What prompt turns a chatbot into an AI research assistant?
Paste this into Claude, ChatGPT, Gemini or Whizi, attach your papers and fill the brackets. It makes the model work like an assistant: it answers from your files, one claim at a time, with proof.

You are my research assistant. My question: [one sentence]. Field: [field]. Use only the attached papers, plus web search for anything published after [year]. Return a table with these columns: claim, paper, page, exact quote, how sure you are. Add a last row for what the papers disagree on. If a claim is not in the papers, write "not found" and stop. Never invent a citation.
Each part closes a gap. The attached papers keep the model off the open web. The page and quote columns make every row checkable. The "not found" line gives the model a way out that is not making something up.
What is an AI research assistant, and what does it do?
An AI research assistant is a tool that does one or more of three jobs: finds papers, reads them for you, or checks what you wrote against them. Most products do the first two. Almost none do the third well.
Finders such as Elicit and Scite search a paper index and summarize what they find. Readers such as Paperguide, ChatGPT, Claude and Gemini take your PDF and answer questions about it. Checkers are the part you do yourself, with help: you ask the model to attack its own table, then you open the sources.
A general chatbot with web search and file upload covers the reader job well and the finder job badly. A dedicated finder does the reverse. Many people use one of each.
Which AI research assistant is free?
Elicit has the most generous free plan for finding papers, and ChatGPT, Claude and Gemini all have free plans for reading them. The table is built from each tool's own pricing or product page, read on 21 November 2026. Where a page was silent, the cell says not stated.
| Tool | Free use | Sign-up | Paid price | Best for |
|---|---|---|---|---|
| Paperguide | 1,500 AI credits a month, 10 papers per table, limited AI searches | Not stated | Plus $8, Pro $19.50 a month billed yearly | Chat with a PDF, literature tables |
| ScholarAI | Not stated for ScholarAI; the page says it has joined Jenni and that Jenni offers a free plan | Not stated | Not stated; a Jenni premium plan is mentioned with no price | Paper search inside a writing tool |
| Scite Assistant | A free trial is offered; no free tier stated | Needed to save history | Not stated | Answers drawn from 323M+ full text articles |
| Elicit | Unlimited search across 138M+ papers, unlimited summaries, unlimited chat with papers | Not stated | Pro $49 a month billed $588 yearly | Screening and extracting from papers |
| ChatGPT | Search, projects, limited uploads, limited deep research | Not stated | Go $8, Plus $20 a month | Reading your PDFs with search |
| Claude | Web search, up to 5 projects, no research mode | Not stated | Pro $20 a month, or $17 billed yearly | Careful reading of long files |
| Gemini | Deep Research listed on the free plan | Not stated | Not stated in US dollars | Research that uses Google apps |
| Whizi | 10 messages a day on Whizi AI, Auto and Claude Haiku, no card; projects with pinned files | Account needed | Pro $29.99 a month | Putting one question to Claude, GPT and Gemini |
Whizi loses in two places. It has no paper index of its own, so Elicit finds papers that Whizi never will. And it caps the text it reads from each pinned file, so a very long paper is truncated where a tool built around full text would read the whole thing.
ScholarAI is worth a note. Its address now opens a page whose banner says ScholarAI has joined Jenni, an AI writing workspace, and the page says ScholarAI's paper search tools power Jenni's research experience. It gives no prices. Check which product you are signing up for.
For finding, start with Elicit free and see how far Basic goes. For reading and checking, use a chatbot you already have. The wider ranking is in best AI for research.
How do you use an AI research assistant? Three worked examples
These are three jobs a student or analyst brings, each with the prompt and what a good answer must contain. They are examples written for this page, not tested output from a named model.
Example 1, a literature scan. Prompt: "I am a nursing student. List the five most cited review papers since 2020 on [topic: nurse staffing and patient falls]. For each give authors, journal, year and the one finding I could quote. Say which ones you are unsure exist." A good answer has five entries with a DOI or a link each, no round numbers without a source, and an honest flag on any entry it cannot verify. Open each link before you trust the entry.
Example 2, one attached paper. Prompt: "Attached is a paper on [topic: remote work and productivity]. Summarize the method in three sentences, list the sample size, and quote the sentence where the authors state their main limitation." A good answer gives the sample size as the paper states it, a quote you can find with Ctrl+F, and no claim about results the paper doesn't make.
Example 3, two papers that disagree. Prompt: "Paper A says [claim A]. Paper B says [claim B]. Both are attached. Explain in a table why they might differ: sample, method, year, definition of the key term. Don't pick a winner unless the papers give you grounds." A good answer lists concrete differences from the files and says plainly when the papers don't settle it.
Which prompts work for finding, reading and checking sources?
These twelve are grouped by the job. Each works first try if you fill the brackets and attach the files it names.
Find
- Map the field: "Give me a map of [topic] in five bullets: the main schools of thought, one landmark paper for each, and the question they disagree on. Mark any paper you are not sure exists."
- Search terms: "I am studying [question]. List 15 search terms and phrases a researcher would use in a database such as PubMed or Google Scholar, grouped by synonym."
- Gap finder: "From the attached review, list five questions it says are still open. Quote the sentence for each."
Read
- Method first: "From the attached paper, state the design, the sample, the measures and the main limitation. Quote each from the text."
- Plain summary: "Explain the attached paper to a first year student in 150 words. Then list three terms they would need to look up."
- Numbers table: "Pull every statistic the attached paper reports into a table: figure, what it measures, the page. Skip anything not in the text."
Synthesize
- Claim matrix: "For the [number] attached papers, make a table: paper, question, method, result, limitation. One row each."
- Disagreement: "Where do the attached papers disagree? Give each disagreement as one sentence and the pages that show it."
- Outline from evidence: "Draft an outline for a [length] review on [question], with each section tied to the attached papers that support it. Leave a section empty if nothing supports it."
Check
- Source audit: "Here is my paragraph: [paste]. For each claim, say which attached paper supports it and quote the line. Flag any claim with no support."
- Red team: "Act as a skeptical reviewer. List the three weakest claims in this table and what evidence would settle each."
- Citation check: "List every citation in your last answer with authors, year and a link. Mark which ones you recalled from memory rather than found on a page."
Can you trust the citations an AI research assistant gives you?
Not without opening them. A 2025 study in JMIR Mental Health had GPT-4o write literature reviews from six prompts and checked every citation. It found 35 of 176 were fabricated, or 19.9%. Of the 141 real ones, 64 carried bibliographic errors, most often a wrong or invalid DOI.
A real match isn't a clean one. The citation can exist, carry a broken DOI and still not say what the model says it does. That is why prompts 6 and 10 ask for the page and a quote. A quote you can search for is a test the model can't talk its way past.
Dedicated tools reduce the problem by working from a paper index, but the pages read for this article give no accuracy figure for any of them. How to fact check AI answers gives the two minute routine, and why AI gets things wrong explains the mechanism.
What mistakes ruin AI research results?
Four mistakes cause most bad results, and each has a quick fix.
- Asking for sources with no files attached. The model recalls citations from memory, and that is where invented ones come from. The fix: attach the papers and say to use only them.
- Accepting a summary of a paper you never opened. The fix: ask for a quote and find it in the PDF.
- Asking one model and stopping. The fix: paste the table into a second model and have it attack the weakest rows.
- Citing the chatbot. The fix: cite the paper, after you have read the passage yourself.
A fifth is a scanned PDF. If the file is an image of text, some tools read nothing from it. Run it through OCR or use a tool that says it handles scans.
Where does AI research help go wrong?
AI research help goes wrong in three places: coverage, depth and honesty about gaps. A model can't cite a paper behind a paywall it can't read, it can miss the newest work, and it may fill a hole with a plausible guess.
Your school or journal may also have rules. Many universities require you to disclose AI use in graded work, so read your own policy before you submit anything. A research assistant is a way to read faster. It is not an author, and it can't be held to account for a claim.
For a medical, legal or policy decision, treat the output as a reading list. Have a qualified person check the sources that matter.
Where does Whizi fit as an AI research assistant?
Whizi is one app instead of a pile of single purpose tools, and it suits you if you want one question put to more than one model. Projects hold up to 10 pinned files of up to 10 MB each, in PDF, Word, Excel, text, Markdown, CSV or JSON, and the project's instructions and file text ride along on every message. Projects work on the free plan too.
Web search is on by default in each chat. The free plan gives you 10 messages a day on Whizi AI, Auto and Claude Haiku, no card. Pro carries Claude, GPT and Gemini in one app, and its first week is $1.99. On Pro you can run side by side Compare, so the red team in prompt 11 can be a different model from the one that wrote the table.
Scanned PDFs get a vision transcription pass, and Word and Excel files made in chat have no plan gate, so a claim matrix can come back as a spreadsheet. The limit: text read from one pinned file is capped at 32,000 characters, and a normal search returns up to 3 results. For a hundred paper review, a dedicated finder is the better first stop. See Whizi pricing.
- Attach the papers and tell the model to answer only from them
- Ask for the page and an exact quote for every claim
- Open each source you plan to cite and find the line yourself
- Paste the result into a second model and ask it to find the weakest rows
- Read your school or journal policy on AI use before you submit
Frequently asked questions
What is the best free AI research assistant?
Elicit is the best free choice for finding papers, since its Basic plan lists unlimited search across more than 138 million papers, as of November 2026. For reading your own files, ChatGPT Free and Claude Free both list web search. Neither page gives an accuracy score.
Can ChatGPT be an AI research assistant?
Yes, for reading and summarizing files you attach. ChatGPT Free lists search, projects and limited uploads, as of November 2026. It is weaker at finding papers, so check every citation it gives you against a real page.
Can AI do research for you?
It can find leads and summarize papers, but it can't do the judging. One 2025 study found 19.9% of GPT-4o's citations were fabricated. Use the output as a reading list and open each source.
Is an AI research assistant allowed in coursework?
That depends on your school or instructor, so check the policy first. Many require you to disclose AI use. Cite the papers you read, not the chatbot that pointed you to them.