这几年，随着大语言模型技术的发展，AI从2023年的人工智障，慢慢变成了如今几乎可以和人类专家媲美的人工智能，其发展速度令人叹为观止，国外有chatgpt，claude等AI模型，国内也有诸如deepseek、kimi、豆包等AI模型。因为AI实在太强了，再加上人类的润色专家可能要花费数天乃至数周才能完成润色的论文，AI可能几个小时甚至十几分钟就能完成润色，就使得很多本身英语水平不是特别好的国内学者会选用AI来帮助润色其科研论文，以求发表在国外顶尖刊物上面。
然而，AI润色一定比人工润色好吗？抱着寻求这个问题的答案，我在论文润色行业的知名机构——AJE官网上，拿一篇真实样例文稿去进行A/B测试，分别用AI工具和人工润色，结果有点出乎意料…首先，这篇示例文稿是公布在aje官网的一篇润色论文，其名为“Emergence of Amoxicillin Resistance and Identification of Novel Mutations of the pbp1A Gene in Helicobacter pylori in Vietnam”，已于2022年发表BMC Microbiology上。
这篇论文的主要内容是关于越南幽门螺杆菌阿莫西林耐药性的研究，专业领域范畴属于基础医学。我选用的AI是Deepseek v4 pro，这是目前国内比较顶尖的一批AI大语言模型。下面举一些具体的润色实例来展示一下两者的表现：
实例1
原文：
Methods: Naïve patients who presenting with dyspepsia undergoing upper gastrointestinal endoscopy were recruited. Rapid urease test and PCR assay were used to diagnose H. pylori infection. Amoxicillin susceptibility was examined by E-test. Molecular detection of mutant pbp1A gene conferring amoxicillin resistance was carried out by real-time PCR followed by direct sequencing of PCR products. Phylogenetic analyses were performed using Tamura-Nei genetic distance model and Neighbor-Joining tree build method.
Deepseek润色：
Methods: Naive patients presenting with dyspepsia who underwent upper gastrointestinal endoscopy were recruited. Rapid urease test and PCR assay were used to diagnose H. pylori infection. Amoxicillin susceptibility was examined by E-test. Molecular detection of the mutant pbp1A gene conferring amoxicillin resistance was carried out by real-time PCR followed by direct sequencing of PCR products. Phylogenetic analyses were performed using the Tamura-Nei genetic distance model and the neighbor-joining tree building method.
人工专家润色：
Naive patients who presented with dyspepsia undergoing upper gastrointestinal endoscopy were recruited. Rapid urease tests and PCR assays were used to diagnose H. pylori infection. Amoxicillin susceptibility was examined by E-tests. Molecular detection of the mutant pbp1A gene conferring amoxicillin resistance was carried out by real-time PCR followed by direct sequencing of the PCR products. Phylogenetic analyses were performed using the Tamura-Nei genetic distance model and the neighbour-joining tree building method.
综合对比这三者可以看到，deepseek确确实实把 presenting、build→building 这类明显有误的语法顺了，但 test/assay/E-test 的单复数没统一，且拼写仍用美式 neighbor；
AJE的人工润色统一成复数并将全文内容改成了英式 neighbour，其他涉及到英式和美式的拼写对比，AJE的全职编辑也统一将其改成了英式（注意：这取决于作者在提交稿件时，告诉编辑，采用美式还是英式英语润色），例如后文的analyse、colonises。如果不对AI提特殊需求的话，这点不只是针对deepseek，它绝对不会帮你改，但英式和美式的英语拼写其实特别体现作者的基本写作素养。
实例2：
原文：
Results: There were 308 patients (46.1% male and 53.9% female, p = 0.190) with H. pylori infection. … The E-test was used to test for susceptibility to amoxicillin … 270 sequences of pbp1A gene fragments were analyzed. There were 77 amino acid substitution positions investigated, spanning amino acids 310 – 596, with the proportion varied from 0.4% to 100%. … Phylogenetic analyses suggested other molecular mechanisms might contribute to the amoxicillin resistance in H. pylori besides alterations in PBP1A.
Deepseek润色：
Results: There were 308 patients (46.1% male and 53.9% female, p = 0.190) with H. pylori infection. … The E-test was used to test for susceptibility to amoxicillin … 270 sequences of pbp1A gene fragments were analyzed. There were 77 amino acid substitution positions investigated, spanning amino acids 310–596, with the proportion varying from 0.4% to 100%. … Phylogenetic analyses suggested that other molecular mechanisms might contribute to amoxicillin resistance in H. pylori besides alterations in PBP1A.
人工专家润色：
Results: There were 308 patients (46.1% men and 53.9% women, p = 0.190) with H. pylori infection. … The E-test was used to determine the susceptibility to amoxicillin … 270 sequences of pbp1A gene fragments were analysed. There were 77 amino acid substitution positions investigated, spanning amino acids 310–596, with the proportion varying from 0.4% to 100%. … Phylogenetic analyses suggested that other molecular mechanisms might contribute to amoxicillin resistance in H. pylori in addition to the alterations in PBP1A.
对比这三者内容不难发现，deepseek把数字范围、varying、that 都顺了，但保留 male/female（人工改 men/women）、test for（人工改 determine the）、analyzed（人工改 analysed）、besides（人工改 in addition to）。综合来说，人工专家修改的更加贴合原作者想表达的意思，deepseek也能修改一些语法错误，但没办法理会作者真正想表达的内容。
实例3：
原文：
Multiple strains of H. pylori can co-infect the same patient (33). The co-infection detected by vacA genotype in our study is presented at a rate of 12.3%. Similar proportion in term of co-infection determination, was obtained when confirmed by random amplified polymorphic DNA (RAPD) fingerprinting (12.5%; p = 0.500) (33), or through vacA and iceA genotyping (11.0%; p = 0.255) (30). Therefore, co-infection needs to be excluded before evaluating the relationship between H. pylori genotypes and other factors such as disease status and clinical symptoms to maintain accuracy.
Deepseek润色：
Multiple strains of H. pylori can co-infect the same patient (33). The co-infection detected by vacA genotype in our study was presented at a rate of 12.3%. A similar proportion in terms of co-infection determination was obtained when confirmed by random amplified polymorphic DNA (RAPD) fingerprinting (12.5%; p = 0.500) (33) or through vacA and iceA genotyping (11.0%; p = 0.255) (30). Therefore, co-infection needs to be excluded before evaluating the relationship between H. pylori genotypes and other factors such as disease status and clinical symptoms to maintain accuracy.
人工专家润色：
Multiple strains of H. pylori can coinfect the same patient (33). The coinfection detected by vacA genotype in our study presented a rate of 12.3%. A similar proportion in terms of coinfection determination was obtained when confirmed by random amplified polymorphic DNA (RAPD) fingerprinting (12.5%; p = 0.500) (33) or through vacA and iceA genotyping (11.0%; p = 0.255) (30). Therefore, coinfection needs to be excluded before evaluating the relationship between H. pylori genotypes and other factors, such as disease status and clinical symptoms, to maintain accuracy.
在这次修改里面，deepseek 修改了时态（is presented→was presented）、in term of改成in terms of、加了冠词，但全文保留带连字符的 co-infection；人工专家则是按领域惯例统一去掉连字符为 coinfection。
在一些很细分的领域，由于一些很地道的表述AI的语料库可能学习的不是特别够，它在修改的时候可能就会忽视，而作为本领域的专家来说，这种地道的表达更加信手拈来。
总结这三处的润色实例，不难发现AJE的专业编辑能够做到更好的全局把控，其中包括英式美式拼写的统一、时态统一、单复数统一等，这点AI目前来说做的有所欠缺；其次，AI有时候不会去纠偏作者错误的表达，可能会顺着作者表达的错意继续润色，但人工专家能够看出来并加以修改；最后也是很重要的一点，在一些细分领域，AI的语料库不一定能够学习到专业的表述习惯，表达可能像一个外行，这也是人工专家的巨大优势。
另外，也是最重要的一点，人工润色，尤其是像AJE这样的机构拥有的同领域母语编辑，最大的价值在于理解而非改写。他们能读懂你的研究逻辑，在保证不改变学术原意的前提下优化表达，知道哪些术语必须保留、哪些限定词不能动、怎么写才是符合学术界公认的逻辑习惯。他们还能保留作者独特的表达风格和写作习惯，让文章读起来是一个真实的学者写的，而不是ai几秒钟生成的流水线产物。
