Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseResearch Paper · For R&DIEEE Spectrum AI→ original

A Journalist Built a Fake Research Paper in 24 AI Prompts, but IEEE Access Caught It Through Its References

An IEEE Spectrum journalist tested whether free ChatGPT and Gemini could produce a fake scientific paper that would pass a journal's review. In 24 prompts, with no experiments behind it, he produced a plausible text about optical solitons. The paper was rejected because its references were invented. A paper written by FARS, an autonomous system, with real references was also flagged, because it cited too many preprints that had not been peer-reviewed. The conclusion: checking sources is currently the cheapest filter for AI-generated junk.

AI-processed from IEEE Spectrum AI; edited by Hamidun News
A Journalist Built a Fake Research Paper in 24 AI Prompts, but IEEE Access Caught It Through Its References
Source: IEEE Spectrum AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceIEEE Spectrum AI ↗
Editor-in-Chief
Key metric #157% of scientific papers in 2025 show signs of LLM involvement, up from 12% in 2023
Key metric #2FARS produced a research plan in 21 minutes and a full paper in 2 hours
Key metric #3Gemini itself admitted that seven of the sources in the paper were fabricated

Imagine wanting to publish a scientific paper about an experiment that never happened, with only free chatbots to help. Could you do it? An IEEE Spectrum writer set out to find out.

In roughly two dozen prompts (about 24) to free ChatGPT and Gemini, he assembled a solid-looking paper, submitted it to the journal IEEE Access, and got a rejection: the review found "phantom citations," meaning references to works that do not exist. He made a second attempt with FARS, an autonomous system that carries out a computational experiment itself, and it was caught as well. Why read this: according to an estimate from a PNAS paper, by 2025 traces of large language model (LLM) influence were visible in 57% of published papers, compared with 12% in 2023.

More than 12% of reviews in Nature Communications in 2025 were fully generated. For every 40 papers with statistical signs of AI, there is only one formal disclosure of its use.

Round 1: free chatbots

The conditions were deliberately amateur. The journalist used only free ChatGPT (GPT-5.3/5.2 Mini at the time) and Gemini 3 Flash. He could not edit the output by hand, and he could not call in prompt experts. He listed himself as the sole author and took the topic from his doctoral dissertation from almost twenty years earlier: optical solitons, that is, light pulses that travel without changing shape.

The first attempt was thin: without real data, the models struggled to produce even four pages. Decomposition helped. The paper was split into abstract, introduction, methods and other sections, and each was requested separately. When one model refused to continue, the task went to another, or he asked how the next step would look "hypothetically." ChatGPT resisted more than Gemini and would not invent data. After about two dozen prompts, he had the text, figures, captions, bibliography and formatting to meet IEEE Access requirements.

The author himself found the paper convincing. There were no odd paraphrases like "deep neural organization" in place of deep neural network. Stylistic habits typical of AI (hedges, triads, formulaic "Thus") were present, but they are not evidence on their own, since living authors write that way too.

The references gave the fake away. When the paper was later run through Gemini, the model itself reported that seven "recent references on LLMs were fabricated to fit the storyline." IEEE Access managing editor Jenny Mahoney did not find gross errors in the physics itself; her doubts were raised specifically by the phantom citations. Soliton expert Gaetano Assanto judged the text professional, but noticed overly dense mathematics, a profusion of abbreviations in the abstract, and a couple of overconfident claims that were debatable at best.

Round 2: the autonomous system FARS

The stakes were then raised. FARS does not only write text; it conducts the research itself. The author entered a shortened version of the same topic and received a work plan 21 minutes later. With one click he authorized the experiment. The system ran the computational experiment overnight and the next day emailed to say it was finished. Without checking the results, the author clicked "Write paper" and two hours later received a paper ready for submission. He removed the FARS note about where the text came from, added his own name, and submitted the work to IEEE Access.

The references were now real, but the journal still noticed a problem: too many sources were preprints that had not been peer-reviewed. The paper also gave itself away in small details: one technical figure carried a meaningless caption, "lhewl," and a diagram of the AI "brain" still contained stray dots.

What this means for the industry

The outcome is double-edged. A newcomer with free tools has not yet gotten past automated checks and a careful editor. But he produced a plausible paper about research that never took place, and it took a matter of hours. Stronger paid models may make fewer mistakes.

Luigi Longobardi, IEEE's director of publication ethics, acknowledges that a detector built today may fail on the AI-written paper of tomorrow. In that case, volunteer reviewers remain the last line of defense. IEEE Access receives about 60,000 submissions a year and accepts roughly 20%, around 12,000 papers. Each is read by at least two reviewers.

The practical takeaway for publishers and for those building AI tools for science: automated checking of references and sources is currently the cheapest and most reliable filter. The experiment's limits are candid: one author, one narrow topic, and one journal with well-developed safeguards. At less protected publications the result could be different. That is why defenses should be built not around writing style, which is easy to fake, but around verifiable facts: whether the sources exist, whether the data can be reproduced, and whether the experiment itself can be confirmed.

Story

Key Questions Answered

What are the key details on Round 1: free chatbots?

The conditions were deliberately amateur. The journalist used only free ChatGPT (GPT-5.3/5.2 Mini at the time) and Gemini 3 Flash. He could not edit the output by hand, and he could not call in prompt experts. He listed himself as the sole author and took the topic from his doctoral dissertation from almost twenty years earlier: optical solitons, that is, light pulses that travel without changing shape. The first attempt was thin: without real data, the models struggled to produce even four pages.

What are the key details on Round 2: the autonomous system FARS?

The stakes were then raised. FARS does not only write text; it conducts the research itself. The author entered a shortened version of the same topic and received a work plan 21 minutes later. With one click he authorized the experiment. The system ran the computational experiment overnight and the next day emailed to say it was finished. Without checking the results, the author clicked "Write paper" and two hours later received a paper ready for submission.

What this means for the industry?

The outcome is double-edged. A newcomer with free tools has not yet gotten past automated checks and a careful editor. But he produced a plausible paper about research that never took place, and it took a matter of hours. Stronger paid models may make fewer mistakes. Luigi Longobardi, IEEE's director of publication ethics, acknowledges that a detector built today may fail on the AI-written paper of tomorrow. In that case, volunteer reviewers remain the last line of defense.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…