost of error is high enough that delegation is never worth it regardless of speed. Legal, medical, financial and regulatory specifics — a plausible answer in these areas can cause real harm, and the appropriate response is a professional, not a language model. Anything about a named person or organisation that could be defamatory or simply wrong. And original analysis you are being paid or graded for, because that is the thing being assessed, not the prose it is written in.
The honest framing is that this is not a restriction on the tool; it is an accurate description of what it is. A prediction engine with no access to truth and no understanding of your situation should not be deciding facts, judging your circumstances, or answering questions where being wrong is expensive. Everything else — the drafting, the restructuring, the summarising, the option generation — is genuinely faster with help, and there is a great deal of it.
Instructor demonstration
We run a real research-and-writing task end to end with the tool doing language work and us doing facts — including deliberately catching a fabricated citation and building the verification routine around it.
- 01
Pick a real piece of work with a factual component
A client proposal, a report, an article, an assignment. It must contain claims that could be checked, because the whole session is about where the checking happens.
- 02
Do the research yourself first, from real sources
Find and read the actual sources before involving the tool. This order matters: established facts go in, and the tool expresses them. The reverse order is how fabrications enter a document.
- 03
Take notes in your own words with sources beside each point
Every fact gets its source noted at the moment you record it. Retrofitting sources later is when things get attributed loosely, and loose attribution becomes a real problem if anyone checks.
- 04
Have the tool summarise a source you have read
Paste the document and ask for a summary plus what it does not address. Because you know the source, you can verify immediately — which is the condition that makes summarising safe.
- 05
Compare the summary against the source and note the gaps
What did it drop, and what did it flatten? Summaries lose nuance predictably, and knowing which nuance matters is your judgement rather than the tool's.
- 06
Ask it to summarise a topic it was not given text for
Observe that the output is fluent and confident with no source in front of either of you. This is the same operation with a completely different risk, and seeing the difference is the point.
- 07
Write your own outline before any drafting
The argument, the sequence, and the evidence for each section. Supplying the structure yourself is what keeps you in control of the content rather than adopting a generic framing.
- 08
Draft one section at a time, supplying your substance
Give it your points and data for that section only. A section is checkable; a whole document at once is not, and unread output is output you do not actually own.
- 09
Edit each section for accuracy and voice immediately
Add the specifics that make it yours and remove the padding that makes it generic. Do this per section rather than at the end, while the source is still in front of you.
- 10
Ask it for a citation and search for the source
Request a reference for one of your claims, then search for the paper. When you find it does not exist — or exists with different authors and a different year — you have seen the failure that matters most.
- 11
Adopt the absolute rule: search before you cite
Every source, every time, thirty seconds. It is producing the shape a citation has, not retrieving one, and the only check that works is looking it up yourself.
- 12
Check any quotation verbatim against the original
A paraphrase in quotation marks with a source attached is a misquotation. If you cannot find the exact wording in the source, it is not a quotation and should not be presented as one.
- 13
Verify the most load-bearing claim first
If the central fact is wrong, everything built on it is wasted. Checking early is cheaper in proportion than checking at the end, which is an argument for verifying as you draft.
- 14
Check every figure and statistic against its original source
Not against the tool's restatement of it. Numbers are where confident error is most damaging, because a wrong figure in a client report is difficult to walk back credibly.
- 15
Ask for objections to your argument
This is a strong use that requires no factual trust: generating counter-arguments you then evaluate. Use it to strengthen the piece rather than to find agreement.
- 16
Do a final read as though you were the sceptical reader
Every claim you would challenge if someone else wrote it, check. This read is the last line, and it works only if you have not already convinced yourself the output is fine.
- 17
Record the verification steps as part of the workflow
Write down what you checked and how. The deliverable of this course is a documented workflow, and the verification steps are its most important part.
- 18
Note what you kept manual and why
Establishing facts, judging your situation, and anything unverifiable. Stating this explicitly is what makes the workflow a workflow rather than an abdication.
Guided practice
Run a real research task end to end with verification built in
Complete a genuine piece of work containing factual claims, using the tool for language and structure while you establish and verify every fact, and document the workflow.
- 01Choose a real piece of work with a factual component that could be checked.
- 02Do the research yourself from real sources before involving the tool at all.
- 03Take notes in your own words with the source recorded beside each point.
- 04Have the tool summarise a document you have read, asking also what it does not address.
- 05Compare the summary against the source and note what it dropped or flattened.
- 06Ask it to summarise a topic without providing text, and observe the difference in risk.
- 07Write your own outline — argument, sequence, evidence — before any drafting.
- 08Draft one section at a time, supplying your substance rather than asking it what to say.
- 09Edit each section immediately for accuracy and voice while the source is in front of you.
- 10Ask for a citation, search for the source, and record whether it exists.
- 11Adopt the rule that every source is searched before it is cited.
- 12Check any quotation verbatim against the original wording.
- 13Verify the most load-bearing claim first, before building further on it.
- 14Check every figure and statistic against its original source rather than a restatement.
- 15Ask for objections to your argument and evaluate them yourself.
- 16Do a final read as a sceptical reader, checking anything you would challenge.
- 17Record every verification step as part of your documented workflow.
- 18Write down what you kept manual and the reason for each.
The standard we hold you to
Research completed from real sources before the tool was involved, with notes carrying a source beside each point; summarising performed on documents already read with gaps compared against the source, and the higher-risk unsourced-summary form identified; an original outline written before drafting with sections drafted and edited one at a time; **a citation requested, searched for, and its existence or absence recorded**; every source searched before citation; quotations checked verbatim; the load-bearing claim verified first; every figure checked against its original source; objections generated and evaluated; a sceptical final read completed; and a documented workflow recording every verification step plus what was kept manual and why.
Common mistakes and how to fix them
Asking the tool what is true and checking afterwards
Fix: Let it work with material you have already established as true. A plausible sentence reads as verified even when it is not, and you will not check afterwards as carefully as you intend to.
Summarising something you have not read
Fix: The output may be a fluent synthesis of plausible material rather than a summary of anything real. The operation looks identical to safe summarising, so notice which one you are doing.
Citing a source you have not searched for
Fix: Search every one, every time. It is producing the shape a citation has, not retrieving one — and a fabricated reference in a document with your name on it is a serious and very public failure.
Presenting a paraphrase as a quotation
Fix: Check the exact wording in the source. If it is not verbatim it is not a quotation, and in academic work a misquotation with a source attached is a significant integrity problem.
Accepting figures from a restatement rather than the original
Fix: Trace every number to its primary source. Wrong figures are the most damaging confident error, because they are hard to walk back credibly once a client has seen them.
Drafting a whole document at once and then editing
Fix: Draft section by section. A section is checkable and a document is not, and unread output is output you cannot honestly claim as your work.
Adopting the tool's framing of a subject
Fix: Outline the argument yourself first. Its framing is generic and its specifics unverified; supplying your own structure keeps the content yours.
Verifying only at the end
Fix: Check as you draft, starting with the load-bearing claim. If the central fact is wrong everything built on it is wasted, and early checking is cheaper in proportion.
Expert notes
The habits that separate someone who can do this from someone who does it well.
- Let the tool work with material you have already established as true. Give it your notes, sources and data and have it reorganise and express them. Asking it what is true and checking later fails because a plausible sentence reads as verified, and you will not check as carefully as you intend.
- Search for every citation before it goes in a document. Thirty seconds, every time, without exception. Having found one fabricated reference yourself changes how you read all subsequent output more effectively than any warning could.
- Draft section by section rather than a document at a time. A section is checkable while the source is in front of you; a whole document invites a skim-read, and unread output is output you cannot honestly stand behind.
- The do-not-delegate list is short and stable: establishing what is true, judging your specific situation, and anything you cannot verify quickly — plus legal, medical, financial and regulatory specifics where the cost of error is too high for any speed gain to justify.
Key terms
- Source-grounded work
- Using the tool only with material already established as true — your notes, sources and data — so it reorganises and expresses rather than invents.
- Fabricated citation
- A reference with a plausible author, journal and year that corresponds to no actual publication. Detected only by searching for the source.
- Verbatim check
- Confirming quoted wording appears exactly in the source. A paraphrase in quotation marks with an attribution is a misquotation.
- Load-bearing claim
- The central fact an argument depends on. Verified first, because if it is wrong everything built on it is wasted effort.
- Primary source
- The original document a figure or claim comes from, as against a restatement of it. The only acceptable basis for verifying a number.
- Section drafting
- Producing one section at a time and editing it immediately. Keeps output checkable, whereas a whole document invites a skim-read.
- Gap analysis
- Asking what a document does not address, which turns a summary into a more useful exercise and forces engagement with the source.
- Verification routine
- A stated, repeatable set of checks applied to delegated output. An intention to check is not a routine, and only a routine reliably happens.
Homework before the next session
Complete a real research task with the workflow
Research from real sources first, draft section by section from your own outline, edit for accuracy and voice, and do a sceptical final read. Keep the notes with sources beside each point.
Run the citation test and record the result
Ask for five references on a topic you know, search for each, and record how many exist as described. This single exercise recalibrates how you read everything these tools produce.
Write your verification routine
A short list of checks you always run: search every source, check quotations verbatim, trace figures to primary sources, verify the load-bearing claim first. Written as steps, not as an intention.
List what you will never delegate
With a reason for each. Include establishing facts, judging your own situation, anything unverifiable, and the high-cost categories where being wrong is expensive regardless of speed.
Assessment rubric
How this session is marked. The certificate for AI Productivity is awarded on the deliverable, not on attendance.
| Criterion | Passing | Excellent |
|---|---|---|
| Source discipline | Used sources in the work. | Research completed from real sources before the tool was involved, with a source recorded beside each point at the moment it was noted. |
| Summarising judgement | Summarised documents. | Summarised only material already read, compared output against the source, used gap analysis, and can explain why unsourced summarising carries different risk. |
| Drafting control | Used the tool to draft. | Wrote the outline personally, drafted section by section supplying own substance, edited immediately for accuracy and voice, and can articulate the ownership boundary. |
| Verification | Checked some facts. | Every source searched before citation, quotations checked verbatim, figures traced to primary sources, load-bearing claim verified first, and a written routine rather than an intention. |
| Delegation boundaries | Knows some tasks are unsuitable. | Can state the do-not-delegate list with reasons, including the high-cost categories, and frames it as an accurate description of what the tool is rather than a limitation on it. |
Session questions
How often are citations actually fabricated?+
Often enough that the rule must be absolute. Run the test yourself: ask for five references on a topic you know and search for each. The result recalibrates how you read everything these tools produce, far more effectively than being told.
Is it acceptable to use AI for an assignment?+
That depends on your institution's rules, which you must check first. As a principle: outlining, supplying substance, verifying and rewriting is using a tool; accepting output you have not read carefully is submitting work that is not yours, and it will not survive scrutiny.
Can I trust a summary of a document I have not read?+
Not safely, because you have no way to detect what was dropped, flattened or invented. If you must work from an unread document, have it summarise the text you paste, and keep the source open beside the summary so any claim can be traced.
What about figures and statistics it gives me?+
Trace every number to its primary source rather than accepting a restatement. Wrong figures are the most damaging kind of confident error, because they are hard to walk back credibly once somebody has relied on them.
How much verification is enough?+
Every specific claim that could matter if wrong, every source before you cite it, every quotation verbatim, and the load-bearing claim first. That is a finite list and it takes far less time than fixing a public error.
This session is part of
AI Productivity
2 weeks · 4 sessions · ₦30,000 · you leave with an ai-assisted workflow you actually use