What prompt makes an AI paper writer produce a usable draft?
Paste this into Claude, ChatGPT, Gemini or Whizi, fill the brackets and attach your sources. It stops the model from inventing references because it can only draw on files you gave it.

Write the [section, such as introduction or literature review] of a [length]-word [type, such as undergraduate research paper] on [topic] for a [course and level]. My thesis is: [one sentence]. Use only the attached sources and cite each claim as [citation style] with the page or section. If the sources do not support a point I need, write "no source" instead of making one up. Finish with a list of every source you used and the claim it supports.
Each bracket closes a gap. The thesis stops a generic essay. The source rule removes the need to recall references. The "no source" line turns a silent invention into a visible gap you can fix. The closing list is what you check in step four.
What does a good AI paper draft look like for a real assignment?
A good draft cites a source for every claim, says where the sources run out, and argues your thesis instead of summarizing the topic. These three assignments show what to ask for and what to check. The answers below describe what must be in the reply. They are written for this page and are not output from a tested model.
- A literature review paragraph. Assignment: three paragraphs on whether later school start times improve teen grades, for a first year psychology course. Prompt: the template above, with five attached studies. A good reply groups the studies into two or three findings, gives a page or section for each claim, flags any study with a small sample, and writes "no source" where you asked about test scores and no attached study measured them.
- A methods section from your notes. Assignment: a 250 word methods section for a survey of 120 students. Prompt: "Write a methods section from these notes only: [paste notes with sample size, recruitment, instrument, analysis]. Add no detail I did not give you. List every detail you had to leave blank." A good reply uses your numbers exactly, names the blanks (the response rate, say) and does not invent a software version or an ethics approval number.
- A paragraph that fails the rubric. Assignment: raise a B paragraph to an A on a rubric that scores "uses evidence to support a claim." Prompt: "Here is my paragraph and the rubric: [paste both]. Score the paragraph against each rubric line, quote the weakest sentence, and rewrite only that sentence. Do not touch the rest." A good reply scores each line, quotes the exact sentence and changes one thing, so you keep your own voice and can see what moved.
Which AI paper writer is free?
NoteGPT's AI Paper Writer is the only one of the top three that says it needs no signup, and Paperguide and Jenni both have free plans with caps. The table is compiled from each tool's own pages, read on 21 November 2026. Where a page was silent, the cell says "not stated".
| Tool | Free use | Signup | Paid price | Sources and references | Best at |
|---|---|---|---|---|---|
| NoteGPT | Free "for now", with a free version promised to stay | None stated: the page says no account or email | Not stated on the page | Pulls real sources from the title, link or DOI you give it | A quick draft with no account |
| Paperguide | 1,500 AI credits a month; 10 papers per table | Not stated; "Start for free" leads to signup | Plus $8 a month, Pro $19.50, both billed annually | Cites the sources you select; picks from your reference manager or public research | Drafting from a paper library |
| Jenni | 10 AI autocompletes a day, 5 AI requests, 10 PDF uploads | Not stated; free signup link | Plus $12 and Pro $29 a month as listed | Unlimited citations on Free | Writing inside one editor |
| ChatGPT | Free plan with limited messages with uploads and limited deep research | Not stated | Go $8, Plus $20 a month | No reference tool stated | Reworking a draft you already have |
| Claude | Free plan with web search and "limited usage" every 5 hours | Not stated | Pro $17 a month billed annually, or $20 monthly | Web search; no reference tool stated | Careful long prose |
| Whizi | 10 messages a day on Whizi AI, Auto and Claude Haiku, no card | Account needed | Pro $29.99 a month | Web search on by default; PDFs attached to a chat | Comparing Claude, GPT and Gemini on one section |
Whizi loses on two counts. NoteGPT writes a paper without an account, and Whizi needs one. Paperguide and Jenni also keep a reference library and citation styles in the editor, which the chat apps in the table do not list.
The pages differ in what they promise. NoteGPT says "Every citation in your paper is pulled from verified articles or journals" and "No fake references". Paperguide says its writer gives "Structured documents with accurate references" and cites the sources you pick. Neither page publishes a test of how often the references hold up. Jenni lists "Unlimited citations" but does not say how it checks them. Treat each of these as a vendor claim until you have opened the references yourself.
Can AI write a research paper with real references?
AI can write the prose of a research paper, but it cannot be trusted to supply the references from memory. A 2025 study in JMIR Mental Health asked GPT-4o for six short reviews, then checked all 176 citations in Google Scholar, Scopus, PubMed and WorldCat. Thirty five did not exist.
The pattern matters more than the total. The rate was lowest on the most studied disorder and nearly five times higher on the least studied. Even among the 141 real references, 64 (45.4%) held errors, most often a wrong DOI. The authors wrote that citation fabrication and bibliographic errors "remain common in GPT-4o outputs". So the model fails hardest on a niche topic, which is where a student most wants help.
The earlier work agrees. Walters and Wilder tested GPT-3.5 and GPT-4 in Scientific Reports in 2023 across 636 citations. They found 55% of the 222 GPT-3.5 works were fabricated and 18% of GPT-4's. Fewer than 10% of citations from either model linked to a real work. Both tested older models, so today's rates will differ, and neither tested a tool that searches first. The rule holds anyway: a reference you did not open is a guess. Best AI for literature review lists the search tools that find real papers first.
The check takes a few minutes a reference. Paste the title into Google Scholar. If it's not there, search the DOI. If it exists, open it and find the claim. If a number or a quote isn't on that page, delete the sentence or fix it. That's the whole routine. How to fact check AI answers covers the same routine for any claim.
What are the best prompts for writing a paper with AI?
The best paper prompts are narrow: one stage, one section, your sources attached. These twelve are grouped by stage. Each one works in Claude, ChatGPT, Gemini or Whizi.
Plan
- Thesis check: "Here is my topic and thesis: [paste]. Name the strongest objection a reviewer would raise and two kinds of evidence that would answer it. Do not write any part of the paper."
- Outline: "Turn this thesis into an outline for a [length]-word paper in [field]. Give each section a heading, the claim it makes and the kind of source it needs. Mark every section where I still need to find a source."
- Search terms: "I am writing on [topic]. Give me 10 search phrases for Google Scholar, mixing the technical term and the plain one, and say what each is likely to find. Do not list any papers."
Draft
- Section draft: use the template at the top, one section at a time.
- Draft from notes: "Write a paragraph from these notes only: [paste]. Keep my claims and my numbers. If a sentence needs a source I did not give, end it with [source needed]."
- Abstract last: "Write a 150 word abstract of the paper below. Cover the question, method, finding and limit in that order. Use only what the text says: [paste]."
Revise
- Rubric audit: "Score this section against each line of the rubric, quote the weakest sentence per line, and say what to change. Do not rewrite it: [paste section and rubric]."
- Cut for length: "Cut this section from [current] to [target] words. Remove repetition first, then weak examples. Keep every citation. Show what you cut as a list."
- Transitions only: "Read these paragraphs. Where one does not lead to the next, write a single bridging sentence. Change nothing else: [paste]."
Cite and check
- Claim and source table: "For each citation in this draft, make a table with the claim, the source, the page or section, and whether the attached file supports it: yes, partly or no. Flag any source not attached: [paste draft]."
- Reference list formatting: "Format these references in [citation style]. Do not add, remove or complete any detail. If a field is missing, write [missing]: [paste]."
- Counter reading: "Read the attached paper and tell me if my summary of it is fair: [paste summary]. Quote the line that most disagrees with me."
What mistakes ruin an AI written paper?
Four mistakes cause most failed papers, and each has a plain fix.
- Asking for references. The model writes plausible ones. The fix: attach the sources and let it cite only those.
- Writing the whole paper in one prompt. Errors pile up where you can't see them. The fix: draft one section at a time and check each before the next.
- Checking that a reference exists but not what it says. The fix: find the exact claim on the cited page. A real paper cited for something it never says is still a fabricated claim.
- Handing in the output as it came. The fix: rewrite it so the argument and the wording are yours, and follow your school's AI policy.
A fifth is quiet: skipping the rubric. A model that hasn't seen the grading criteria writes to no one in particular, so paste the rubric into the prompt.
Should you use a paper tool or a general chatbot?
Use a paper tool if you want the editor, the citation styles and a reference library in one place, and use a chatbot if you already have your sources and want the best prose. The table shows the trade. Paperguide and Jenni sell the editor. ChatGPT, Claude and Whizi sell the model.
A paper tool also hides which model wrote the text. NoteGPT, Paperguide and Jenni name no model on the pages we read, so you can't say what produced a sentence. In a chatbot you pick the model and can swap it. That matters when you're checking a claim, because a second model from another company is the cheapest second opinion you have.
The cost gap is small. Paperguide's paid plans start at $8 a month on annual billing, Jenni's Plus plan is listed at $12, and ChatGPT Go is $8. Pick on workflow, not price.
Where does Whizi fit as an AI paper writer?
Whizi is one app instead of a pile of single purpose paper tools, and it suits you if you want to try a section on two models. You can attach PDF, Word and Excel files to a chat, and scanned PDFs get a vision transcription pass, so a photocopied article is not a dead end. 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, with no card. Pro carries Claude, GPT and Gemini in one app, and its first week is $1.99. Pro and Powerhouse also have side by side Compare, which sends one prompt to two models. Put prompt 10 to both and read only where they disagree. A finished section can come back as a Word file, and Word and Excel files made in chat have no plan gate. See Whizi pricing for what each plan holds.
Whizi isn't a scholarly database and it doesn't keep a reference library. For finding papers and managing citations, use Google Scholar, your library and a reference manager, then bring the PDFs here.
What can an AI paper writer not do?
It cannot take responsibility for your paper, and the rules say you must. Nature Portfolio's AI policy says "Accountability for scholarly content, evaluation, and editorial decisions cannot be delegated to AI systems", and it asks authors to "Remain fully accountable for originality, accuracy, and integrity" and to "Disclose AI use for trust and transparency". That page governs journal submissions. Your course may be stricter, so read its policy before you draft.
It also cannot promise that detectors will pass your paper. NoteGPT's own FAQ admits "nothing is foolproof." Write the argument yourself and the question stops mattering. Why AI gets things wrong explains the mechanism behind the invented references.
- Write the thesis and outline yourself before any draft
- Attach your own sources and tell the model to cite only them
- Draft one section at a time and check it before the next
- Find every claim on the page its reference points to
- Rewrite in your own words and follow your school's AI policy
Frequently asked questions
Is there a free AI paper writer with no signup?
Yes. NoteGPT's AI Paper Writer says it is free with no account and no email, as of November 2026, though it adds "For now". Paperguide and Jenni both have free plans, and their pages show a signup link. Open every reference any of them gives you.
Can ChatGPT write a research paper?
It can write the prose, but not reliable references from memory. ChatGPT's free plan lists limited messages with uploads and limited deep research, as of November 2026. Attach your sources and tell it to cite only those, as in the template above, and check each reference.
Can AI write a paper that cites real sources?
Only if the sources are real and supplied. In a 2025 study GPT-4o invented 19.9% of 176 citations, and 29% on the least studied disorder. A tool that searches or reads your files does better, but you still have to open each reference.
Is it cheating to use an AI paper writer?
It depends on your school's policy, so read it first. Policies differ by school, and journals such as Nature Portfolio require you to disclose AI use. Rewriting the paper in your own words and stating the help is the safe line.