Skip to content

An AI Research Assistant Promises Citation-Backed Drafts. A Study Counted Fabricated References.

Illustration: a researcher at a desk looking at a floating holographic panel of documents and citation graphs

Of 636 citations in 84 AI-generated literature reviews, 55 percent of the GPT-3.5 ones pointed at papers that do not exist. That is the number to hold next to a sales pitch.

In a webinar on the AnswerThis channel, co-founder and CMO Ryan McCarroll shows an AI research assistant that searches the literature, drafts citation-backed text and summarises papers. The pitch is speed: literature work that took weeks, in minutes. The claim comes from the vendor. This post sets it against the studies and the rules that followed.

What the company sells

AnswerThis describes itself as "an all-in-one AI research assistant that supports your entire workflow, from finding research gaps and collecting papers to summarizing, analyzing, and drafting citation-backed content for your research paper, dissertation, or thesis".

The demo is about finding gaps in the literature, mapping citations between papers, and a first draft of a literature review with references attached. The company says that work which used to take months now takes minutes. Its Y Combinator profile describes the platform as used by over 200,000 researchers.

The company's own pages do not agree on how many people use it. The homepage speaks of "1 Million+ researchers" and a "300M+ paper" database, while the pricing FAQ says 250M+ papers. A million on one page, two hundred thousand on another.

It is a paid product. The pricing page lists a free tier, a Pro tier at 35 dollars per seat per month and institutional licences. Annual-billing tiers run from 0 to 100 dollars.

A citation-map feature, a benchmark against Google Scholar or PubMed, and the claim of "more than 20 citations in minutes" appear only in the vendor's own pages. They are vendor claims, not measured results.

All of it comes from the company.

What the studies counted

A fake reference looks just like a real one.

A peer-reviewed study in Scientific Reports (Walters & Wilder, 2023) examined 636 bibliographic citations in 84 AI-generated reviews. It found that 55 percent of the GPT-3.5 citations and 18 percent of the GPT-4 citations were fabricated. None of them existed. Of the real ones, 43 percent (GPT-3.5) and 24 percent (GPT-4) contained substantive errors.

A second study in the Journal of Medical Internet Research (Chelli et al., 2024) asked the models to reproduce systematic reviews. It measured hallucinated references at 39.6 percent for GPT-3.5, 28.6 percent for GPT-4 and 91.4 percent for Bard. LLMs, the authors concluded, should not be used "as the primary or exclusive tool for conducting systematic reviews".

AnswerThis sells the citation. In both studies, the citation is the part that failed.

Rules instead of bans

Publishers wrote rules. ICMJE states that chatbots and other AI tools cannot be listed as authors. They cannot answer for accuracy, integrity and originality. Authors must review and edit AI output, which the guidance notes "can be incorrect, incomplete, or biased". COPE says the same: AI tools "cannot meet the requirements for authorship". Over 200,000 researchers use a tool that cannot be listed as an author.

The universities are blunter. Erasmus University Rotterdam tells staff and students not to use GenAI to look up or check facts. A plausible answer is not a reliable one. Amsterdam's policy states that GenAI output is not a reliable academic source.

Ghent University's framework lists invented references among the known limits. Antwerp's guidelines note that AI texts often lack proper sourcing or carry invented citations. Harvard's FAS guidance requires any use of AI to be acknowledged and cited. The student answers for the result.

The rules differ. They agree that the tool writes and the person signs.

What to do with a tool like this

Treat these references as leads, not sources. Open every citation before it goes into your bibliography. Confirm the DOI, the journal and the year. Keep that check inside the workflow. A bibliography is where a fake becomes a plagiarism case.

Check your institution's rules before you start. They differ on acknowledgment, permitted use and the responsibility clause. Ask the vendor for the audit trail. A tool that shows where a citation came from is a different product from one that only writes the sentence.

Original video: Academic AI Research Tool, Boost Your Research Productivity In One Click

Sources