记者本杰明·斯库斯（Benjamin Skuse）开展了一项实验，分别使用免费大语言模型和全自动科研系统FARS生成了两篇关于光学孤子的虚假学术论文，并尝试将其提交至IEEE Access期刊 1。该实验旨在测试传统同行评审体系对AI生成内容的检测能力，结果揭示了当前学术出版面临的严峻挑战 1。在背景数据方面，估计显示2025年所有已发表的学术论文中，有57%呈现出可检测的大语言模型影响，这一比例较2023年的12%大幅上升 1。此外，2025年超过12%的《自然·通讯》审稿意见被证实完全由AI生成，而在每40篇有统计证据表明使用AI的论文中，仅有一篇进行了正式披露 1。
针对第一篇由人工提示生成的论文，自动检测工具成功拦截了其中包含的12个幻觉引用（ghost citations），另有24个引用因无法验证而被标记 1。第二篇由安纳莱玛公司（Analemma）开发的FARS系统生成的论文则因过度依赖未经同行评审的预印本以及图表中出现“lhewl”等无意义术语而被拒稿 1。值得注意的是，FARS系统在417小时内创建了166篇论文，其单篇生产流程包括21分钟完成研究提案、次日完成实验以及最终耗时2小时撰写论文 1。目前，安纳莱玛公司在被百度收购后已停止提供相关在线服务 1。
这一事件反映了AI技术对以IEEE Access为代表的高容量出版机构的冲击，该期刊每年收到约60,000份投稿，接受率约为20%（即约12,000篇），每篇接收论文至少需经过两名独立审稿人的审查 1。自2018年以来，IEEE Access为每篇文章发送的平均审稿邀请数量增加了50%，显示出同行评审环节正承受着巨大的资源压力 1。Retraction Watch联合创始人伊万·奥兰斯基（Ivan Oransky）对此评论道：“不发表就出局”的文化是干柴，论文工厂是汽油，而AI则是助燃剂 1。此次实验不仅展示了现有检测工具的局限与效力，也引发了在AI能够自主完成从文献综述到实验执行全过程的背景下，如何重新定义“合法科学”及调整未来科研与出版策略的思考 1。
Journalist Benjamin Skuse conducted an experiment to test the defenses of academic publishing against artificial intelligence, attempting to submit two fabricated papers on optical solitons to the journal IEEE Access 1. The first paper, generated using free large language models with human prompts, was intercepted by automated detection tools after containing 12 hallucinated "ghost citations" and 24 unverifiable references 1. The second submission, created entirely by FARS (Fully Automated Research System), was rejected due to its heavy reliance on unreviewed preprints and nonsensical terms like "lhewl" appearing in charts 1. This case study highlights a broader crisis in scholarly communication, where estimates suggest that 57% of all published academic papers in 2025 show detectable LLM influence, a sharp rise from 12% in 2023 1. Despite this prevalence, only one out of every 40 papers showing statistical evidence of AI usage formally discloses it 1. Furthermore, over 12% of reviewer comments for Nature Communications in 2025 were found to be fully generated by AI 1.
The scale of the challenge facing journals like IEEE Access is immense, as they receive approximately 60,000 submissions annually with a 20% acceptance rate, meaning around 12,000 papers are published each year 1. Each accepted article undergoes review by at least two independent peers, yet the workload has intensified; since 2018, IEEE Access has increased the average number of review invitations per article by 50% 1. The efficiency of AI systems exacerbates this pressure: FARS developed by Analemma (which ceased online services after being acquired by Baidu) produced 166 papers in just 417 hours 1. A single FARS-generated paper required only 21 minutes for the research proposal, completed experiments the following day, and took two hours for final writing 1. Retraction Watch co-founder Ivan Oransky characterized the situation by stating that the "publish or perish" culture is dry kindling, paper factories are gasoline, and AI acts as the accelerant 1.
评论
还没有评论，欢迎留下第一条。