Skip to content
AIupick

Guide for researchers · 6 jobs

AI for research: find papers, read faster, check every claim

AI can now search, screen and summarise the literature in minutes. The hard part is the same as before: knowing what the evidence actually says, how strong it is, and whether you could defend each step to a reviewer. Each job below leaves a record you can check.

  2. 02

    Screen thousands of abstracts, in a way you can defend

    Screening a few thousand abstracts by hand takes weeks, and screening by AI alone is hard to defend to a reviewer. Write the criteria first, let AI screen against them, check its decisions on a sample, and report how you checked.

    Example

    2,300 abstracts on exercise and depression. AI screening leaves 140 to read in full. On a random 100 you also screened yourself, you agree on most; every disagreement is an "unsure", and you read all of those.

    How to do it

    1. 1Write the inclusion and exclusion criteria before you search, and keep the date.
    2. 2In Elicit, choose Systematic Review, enter your question on the Setup page, then gather and screen.
    3. 3Screen a random sample of 50 to 100 abstracts yourself and compare your decisions with the AI's.
    4. 4In your methods, report the criteria, the tool, the size of the sample and how often you agreed.

    Built in

    Prompt to copy

    Screen the abstracts below against these criteria.
Include: [population, intervention, design, years]. Exclude: [criteria].
For each abstract: include, exclude or unsure; the criterion that decided it; and the sentence of the abstract it rests on. When the abstract does not say, answer unsure.

    Where it goes wrong

    • Criteria written after seeing the results bend toward what you found. Fix them first.
    • "Unsure" is a valid answer. Read every unsure abstract yourself.
  3. 03

    Ask a stack of PDFs where they disagree

    The useful questions across many papers are about differences: where they disagree, which methods differ, what nobody measured. Ask for those, with a quote and a page for every claim, and ask what the sources do not answer.

    Example

    Twelve papers on screen time and sleep in teenagers. They disagree because some measured all screen time and others only screen time after 9 pm. That difference, with a quote from each side, is the paragraph your review was missing.

    How to do it

    1. 1Add the PDFs to a notebook in Gemini Notebook, or use Chat with PDF in SciSpace for one paper at a time.
    2. 2Ask the comparison questions in the prompt on this page.
    3. 3Open the cited passage for every claim you will use.
    4. 4Make a Mind Map to see the topics across the set and spot what is missing.

    Built in

    Prompt to copy

    Across these sources: where do they disagree, and why (different populations, methods, measures or years)?
For each point, quote the passage and give the source and page.
Then list the questions these sources do not answer.

    Where it goes wrong

    • An answer that blends two papers can give one paper's result to another. The quote shows which is which.
    • "Not in the sources" is a useful answer. Be wary if it never says it.
  4. 04

    Transcribe interviews and use AI as a second coder

    AI transcribes and suggests codes fast, but a codebook it invents fits the data too neatly. Code a sample yourself first, then let AI apply your codebook and look at where you disagree.

    Example

    Twenty interviews with nurses about night shifts. You code three by hand into a codebook of eight codes; AI applies it to the other 17. On the one interview you both coded you disagree on six passages, which shows two codes overlap and need sharper definitions.

    How to do it

    1. 1Check that your ethics approval and consent forms allow the tool, and remove names before uploading.
    2. 2Transcribe, then correct the transcript against the recording for every passage you will quote.
    3. 3Code two or three interviews yourself and write the codebook.
    4. 4Let AI apply your codebook to the rest with the prompt on this page, and compare its coding with yours on one interview.

    Prompt to copy

    Apply this codebook to the interview below: [codes with definitions].
For each passage: the code, a one-line reason, and the line numbers.
List passages that fit no code, and passages that fit two codes, separately. Do not create new codes.

    Where it goes wrong

    • Transcripts drop hesitations and mishear names and technical terms. Check every quote you publish.
    • Agreement looks high when codes are vague. Read the disagreements, not the percentage.
  5. 05

    Run a first analysis you can rerun

    An analysis you cannot rerun is an anecdote. Ask for code, not answers, and write down the planned analysis before you see results, so the AI does not keep testing until something is significant.

    Example

    Planned in advance: a t-test of exam scores between two teaching methods. The code reports 412 rows, nine with missing scores, then the test with its effect size and interval. When the AI offers to "also check by gender", you record that as a separate, exploratory question.

    How to do it

    1. 1Write down the question and the analysis you planned before you open the data.
    2. 2Ask for code that does exactly that, with the prompt on this page, and run it yourself.
    3. 3Check the row counts after each step.
    4. 4Ask for the assumptions of each test and check them before you report a result.

    Prompt to copy

    Write [R or Python] code that loads [file], reports rows, columns, missing values and outliers, then runs [planned analysis] for [question].
Comment each step. Do not run any other test. Report the effect size with its confidence interval, not only the p-value.

    Where it goes wrong

    • Ask "is anything significant?" and it will find something. Stick to the plan.
    • Code that runs is not code that is right. Check one result by hand or against a known example.
  6. 06

    Keep up with new papers without searching again

    The literature moves while you write. Instead of repeating your search every month, set a weekly routine that searches for you and reports only what is new.

    Example

    A routine for "language models in clinical documentation, randomised or controlled studies". Most weeks it reports a handful of new papers; you add the one that matters to your reading list with one line on why.

    How to do it

    1. 1In Elicit, open Routines and describe what to watch in the words of your inclusion criteria.
    2. 2Pick a weekly schedule.
    3. 3Each week, skim the report and add relevant papers to your reading list with one line on why.
    4. 4Every few months, run your main search again yourself; a routine can miss papers that use new terms.

    Built in

    Prompt to copy

    Every week, search for new papers on [topic] that [criteria]. Report only papers published since the last run: citation, one-sentence finding, design and sample size.

    Where it goes wrong

    • A routine repeats your blind spots. Change the wording now and then.
    • A weekly list nobody reads is noise. Keep it short enough to read in five minutes.

Rules for researchers

  • Check that your ethics approval and data agreements allow the tool before uploading participant data.
  • Where chats are used for training by default, switch that off before uploading unpublished work.
  • Cite the papers you read, never an AI's summary of them.
  • Report how you used AI where your journal or funder asks for it.

Go deeper

Esc
↑ ↓ to moveEnter to open