AI for Academic Fraud Detection — 1,305 Distinguished Young Scholar Papers Screened Overnight

I don’t know if you’ve been following this story: there is far more fabrication in the flagship papers of Distinguished Young Scholar teams than I ever expected. These teams carry a huge share of the country’s research funding — they’re the backbone of Chinese science — and yet so many of the flagship papers posted on the official site are fake. I was genuinely shocked when it first came across my feed.

I got very curious right away: what kind of person gets awarded Distinguished Young Scholar status? How many are there in this country? And how much research funding does the state give them, such that they’d take this kind of risk and commit academic fraud? And the part that matters more — academic fraud this easy to spot somehow has no review system at all. Do the journals really have no checking mechanism at publication time?

Turning Geng into a skill

I took a close look at how Geng actually found these fabrications. He has his own detection logic — repeated data, reused images — and all of it is genuinely easy to spot.

So I started thinking: could I use AI to turn Geng into a skill, run it across every Nature and Nature-family paper by 2025’s Distinguished Young Scholars, and see how many fabricated papers are actually in there?

Step one is the list. The state doesn’t actually publish a master list of Distinguished Young Scholars — it only publishes smaller lists separately. So I had AI aggregate and deduplicate those, which got me to a few hundred names. Then I searched their publication records across the whole Nature family of journals, and ended up locking in 187 awardees with output there and over 1,300 papers.

Because Nature family journals are English-language, author names are written in pinyin. To make sure a given name really was the right person, I ran a triple confirmation pass: confirm the pinyin matches, confirm the university matches, confirm the research area matches — only then does the paper get recorded under that name. Then I pulled down every paper’s source, all its data, and all the PDFs. 1,305 of them.

Four fraud fingerprints

For those 1,305 papers, I designed four detection mechanisms:

• Constant difference between two columns — Geng’s point: if someone dragged rows and columns in Excel to compute values, you’ll find the difference between different data groups is a fixed constant. This is a laughably primitive way to fabricate data — anyone who knows what a random number is wouldn’t do it
• Decimal-place locking — when altering data, they only change the integer digits, so the final decimal places all stay the same, and a whole swath of data has identical decimals
• Reused images — hash comparison; only an exact match reports positive
• Row-level duplication — different sample sets with completely identical data

Nothing more complex than that. A hit only produces a “positive report” — it may be fabrication; it is not a verdict.

Result: 2 anomalies plus 2 questionable out of 578

Of the 1,305, the papers I actually got to analyze numbered only 578 — 44% data coverage. The main reason: plenty of papers publish no spreadsheet data at all, and some park it behind external links or in odd formats that a simple AI pass can’t fully read.

Among those 578 papers, two show objective data anomalies that are close to unambiguous, and two more have data that looks questionable. Unfortunately, all four are in the life sciences — four suspect cases out of just over 300 life sciences papers. Counting the papers Geng surfaced, the life sciences are where the rate runs highest. Reportedly, around 2010 there was a wave of large-scale data fabrication and large-scale retractions in life sciences papers.

AI gives you the lead; a human makes the call

Having run the whole screen, my takeaway is that AI only gets you so far here: the machine can hand you a lead, but calling it real fraud in the end still takes an enormous amount of human effort — it takes domain expertise to see what’s genuinely wrong inside the data.

I’ve open-sourced this skill on my GitHub, so you can download it and run more checks — because this is only the result for 2025’s new Distinguished Young Scholars.

I’d also like to see the academic world build public oversight mechanisms, even if it’s just individual institutions setting up their own self-screening. Let the research funding go to people with real ability doing real research, making real contributions to the country’s scientific development — instead of all that money flowing to fraudsters.

Source: EP0037_audio.mp3 · ASR model gemini-2.5-pro (chunked parallel) · full text of the original recording

[00:00] I don’t know if you’ve been following this news lately — there’s so much fabrication in the flagship papers of Distinguished Young Scholar teams. I really didn’t expect it. These teams carry a huge share of the country’s research funding, they’re the backbone of Chinese science, and then so many of the flagship papers posted on your own official site are fake. I was genuinely shocked when it first came across my feed. Geng, by the way, has basically lost his reach on some platforms by now. I got very curious right away, because I’d never paid attention to this field before: what kind of person actually gets awarded Distinguished Young Scholar status? How many does this country have? And how much research funding does the state give them, such that they’d take the risk and commit

[00:25] academic fraud like this. And the part that matters more: academic fraud this easy to spot somehow has no review system at all. Do the journals really have no checking mechanism when they publish? So I took a close look at how Geng actually found these fabrications. He has his own detection logic, and it includes repeated data, reused images — all of it genuinely easy to spot. So I started thinking: could I use AI to turn Geng into a skill, and take all of 2025’s Distinguished Young Scholars’

[00:50] papers in Nature and the Nature-family journals and run them all through it, to see how many fabricated papers are actually in there. So first I went and looked at the numbers. To start with, the state doesn’t publish a master list of Distinguished Young Scholars — it publishes smaller lists separately. So I had AI aggregate and deduplicate them, and once I had a list of a few hundred names, I ran their publication records across the whole web, across all of Nature’s sub-journals, all of them, and ended up locking in 187

[01:15] awardees with output in the Nature family of journals, and over 1,300 papers by them. That spans different fields, and their universities. Going by name and university — because you know Nature-family journals are English-language, so the authors are all written in pinyin. To make sure a given person really is that person, I ran another round of comparison and confirmation: confirm the pinyin matches, and confirm the university and the research area match, and only then does it get recorded

[01:40] under that name. Right now you can see, for every paper, its exact address, plus all of its data and the paper’s PDF — all of it downloaded. 1,305 papers. And for those 1,305, what mechanism did I design? The first one is what Geng calls constant difference between two columns: if someone manually dragged rows and columns in Excel to compute the values, you find that across different data groups, the difference between the groups is

[02:05] a fixed constant. That’s a laughably primitive way of cooking data, right? Like Geng said, anyone who knows what a random number is wouldn’t do it that way. Then there’s decimal-place locking: when they alter the data they only change the integer digits, so the final decimal places all stay untouched, and a whole swath of data has identical decimals. That simple a way, that simple a way of faking it. Or else it’s reused images, or row-level duplication — different sample sets whose data is completely identical. I only did

[02:30] those four checks, nothing more complex. For reused images I compared hash values, so only an exact match produces what’s called a positive report — meaning it might be fabrication. And once I’d screened all of them out, what I found is that it’s actually pretty hard to do this with AI. The machine can hand you a lead, but calling it in the end — whether this really is fraud — still takes a human spending an enormous amount of

[02:55] effort. So let’s take a look: out of those 1,305, I actually only got, only compared 40-something percent, 44% of the data. Why? I only got 578 papers with their detailed data. One big reason is that plenty of papers publish no spreadsheet data at all, and another chunk of them —

[03:20] there’s simply no normal way to get their data. It might be parked behind an external link, or in some special format, where a simple AI approach just can’t do a complete analysis of the data. So the number of papers I really analyzed is only 578. And among those 578 papers, two cases are close to unambiguous — you can see objective data anomalies in them. And two more pieces of data, two papers’ data, are questionable.

[03:46] Unfortunately, all four cases are in the life sciences. I don’t know why either. Is it that the Distinguished Young Scholars in life sciences don’t really get how computers work, don’t know how to use a random number generator, or what? It really is — and counting the papers Geng surfaced too, the life sciences are where the rate runs high. I’ve already open-sourced this skill on my GitHub, so you can download it and use it to run more checks.

[04:11] Because this is only the result for 2025’s newly minted Distinguished Young Scholars. And, let me see, in life sciences four suspect cases already came out of just over 300 papers. The further back you go, the higher the fabrication rate — especially, reportedly, around 2010. In that period, life sciences papers had data fabrication on a massive scale, mass fabrication and mass retractions. So even with AI behind you, academic fraud detection is not

[04:36] an easy thing to do. It takes a lot of human verification, and the humans need domain expertise to find the parts that are, so to speak, unreasonable. So I’m also hoping the academic world gets public oversight mechanisms — even just individual institutions setting up their own self-screening — so that the academic world really, really has people with real ability doing real research getting our research funding, making real contributions to this country’s

[05:01] scientific and technological development — instead of all that money flowing to the fraudsters who fake it.