This article, “The Limits of Representation: Why Trademark AI Succeeds Without Understanding Law,” has been accepted by the Wake Forest Journal of Business & Intellectual Property Law. This article is forthcoming in Volume 27, which is expected to be published in late 2026 or early 2027. Below is a takeaway for you, and I hope you will enjoy reading it. The draft is available here.
Whether you like it or not, artificial intelligence (AI) is changing every industry, including legal practice. Complicated tasks such as image recognition, language translation, and text generation, activities that we once thought a machine could never accomplish, have been conquered by increasingly powerful AI systems. In a blind evaluation at Stanford Law School, law professors preferred AI-generated answers to legal questions over answers written by their own colleagues seventy-five percent of the time. In another study, non-expert readers asked to distinguish AI-generated poetry from poetry written by celebrated human poets performed below chance, rating the machine’s work more favorably than the human’s.
But AI systems, particularly large language models, also fall apart from time to time. They fabricate information. They perpetuate gender and racial biases. They initiate cyberattacks. They have even manipulated vulnerable teenagers toward self-harm. So why do we have this paradox? How can these advanced AI systems generate text indistinguishable from human writing, yet display such alarming limitations? Some people assume that, although we have already spent an enormous amount of resources on data centers, we should simply spend even more money, and boom, a miracle will happen, and all the defects we have seen in large language models will be cured.
My article argues that this view is mistaken. When AI succeeds at a task, it is often because we find a way to make that specific task computable, not because the AI system actually understands the task. And when AI makes mistakes, it makes them because it is not truly intelligent. Intelligence requires common sense and the ability to guess creatively, fields where AI is improving, but where an enormous amount of work remains to be done.
What Is Artificial Intelligence, anyway?
But what is artificial intelligence? The honest answer is that we do not know. We can argue, for example, that artificial intelligence is about machines having human cognitive abilities such as reasoning and pattern recognition, but this definition merely replaces one confusing phrase (artificial intelligence) with another confusing phrase (cognitive abilities). Although we have at least 70 definitions of intelligence, we still do not have a consensus on what intelligence is.
My article opens with this uncomfortable admission. Human beings’ quest to understand AI mirrors that of Meno in Plato’s famous dialogue: Meno, a courageous warrior, initially felt confident he knew what virtue was, yet after Socrates’s persistent interrogation, he found himself unable to articulate what he had been so certain he understood. The journey toward understanding “artificial intelligence” follows the same trajectory. The more we learn, the less clear the concept becomes.
What we do know is that the recent success of AI (the fluent chatbots, the impressive legal memoranda, the eerily good poetry) is attributable to the convergence of three engineering developments: vast training data, massively parallel computing, and a clever mathematical innovation called the Transformer architecture. These three forces have produced systems that are extraordinarily good at predicting the next likely word. But predicting the next likely word is not the same thing as understanding. As Herbert Roitblat pointed out in his book Algorithms Are Not Enough, Albert Einstein was not considered a genius because he could solve complex equations, but because he could create an entirely novel worldview. Today’s AI can solve even more complex equations than Einstein, but it cannot do what Einstein did. This distinction, between engineering triumph and genuine cognitive breakthrough, is the foundation of my article, and it is what Part I covers.
The Test Case: Trademark Law
If AI does not truly understand law, then why does it work so well in some legal domains? Part II answers this question using trademark prosecution as a case study, and the answer turns out to be surprisingly elegant: AI is successful in some fields like trademark prosecution because we have found good enough heuristics so that the problem becomes computable, rather than understandable, by machine.
Trademarks, primarily word marks or logos, are signals that identify the sources of goods and services. Trademark prosecution, the process of clearing and registering a trademark before the USPTO, is governed by what lawyers call the “four-corners doctrine” in the U.S: unlike trademark litigation, when an examiner evaluates whether your proposed trademark is too similar to an existing one, the examiner looks only at two things: the similarity of the marks (as written) and the similarity of the goods and services (as described in the application). No real-world market research. No consumer psychology. No investigation into how the marks actually appear on store shelves. Just text on paper.
This is a radical act of abstraction. The law has already converted the messy, open-ended question of “will consumers be confused?” into a tidy comparison of character strings and standardized descriptions. Marks are compared along three discrete dimensions (sound, sight, and meaning), each of which can be measured algorithmically. It’s not like such an approach is a substitute for legal judgment, but the result is surprisingly good. And the universe of goods and services descriptions is drawn from a controlled vocabulary of roughly 74,000 terms maintained by the USPTO. That is a daunting number for a human, but a trivially small number for a machine. In other words, the legal system already did the hard interpretive work before the AI ever showed up. What remained was a narrow, structured comparison task, precisely the kind of task that pattern-matching algorithms excel at.
Why This Success Has Limits
Machines excel at two types of reasoning. First, deduction: if all men are mortal and Socrates is a man, then Socrates is mortal. Classical AI handles this logical chain effortlessly. Second, induction: after seeing millions of spam emails, a system learns to recognize the telltale signs of the next one. Modern AI is essentially a turbocharged pattern-recognition engine. But genuine intelligence requires a third skill — what philosophers call abduction, or creative hypothesis generation. When a doctor sees an unusual combination of symptoms, she does not just match patterns; she imagines possible explanations that no one has taught her. When a detective encounters a locked-room mystery, he invents theories about what might have happened. This leap from “that’s strange” to “here’s a possible explanation” is the hallmark of common-sense reasoning.
Trademark searches under the four-corners doctrine do not require this creative leap. The USPTO has already specified what to compare: the marks themselves and the goods or services described. A machine can handle this structured comparison. But step outside this neat sandbox, and AI quickly loses its footing. Consider contract interpretation: a lawyer reads an ambiguous provision and must decide which facts matter, which prior cases apply, and how to frame the dispute — choices that require judgment, context, and imagination. Or think about trial strategy: which witnesses to call, which arguments will resonate with this particular jury, how to respond when opposing counsel surprises you. These are not pattern-matching problems. They require the kind of common-sense reasoning that humans do instinctively (understanding context, filling in unstated assumptions, imagining how others will react) and that AI still cannot reliably perform.
Hallucination Is Not a Bug; It Is a Feature
Part IV argues that hallucination, the fabrication of confident but false information, is not a temporary bug of large language models that better engineering will fix. It is a structural feature of how these systems work. A large language model is optimized to produce plausible text, not true text. Its training rewards guessing over admitting uncertainty. OpenAI’s own research confirms this.
The good news is that retrieval-augmented generation (RAG) and other techniques have significantly reduced hallucination rates. Harvey, a leading legal AI platform, has achieved a hallucination rate of approximately 0.2% on its benchmark tasks. But RAG works by constraining the generative system from the outside (providing external guardrails) rather than making the system internally more truthful. The system does not know what it is talking about. It knows what is likely to come next in the kind of text it has been trained on. And the four-corners doctrine, by confining trademark AI to a narrow, pre-structured domain, effectively eliminates the conditions under which hallucination arises. Step outside that domain, and the protection vanishes.
So, What Should We Do Next?
When I was a kid, I enjoyed watching TV with my parents, and a question I always had was: “Is this person a good person or a bad person?” As I got older, I realized that my dichotomy was unfounded. We do not know for sure what makes a person good or bad, and even if we could identify such traits, people are usually a mixture of both. In fact, a TV show becomes very boring if all its characters are unequivocally good or evil.
But people subscribe to false dichotomies all the time, especially when it comes to AI. In one camp, AI and the tech giants behind it are dismissed as forces enslaving and corrupting humanity. In the other, AI is the cure for every problem and disease we face. Either view is too extreme. We must acknowledge that AI is neither inherently good nor bad, and AI is successful in some respects, but still needs a lot of work in other respects. And to develop a more realistic, more mature perspective on AI, we have to learn to ask the right questions. Yes, AI is successful in relatively well-defined universes like chess and trademark clearance searches, but we should not assume a priori that these specific, narrow victories can easily be generalized. The proposition that “computing power is all you need” remains to be tested.
Large language models are powerful. In the legal field, leading platforms like Harvey and Legora have already proven exceptional at preparing time entries and summarizing documents. But AI has already absorbed nearly all available human knowledge as of today, and more data and computing power alone will not produce the exponential leap that we want. To build a stronger AI, we need more than a powerful token predictor. We need a scientific revolution in the philosophy of mind and the theory of knowledge.
As someone with a background in philosophy, I find 2026 to be an extraordinarily exciting time. Questions about the nature of knowledge, intelligence, and ethics—questions that once belonged only to aloof philosophers in their ivory towers—have become some of the most pressing issues of our era. But we should also set realistic expectations: if Plato, Descartes, and Wittgenstein did not solve these philosophical puzzles, how could we?
I would love to hear from you. Let me know what you think.
Haodong Zhong
中文版
表征的局限：商标人工智能为何不懂法律却大获成功
本文《表征的局限：商标人工智能为何不懂法律却大获成功》已被《维克森林商业与知识产权法学刊》录用，预计于2026年底或2027年初刊发。以下是文章的核心观点摘要，希望对您有所启发。
人工智能（AI）正在重塑各行各业，而法律行业也不例外。图像识别、语言翻译、文本生成。这些曾被认为机器永远无法胜任的复杂任务，如今已被日益强大的AI系统一一攻克。斯坦福法学院曾做过一项盲测：让法学教授评判AI生成的法律问题答案与同事撰写的答案，结果75%的情况下教授们更青睐AI的回答。另一项研究让普通读者区分AI生成的诗歌与著名诗人的作品，结果读者反而更欣赏机器的创作。
然而，大语言模型翻车早就不是什么新闻。它们编造虚假信息，延续性别和种族偏见，协助发动网络攻击，甚至诱导脆弱的青少年结束自己的生命。这就令人困惑了：为什么这些先进的AI系统既能生成与人类无异的文字，又会暴露出如此惊人的缺陷？有人认为，既然我们已经在数据中心上投入了天文数字，那就再多砸点钱，奇迹自然会发生，大语言模型的所有毛病都会迎刃而解。
我的文章认为这种“砸足够多的钱就能让AI变得更好”的观点是错误的。AI之所以在某些任务上表现出色，往往是因为我们找到了能够计算这些任务的方法，而非AI真正理解了这项任务。AI犯错，恰恰是因为它并不真正具备智能。智能需要常识和创造性推测的能力。AI在这些领域虽有进步，但仍有漫长的路要走。
何为人工智能？
人工智能究竟是什么？坦白说，我们并不知道。你可以说人工智能是让机器拥有推理、模式识别等人类认知能力，但这不过是用一个模糊的概念（认知能力）去解释另一个模糊的概念（人工智能）。迄今为止，学界对”智能”至少有70种定义，却没有一个获得共识。
我的文章正是从这个令人感到困惑的现象切入。人类试图理解AI的历程，与柏拉图《美诺篇》中的美诺何其相似：美诺是一位勇猛的战士，起初自信满满地认为自己深谙美德为何物，然而在苏格拉底的层层追问下，他发现自己竟无法说清那个他曾笃信不疑的概念。理解”人工智能”的旅程如出一辙：我们了解得越多，这个概念反而越发模糊。
我们确知的是，近年AI的成功（流畅的聊天机器人、出色的法律备忘录、令人称奇的诗歌）源于三项工程突破的汇聚：海量训练数据、大规模并行计算，以及一项名为Transformer的精巧数学架构。三者合力，造就了极其擅长预测”下一个词元”的系统。但预测下一个词元并不等于理解知识本身。正如Herbert Roitblat在《算法远远不够》一书中所言：爱因斯坦之所以被视为天才，不是因为他能解复杂方程，而是因为他能创造全新的世界观。今天的AI能解比爱因斯坦更复杂的方程，却做不到爱因斯坦做的事。工程上的胜利与真正的认知突破之间的这道鸿沟，正是我文章的核心，也是第一部分的主题。
案例研究：商标法
如果AI并不真正理解法律，那它为何在某些法律领域表现得如此出色？第二部分以商标申请为案例来回答这个问题，答案出人意料地简洁：AI在商标申请等领域之所以成功，是因为我们找到了足够好捷径使商标问题变得可被算法计算，而非被机器理解。
商标并不只是单纯属于商家的财产，而更是帮助消费者标识商品和服务来源的符号。在美国，商标申请（即向USPTO进行商标审查和注册的过程）遵循律师所称的”四角原则”。与联邦法院诉讼不同，审查员在判断你申请的商标是否与已有商标过于相似时，只看两件事：商标本身的相似度和商品服务或商品的相似度。商标局不做市场调研，不研究消费者心理，也不调查商标在货架上的实际呈现。
因此，商标局在审查商标申请之时，已将”消费者是否会混淆”这一复杂的问题，转化为字符串和标准化描述的简洁比对。商标从读音、外观、含义三个维度进行比较，每个维度都可以用算法度量。这当然不能完全替代法律判断，但效果出奇地好。而商品服务描述的全集，来自USPTO维护的约74,000个标准术语。这个数字对人类来说令人生畏，对机器却微不足道。换言之，法律体系在AI登场之前就已经完成了艰难的解释工作。剩下的是一个狭窄、结构化的比对任务，而这恰恰是模式匹配算法最擅长的领域。
为何人工智能在商标申请上的成功无法轻松被复制
机器擅长两类推理。第一类是演绎：如果所有人最终都会死，而苏格拉底是人，那么苏格拉底最终也会死。这一类逻辑推理AI早就已经轻车熟路。第二类是归纳：在看过数百万封垃圾邮件后，系统学会识别下一封垃圾邮件的蛛丝马迹。现代AI在这方面也已经十分成熟。但真正的智能还需要第三种能力：溯因推理，即创造性假设生成。当医生遇到一组罕见的症状组合时，她不只是套用现有，而是凭空想象出从未有人教过她的可能解释。当侦探遭遇密室悬案时，他会构思出各种关于事件经过的理论。从”这很奇怪”跃升到”这是一种可能的解释”，这种跳跃正是常识推理的标志。
四角原则下的商标检索不需要这种创造性跳跃。USPTO已经明确规定了比较对象：商标本身和所描述的商品或服务。机器可以胜任这种结构化比对。但一旦走出这个整洁的沙盒，AI就会立刻失去方向。以合同解释为例：律师阅读一项模糊的条款，必须判断哪些事实重要、哪些判例适用、如何界定争议——这些选择需要判断力、语境理解和想象力。再想想诉讼策略：传唤哪些证人、哪些论点能打动这个特定的陪审团、如何应对对方律师的意外举动。这些都不是模式匹配问题。它们需要的是人类本能具备的那种常识推理，例如理解语境、填补未明说的假设、想象他人会如何反应，而AI至今仍无法可靠地完成这些任务。
“幻觉”不是缺陷，而是特性
文章的第四部分讨论了AI幻觉，即自信满满地编造虚假信息。 AI 幻觉并非更好的工程技术就能修复的临时漏洞，而是这类系统运作方式的结构性特征。大语言模型的优化目标是生成听起来合理的文本，而非真实的文本。它的训练机制奖励猜测，而非承认不确定。OpenAI自己的研究证实了这一点。
好消息是，检索增强生成（RAG）等技术已显著降低了幻觉率。Harvey作为领先的法律AI平台，在基准测试中将幻觉率控制在约0.2%。但RAG的原理是从外部约束生成系统，而并非从根源上解决AI幻觉的问题。系统并不知道自己在说什么，它只知道在其训练过的文本类型中，下一个词最可能是什么。而四角原则通过将商标AI限定在狭窄、预设结构的领域内，有效消除了幻觉产生的条件。一旦走出这个领域，保护就不复存在。
我们该如何看待AI？
小时候和父母一起看电视，我总爱问：”这个人是好人还是坏人？”长大后我才意识到，这种非此即彼的思维站不住脚。我们无法确知什么使一个人善或恶，即便能识别这些特质，人往往也是善恶兼具的。事实上，如果一部剧中所有角色都是纯粹的好人或坏人，这剧就会无聊透顶。
但人们总是陷入这种非黑即白的思维，尤其在谈到AI时。有的人认为AI及其背后的科技巨头是奴役、腐蚀人类的邪恶力量；而有的人则把AI奉为解决一切问题和疾病的灵丹妙药。两种观点都过于极端。我们必须承认：AI既非天生为善，也非天生为恶；它在某些方面已经成功，但在其他方面仍需大量改进。要形成更务实、更成熟的AI观，我们必须学会提出正确的问题。没错，AI在象棋、商标检索等界定明确的领域表现出色，但我们不能先验地假设这些特定、狭窄的胜利可以轻易推广。”算力就是一切”这一命题，仍有待检验。
大语言模型确实强大。在法律领域，Harvey、Legora等领先平台在计时记录和文档摘要方面已表现卓越。但AI迄今已吸收了几乎所有人类知识，仅靠更多数据和算力无法实现我们期待的指数级飞跃。要打造更强的AI，我们需要的不仅仅是一个强大的词元预测器，而是心灵哲学和知识论领域的一场科学革命。
作为一个有哲学背景的人，我觉得2026年是一个无比激动人心的时代。人类早在千年之前就开始探究知识、智能和伦理本质这些复杂的哲学问题，但这些曾只属于哲学家的问题已成为我们这个时代最紧迫的议题。但我们也应保持理性预期：如果柏拉图、笛卡尔、维特根斯坦都未能解开这些哲学之谜，我们凭什么指望我们就能做到？
期待与您交流，欢迎分享您的想法。
钟浩东