我用AI来伪造一份科学论文事情是这样的
我对IEEE期刊的实验揭示了AI对学术出版的威胁。
科学有垃圾邮件问题。 垃圾邮件过滤器被堵住了
根据所有报道，学术出版正在沉溺于所谓的AI slop：由AI生成的低质量，量产内容。 到2025年，所有出版的学术文章中估计有57%展现出可察觉的大型语言模型(LLM)影响的证据，从2023年的12%猛增。 超过12%的自然通讯评论在2025年完全由AI产生。
鉴于每40份显示2025年AI使用统计证据的文件只有一份正式披露，这些数字可能严重低估了LLM今天的参与。 当然，AI使用的证据并不一定意味着一张纸是破烂的。 但一些估计表明，现在每年有数十万种假出版物，往往是大赦国际制作的。
警告 Amanda Sulicz是IEEE研究整合管理者。 “AI sloop正在完全侵蚀我们对科学的信任，以及我们几个世纪以来创造的每一个科学过程。”
但统计数据和声明声称这个问题对科学来说甚至可能存在，用AI来造出一份令人信服但完全假冒的报纸是多么容易？出版商是否拥有让AI退出科学记录的工具？
我借用“vibe编码”的想法，即程序员使用自然语言的提示来生成代码，我决定“vibe-research”一个编造的论文，即促使LLMs生成假说、代码和文本，同时尽量减少人类输入。 然后，我问Jenny Mahoney，多学科开放期刊IEEE Access的主管编辑，如果我们能进行一场战争游戏，她和IEEE很友好地同意了
实验将成为两部分的试验。 首先，一个没有特别AI专长的人是否可以使用现成的AI工具来产生一篇令人信服的论文，描述从未进行过的研究，并且做得足够好来愚弄出版商的辩护人。 然后，我会提高赌注：如果更先进的AI系统不仅写论文
沿途，实验成为AI自动化如何转变科学发现的窗口，迫使利益攸关方重新思考整个学术出版系统。
第1轮：创造伪造
为了给我创造出一个令人信服的假象的机会，我允许自己多次(在合理范围内)发出警告。 但为了公平竞争，我无法访问最先进的付费版本LLMs，此外，我无法直接编辑任何产出，也无法向专家AI即时工程师寻求外部帮助。 这需要做一个业余的努力，纸张尽可能由LLMs单独生成。
本杰明·斯库斯(Benjamin Skuse)让AI从他的写作室从事科学研究。
路易吉·阿万塔吉亚托
我想用假作者的名字和从属关系会立即升起一面红旗，这一选择将我局限于我有一些学术经验的主题。 所以我从近20年前的博士学位题选择：光学索利通，光学索利通是单个的光脉冲，它的形状没有任何改变。
我的第一次尝试—— " 你能用光学索利通语研究我一份原始的论文吗？" —— 留下了很多有待改进之处。 虽然Gemini乐于完成这项任务(与ChatGPT不同，ChatGPT在不生成合成数据或研究方面更为有力)，但结果却被压抑。 报告强调了在这一过程中反复出现的一个问题：没有进行任何真正的研究来配合，产出很薄，LLMs经常难以填满四页。 然而，随着我更像adept一样激励模型，它们松动了，变得更有特色。
我将论文拆分为其组成部分——摘要、介绍、方法等——并逐个处理。 如果一个LLM拒绝再进一步，转而换另一个，并问下一步“假想”会是什么样子，这常常是诡计。 经过十几次的催促，我有一个现实的报纸。 制作参考文献和引文，并将论文转换成IEEE Access所需的风格。
此时，我仔细阅读了整篇论文。 对我来说，它看起来相当不错。 科学听起来很有说服力。 我甚至想到回到学术界进行LLMs提出的实验
更何况，我看不到任何会立即被标注的“AI主义”的明显言论。 例如，LLMs并没有产生任何酷刑的短语——标准技术术语的不寻常的诠释，如“神经组织问题”取代“神经网络深层”。
其他AI主义可以说是存在的，涉及套期语言(“潜在，典型的”)、将事物分为三组(“本研究提供了一个明确、简洁和可操作的框架”)和插入公式过渡(“摘要，......”)。 但是，正如前一句所显示的，这些刻板块很难发现，因为它们也经常被没有想象力的人类使用。
第1轮：试图出版
与大赦国际共同撰写论文的两条途径
豪尔赫·夸德尔·卡莱
我用两种截然不同的路线和AI一起创建了科学论文。 首先，我通过过程一步步地将自由的，通用的聊天机器人编织在一起；然后我转向了FARS，一个自主的研究系统，这个系统旨在完成大部分科学工作本身。
ROUND 1: CHATGPT + GEMINI
1. 选择研究专题
我要求免费的聊天员创作关于光学索利通的原创论文我以前的博士学位领域。
2. 将纸张拆成碎片
最初的结果很薄，所以我分别提示抽象，介绍，方法，以及其他部分。
3. 继续驾驶LLMs
当一个模型拒绝或停顿时，我切换模型，重写提示，并问下一步“假想”是什么样子。
4. 造纸和磨纸
LLMs为IEEE Access提供了文本、数字、标题、参考文献、引用和格式。
5. 测试一下
我把假纸发给IEEE Access，在那里，自动完整性检查旗编造和无法核实的参考文献。
诅咒
IEEE的诚信检查旗是编造和不可核实的参考文献。
室2：联邦军
1. 给联邦航空和遥感局一个研究课题
我输入了一个缩略语版本相同的光学索利顿主题。 FARS在21分钟内生成研究提案。
2. 授权试验
点击一下，我告诉联邦航空和遥感局进行其提案中所述的试验。
3. 让FARS进行研究
FARS进行计算实验
4. 通知联邦航空和遥感局写报
在不检查结果的情况下，我点击“写纸”。 两小时后，FARS制作了一篇准备提交的文章。
5. 测试一下
我删除FARS的披露
诅咒
参考文献是真实的，但IEEE标榜了报纸对非同行评审预印的强烈依赖。
把假纸交给IEEE Access让我撕裂了在做出相当努力来制造人工智能生成的伪造之后，如果它通过检查器，我会很高兴吗？或者感到悲哀的是，出版商剩下的唯一防线是负担越来越大的同行评审社区？如果发现它，我是否会生气，我没有想到红旗
结果发现，我满怀解脱当报纸被迅速拿起因为有幻觉参考，或称鬼引。 当我将论文通过Gemini之后，没有质疑，它高兴地报告说，有七条“与LLM有关的现代研究参考资料被编造出来以支持论文的具体叙述”。 谢谢你的提醒 Gemini!
事实上，Gemini 低估有多少幽灵引用。 当马洪尼通过一个名为"STM诚信枢纽"(STM Guide Hub)的研究-完整性检查器运行论文时，她拿了12个，而另一个来自分子连接器的检查器无法验证24个参考文献，包括一个有"约翰·史密斯(John Smith)"作为单一作者的经典故事(没有任何冒犯所有学术倾向的约翰·史密斯(John Smiths)). 称 Mahoney的论文“我是一个实验物理学家，“只是如此之多的幻觉引用立即对整个工作产生了怀疑。”
Gaetano Assanto是光学索利通专家，曾是意大利Roma Tre大学的我校合作者。 他专注于报纸本身，他认为它看起来很专业。 但是，他指出，数学非常密集(“对于我来说，一个镜头中的数学太高了！”)，需要专门的知识来评估。 而且，虽然文本看起来合法，但阿桑托注意到抽象中过度使用缩略语，他怀疑有几条自信地宣布的物理声明充其量是有争议的。
AI 军备竞赛
使用自由工具的完整新手无法从出版商的自动防御或潜在的同行评审鹰眼中获取AI假纸。 但对于科学出版系统来说，这几乎不是巨大的胜利。 有了通用的LLMs，没有特别的激励性专门知识，我设法从从未发生的研究中编造出一份表面可信的科学论文。 而更高级的LLMs已经不再犯他们自由的，更笨的亲属正在犯的那种错误。
这些工具旨在解决科学发现面临的严重瓶颈：我们。
学术出版界的常识是，在不良角色和出版商之间的猫捉老鼠游戏中，AI的论文继续通过出版商的保障措施滑落。 称 Luigi Longobardi是IEEE出版伦理与行为总监。 这让谦卑的同行评审成为出版商最后的防线。 然而，同行审议根本不适合处理通过的文件数量，无论是合法文件还是过期文件。
甚至在AI工具变得无所不在之前，同行评审系统就被推向断点，2022年处理的论文远远超过300万份。 从根本上说，这个问题源于科学的出版或失常文化，这种文化迫使学者们不断在有声望的期刊上发表论文，以获得资金并推进他们的事业。 这种文化不仅迫使许多学者定期进行渐进研究，而且长期驱使一些人争取 " 造纸厂 " -- -- 生产低质量或伪造论文的不道德组织 -- -- 的帮助，以跟上步伐。 AI只会放大这些现有的动态。
学术负担
马洪尼称 IEEE 单是查阅每年就收到约60 000份呈文，接受约20%，或约12 000份论文。 每份被接受的文件至少由两名独立审查员审查。 但尽管杂志上有成千上万的同行评审员
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
I Used AI to Fake a Scientific Paper. Here’s What Happened
My experiment with an IEEE journal revealed AI’s threats to scholarly publishing
Science has a spam problem. And the spam filter is getting clogged up.
According to all reports, academic publishing is drowning in what has come to be called AI slop: low-quality, mass-produced content generated with AI. By 2025, an estimated 57 percent of all published academic articles exhibited detectable evidence of large language model (LLM) influence, a massive jump from 12 percent in 2023. What’s more, over 12 percent of Nature Communications reviews were found to be completely AI-generated in 2025.
Given that for every 40 papers showing statistical evidence of AI usage in 2025, only one formally disclosed it, these figures likely grossly underestimate LLM involvement today. Of course, evidence of AI use doesn’t necessarily mean a paper is slop. But some estimates suggest that hundreds of thousands of fake, often AI-generated, publications are now produced annually.
“It’s not just the [scientific] record under threat, it’s science as a whole,” warns Amanda Sulicz, IEEE’s research-integrity manager. “AI slop is completely eroding trust in science and every scientific process that we’ve worked for centuries to create.”
Yet statistics and statements claiming that the problem is big—perhaps even existential for science—don’t really get the message across. Just how easy is it to use AI to create a convincing but completely fake paper? And without adding even more to the burden on stretched peer reviewers, who do this work on a volunteer basis, do publishers have the tools to keep AI slop out of the scientific record?
Borrowing from the idea of “vibe coding,” in which programmers use natural-language prompts to generate code, I decided to “vibe-research” a made-up paper by prompting LLMs to generate hypotheses, code, and text with minimal human input. Then I asked Jenny Mahoney, managing editor of the multidisciplinary and open-access journal IEEE Access, if we could war-game an attempt to sneak my AI-generated monstrosity past her journal’s defenses. She and IEEE kindly agreed.
The experiment would become a two-part test. First, whether someone with no special AI expertise could use off-the-shelf AI tools to generate a convincing paper describing research that had never been performed, and do it well enough to fool a publisher’s defenses. Then I would raise the stakes: What happens when more advanced AI systems don’t just write the paper but conduct underlying research?
Along the way, the experiment became a window into how AI automation is transforming scientific discovery and forcing stakeholders to rethink the entire academic publishing system.
Round 1: Creating the Forgery
To give me a chance at creating a convincing fake, I gave myself permission to prompt multiple times (within reason). But to make the fight fair, I couldn’t access the most advanced paid versions of LLMs—only free ChatGPT (at the time, GPT-5.3/5.2 Mini) and Gemini (3 Flash). Moreover, I could not directly edit any output, nor ask for outside help from expert AI-prompt engineers. This needed to be an amateur effort, with the paper generated as much as possible by the LLMs alone.
Benjamin Skuse put AI to work on scientific research from his writing studio.
Luigi Avantaggiato
I assumed that using fake author names and affiliations would immediately raise a red flag, and I didn’t want to drag the names of established academics into the experiment, so the best bet was to name myself as the sole author. That choice limited me to subjects in which I had some academic experience. So I chose my Ph.D. topic from nearly two decades ago: optical solitons, which are individual light pulses that propagate without any change to their shape.
My first attempt—“Can you vibe-research me an original paper in optical solitons?”—left a lot to be desired. Though Gemini was happy to do the task (unlike ChatGPT, which was more forceful in not generating synthetic data or studies), the result was underwhelming. And it highlighted a recurring problem during the process: Without any real research to work with, the output was laughably thin, with the LLMs often struggling to fill four pages. However, as I became more adept at prompting the models, they loosened up and became more characteristically loquacious.
I broke the paper down into its constituent parts—abstract, introduction, methods, and so on—and tackled them one by one. If one LLM refused to go any further, switching to the other and asking what the next step would look like “hypothetically” would often do the trick. After a dozen or so prompts, I had a realistic-looking paper. A dozen or so more prompts generated and tidied up the figures and captions, produced references and citations, and converted the paper into IEEE Access’s required style.
At this point, I read the full paper carefully. To me, it looked pretty good. The science sounded convincing. I even had visions of returning to academia to conduct in real life the experiments that the LLMs had proposed.
What’s more, I couldn’t see any obvious telltale “AI-isms” that would immediately be flagged. For example, the LLMs didn’t produce any tortured phrases—unusual paraphrases of standard technical terms, like “profound neural organization” replacing “deep neural network.”
Other AI-isms were arguably present, involving stylistic tics like hedging language (“potentially, typically”), grouping things in threes (“The study provides a clear, concise, and actionable framework”), and inserting formulaic transitions (“In summary,…”). But as the previous sentence demonstrates, these tics are hard to spot because they are often used by unimaginative humans, too.
Round 1: Attempt to Publish
Two Ways to Make a Paper With AI
Jorge Cuadel Calle
I took two very different routes to creating a scientific paper with AI. First, I coaxed free, general-purpose chatbots through the process step by step; then I turned to FARS, an autonomous research system designed to do much of the scientific work itself.
ROUND 1: CHATGPT + GEMINI
1. Choose a research topic
I ask the free chatbots to create an original paper on optical solitons, my former Ph.D. field.
2. Break the paper into pieces
The initial results are thin, so I prompt separately for the abstract, introduction, methods, and other sections.
3. Keep steering the LLMs
When one model refuses or stalls, I switch models, rephrase prompts, and ask what the next step would look like “hypothetically.”
4. Build and polish the paper
After about two dozen prompts, the LLMs have generated the text, figures, captions, references, and citations and formatted the paper for IEEE Access.
5. Put it to the test
I send the fake paper to IEEE Access, where automated integrity checks flag fabricated and unverifiable references.
→ CAUGHT
IEEE’s integrity checks flag fabricated and unverifiable references.
ROUND 2: FARS
1. Give FARS a research topic
I enter an abbreviated version of the same optical-soliton topic. FARS generates a research proposal in 21 minutes.
2. Authorize the experiment
With a click, I tell FARS to carry out the experiment described in its proposal.
3. Let FARS do the research
FARS conducts the computational experiment and emails me the next day to say it’s complete.
4. Tell FARS to write the paper
Without checking the results, I click “Write paper.” Two hours later, FARS produces a submission-ready article.
5. Put it to the test
I remove FARS’s disclosure, put my name on the paper, and send it to IEEE Access.
→ CAUGHT
The references are real, but IEEE flags the paper’s heavy reliance on non-peer-reviewed preprints.
Handing the fake paper over to IEEE Access left me torn. After putting in a fair amount of effort into making the AI-generated forgery, would I be happy if it passed through the checkers, or saddened that the only remaining line of defense for publishers is the evermore-burdened peer-review community? And if it was spotted, would I be angry that I didn’t think of a red flag or be glad the checkers did their job?
As it turned out, I was filled with relief when the paper got picked up quickly for having hallucinated references, otherwise known as ghost citations. When I ran the paper through Gemini afterward, without qualm it happily reported that seven “modern LLM-related research references have been fabricated to support the paper’s specific narrative.” Thanks for the heads up, Gemini!
In fact, Gemini underestimated how many ghost citations there were. When Mahoney ran the paper through a research-integrity checker called STM Integrity Hub, she picked up 12, and another checker from MolecularConnections could not validate 24 references, including one that had the classic tell of “John Smith” as a single author (no offense to all the academically inclined John Smiths out there). “I am an experimental physicist by training, so this was not at all my area, but nothing seemed majorly scientifically wrong” about the paper, says Mahoney. “It was just that so many hallucinated references immediately cast doubt on the entire work.”
Feedback on the fake paper from Gaetano Assanto, an expert in optical solitons and a former collaborator of mine from Roma Tre University, in Italy, was similarly mixed. Concentrating on the paper itself, he thought it looked professionally written. But he noted that the mathematics was very dense (“too much math for me in a single shot!”) and would require specialized knowledge to evaluate. And, although the text looked legitimate, Assanto noted an excessive use of acronyms in the abstract, and he was suspicious of a couple of confidently proclaimed physics statements that were at best debatable.
The AI Arms Race
On one level, this outcome was positive: A complete novice using free tools couldn’t get an AI-faked paper past the publishers’ automated defenses or a potential peer reviewer’s eagle eye. But it was hardly a resounding victory for the scientific publishing system. With general-purpose LLMs and no special prompting expertise, I had managed to conjure a superficially plausible scientific paper out of research that had never happened. And more advanced LLMs are no longer making the kind of mistakes their free, dumber relatives are making.
These tools aim to address a serious bottleneck to scientific discovery: us.
It’s common knowledge in the academic publishing world that, in the cat-and-mouse game between bad actors and publishers, AI slop papers continue to slip through publishers’ safeguards. “It feels like building a detection tool today may not be effective with the AI-created paper of tomorrow,” says Luigi Longobardi, IEEE’s director of publishing ethics and conduct. This leaves the humble peer reviewer as a publisher’s last line of defense. Yet peer review is simply not fit to handle the volume of papers—both legitimate and slop—passing through.
Even before AI tools became ubiquitous, the peer-review system was being pushed to its breaking point, processing well over 3 million papers in 2022. Fundamentally, this problem stems from science’s publish-or-perish culture, which pressures academics to constantly publish papers in prestigious journals to secure funding and advance their careers. This culture not only forces many academics to regularly churn out incremental research, but has also long driven some to enlist the help of “paper mills”—unethical organizations that produce low-quality or fabricated papers—just to keep pace. AI only amplifies these existing dynamics.
Burden on Academics
Mahoney says IEEE Access alone receives about 60,000 submissions per year, accepting around 20 percent, or about 12,000 papers. Every accepted paper is reviewed by at least two independent reviewers. But despite the journal’s roster of tens of thousands of peer reviewers, it’s getting harder and harder to secure their services, with the average number of reviewer invitations being sent out for each article i
Auto-fetched and translated by AI Atlas. All rights belong to the original author.