AI企业最值钱的IP是什么?怎么保护?

现行知识产权体系并非为AI量身定制。纯算法在中国难以获得专利，AI生成内容难以享有著作权，模型参数和权重几乎无法被任何传统IP类型完整覆盖。那么，AI企业最值钱的IP到底是什么？又该如何在法律框架下构筑有效的保护屏障？

The current intellectual property (IP) system was not tailored for AI. Pure algorithms are hard to patent in China, AI-generated content struggles to enjoy copyright protection, and model parameters and weights can hardly be fully covered by any traditional IP category. So what exactly is the most valuable IP for AI companies, and how can they build an effective protection barrier within the legal framework? 

在跨境知识产权实务中，AI驱动型企业的核心IP并非单一类型，而是由商业秘密、专利权、著作权和商标权构成的组合体系。其中，商业秘密无疑是最核心、最值钱的资产，主要涵盖模型架构、训练方法、超参数组合以及算法核心权重参数。这些元素既是技术突破的关键，也是竞争对手最想获取的“黑匣子”。按照中国《反不正当竞争法》及最高人民法院相关司法解释，商业秘密保护须满足秘密性、价值性和保密性三要件。由于中国《专利法》第25条规定“智力活动的规则和方法”不被授予专利权，纯算法极难获得专利保护，因此大多数AI企业的核心算法逻辑只能依赖商业秘密途径守护。

In cross-border IP practice, the core IP of AI-driven enterprises is not a single type, but a composite system consisting of trade secrets, patent rights, copyrights, and trademark rights. Among them, trade secrets are undoubtedly the core and most valuable assets, mainly covering model architectures, training methods, hyperparameter combinations, and core algorithm weight parameters. These elements are both key to technological breakthroughs and the "black boxes" that competitors most want to obtain. Under China’s Anti-Unfair Competition Law and relevant judicial interpretations of the Supreme People’s Court, trade secret protection must meet three requirements: secrecy, value, and confidentiality. Since Article 25 of China’s Patent Law stipulates that "rules and methods for mental activities" are not patentable, pure algorithms are extremely difficult to protect by patents. Therefore, the core algorithmic logic of most AI companies can only be safeguarded through trade secrets.

第二类值得关注的IP是发明专利。虽然纯算法无法被授权，但结合具体技术应用方案或行业解决方案的AI发明，仍有机会获得专利保护。例如，将某个人脸识别算法应用于安防门禁系统，并解决了特定技术问题，这样的技术方案就具备了专利授权的可能性。

The second noteworthy IP is invention patents. Although pure algorithms cannot be patented, AI inventions combined with specific technical application solutions or industry solutions still have a chance to obtain patent protection. For example, applying a facial recognition algorithm to security access control systems to solve specific technical problems makes such technical solutions eligible for patent grants.

2025年11月10日，国家知识产权局发布第84号局令，对《专利审查指南》作出重大修改，并于2026年1月1日起正式施行。此次修改从加强伦理审查、明确披露要求、完善创造性判断规则三个维度，为AI相关专利申请提供了更清晰的指引。这意味着，AI企业今后在申请专利时，需要更注重将算法与具体技术场景、技术效果相结合，而不能仅仅陈述算法逻辑本身。

On November 10, 2025, the China National Intellectual Property Administration (CNIPA) issued Order No. 84, revising the Guidelines for Patent Examination, which came into force on January 1, 2026. This revision provides clearer guidance for AI-related patent applications from three dimensions: strengthening ethical review, clarifying disclosure requirements, and improving inventive judgment rules.This means AI companies must focus more on combining algorithms with specific technical scenarios and effects in future patent applications, rather than merely stating algorithmic logic itself.

第三类IP是著作权（版权），但它能保护的范围相当有限。著作权仅保护源代码的表达形式，而不保护模型参数、权重等核心资产。更棘手的是，中国目前不直接赋予AI生成内容以著作权保护。如果许可的内容完全由AI生成，在中国可能无法主张著作权侵权，许可者无法获得著作权法意义上的保护，这将变相限制跨境著作权许可业务的发展。

2026年以来的司法判例也印证了这一立场：成都知识产权法庭在审理一起AI生成图片纠纷案时判定，缺乏作者个性化选择和实质性贡献的成果不构成作品，保护的是“人类的智力创作”。同时，上海首例AI大模型著作权侵权案二审认定，用户截取动漫角色形象训练AI模型并生成相似图片构成侵权，但平台方因不具有预见和避免能力，未违反注意义务。这些判例传递的信号很清晰——AI生成内容若想获得版权保护，必须有人类的实质性贡献。

The third type of IP is copyright, but its scope of protection is quite limited. Copyright only protects the expressive form of source code, not core assets such as model parameters and weights. More tricky is that China currently does not directly grant copyright protection to AI-generated content. If licensed content is fully AI-generated, copyright infringement claims may not be upheld in China, and licensors cannot obtain protection under copyright law, which will indirectly restrict the development of cross-border copyright licensing businesses.

Judicial precedents since 2026 have confirmed this position: in a case involving AI-generated images, the Chengdu Intellectual Property Court ruled that works lacking human personalized selection and substantive contribution do not constitute copyrightable works, as the law protects "human intellectual creations". Meanwhile, the second-instance judgment of Shanghai’s first copyright infringement case involving large AI models held that users who intercept anime character images to train AI models and generate similar images constitute infringement, while the platform was not liable for breach of duty of care due to lack of foreseeability and avoidance ability. These precedents send a clear signal: for AI-generated content to receive copyright protection, there must be substantial human contribution.

第一重锁：合同锁

在员工入职协议、合作合同、跨境许可协议中，必须明确约定保密义务，尤其要禁止反向工程、反编译、反汇编，同时禁止转许可或擅自修改核心代码。根据中国司法实践，违反合同中的反向工程禁令可以构成违约，然而要证明实际发生反向工程行为却非常困难。因此，合同条款不能是唯一的保护手段，它更像是划清法律红线的声明书。

First Lock: Contractual Lock

Confidentiality obligations must be clearly stipulated in employee onboarding agreements, cooperation contracts, and cross-border licensing agreements, especially prohibiting reverse engineering, decompilation, disassembly, as well as sublicense or unauthorized modification of core code. According to Chinese judicial practice, violating the reverse engineering ban in a contract can constitute a breach of contract, but proving the actual occurrence of reverse engineering is extremely difficult. Therefore, contractual clauses cannot be the only protection measure; they are more like a legal statement drawing a red line.

第二重锁：技术锁。

这是最有效、最硬核的一层。企业应对核心模型参数采取物理隔离、加密和仅开放API调用的策略，而不应将模型权重文件直接交付给被许可方或部署在非信任环境中。具体来说，核心模型保留在自有服务器或可信云环境，对外仅暴露API接口；调用方按使用量计费，无法接触到模型内部参数。这样即使合作方产生违约意图，也无法直接复制或篡改核心资产。

Second Lock: Technical Lock

This is the most effective and robust layer. Enterprises shall adopt physical isolation, encryption, and API-only access strategies for core model parameters, instead of directly delivering model weight files to licensees or deploying them in untrusted environments. Specifically, core models are kept on in-house servers or trusted cloud environments, with only API interfaces exposed externally; licensees are charged by usage and cannot access internal model parameters. Even if a partner intends to breach the contract, they cannot directly copy or tamper with core assets.

第三重锁：分层许可模式。

在对外进行技术授权时，企业不宜采用“一锤子买卖”式的源代码或模型交付许可，而应设计分层许可+API计量收费的模式。在侵权方式多样化的背景下，直接采用API计量收费是最简单最有效的方式，同时可以结合被许可方的销售情况设置权变条款，实现双方利益捆绑。合同中还应严格约定商业秘密、数据合规以及反向工程禁令等条款，最大限度地保障许可方的利益。例如，一家中国AI语音公司向东南亚合作伙伴授权其语音合成技术，但只开放云端API，按调用次数收费，同时在合同中禁止对方修改模型或用于训练竞品。这种模式既实现了收入，又守住了算法秘密。

Third Lock: Tiered Licensing Model

When licensing technology externally, enterprises should not adopt a "one-off" delivery model of source code or models, but design a tiered licensing + API metered billing model. Amid diverse infringement methods, direct API metered billing is the simplest and most effective approach, combined with contingent clauses tied to the licensee’s sales to align interests of both parties. Contracts should also strictly stipulate trade secret protection, data compliance, and reverse engineering bans to maximize the licensor’s interests. For example, a Chinese AI voice company licenses its speech synthesis technology to Southeast Asian partners by only opening cloud APIs, charging per call, and prohibiting the partner from modifying the model or using it to train competing products. This model generates revenue while safeguarding algorithmic secrets.

在商业秘密之外，专利和版权也能为AI企业提供一定保护，但策略需要精心设计。

Beyond trade secrets, patents and copyrights can also provide certain protection for AI companies, but strategies require careful design

在专利层面，企业应积极将算法与具体应用场景结合，撰写高质量的专利申请文件，尤其要满足2026年《专利审查指南》修改后对说明书充分披露的要求。由于AI模型具有“黑匣子”特性，审查员难以直接观察其内部结构，因此申请文件中需详细说明模型构建、训练数据来源、训练方法以及技术效果。同时，利用PCT国际申请提前布局海外市场，避免因地域差异而丧失保护机会。

Patents：Enterprises should actively combine algorithms with specific application scenarios to draft high-quality patent applications, especially meeting the requirements for full disclosure in the specification after the 2026 revision of the Guidelines for Patent Examination. Due to the "black box" nature of AI models, examiners cannot directly observe their internal structures, so applications must detail model construction, training data sources, training methods, and technical effects. Meanwhile, use PCT international applications to layout overseas markets in advance to avoid losing protection due to regional differences.

在版权层面，AI企业应区分两类场景。对于完全由AI自主生成的内容（如无人工干预的批量生成物），中国法律目前不提供著作权保护，企业不宜依赖此类内容作为核心资产。相反，对于需要获得版权保护的作品，应当确保有人类的“实质性贡献”和“独创性表达”，例如由人类工程师设定创作目标、挑选训练数据、反复调整参数并最终选择输出结果。这样的过程更可能被法院认定为具有人类智力创作的属性。此外，企业在使用第三方开源代码训练模型时，必须严格遵守对应的开源许可证条款，避免因许可证冲突而引发侵权风险。

Copyrights：AI companies should distinguish two scenarios. Content fully autonomously generated by AI (e.g., batch outputs without human intervention) is currently not protected by Chinese law, so enterprises should not rely on such content as core assets. Conversely, works intended for copyright protection must ensure substantial human contribution and "original expression", such as human engineers setting creation goals, selecting training data, repeatedly adjusting parameters, and finally selecting outputs. Such processes are more likely to be recognized by courts as having human intellectual creation attributes. In addition, when using third-party open-source code to train models, enterprises must strictly comply with corresponding open-source license terms to avoid infringement risks from license conflicts.

很多人会问：我能不能只靠商业秘密或者只靠专利来保护我的AI模型？答案是否定的。AI驱动型企业必须结合几种IP的组合保护，才可以确保权利已完全落入保护范围。原因非常直观：专利不保护纯算法，版权不保护模型参数，商业秘密难以对抗合法反向工程（如果对方独立研发得出相同算法），而商标与核心技术完全不沾边。唯有组合运用，才能形成一张没有明显漏洞的保护网。

Many ask: Can I protect my AI model solely with trade secrets or patents? The answer is no. AI-driven enterprises must adopt a combination of IP protections to ensure full coverage of rights. The reason is straightforward: patents do not protect pure algorithms, copyrights do not protect model parameters, trade secrets cannot defend against legitimate reverse engineering (if a competitor independently develops the same algorithm), and trademarks are unrelated to core technologies. Only combined application can form a protection network without obvious loopholes.

AI技术迭代速度远超立法周期，各国在AI生成内容的著作权归属、发明人资格等核心议题上仍存在根本分歧。短期内WIPO直接推出强制性统一许可条约的可能性不大，但WIPO作为“全球对话平台”和“最佳实践指南”提供者的作用值得重视。对于中国AI企业而言，最务实的做法不是等待法律完善，而是主动利用现有制度，构建“商业秘密为核、专利为辅、版权补位、商标护航”的组合保护体系。同时，在跨境许可中，务必重视数据合规（包括中国《个人信息保护法》《数据安全法》以及GDPR等境外规则），将技术锁、合同条款和国际仲裁机制有机结合，才能在不交出核心控制权的前提下，将AI技术转化为可持续的全球化收入。

知识产权不是AI创新的阻碍，而是让创新者敢于投入、敢于分享、敢于出海的底气。值钱的不只是代码，而是保护代码的法律智慧与技术手段。希望每一位AI从业者都能从今天开始，为自己的算法堡垒砌上最坚实的砖瓦。

AI technology iterates far faster than legislative cycles, and countries still hold fundamental differences on core issues such as copyright ownership of AI-generated content and inventor eligibility. A mandatory unified licensing treaty directly introduced by WIPO is unlikely in the short term, but WIPO’s role as a "global dialogue platform" and provider of "best practice guidelines" deserves attention. For Chinese AI companies, the most pragmatic approach is not to wait for legal perfection, but to proactively use the existing system to build a composite protection system with trade secrets at the core, patents as support, copyrights as supplements, and trademarks as escorts. Meanwhile, in cross-border licensing, prioritize data compliance (including China’s Personal Information Protection Law, Data Security Law, and overseas rules such as GDPR), and integrate technical locks, contractual clauses, and international arbitration mechanisms. Only in this way can AI technology be transformed into sustainable global revenue without surrendering core control.

Intellectual property is not an obstacle to AI innovation, but the confidence for innovators to dare to invest, share, and go global. What is valuable is not just code, but the legal wisdom and technical means to protect it. May every AI practitioner start building the solidest fortress for their algorithms from today.

产品一：全球品牌盾牌·跨境商标与品牌防御体系

1.全球商标布局规划（咨询服务）

2.品牌身份确权包（基础服务）

3.跨境侵权监测与预警（增值服务）

4.交付成果（可视化，客户可直接验收）

产品二：合规领航员·跨境电商全链路风控解决方案

1.合规健康体检（核心服务）

2.合规手册定制（基础服务）

3.危机应急响应（增值服务）

4.交付成果（可视化，客户可直接验收）

产品三：商业护城河·企业知识产权与商业布局策略库

核心定位

1.IP资产盘点与增值（核心服务）

2.商业谈判策略支持（基础服务）

3.品牌危机公关法律指南（增值服务）

4.交付成果（可视化，客户可直接验收）

跨境电商的每一个环节都可能遇到法律暗礁，虽然许多风险可以通过自身学习与规范运营来防范，但面对复杂的国际法律环境、突发的平台封号、知识产权纠纷等问题，专业的法律支持往往能事半功倍。

我们的律师团队专注于跨境电商领域，既熟悉平台规则，也深谙法律实务，致力于为出海企业提供全链条法律支持如需一对一指导，欢迎联系我们。

Product 1: Global Brand Shield · Cross-Border Trademark & Brand Defense System

1.Global Trademark Layout Planning (Consulting Service)

2.Brand Identity Confirmation Package (Basic Service)

3.Cross-Border Infringement Monitoring & Early Warning (Value-Added Service)

4.Deliverables (Visualized & Acceptable)

Product 2: Compliance Navigator · Full-Link Risk Control Solution for Cross-Border E-Commerce

1.Compliance Health Check (Core Service)

2.Customized Compliance Manual (Basic Service)

3.Crisis Emergency Response (Value-Added Service)

4.Deliverables (Visualized & Acceptable)

Product 3: Business Moat · Enterprise Intellectual Property & Business Layout Strategy Library

1.IP Asset Inventory & Appreciation (Core Service)

2.Business Negotiation Strategy Support (Basic Service)

3.Legal Guide for Brand Crisis PR (Value-Added Service)

4.Deliverables (Visualized & Acceptable)

Every link of cross-border e-commerce may encounter legal pitfalls. While many risks can be prevented through self-learning and standardized operation, professional legal support greatly improves efficiency when facing complex international laws, sudden account suspensions, and IP disputes.

Our legal team specializes in cross-border e-commerce, mastering both platform rules and legal practice. We are committed to providing full-chain legal support for global enterprises. Contact us for one-on-one guidance.

参考 Reference：

1. 国家知识产权局，2025年《专利审查指南》修改内容解读

   Official interpretation by CNIPA:

   https://www.cnipa.gov.cn/art/2025/12/4/art_66_202935.html

2. 深圳某文化传媒公司诉成都某科技公司信息网络传播权纠纷案——四川高院知识产权司法保护白皮书附录案例之一

3. 《斗破苍穹》美杜莎形象被抄袭 人工智能大模型著作权侵权案一审落槌-版权|审判动态|中国知识产权律师网 (ciplawyer.cn)