Research with AI without inheriting its mistakes
Using a model to find things, with a verification loop that does not cost more than the research did.
Before you start
- An assistant with web search
- A topic you need to get right
What you will be able to do
- Separate the finding step from the verifying step
- Spot a fabricated citation before it reaches your draft
- Keep an audit trail of what was checked and what was not
Used carelessly, an assistant is a very efficient way to acquire confident misinformation. Used deliberately, it is the fastest way to find out what you should be reading.
The difference is entirely in whether you let it be the source or only the scout.
Ask for the landscape, not the answer
Terminology, positions and who disagrees — all things you can verify quickly.
Open with orientation questions: what is this field called, what are the main positions, who disagrees with whom, what would I search for to find the primary sources.
These are low-risk because they are checkable in one search each, and they solve the real beginner problem — you cannot search well for something whose vocabulary you do not know yet.
Treat every citation as unverified
Fabricated references are well-formed, plausible and common.
Invented citations do not look invented. They have real-sounding authors, credible journals, plausible years, and sometimes a real author paired with a paper they never wrote.
Search each one by title. If it does not exist, do not ask the model to correct it — it will produce another. Go to a real index instead.
- Accepting a citation because the tool showed a link. A link proves a page exists, not that it says what the summary claims.
Read the source for anything load-bearing
Summaries drift most on the caveats, which is where the meaning usually lives.
For any claim your work rests on, open the actual source. Summaries are reliable on the headline and unreliable on the qualifications — sample size, population, the conditions under which the finding held.
That is precisely the material that turns a defensible claim into an overstatement, and it is the first thing a summary drops.
Tip Keep a "could not confirm" list
The things you failed to verify are information, not failure.
Maintain a short list of claims you could not stand up. Some will be real but obscure; some will be fabrications; a few will be genuinely contested, which is worth knowing and often worth saying in the piece.
Without the list, unverified claims quietly graduate into your draft on the strength of having been read three times.
Leads from the model, facts from the source, and a note of what you could not confirm. That last file is more useful than it sounds.
Common questions
Was this guide useful?
93% of readers found this useful
Read next
Edit AI drafts so they stop sounding like drafts
Readers cannot always name what makes text feel machine-written, but they notice. These are the specific tics, and most of them ar…
Write a blog post with AI without it sounding like AI
Most AI-written posts read as AI-written because of how they were prompted, not because of the model. This walks through a process…
Estimate what an AI feature will cost you per month
Per-token pricing looks trivially cheap and routinely surprises people at the invoice. The gap is almost always retries, context a…
Know when an agent is the wrong answer
A surprising share of agent projects would work better, cheaper and more reliably as a script with one model call in it.
