Can AI help write a research paper? Yes, as a drafting aid. The author answers for every claim and citation.
Short answer. Yes, as a drafting and editing aid, within the rules of the venue you submit to. The International Committee of Medical Journal Editors says chatbots should not be listed as authors, because they cannot be responsible for accuracy, integrity and originality, and that authors must disclose how AI was used and carefully review its output. IEEE requires AI-generated text, figures or code to be disclosed in the acknowledgments. The main practical risk is citations: a 2023 study found that 55 per cent of GPT-3.5 and 18 per cent of GPT-4 citations were fabricated, and a 2026 study found fabricated citations in about 1 per cent of papers accepted at NeurIPS 2025. Verify every reference.
01
What the rules say
Journals and conferences differ, and a funder or university may add its own rules. Check the target venue’s current policy when you submit. Two published policies show the pattern.
Body
What it asks of authors
ICMJE, the medical journal editors’ recommendations
Disclose at submission whether AI-assisted technologies were used and describe how, in the cover letter and in the work. Chatbots should not be listed as authors because they cannot be responsible for accuracy, integrity and originality. Humans are responsible for the material, should carefully review and edit AI output, which can be incorrect, incomplete or biased, and should be able to assert there is no plagiarism. Referencing AI-generated material as the primary source is not acceptable.
IEEE publications
Disclose AI-generated content, including text, figures, images and code, in the acknowledgments, naming the AI system and the sections concerned. Using AI for editing and grammar is generally outside the policy, though disclosure is recommended. When using AI for editing, leave out the reference section, or the tool may make unwanted changes, and always double-check the result.
02
The citation problem
The 2023 study found the newer model, GPT-4, invented fewer references than GPT-3.5, yet the 2026 findings show that fabricated citations have not gone away. Crossref offers free, open APIs for the metadata of registered DOIs, so a script can confirm that a reference exists and that its title and authors match. It cannot confirm that the paper says what your sentence claims. Only reading it can.
Study
Finding
Walters and Wilder, Scientific Reports, 2023
In 84 short literature reviews on 42 topics, containing 636 citations, 55% of GPT-3.5 citations and 18% of GPT-4 citations were fabricated. Of the real citations, 43% and 24% contained substantive errors.
Ansari, 2026
100 AI-generated hallucinated citations in papers accepted at NeurIPS 2025 passed review by three to five expert reviewers. They appeared in 53 papers, about 1% of those accepted.
Badalova and Mayr, 2026
Tools that check references give useful early warnings in many cases, but are limited by extraction errors, incomplete metadata, limited database coverage and inconsistent results.
03
What a person must still check
Every reference: that it exists, that the authors and year are right, and that it says what your sentence says.
Every number, method description and claim of novelty. The model does not know what you actually did.
Unattributed copying. Authors are asked to be able to assert there is no plagiarism, including in AI-produced text and images.
Confidential or unpublished data. Do not paste it into a public service. IEEE tells reviewers not to process manuscripts through public AI platforms.
The reference list after AI editing, since tools can change it unasked.
The venue’s disclosure rule, each time you submit.
04
How a lab or college would set this up
Agree which uses are allowed: for example language editing, restructuring and summarising papers you supply, but not generating citations from memory.
Keep drafts, prompts and outputs on lab machines, not on a shared online account.
Build the reference list from a reference manager or database, never from a model’s recall.
Verify each DOI automatically, then read each cited paper that carries a claim.
Write the disclosure statement at the start, naming the tool and the sections it touched.
Have the corresponding author sign off that every co-author has checked the final text.
05
What ANVE does today
ANVE’s research maker is Scholar-Atelier. It drafts and revises papers, learns a writing style, and exports Word or PDF, including IEEE and ACM formats. It looks sources up online only when asked. Models run on the box; cloud models are optional, added by an administrator, and their answers are labelled, so drafts and unpublished data go out only if the office chooses that.
None of that removes the author’s duty to disclose AI use and to verify the references. Scholar-Atelier runs today in ANVE’s development build; no sealed unit exists yet.
06
Questions people ask
Can AI be listed as an author?
Not under the ICMJE recommendations, which say chatbots cannot be responsible for the accuracy, integrity and originality of the work. Humans are responsible for anything submitted.
Do I have to say I used AI?
Many venues require it. ICMJE asks authors to disclose AI-assisted technologies at submission. IEEE requires disclosure of AI-generated content but treats routine editing and grammar help as generally outside the rule. Check your venue.
Can AI find real references for me?
It can help you search and can summarise papers you give it. References a model recalls from memory are the risky ones, so look each up in Crossref or the publisher’s site and read it.
Is a local model safer for unpublished work?
It keeps drafts and data on your own machine instead of a third-party service, which helps with confidentiality. It does not remove the accuracy or disclosure duties.
07
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
