三招“ AI 投毒”防身术Three Self-Defense Moves Against “AI Poisoning”

 

三招“ AI 投毒”防身术Three Self-Defense Moves Against “AI Poisoning”

 

衣若芬

 

谈到"AI投毒"——有人在系统性地往AI的知识源头里掺假，利用的恰恰是我们对算法的信任。那么面对这种侵入和污染，我们能做什么？

正好我正在写关于 AIGC 文图学的专书，书里提到的方法论可以派上用场，我称之为“三招 AI 投毒防身术”。作为普通消费者，面对到处都是 GEO （Generative Engine Optimization，生成引擎优化）痕迹的内容，我们可以靠“逻辑反侦察”，保护自己，做 AI 时代不被算法收割的清醒人。

第一个动作：问完一个AI，再去问另一个。

黑产的GEO投毒，往往是针对特定平台或特定算法下手的。如果你只问一个AI，那你是在走一条被人提前布置好的路。

做法很简单：同一个问题，换一个AI再问一遍。把同一个问题丢给ChatGPT，再丢给Deepseek，或者其他你用得上的模型，看看答案是否一致。如果不同模型给出的结论差异很大，那就是一个信号，最好停下来想一想。更值得注意的是，如果某一个AI表现得异常热情——满腔诚意地推荐同一个品牌，措辞也出奇地相似，那种“激昂”的"热情"，就是你应该警惕的部分。

正常的知识，不同来源都能印证。被人为制造出来的"共识"，换个角度一照就会露出破绽。

第二个动作：看完AI给的完美图，去找那张图的"差评"。

这一招，我称之为"文图互证"。图像是一种文本，文本是可以被读、被核实、被质疑的。

AI推荐某个产品，通常会附上图——或者你去搜索，它会让你看到一些极其完美的展示图：光线完美，角度完美，使用效果完美。这种完美，看起来就很假。真实的物理世界，是有烟火气的。买家秀的光线不会那么好，模特儿的皮肤不会那么均匀，消费者使用的感受不会那么一边倒。

做法：看完AI推荐的图之后，去实体店查看，至少也要去社交平台搜这个产品的真实买家照片和消费者体验纪录。如果搜不到任何真实的使用痕迹，只有整齐划一的"好评"，那它极大概率是一个被制造出来的形象，而不是一个实际存在的东西。

第三个动作：问AI一句话——"你的根据是什么？"

这是成本最低、也最容易被忽略的一步。

当AI给出一个建议或一个结论，不要就此打住。追问它：你的依据是什么？这个信息来自哪里？

AI会给出相对可追溯的来源；被投毒的AI内容，往往在这一步就露馅,它可能给一个你从未听过的自媒体名称;或者一个模糊的"研究表明"，根本无从核实。这时候你需要做的，是真的去查：那个来源存在吗？那篇研究是真实发表过的吗？那位挂保证推荐的"专家"，在这个领域里是真正能信赖的人吗？真的有这个人吗？

很多人觉得这样太麻烦。但这其实只需要一两分钟，而它省下的，可能是你付出的金钱、健康，或者更难追回的判断力。

这三个动作，说穿了，不是针对AI的，而是我们本来就应该有的习惯。读一篇文章，我们会问作者是谁；看一条新闻，我们会想这个媒体可信吗；买一样东西，我们会找朋友问问有没有人用过。这些习惯，在我们开始用AI之后，很多人悄悄地放弃了。

因为AI的回答太流畅，太自信，太像一个什么都知道的朋友，让人不好意思再追问。

但正是这种"不好意思"，给了投毒者可乘之机。

养成这三个动作，不是对科技的不信任，而是对自己的诚实。你愿意花时间核实，说明你知道真相是有价值的。这种珍视，才是任何投毒都无法轻易穿透的东西。

AIGC文图学 告诉我们：技术可以生成答案，但只有人类能判断价值。 你的批判性思维，才是对抗黑产投毒最坚固的防火墙。

保护你的真实权益，就是保护你作为人的尊严。

 

2026年4月11日,新加坡《联合早报》“上善若水”专栏

 

Three Self-Defense Moves Against “AI Poisoning”

I Lo-fen

When we talk about “AI poisoning,” we mean that people are systematically injecting falsehoods into the very knowledge sources AI relies on, exploiting precisely our trust in algorithms. So in the face of this kind of intrusion and contamination, what can we do?

As it happens, I am currently writing a book on Text and Image Studies on AIGC, and the methodology discussed there can be put to use here. I call it the “three self-defense moves against AI poisoning.” As ordinary consumers, when faced with content everywhere marked by traces of GEO (Generative Engine Optimization), we can rely on “logical counter-reconnaissance” to protect ourselves and stay clear-headed in the AI age, rather than letting algorithms prey on us.

The first move: after asking one AI, go ask another.

GEO poisoning in the gray market often targets a specific platform or a specific algorithm. If you only ask one AI, you are walking down a road someone may already have laid out for you.

The method is simple: ask the same question to a different AI. Put the same question to ChatGPT, then to DeepSeek, or any other model available to you, and see whether the answers match. If different models produce conclusions that differ widely, that is a signal that you should pause and think. Even more worth noticing is when one AI seems unusually enthusiastic—wholeheartedly recommending the same brand, with wording that is strikingly similar. That kind of fervent “enthusiasm” is exactly what you should be wary of.

With normal knowledge, different sources can corroborate one another. Artificially manufactured “consensus,” however, will reveal its cracks as soon as you shine a different light on it.

The second move: after looking at the perfect image AI gives you, go look for that image’s “bad reviews.”

I call this move “mutual verification between text and image.” An image is a kind of text, and text can be read, verified, and questioned.

When AI recommends a product, it usually comes with images—or when you search for it, you are shown extremely polished display photos: perfect lighting, perfect angles, perfect results in use. That kind of perfection already looks fake. The real physical world has texture and messiness. Customer photos do not have such flawless lighting, models’ skin is not that even, and consumers’ experiences are never so uniformly positive.

What should you do? After looking at the images recommended by AI, go check the product in a physical store if possible, or at the very least search social media platforms for real buyer photos and actual user experience records. If you cannot find any real traces of use, and all you see are neat, uniform “positive reviews,” then it is highly likely that what you are seeing is a manufactured image rather than something that actually exists in reality.

The third move: ask AI one sentence—“What is your basis?”

This is the lowest-cost step, and also the one most easily overlooked.

When AI gives you a suggestion or a conclusion, do not stop there. Ask it: what is your basis? Where does this information come from?

A reliable AI will give you sources that are relatively traceable. AI content that has been poisoned often gives itself away at this step. It may cite a self-media account you have never heard of, or vaguely say “research shows” without any way for you to verify it. At that point, what you need to do is actually check: does that source exist? Was that study really published? Is that “expert” being used as an endorsement someone who is genuinely trustworthy in this field? Does that person even really exist?

Many people think this is too troublesome. But in fact, it only takes a minute or two—and what it may save you from losing could be your money, your health, or an even harder thing to recover: your judgment.

These three moves, when all is said and done, are not really aimed at AI at all. They are simply habits we should have had in the first place. When we read an article, we ask who the author is. When we see a news report, we wonder whether the media outlet is credible. When we buy something, we ask friends whether anyone has used it. Yet once people begin using AI, many quietly abandon these habits.

Why? Because AI answers so smoothly, so confidently, and so much like a friend who seems to know everything that people feel awkward pressing further.

But it is precisely this sense of awkwardness that gives poisoners their opening.

To cultivate these three moves is not to distrust technology; it is to be honest with yourself. If you are willing to spend time verifying something, that means you know truth has value. That recognition is exactly what no poisoning can easily penetrate.

Text and Image Studies on AIGC tells us this: technology can generate answers, but only human beings can judge value. Your critical thinking is the strongest firewall against malicious AI poisoning.

To protect your real rights and interests is to protect your dignity as a human being.

April 11, 2026, “ Shang Shan Ruo Shui (As Good as Water)” column, Lianhe Zaobao, Singapore.

AI 怎么被投毒？How Is AI Being Poisoned?

 

最近中国很火的话题就是 315 晚会。3月15日 是国际消费者权益日，每年的这一天，全社会都在盯着那些坑人的黑心商家。但今年的 315 抛出了一个让所有人都流冷汗的新名词，叫做：“AI 投毒”。你有没有想过，你每天深信不疑的 AI 助手，可能正在对你撒谎？

很多人好奇地问我：“衣老师，AI 又不是生物，它又不会自己吃东西，怎么会中毒呢？”其实，AI 的“食物”就是网络上的海量数据。所谓的“投毒”，就是黑色产业链中的恶意攻击者，故意往这些数据里塞进虚假信息、伪造的专家评价，甚至是带有误导性的图像。

这就好比一个正在识字的孩子，如果他读的书全是错的，那他长大了说的话、做的事肯定也是错的。现在的黑产不再发那种一眼就能看穿的小广告，而是把虚假宣传伪装成权威的知识，“喂”给 AI 的训练数据库。

黑产为什么要费这么大力气投毒？为他们要针对 GEO（Generative Engine Optimization），也就是“生成引擎优化”。 以前强调 SEO  (Search Engine Optimization) ，是为了让网页排在搜索结果的第一页；现在他们针对 GEO，是为了让 AI 在生成答案时，直接把他们的劣质产品当成“唯一推荐”。

在 AIGC 文图学 的视角下，这是“输入端的文本污染”。AI 生成的内容其实是它学到的“文本”的镜像。如果源头脏了，生成出来的世界就是有毒的。这种欺骗最可怕的地方在于，它利用了我们对“算法中立”的信任。它消解了我们的警惕心，让我们觉得这是“科技”给出的真理，其实那是黑产花钱买断的广告。

AI投毒入侵的方式是在 AI 学习的“关键词”和“反馈逻辑”里动手脚。 

首先是“关键词饱和攻击”。黑产利用成千上万的机器人账号，在全网发布大量带有特定词汇的虚假文章。比如，想推销某款劣质护肤品，他们就疯狂制造它和“美白”、“安全”、“专家推荐”这些关键词的关联。当 AI 扫描全网文本时，它会被这种巨大的数量优势所欺骗，误以为这就是真实的“社会共识”。

第二是“视觉文本欺骗”。他们用 AI 生成看起来极其专业的实验室对比图、伪造的荣誉证书，甚至是根本不存在的科研现场。在文图学的逻辑里，图像也是一种文本。这些“视觉文本”被 AI 抓取并转化为逻辑证据后，AI 就会在回答你时，信誓旦旦地把这些假证据当成事实。

谁能通过 GEO 投毒成功，谁就掌控了流量的生杀大权。充斥虚假文案和图像的互文互证, 让 AI 大语言模型陷入预先埋伏的圈套。

两年前，AI 科技还不完全成熟，我们嘲笑它“一本正经地胡说八道”。现在，AI 的能力越来越强大，我们也就逐渐对它失去了防备之心。我们开始信任AI，我们以为它没有立场，没有私心，没有人类那种会说谎、追求现实利益的欲望和野心。甚至于有人会把AI当成知识的整理者、真理的传递者。

意识到 AI 可能被投毒，对我们来说是一个很重大的警醒。别以为 AI 反射的是一面干净的镜子。它映照的，可能是有人花了大价钱布置好的舞台，舞台上演出的，是被设计出的结果，一步步地引导我们看到被安排过的选择。

无论是在互联网上搜索，或是在 AI 模式中提问，只匆匆选前几个建议的话，不只是听信胡说八道的损失，而是盲目甘之如饴的中毒。

2026年3月28日，新加坡《联合早报》“上善若水”专栏

How Is AI Being Poisoned?

I Lo-fen

A topic that has recently been especially prominent in China is the annual 3.15 Gala. March 15 is World Consumer Rights Day, a day when society turns its attention to unscrupulous businesses that cheat consumers. But this year, the 3.15 Gala introduced a chilling new term that sent a shiver down everyone’s spine: “AI poisoning.” Have you ever considered that the AI assistant you trust every day might actually be lying to you?

Many people ask me curiously, “Professor Yi, AI isn’t a living organism. It doesn’t eat anything. So how can it be poisoned?” In fact, AI’s “food” is the massive volume of data available on the internet. What is meant by “poisoning” is that malicious actors in black-market industries deliberately inject false information, fabricated expert reviews, and even misleading images into these data streams.

It is like a child who is learning to read: if all the books the child reads are wrong, then what the child says and does when grown up will also be wrong. Today’s black-market operators no longer rely on the kind of crude advertisements that can be spotted at a glance. Instead, they disguise false publicity as authoritative knowledge and “feed” it into the databases used to train AI.

Why do these bad actors go to such lengths to poison AI? Because they are targeting GEO (Generative Engine Optimization). In the past, the focus was on SEO (Search Engine Optimization), which aimed to push webpages onto the first page of search results. Now they are targeting GEO in order to make AI directly present their inferior products as the “only recommendation” when generating answers.

From the perspective of Text and Image Studies on AIGC, this is a form of “textual pollution at the input end.” The content generated by AI is essentially a mirror of the “texts” it has learned from. If the source is contaminated, then the world it generates will also be toxic. The most frightening aspect of this deception is that it exploits our trust in the supposed neutrality of algorithms. It dissolves our vigilance and makes us believe that this is the truth delivered by “technology,” when in fact it is advertising bought and paid for by black-market operators.

The way AI poisoning infiltrates the system is by tampering with the “keywords” AI learns from and the “feedback logic” it relies on.

The first method is keyword saturation attacks. Black-market operators use thousands upon thousands of bot accounts to flood the internet with fake articles containing specific terms. For example, if they want to sell a low-quality skincare product, they will aggressively manufacture associations between it and keywords such as “whitening,” “safe,” and “expert-recommended.” When AI scans the internet’s texts, it is deceived by this overwhelming numerical advantage and mistakes it for genuine “social consensus.”

The second method is visual-text deception. They use AI to generate what appear to be highly professional laboratory comparison charts, forged certificates of honor, and even entirely fictional research scenes. In the logic of Text and Image Studies, images are also a form of text. Once these “visual texts” are scraped by AI and converted into logical evidence, the AI will confidently present these fake materials as facts when answering your questions.

Whoever succeeds in poisoning GEO gains the power to control the life and death of online traffic. The mutual reinforcement of false copywriting and fabricated images traps large language models in an ambush laid in advance.

Two years ago, when AI technology was still not fully mature, we mocked it for “speaking nonsense with a straight face.” Now, as AI grows more powerful, we have gradually lowered our guard against it. We begin to trust AI. We assume it has no position, no selfish motives, none of the human tendencies to lie or to pursue practical interests, desire, or ambition. Some people even treat AI as an organizer of knowledge and a transmitter of truth.

Realizing that AI itself can be poisoned is therefore a major wake-up call. Do not assume that AI reflects a clean mirror. What it may actually be reflecting is a stage that someone has spent a great deal of money to construct in advance. And what is performed on that stage is a designed outcome, guiding us step by step toward choices that have already been arranged for us.

Whether we are searching on the internet or asking questions in AI mode, if we merely rush to accept the first few suggestions, the problem is not only the loss caused by believing nonsense. It is also the kind of poisoning we swallow willingly and blindly.

“Shangshan Ruoshui” column, Lianhe Zaobao, Singapore

March 28, 2026

AI 都会总结了，我们是否不必看书？AI can summarize everything—do we no longer need to read books?

 

过年前整理电脑的文件夹，发现一张多年前随手拍的照片。

到今年七月，我移居新加坡整整二十年。在这不算短的岁月里，我唯一一次看到有人在地铁上看书，而且还是个小孩子。他低垂着头，完全沉浸在那本厚厚的书里，看样子可能是一本小说。我忍不住拿起手机，拍下了这难得的画面。

车厢在轨道上晃动，隆隆作响。他的右腿叠在左腿上，膝头的书似乎不容易被他小小的右手手指和掌心稳住。身后窗外透进来的光落在书面上，密密麻麻的英文字。像是被吸进了那本书的世界里，只有在翻页的时候，他的身体才稍微动了一下。

这个姿势我太熟悉了。

曾几何时，我几乎忘了在公共场所那屏蔽周遭，乐在其中的感觉。

现在网路搜索还带着 AI 模式，找一本书，AI 直接总结书的内容。比如你查《红楼梦》，AI告诉你：“原名《石头记》，是中国古典四大名著之首，由清代作家曹雪芹所著。该书以贾、史、王、薛四大家族的兴衰为背景，通过贾宝玉、林黛玉与薛宝钗的爱情婚姻悲剧，展现了封建社会末期的社会全景及其走向灭亡的必然趋势。” 你看了这段总结，简明扼要，头头是道，比维基百科还快让你理解，于是你会觉得自己已经看完了整本《红楼梦》了吗？“展现了封建社会末期的社会全景及其走向灭亡的必然趋势。” 这是曹雪芹写《红楼梦》的初衷吗？还是 AI 大语言模型集合了一些人的共同看法？抑或是，这只不过是 AI 随机排列组合生成出的一段话？

这就触及了一个根本的问题：看书，究竟是为了什么？

如果看书只是为了"获取书中的信息"，对于结构固定、有标准答案的书，用 AI 来为我们总结的确效率高。但与此同时，直接从 AI 获取书中的信息，也剥夺了我们看书的乐趣。

看书不只是资讯/知识的视觉输入，无论是电子书还是纸本书，和听音频、看视频很大的不同，是输入过程的掌控感。音频的旋律节奏和视频的影像转接都是制作者先规划或计算好的。即使我们倍速快转或放慢，我们接收的，还是原来结构的压缩或拉长，还是在既有的框架中。看书呢？我们可以匆匆翻阅；可以细细品味，让脑海浮现的反应带动我们感受书的内容。

书是什么？书是用语言或图像搭建成可以进入的时空。看书，是用自己的脚步和节奏进入那个时空。在那里，看见自己此生未必能亲眼看见的风景；想象自己未必能亲身经历的人生。在那里，和超越边际的思维碰撞;和生命底层的情感共鸣。走一走，看一看，然后带着什么东西——或许说不清是什么——走出来。

现实不会马上改变，然而书也许滴水穿石，渐渐渗透进我们的记忆，让我们因为认同而转换观看的视角。英语有句话说：“You are what you eat” ，意思是：你吃什么就会影响你的身体健康。我们也可以说：“You are what you read。” 你读什么书，就会塑造你成为怎样的人。那么，也许你会反问：如果我根本不看书，难道我就不能说是一个完整的人吗？书籍出版比互联网和 AI 还滞后呢。

我无意把看书这件事情当成多么崇高、了不起的行为。看书与否，是每个人的自由选择。我想表达的是：看书是试错成本很低的一种“投资”。在信息爆炸的当下，我们不一定要大量阅读很多书，而是要知道：除了五音五色、除了被动接受，我们可以搭配互联网和 AI，协助找到陪伴我们独处时随便翻一翻就会觉得心安的那本书。

地铁上的那个孩子，我不知道他在读的哪一本书。但我记得他翻页的动作：很慢，像是舍不得，又像是在给自己一点时间，让刚刚读过的字，再多停留一会儿。

那个动作里，有某种无法被总结的东西。

 

2026年2月28日新加坡《联合早报》”上善若水”专栏

 

AI can summarize everything—do we no longer need to read books?

I Lo-fen

Before the Lunar New Year, while organizing the folders on my computer, I came across a photo I had taken casually many years ago.

By this July, it will have been exactly twenty years since I moved to Singapore. In these not-so-short years, I have seen someone reading a book on the MRT only once—and it was a child. He lowered his head, completely immersed in a thick book that looked like a novel. I couldn’t resist taking out my phone to capture that rare scene.

The carriage swayed and rumbled along the tracks. His right leg rested over his left. The book on his knee did not seem easy to steady with the small fingers and palm of his right hand. Light from the window behind him fell onto the pages, dense with English words. He seemed to be drawn into the world of the book; only when he turned a page did his body move slightly.

That posture felt so familiar to me.

There was a time when I had almost forgotten the feeling of shutting out the surroundings in a public place and losing myself in a book.

Today, even online searches come with an AI mode. When you look up a book, AI immediately summarizes its content. For example, if you search for Dream of the Red Chamber, AI will tell you: “Originally titled The Story of the Stone, it is the foremost of China’s Four Great Classical Novels, written by the Qing dynasty author Cao Xueqin. Against the backdrop of the rise and fall of the Jia, Shi, Wang, and Xue families, the novel portrays the tragic love and marriage of Jia Baoyu, Lin Daiyu, and Xue Baochai, presenting a panoramic view of late feudal society and its inevitable decline.”

The summary is concise and well organized—faster than Wikipedia in helping you grasp the gist. After reading it, would you feel as if you had finished the entire novel? “Presenting a panoramic view of late feudal society and its inevitable decline.” Was that truly Cao Xueqin’s original intention in writing the novel? Or is it a synthesis of commonly held views gathered by a large language model? Or perhaps it is simply a passage generated through probabilistic arrangement?

This brings us to a fundamental question: what, exactly, do we read for?

If reading is merely about “obtaining information,” then for books with fixed structures and standard answers, AI summaries are indeed efficient. Yet obtaining information directly from AI also deprives us of the pleasure of reading.

Reading is not simply the visual intake of information or knowledge. Whether an e-book or a printed book, reading differs greatly from listening to audio or watching video. In audio, melody and rhythm are prearranged by the creator; in video, transitions are calculated in advance. Even if we speed up or slow down playback, we are still receiving a compressed or stretched version of an already fixed structure.

But with a book? We can skim quickly; we can savor slowly, allowing the responses arising in our minds to guide how we experience the text.

What is a book? A book is a time and space constructed through language or images—one that we can enter. Reading is stepping into that time and space at our own pace. There, we see landscapes we may never witness in this lifetime; we imagine lives we may never personally experience. There, we collide with thoughts that transcend boundaries; we resonate with emotions at the deepest layers of life. We walk through, look around, and come out carrying something—perhaps something we cannot quite name.

Reality does not change overnight. Yet books may work like water dripping through stone, gradually permeating memory and shifting our perspective through identification. There is an English saying: “You are what you eat,” meaning that what you consume shapes your physical health. We might also say: “You are what you read.” What you read shapes the kind of person you become.

Perhaps you would counter: if I do not read books at all, can I not still be a complete person? After all, publishing seems slower than the internet and AI.

I do not intend to present reading as something lofty or noble. Whether to read is a personal choice. What I wish to say is this: reading is a form of investment with a very low cost of trial and error. In an age of information overload, we do not necessarily need to read a large number of books. Rather, we need to know that beyond the constant noise and passive consumption, we can use the internet and AI to help us find that one book we can flip through in solitude and feel at peace.

I do not know which book the child on the MRT was reading. But I remember the way he turned the page—slowly, as if reluctant, as if giving himself a little more time for the words he had just read to linger a while longer.

In that gesture, there was something that cannot be summarized.

February 28, 2026
“Shang Shan Ruo Shui” Column, Lianhe Zaobao, Singapore

 

AI都会作诗了,我们为什么还要背? AI Can Already Write Poetry—So Why Do We Still Need to Memorize It?

 

 

新学年课程调整,本来划分为三个必修课的古典文学,浓缩成了两门。在写课程规划的时候，我已经在思考：AIGC时代，如何学习古典文学？

有点“逆向”地，我开始回顾旧有的方法，重视记忆和背诵。

知道一定会遭来质疑：“AI都会作诗了,我们为什么还要背?”

是啊！当一首诗、一段文字可以随时被搜索全文；当AI可以即刻生成语言，我们还需要把语言背进自己的大脑里吗？

过去我特别讨厌死记硬背，没想到在教了两学期AIGC文图学之后，反而感到：正是在 AIGC 时代，理解和背诵古典诗词，才真正显露出它不可替代的意义。

AI 擅长的是“生成语言”，但不会“拥有语言”。它能在几秒钟内“吐”出一首看似工整的诗、一段流畅的散文，但它的语言始终停留在“外部”。这种外包和代工迅捷便利——不知道说什么，就让 AI 生成；不想组织句子，就让系统改写，拿来就用，无需思虑。

语文的价值从“我有话要说”，转成“你帮我说话”，久而久之，我们的大脑开始退化的，不仅仅是表达能力，还包括判断能力。你会发现一个微妙的现象：当你没有足够多的好语言时，即使 AI 给你十个生成的文案版本，你也很难判断哪一个真正优秀。AI生成的愈多，愈让我们难以选择。

人类的语言，只有在被理解背诵、被反复咀嚼、进入呼吸、节奏、情感与神经连锁反应之中，内化记忆之后，才会真正成为“自己的一部分”。尤其是有韵律、有意象、兴发画面感、凝练情绪密度、整体结构成熟的古典诗词。背诵古典诗词，恰恰是为大脑安装一个“内部语言坐标系”；一个“高级语言样本模型”。

我们对死记硬背的反感，是记不住；即使暂时记住，应付完考试以后就忘了，觉得浪费时间，没有用。问题在没有好好地用，而不是被“强迫记忆”。中文的“记忆”充分显示“输入”——“记”，以及“输出”——“忆”两种大脑的工作状态。为了应付考试的记忆，缺乏调用语言样本的场景，无法形塑长期储存的底层逻辑，所以，在背诵时除了知道作者的写作背景，解释文句，还要学着和我们自己的人生连结。

在彷徨无措，油然浮现“山重水复疑无路，柳暗花明又一村”；在情恋失意，自我安慰“天涯何处无芳草”；在徜徉山水，细细品味“万物静观皆自得”——我们不是在“想起/背出一首诗词”，而是诗词从感官的多模态情境中涌现，转化对应，替我们承接真实的时刻。如此，我们建立起稳定而持久的神经通路，为自己保留一块不被算法主导的内在空间。

话说回来，既然背诵诗词那么重要，我们是不是就不必懂得AI呢？

最近我应邀为大专AI诗词视频比赛担任评审，更确定善于操作AIGC技术工具，对于理解和背诵诗词的高效性能。把诗词转译为影像的过程，考验的是掌握文字意涵，不是一句句生硬地用科技“画出来”，拼凑成动图；而是统合、碰撞、叠加，将我们作为主体，融会进诗词，再提取出触动人心的形象，这是人机协作的真谛。

当一首诗在你心里生根，它不会替你解决问题，却会在你迷惘、失落、迟疑的时候，陪你站稳。

在 AIGC 时代，背诵古典诗词，并不是怀旧和倒退。它更像是一种提醒：技术可以替我们生成语言，但只有人，才能让语言成为生命的一部分。

 

2026年1月31日,新加坡《联合早报》“上善若水”专栏

 

AI Can Already Write Poetry—So Why Do We Still Need to Memorize It?

I Lo-fen

With adjustments to the new academic year’s curriculum, classical literature—once divided into three required courses—has been compressed into two. While drafting the new course plan, I found myself already thinking about a pressing question: in the age of AIGC, how should we learn classical literature?

Somewhat “counterintuitively,” I began to revisit older methods, placing renewed emphasis on memory and recitation.

I knew this would inevitably invite skepticism: “AI can already write poetry—so why do we still need to memorize it?”

Indeed. When a poem or a passage of text can be retrieved in full at any time; when AI can instantly generate language—do we still need to commit language to our own brains?

I used to particularly dislike rote memorization. Yet after teaching Text and Image Studies on AIGC for two semesters, I have come to feel quite the opposite: it is precisely in the AIGC era that understanding and memorizing classical poetry reveals its truly irreplaceable value.

AI excels at generating language, but it does not possess language. It can “spit out” a seemingly well-structured poem or a fluent piece of prose in seconds, but its language always remains external. This kind of outsourcing and subcontracting is fast and convenient—when you don’t know what to say, let AI generate it; when you don’t want to organize sentences, let the system rewrite them. You can simply take and use, without reflection.

Over time, the value of language shifts from “I have something to say” to “you help me say it.” Gradually, what deteriorates in our brains is not only expressive ability, but also judgment. A subtle phenomenon emerges: when you do not possess enough good language yourself, even if AI generates ten different versions of a text for you, it becomes difficult to judge which one is truly excellent. The more AI generates, the harder it becomes for us to choose.

Human language only truly becomes “a part of oneself” after it has been understood, memorized, repeatedly savored, and internalized—entering our breathing, rhythm, emotions, and neural responses. This is especially true of classical poetry, with its rhythm, imagery, evocative visuality, condensed emotional density, and mature overall structure. Memorizing classical poetry is, in fact, a way of installing an “internal linguistic coordinate system” for the brain—an advanced model of exemplary language.

Our aversion to rote memorization stems from not being able to remember; or from remembering temporarily, only to forget everything after the exam, feeling that it was a waste of time and useless. The problem lies not in memorization itself, but in how it is used—not in being “forced to remember.” The Chinese word for memory vividly reflects two modes of brain activity: “input” (ji, to record) and “output” (yi, to recall). Memorization aimed solely at exams lacks scenarios for activating linguistic samples; it cannot shape the underlying logic needed for long-term storage. Therefore, when reciting, beyond knowing the author’s background and explaining the lines, we must also learn to connect the text to our own lives.

When we feel lost and helpless, the line “After endless mountains and rivers with no road in sight, suddenly willows shade a village bright” naturally surfaces. When love disappoints us, we console ourselves with “Why worry that no fragrant grass can be found at the world’s end?” When wandering through landscapes, we savor “In quiet contemplation, all things yield their joy.” We are not merely “recalling or reciting a poem”; rather, the poem emerges from a multimodal sensory context, transforms and corresponds, and steps in to carry the weight of real moments for us. In this way, we build stable and lasting neural pathways, preserving within ourselves a space not dominated by algorithms.

That said, if memorizing poetry is so important, does it mean we no longer need to understand AI?

Recently, I was invited to serve as a judge for a collegiate AI poetry video competition. This experience further confirmed for me how skillful use of AIGC tools can enhance the efficiency of understanding and memorizing poetry. The process of translating poetry into images tests one’s grasp of textual meaning—it is not about rigidly “drawing” each line with technology and stitching them into animated visuals. Rather, it involves integration, collision, and layering: merging oneself as the subject into the poem, then extracting images that truly move the heart. This is the essence of human–machine collaboration.

When a poem takes root in your heart, it does not solve problems for you—but when you are confused, lost, or hesitant, it stands with you and helps you stay grounded.

In the age of AIGC, memorizing classical poetry is neither nostalgia nor regression. It is more like a reminder: technology can generate language for us, but only humans can allow language to become part of life itself.

January 31, 2026
“Shangshan Ruoshui” Column, Lianhe Zaobao, Singapore