I Used AI to Fake a Scientific Paper. Here’s What Happened

I have a rule for building in public: if you can’t explain the value without a buzzword, the value isn’t there. For years, the academic world has operated on the ultimate value metric—the peer-reviewed paper. It is the currency of credibility. But right now, that currency is being debased faster than a failed memecoin.

We are currently witnessing a massive influx of what the industry calls “AI slop.” These aren't just student essays generated the night before a deadline; these are entire scientific manuscripts, complete with citations and figures, produced by Large Language Models (LLMs). By 2025, data suggests that nearly 60 percent of published academic articles show signs of LLM influence. That is a staggering jump from just 12 percent a couple of years ago. We aren't just talking about grammar checks here; we’re talking about the systematic fabrication of knowledge.

The “Vibe-Research” Experiment

As a founder, I’m always interested in the “low-bar” entry point. How easy is it for a complete amateur to break the system? To find out, a test was conducted involving “vibe-researching” a paper. The goal: use free, off-the-shelf tools like ChatGPT and Gemini to generate a scientific paper on a complex topic—optical solitons—without doing a single minute of actual lab work. The partner in this crime? IEEE Access, which agreed to “war-game” the submission to see if their filters could catch the fraud.

The process was surprisingly simple. You don't need to be a prompt engineer. You just need to be persistent. When a chatbot refuses to generate synthetic data, you pivot. You ask it what a “hypothetical” result would look like. You break the paper into chunks—abstract, intro, methodology—and stitch them together. Within a few dozen prompts, a professional-looking, four-page paper was ready for submission.

Round 1: The Chatbot Forgery

The first attempt was a pure forgery. The science sounded plausible to a layperson, and even a former academic collaborator noted the writing was professional. However, it failed. Why? Because LLMs are still world-class liars. The paper was flagged almost immediately for “ghost citations.” The AI had simply invented references that didn't exist, attributing them to real-sounding but non-existent scholars. Gemini even admitted to the fabrication when asked about it later.

This is the first lesson for builders in the AI space: Detection follows the path of least resistance. Right now, the path of least resistance for catching AI slop is checking the bibliography. But that is a temporary victory. As models get better at retrieving real-time data, the “hallucinated reference” will vanish, and the gatekeepers will be left staring at a perfectly formatted, perfectly cited lie.

The Rise of the AI-Science Engine

If Round 1 was an amateur forgery, Round 2 was a glimpse into a terrifyingly efficient future. The experiment shifted from using generic chatbots to using a specialized tool called FARS (Fully Automated Research System). Unlike ChatGPT, which is a generalist, these agentic systems are designed to orchestrate the entire scientific process.

In 21 minutes, FARS generated a research proposal. The next day, it finished a computational experiment. Two hours after that, it spat out a submission-ready article. This wasn't just text; it was “actual” research performed by an agent. When this paper was sent to IEEE, it didn't fail because of fake citations. The citations were real. It failed because it relied too heavily on “preprints”—papers that haven't been peer-reviewed yet—and because a technical figure contained a nonsense label (“lhewl”) and tell-tale AI visual artifacts.

We are moving from a world where AI fakes science to a world where AI does science, but without any human steering. This creates a massive philosophical rift. If an AI performs a valid simulation and writes a paper about it, is that “legitimate”? Or is it just high-speed noise?

The Burden on the Human Filter

The real victims here aren't the publishers; they’re the peer reviewers. These are volunteers who are already overworked. In 2022, the system processed over 3 million papers. Now, they are being asked to act as forensic AI detectors on top of their actual jobs. The number of reviewer invitations sent out per article has jumped 50 percent since 2018. The system is at a breaking point.

The “publish or perish” culture is the kindling, and AI is the accelerant. For years, we've incentivized researchers to prioritize quantity over quality. Now, we've given them a machine that can produce infinite quantity for zero cost. We've effectively DDoSed the scientific method.

What This Means for Builders

If you are building in the AI or crypto space, this is a cautionary tale about trust. When the cost of production drops to zero, the value of the output usually follows. We are entering an era where “proof of work” needs to move from the digital realm into the physical realm for science to survive.

Some institutions are already looking at blockchain-based “trust markers.” Imagine a world where laboratory equipment automatically logs data to a decentralized ledger, proving that a specific researcher was in a specific lab at a specific time. We are going to need these “on-chain” proofs of physical reality to distinguish between a discovery and a hallucination.

The Takeaway

We cannot ban AI in science. That’s like trying to ban calculators in a math class. It’s a tool that can accelerate discovery in ways we can’t imagine. But we have to change what we value. If the “paper” is no longer the unit of knowledge because an AI can churn one out in two hours, we need a new unit.

The future of science isn't just about the result; it's about the provenance. We need to stop asking “Is this paper good?” and start asking “Can this work be reproduced, and who is actually standing behind it?” If we don't fix the incentive structure soon, the scientific record will become nothing more than a giant, automated echo chamber.

Read the original at IEEE Spectrum →