论学术圈的AI审稿

On the AI Reviewer in Academia

2025年，人工智能领域的一个重要学术会议AAAI（人工智能促进协会）发布公告，开始实验性地在其论文审稿的流程中引入AI审稿人。在这个实验之中，AI将会作为一个辅助性的审稿人，协同其他的人类审稿人一起为投稿稿件提供评估意见，并且不会给出拒稿或收录的意见，最终裁决权依然在人类的手中。这毫无疑问是学术界生态应对AI冲击的一个积极的早期尝试。

In 2025, AAAI (the Association for the Advancement of Artificial Intelligence), a major academic conference in the field of artificial intelligence, announced that it would begin experimentally introducing AI reviewers into its paper review process. In this experiment, AI will serve as an auxiliary reviewer, collaborating with human reviewers to provide evaluation feedback on submitted manuscripts, but will not make decisions on acceptance or rejection. The final decision-making authority will remain in human hands. Undoubtedly, this represents a proactive early attempt by the academic community to respond to the challenges posed by AI.

实际上，与其说这是一次开创性的实验，更不如说这是一种对大趋势的官方追认。虽然很少有圈内人会公开承认，但各个会议、期刊的审稿人用AI来生成评估意见、甚至直接撰写审稿报告，早就已经成为了一个普遍的现象。期刊和会议当然都会公开声明禁止审稿人使用AI来进行审稿。但一方面，审稿对审稿人而言本就属于义务劳动；另一方面，审稿本身缺乏质量把控，很多时候人的审稿意见反而还不如AI撰写的报告。因此，相比于强硬地拒绝AI的参与，审慎地引入AI审稿人是符合趋势的明智决策。

In fact, rather than calling this an innovative experiment, it is more accurate to say that it is an official recognition of an emerging trend. Although few insiders will openly admit it, it has already become common practice for reviewers at various conferences and journals to use AI to generate evaluation comments or even to directly write review reports. Of course, journals and conferences all officially declare that reviewers are prohibited from using AI in the review process. However, on the one hand, reviewing is essentially voluntary labor for the reviewers; on the other hand, there is a lack of quality control in the review process itself, and in many cases, the reports written by human reviewers are actually not as good as those generated by AI. Therefore, compared to rigidly rejecting the involvement of AI, cautiously introducing AI reviewers is a wise decision that aligns with the current trend.

AAAI在这一次初步实践中，明确提到了AI不会对收录或拒稿给出建议，最终的裁决权依然在人类手中。在当下的时间点，这个选择是无可非议的。AI的能力虽然在过去几年之中飞速进步，但到目前为止相比于领域内的顶尖专家还有相当的差距，因此只能作为参考辅助而不能作为决策主体。这个理由在现在也是完全立得住脚的。

In this initial experiment, AAAI has explicitly stated that AI will not make recommendations regarding acceptance or rejection, and the final decision-making authority will still rest with humans. At this point in time, this choice is entirely reasonable. Although AI capabilities have advanced rapidly in recent years, there is still a considerable gap compared to top experts in the field, so AI can only serve as an auxiliary reference rather than as a primary decision-maker. This rationale is completely valid for the present.

可以看出，问题的关键便在于AI的能力。当AI的能力还在人类顶尖专家之下时，保留人类裁决权的理由就依然是充分的。但如果AI的能力已经超越人类顶尖专家了呢？在这个假想之中，我们假设一个期刊或者会议都能够建立五个AI的代理系统，第一个代理系统负责评价稿件在当下的整体贡献，第二个系统负责检查稿件结论的正确性，第三个系统负责评估稿件的技术难度，第四个系统负责预估稿件可以产生的影响力，第五个系统则负责撰写整体的修改意见，并且这五个系统都拥有顶尖专家以上的水准。那么请问此时引入人类审稿人的意义是什么？引入这个极端的设想（或许也没那么极端，三四年后可能就会成为现实），我们也就能够更好地从底层审视同行评审制度的存在逻辑。

It is clear that the key issue lies in the capabilities of AI. As long as AI’s abilities remain below those of top human experts, there are ample reasons to retain human decision-making authority. But what if AI’s capabilities have already surpassed those of the top human experts? Suppose, hypothetically, that a journal or conference could deploy five AI agent systems: the first evaluates the overall contribution of the paper, the second checks the correctness of the conclusions, the third assesses the technical difficulty, the fourth estimates the potential impact, and the fifth drafts comprehensive revision suggestions—and all five systems operate at a level above top experts. In that case, what would be the purpose of involving human reviewers? By considering this extreme scenario (which may not be so extreme, in three or four years it could become reality), we can better examine the fundamental logic behind the existence of the peer review system. 

不妨思考这样一个基础的问题：当我们做出了一个学术成果，我们为什么要投稿到AAAI、NeurIPS、ICML、Nature、Science等顶尖会议或者期刊？如果说是为了能够把成果向全球学术界展示，那么现在已经有Arxiv等平台可以让所有人快速获取最新的研究成果，甚至把成果放在个人主页上都能获得理论上的连接全球同行的渠道。另一方面，如果说是为了经过同行评审而获得一种质量上的保障，那现在完全可以做到把所有新成果发布到网络上，接受所有人的公开评估并进行充分交流讨论，这个过程中足以更公正地筛选出具有重要价值的工作，根本不需要3-5个匿名审稿人来给予认可。但是这么多年过去了，我们发现学术界依然将各个领域中具有高度封闭性的这些协会、组织、期刊、会议的认证当作是硬通货，丝毫没有变得更加开放的趋势。从这个观察中不难得出结论：所谓的成果发表和同行评议这个机制，其目标早已不是为了推动学术进步（这充其量只是一个门面）。它的本质是一种相当纯粹的权力游戏，而组织的背书则是一种能够快速衡量“可交换价值”的符号。

Let’s consider a fundamental question: when we have produced a scholarly result, why do we choose to submit it to top conferences or journals like AAAI, NeurIPS, ICML, Nature, or Science? If the goal is to present our work to the global academic community, platforms like Arxiv now allow anyone to quickly access the latest research, and even posting results on a personal homepage can, in theory, connect us with peers worldwide. On the other hand, if the purpose is to obtain some quality assurance through peer review, it is now entirely possible to publish all new findings online, receive open evaluations from everyone, and engage in ample discussion—this process would be sufficient to fairly filter out work of genuine value, without the need for recognition from three to five anonymous reviewers. 

Yet after so many years, we find that the academic community still treats the certification from these highly exclusive associations, organizations, journals, and conferences as hard currency, showing no signs of becoming more open. From this observation, it’s not hard to conclude: the mechanism of publication and peer review is no longer primarily about advancing scholarship (at most, that serves as a façade). At its core, it is a rather pure power game, and institutional endorsement has become a symbol that allows for the rapid measurement of “exchangeable value.” 

具体而言，一个学者产出了多少成果不是最重要的，其在多大程度上得到这些组织的背书更能反映其地位；有了这些支撑，高校也便能够更好地评估学者的声誉，从而决定是否招揽其作为本校的职工；而高校的声誉则又反过来通过其学者的声誉、以及他们的成果获得的组织背书，在各种“高校排名”中力争上游。更通俗地说，高校的声誉通过其职工和他们身上附带的各种组织符号来决定；学者的声誉由其工作高校和身上附带的组织符号决定；而组织的声誉又通过其所能够关联的学者与高校声誉决定。这形成了一个完美的“循环论证”，它们共同撑起了学者在社会中的稳固权力与地位。这个描述并不否认很多学者的工作的确带来了切实的知识进步，但需要注意的是：我们没有任何对比实验能够说明是否另一种组织模式能够带来更好的进步，同时也要注意到：最近二三十年以来的知识和技术进步，学术界在其中越来越处于一个从属地位。

Specifically, the number of research outputs a scholar produces is not the most important factor; rather, the extent to which their work is endorsed by these organizations better reflects their status. With this backing, universities can more effectively evaluate a scholar’s reputation and decide whether to recruit them as faculty. In turn, a university’s reputation is determined by the reputation of its scholars and the organizational endorsements attached to their work, which help the institution climb various university rankings. Put simply, a university’s reputation is determined by its faculty and the organizational credentials they carry; a scholar’s reputation is determined by their university affiliation and their own organizational credentials; and the reputation of the organizations themselves is determined by the scholars and universities with which they are associated. This forms a perfect “circular logic,” collectively supporting the stable power and status of scholars in society. 

This description does not deny that the work of many scholars has indeed led to genuine advances in knowledge. However, it is important to recognize that we have no comparative experiments to determine whether another organizational model might lead to greater progress, and it should also be noted that, in the past two or three decades, academia has played an increasingly subordinate role in driving knowledge and technological advancement. 

从这个角度来讲，同行评议制度的必要性也就清晰可见了。这个制度从本质上就并非为了保证公平性或者发表水平，而是这个共同体牢牢把握内部权力的必要手段。尽管当前同行评议在审查文章创新性、方法严谨性、可重复性等方面早已遭受了普遍的质疑，这个制度依然顽强地存在于学术运作的必要流程之中。因此，AAAI明确地声明AI不会有裁决权力，除了能力的差距外，另外一个更关键的、也是官方不会公开声明的理由则是：如果AI具有裁决权力，这便意味着官方变相承认AI已经有能力匹配会议方所能组织的最好的专家团队的水平，这立马就会引来一个棘手的疑问：你这个组织存在的理由是什么？这种疑问会开始破坏这个循环论证游戏的根基，从而威胁到共同体内部的所有人在社会之中的精英地位。在当下，这个循环游戏之所以能够玩得下去，就是因为我们的社会整体上依然相信这样的一个叙事：处在这个权力循环之中的人有能力引领社会的发展。一旦这个叙事被打破，利益链条的崩溃也会异常迅速。

From this perspective, the necessity of the peer review system becomes clear. The true essence of this system is not to ensure fairness or publication standards, but rather to serve as a crucial mechanism by which the academic community maintains internal power. Although peer review has long been widely questioned for its effectiveness in evaluating the novelty, methodological rigor, and reproducibility of articles, it remains a tenaciously entrenched part of academic operations. 

Therefore, AAAI’s explicit statement that AI will not have decision-making authority is not only due to the current gap in capabilities, but also—more crucially, and never officially acknowledged—because granting AI such authority would be a tacit admission that AI is already capable of matching the best expert panels the conference can assemble. This would immediately raise a difficult question: what is the purpose of your organization’s existence? Such a question would begin to undermine the very foundation of this circular logic, thereby threatening the elite status of everyone within the academic community. At present, this circular game continues only because society as a whole still believes the narrative that those within this cycle of power are capable of leading social progress. Once this narrative is broken, the collapse of the associated chain of interests will be exceptionally swift.

因此，虽然我很乐意见到AAAI开始试探性地引入AI审稿的机制，但基于以上的逻辑，AI审稿在学术界之中的推进也必然会是缓慢的、扭捏的。AAAI处在AI革命的前沿，采取一些不会威胁到其切身利益的开放性姿态，也符合其本身的定位；学术圈中更多的小共同体则会有意地抵制AI这个强大的“叙事消解器”。但AI的发展节奏不会以他们的意志为转移，数年之内发展出足以超越顶尖人类个体的通用人工智能是一个大概率事件，社会也会逐渐形成AI拥有人类无法匹敌的工具性能力这一共识。学界的精英们必然需要痛苦地接受自己角色的转变：学术研究只是一种生活的选择，不是什么精英的凭证。

 Therefore, while I am quite pleased to see AAAI begin to tentatively introduce AI review mechanisms, based on the above logic, the advancement of AI-based reviewing within academia will inevitably be slow and hesitant. As a leader at the forefront of the AI revolution, AAAI’s adoption of some open policies that do not threaten its own core interests is consistent with its positioning; meanwhile, many smaller communities within academia will intentionally resist AI as a powerful “narrative disruptor.” However, the pace of AI development will not be determined by their will. It is highly likely that within a few years, general artificial intelligence will develop to the point where it surpasses even the best human experts, and society will gradually reach a consensus that AI possesses instrumental capabilities far beyond human reach. The academic elite will inevitably have to painfully accept the transformation of their own roles: academic research will become merely a lifestyle choice, not a certificate of elite status. 

所幸的是，完成这一角色转变后，真正的内在自由才可能随之到来。学术研究是一个“无限游戏”，而不是一个追求认证、背书、话语权的“有限游戏”。

Fortunately, after this transformation, true inner freedom will become possible. Academic research is an “infinite game,” not a “finite game” of seeking certification, endorsement, or discourse power.