04 · AI for Writing & Research Tasks¶
Writing and research are two of the highest-volume use cases for AI tools, and they're often lumped together — but they carry different risks and call for different habits. Writing assistance is largely about quality and voice; research assistance is largely about accuracy and sourcing. This module treats them separately and gives you a checklist for each.
1. AI for writing tasks¶
| Sub-task | What AI tools generally do well | What to watch for |
|---|---|---|
| First drafts | Fast, unblocks a blank page | Generic phrasing; needs a real edit pass, not just acceptance |
| Editing for clarity/concision | Reliable improvement on wordy or unclear text | Can flatten a distinctive personal voice if over-relied on |
| Tone adjustment | Strong at shifting formal/casual register on request | Verify the adjusted tone still fits the actual audience |
| Outlining | Good at turning rough notes into a logical structure | Check the outline reflects your actual argument, not a generic template |
| Grammar/proofreading | Reliable for mechanical correctness | Not a substitute for a substantive edit |
A useful habit: use AI heavily for the first 60% of a piece of writing (structure, first draft, mechanical polish) and do the last 40% yourself (making sure it says what you actually mean, in a voice that sounds like you). The ratio shifts toward more human effort as the stakes rise.
2. AI for research tasks¶
Research is where the accuracy risk is highest, because research output is often quoted, cited, or acted on directly.
| Sub-task | What AI tools generally do well | What to watch for |
|---|---|---|
| Summarizing a document you provide | Reliable — it has the actual source text | Verify the summary doesn't drop an important caveat or condition |
| Explaining a general topic | Good starting orientation | Treat as a starting point, not a citable source |
| Finding and citing external sources | Only as reliable as the tool's ability to actually search and link real sources | If a tool cannot browse live sources, do not trust specific citations, quotes, or statistics it produces from memory |
| Synthesizing across multiple documents | Strong if all documents are provided directly | Check that claims are attributed to the right document when multiple sources are mixed |
| Comparative analysis (e.g., "compare these three reports") | Good structuring aid | Verify the underlying facts pulled from each report, not just the comparison logic |
3. The source-checking rule¶
This is the single highest-leverage habit in this module:
Any AI-generated claim you plan to repeat, cite, or act on must be traceable to a real, checkable source before you use it — either a document you gave the tool directly, or an independent source you check yourself.
| Situation | Rule application |
|---|---|
| You uploaded the source document yourself | Lower risk — verify the summary matches, but the source is known |
| The tool searched the live web and shows links | Medium risk — click through and confirm the link actually supports the claim |
| The tool produced a specific fact, quote, or citation "from memory," with no link | High risk — treat as unverified until you find it independently; do not repeat it as-is |
Worked example¶
A graduate student is writing a literature review section. She gives an AI writing assistant three PDF papers she's already read and asks it to draft a paragraph comparing their methodologies — a research task done well, because the source material is directly provided and checkable. She reads the draft against the actual papers and catches one place where the assistant slightly overstated how similar two methods were; she corrects the phrasing.
Separately, she asks the same tool (without web search enabled) "what did [a specific well-known researcher] conclude in her 2019 paper on this topic?" The tool answers confidently and specifically. Because this claim came from the tool's memory rather than a provided or linked source, she treats it as a lead, not a fact — she goes and finds the actual 2019 paper before citing anything about it. It turns out the paper's actual conclusion was more nuanced than the AI's summary suggested.
How It Actually Works¶
Writing assistance and research assistance stress different parts of the same underlying model, which is why they call for different verification habits. For writing tasks — rephrasing, tightening, matching a tone — the model is doing something it's structurally well-suited to: transforming one sequence of text into another sequence of text, guided entirely by patterns learned from enormous volumes of human-written prose. There's no external fact against which a rewritten sentence needs to be checked; "does this read well and preserve meaning" is a judgment the model's training data (itself full of edited, polished prose) equips it to approximate reasonably well.
Research tasks stress a different, weaker part of the same mechanism: the model's "knowledge" of specific facts, dates, and sources is not a lookup table — it's a byproduct of patterns compressed into billions of numeric weights during training. A fact that appeared often, consistently, and recently in training data is usually reproduced accurately; a fact that's rare, was inconsistently reported, or postdates the model's training cutoff is not reliably retrievable at all — the model instead generates the most statistically plausible answer, which can be indistinguishable in tone from a well-supported one. This is why "cite your source" is a much harder ask than it sounds: unless the tool is explicitly doing retrieval (searching the live web or a document store and quoting from what it found), any citation the model produces on its own is itself just a plausible-sounding piece of generated text — it may name a real, wrong, or entirely fabricated source with equal confidence. Tools that show real citations are running a retrieval step underneath the generation, not asking the base model to "remember" where a fact came from.
Exercise¶
Pick a real writing or research task you have coming up. If it's writing: draft it with an AI tool, then explicitly identify which 60% you let the tool drive and which 40% you did yourself, per section 1. If it's research: give the tool a real source document and ask it to summarize or synthesize a specific claim from it, then verify that claim against the source directly and note any discrepancy, however small. Either way, write one sentence stating which category (provided-source, live-search, or from-memory) any specific fact in your output falls into, using the table in section 3.