又是一年毕业季，在多个社交平台上，关于论文AI率“亮红灯”的帖子引发热议。
有网友反映，毕业论文中原创内容经一些系统检测后，常被指出AI生成内容比例过高。
不少大学生分享了降低AI率的经历：有人反复修改结果导致语句生硬；有人被迫删减精彩段落；更有甚者调侃需故意“写笨”以规避误判。
A new academic dilemma has emerged as Chinese university students report unexpected "high AI-generated content" alerts when using popular plagiarism detection systems to check their graduation theses that they wrote without AI assistance. Many have resorted to awkward phrasing modifications, deletion of quality paragraphs or even intentional "dumbing down" of language to bypass AI detection algorithms.
据大河报消息，将朱自清的名篇《荷塘月色》与刘慈欣《流浪地球》的片段上传至某常用论文检测系统后，结果显示，这两篇经典作品AI生成内容总体疑似度竟分别达到了62.88%和52.88%。
In a startling experiment reported by media, excerpts from Zhu Ziqing's literary classic Moonlight Over the Lotus Pond and Liu Cixin's science fiction The Wandering Earth were submitted to a commonly used thesis detection system. The results showed 62.88 percent and 52.88 percentAI-generated content suspicion rates, respectively, for these two authors' original works.
面对这样的检测数据，不少网友担心自己的论文也会被误判AI率过高，从而影响毕业。
多所高校检测毕业论文AI率
有大学生花钱“求通过”
据了解，多所高校曾发文对毕业论文AI生成内容进行规范。
面对AI率检测的压力，有大学生开始向声称可以提供“降AI率”服务的店家寻求帮助。
这些服务的常用手法包括替换同义词、调整句子顺序，或加入不必要的修饰和口语化表达。部分商家甚至会先用自动化工具初步修改，再少量人工调整。
但这种“修改”往往使论文语言生硬，学术性降低，对降低AI率效果有限，有时反而弄巧成拙。更严重的是，学生“花钱包过”，还可能面临诈骗和隐私泄露风险。
Meanwhile, a shadow market has emerged offering "AI-reduction" services ranging from mechanical synonym swaps to artificial colloquialization. However, these quick fixes often degrade academic rigor while exposing users to fraud and data breaches.
专家提醒：AI论文反识别或存在误差
“当前AI率检测是AI技术融入学术领域的初期表现，其标准与技术均在发展和完善过程中。”中南大学讲师关耳（化名）分析道。她认为，各高校的探索实践也反映出，在学术层面，一个普遍适用的指导性标准尚未完全形成。
“除了标准不一样让人头疼，AI检测技术本身到底准不准，也是个大问题。毕竟，这直接关系到学生辛辛苦苦写的原创内容，会不会被机器错怪成AI写的。” 关耳觉得“目前的AI检测技术用起来还有些问题”，给出的结果最好只当个参考，不能完全以此来判断一篇论文的好坏。
"Current detection tools represent early-stage attempts to integrate AI governance into academia, not definitive quality assessments," cautioned Central South University lecturer Guan Er (pseudonym). She noted overreliance on flawed detection metrics risks distorting academic values. Unlike clear-cut plagiarism cases, most institutions currently treat high AI scores as teaching opportunities rather than misconduct, reflecting awareness of the technology's limitations.
据央视新闻此前报道，目前国内有多个团队正在从不同方向发力，进行AI论文反识别的检测研究。一些学生和专家提醒，AI工具是根据它捕获的海量信息进行的智能分析，但这些信息的真实性和准确性可能存在缺陷，从而得出错误的结论。
关耳强调，一篇论文怎么样，关键还是看学生有没有自己的思考和原创的东西。
As multiple research teams develop next-generation detection systems, educators emphasize cultivating authentic critical thinking. "The essence of scholarship remains human insight," Guan said.
据她观察，目前多数高校对于AI生成内容比例超标的情况，处理方式与对待查重率超标这类明确的学术不端问题尚有不同，更多的是引导修改，而非直接等同于《学位法》中严厉处罚的代写、剽窃行为。“这种审慎的处理，也说明学术界已认识到AI检测技术的工作原理存在一定局限性”。
堵不如疏，专家看来，强化师生的AI素养非常重要。近两年，众多高校都把人工智能作为通识必修课，不断推动师生AI素养的提升。
来源：中国青年报
