EPO报告：人工智能的专利保护策略
每日IP英文第84天：
2018年5月30日，EPO召开了一个人工智能（AI）专利保护研讨会，就人工智能对专利制度的影响展开讨论。 它吸引了350多位工业界，学术界，用户协会，专利律师事务所，司法机构，国家专利局和政府机构的代表。
会议主题主要包括：
保护AI创新的策略
撰写和申请人工智能专利申请的挑战
对未来专利法的一些考虑
目前，我国很多企业的AI技术已经是世界领先水平，迫切需要高水平的专利保护。而这次会议无疑提供了最为领先的经验和观点，因而，我摘录了会议报告的核心部分，并给出了相应的Google翻译，供诸位参考。
点击文末“阅读原文”，可以获取报告PDF原文。
Google翻译：
II 保护人工智能发明的一般策略 - 专利是最佳选择
改变游戏规则，促成更多合作？
人们普遍认为，人工智能是一个改变游戏规则的人，当谈到相关知识产权如何受到保护时，我们正在进入新的领域。由于我们不再在孤岛中发明，而是在合作中，知识产权从业者不得不采取新的方法。同样，专利局需要技术上混合的审查部门（欧洲专利局已经可以）。
正如开幕主题演讲中所提到的，AI为更多的协作创新和更有思想的专利保护策略提供了机会。此外，该技术显然是全球性的，知识产权从业者必须了解不同国家的不同专利制度。
关于创新的所有权，人工智能有助于发明或甚至“发明”的前景，可以想象从发明者所有权到发明所来自的公司和组织的投资保护系统的转变。
与AI相关的创新的IP保护类型
从创新的角度来看，为了社会的利益，创新者应该尽可能多地激励人工智能创新 - 例如算法和他们的培训方式 - 而不是选择商业秘密。
源代码受版权保护，并未与专利相关联。因此，版权和专利保护的混合被认为是一种可能性。这将为生成代码提供协作基础（“低端算法堆栈”）;专利申请应考虑“高端堆栈”。开放代码也有一些变化。
另一个想法是，如果要保护“基本”算法，那么应该像德国专利法第24（2）条那样强制交叉许可，以保证对算法的访问，然后将其用于改进专利。对于基本思想，算法的创建者应该有可能具有一种“达成”声明，即使它没有在特定应用程序中“测试”。
专利申请统计数据显示，在与AI相关的技术领域，申请对算法产生了技术影响，这符合EPO专利数学方法申诉委员会判例法中规定的要求。
专利是最佳选择，但有些变化建议
专利被视为鼓励人工智能领域创新的最佳知识产权方式，因为它们提供了额外的保护措施，并且是额外的贸易资产。然而，鉴于AI技术的发展速度，专利的20年寿命是值得怀疑的。 （注意：如果专利所有人未支付续展费用，专利保护失效，即他们不必在整个20年内保持专利保护，而在大多数技术领域，通常不会这样做。）同样， 18个月的保密期，直到申请的公布可以缩短，因为它可能经常导致不必要的平行研究和开发。
III EPO如何应对专利申请中的AI挑战
双障碍方法和对技术特征的贡献
人工智能不仅仅是机器学习（ML）和涵盖高等数学，它提出了数学模型的专利保护资格问题。在评估可专利性时，欧洲专利局对“混合型发明”应用了双重障碍的方法，并提出问题“AI和ML方法（步骤）是否有助于实现本发明的技术特性？”。
还考虑了数据的性质，区分了功能和认知数据。
如何显示技术效果？
在回答听众提出的关于包含测试以验证技术效果的问题时，建议尽可能多地在技术规范中包含有关技术效果的信息。
申请人/律师对话表明，有时需要经常与发明人互动以确定技术特征和相应的技术效果。
清楚
由于权利要求中的流行语和营销术语，审查员可能会提出缺乏明确性的异议。他们意识到仔细考虑保护范围的重要性。
专利律师建议定义人工神经网络功能，如输入/输出数据和数据网络架构。
欧洲专利公约（EPC）下的本领域技术人员
本领域技术人员的概念出现在EPC第56和83条中（分别是创造性和公开的充分性）。根据欧洲专利局的审查指南，技术人员也可能是一个团队。技术人员知道在应用领域中使用的概念和术语，并且具有用于日常工作和实验的手段，如果它们的使用在所涉及的领域中是常见的，则可以包括AI工具。
参考资料：审查指南和判例法
2018年的审查指南更新（2018年10月开始生效，2018年11月生效）将以人工授精为例，并根据欧洲专利局上诉委员会的决定提供有关CII技术性的详细信息。
专利律师强调，最新的判例法知识至关重要，正如“检查指南”（F-IV，3.9）中所述的与不同类型的索赔格式有关的考虑因素也是如此。
IV 起草相关发明专利申请的当前和未来挑战 - 提出了一些大胆的建议
确定了三种类型的AI专利建议
为了讨论的目的，小组成员确定了三种可能的AI专利类型：
- “核心AI”，其挑战在于它通常与算法本身有关，因为数学方法不具有可专利性;
- 训练有素的模型/机器学习，声称变化和范围可能是一个问题;
- AI作为应用领域的工具，通过技术效果定义。
欧洲和美国的专利性要求
在EPC下，作为一种数学方法的AI在被声明时被排除在可专利性之外。但是，如果索赔涉及涉及技术手段（例如计算机）或设备的方法，则其主题事项具有整体的技术特征，因此不排除可专利性。在美国，资格提出了挑战，因为抽象的想法不具有可专利性。然而，目前关于如何定义“抽象”的指导很少。此外，仅仅使用计算机来实现抽象概念并不足以通过美国的资格障碍。因此，对于美国的AI应用而言，资格比欧洲更为严重。
专利AI算法就是这样
该小组对允许AI算法专利的提案存在分歧。目前它们不具有可专利性，但算法代表了AI的核心。
一种解决方案（由于其面临的挑战，通常受到美国对话者的青睐）将是立法使AI算法具有可专利性。
请求保护训练模型/机器学习
可能需要比较示例和参数范围，并且来自诸如工业化学的其他领域的创造性实践可能是相关的。有人建议，EPO可能对特定数据集所赋予的技术性更加宽容，并允许“模型的第二次使用”类比于药剂学中的第二次医疗使用声明。如果通过不同方式达成，则不应将其视为等效。
AI在应用领域，通过技术效果定义
在诸如自动驾驶车辆和医疗保健之类的领域中，AI可能被声称为嵌入在较大声明中的工具，其被定义为提供分辨率的功能特征。在这种情况下，应该评估整个索赔的创造性，而不仅仅是功能定义的AI工具。
AI初创企业
人工智能创新越来越多地来自资源有限的初创企业。需要更快的搜索和专利以及更“可访问”的专利制度，以便公司能够将其IP货币化。
V.人工智能相关专利的授权后方面 - 仍然很少有判例法，但存在一些指导
日本
对CII / AI发明的索赔必须以某种格式起草，以满足可专利性标准。在执法方面，找到并提供足够的证据是一项挑战，因为侵权产品或方法的运作方式并不一定可见。为此，法院可以发布“文件生产订单”以揭示基础系统。已经有一些确定的案例。
美国
关于美国专利法第101节实施的不确定性，以及专利保护资格的标准对美国所有申请人来说都是一个主要问题。美国专利和商标局的新专利局局长宣布，他将专注于寻找可持续的解决方案。第二个问题是专利审判和上诉委员会发布的专利有效性决定。然而，据说在过去的三个月里，它的决定变得更加专业化。
欧洲
算法在AI中发挥了重要作用。由于它们类似于数学方法，因此被排除在专利性之外，其中声称“如此”。但是，如果索赔涉及涉及技术手段（例如计算机）或设备的方法，则其主题事项具有整体的技术特征，因此不排除可专利性 - 即通过了资格障碍。在执法方面，侵权证据是一项挑战。由于通常无法确定AI的工作原理，因此很难证明侵权产品使用相同的方法。化学中的情况有很多相似之处。
结论
证明侵权是另一个难题，因为有时很难提供证据。在执法方面，证据的可用性和适当快速反应的能力是非常重要的，法律制度的可预测性也是如此。观众提出了专利池并扭转了举证责任。
原文如下：
II. General strategies for protecting AI inventions – patents the best option
A game changer leading to more collaboration?
There was general agreement that AI was a game changer and that we were moving into new territory when it came to how the associated IP would be protected. As we were not inventing in silos any more, but collaboratively, IP practitioners had to take a new approach. Similarly, patent offices would need technically mixed examining divisions (which is already possible at the EPO).
As had been mentioned in the opening keynotes, AI provided an opportunity for more collaborative innovation and more thoughtful patent protection strategies. Moreover, the technology was clearly global, and IP practitioners had to be aware of the different patent systems in different countries.
Regarding ownership of the innovations, with the prospect of AI contributing to inventions or even “inventing”, a shift was conceivable from inventor based ownership to investment protection systems for the companies and organisations from which the inventions came.
Types of IP protection for AI-related innovations
From the perspective of innovation for the benefit of society, there should be as much incentive as possible for innovators to disclose AI innovations – such as the algorithms and how they were trained – and not to choose the option of trade secrets.
Source code was protected by copyright and not linked as such to patents. A mixture of copyright and patent protection was thus seen as a possibility. That would provide a collaborative basis for producing code (“lowend stack of algorithm”); patent applications should be considered for the “highend stack”. There was also some movement towards open code.
Another idea was that, if “fundamental” algorithms were to be protected, then there should be compulsory cross-licensing as in Section 24(2) of the German Patent Act to guarantee access to the algorithms, which would then be used in improvement patents. It should be possible for the creator of an algorithm to have a type of “reachthrough” claim for the basic idea, even if it has not been “tested” in a specific application.
The patent filing statistics showed that, in AI-related technical areas, applications claimed a technical effect for algorithms, which was in line with the requirements laid down in the case law of the EPO boards of appeal for patenting mathematical methods.
Patents the best option, but some changes suggested
Patents were seen as the best IP way to encourage innovation in the area of AI, as they provided an extra measure of protection and were additional trade assets. However, patents’ 20-year lifetime was questionable in view of the speed of evolution of AI technology. (Note: if a patent proprietor fails to pay a renewal fee, patent protection lapses, i.e. they do not have to maintain the patent protection for the entire 20 years, and in most technical fields it is usual not do so.) Similarly, the 18 month secrecy period until publication of the application could be shortened, as it might often lead to unnecessary parallel research and development.
III. How the EPO deals with the challenges of AI in patent applications
The two-hurdle approach and the contribution to the technical character
AI was more than just machine learning (ML) and covered advanced mathematics, which raised the question of mathematical models’ eligibility for patent protection. When assessing patentability, the EPO applied the two-hurdle approach for “mixed-type inventions” and asked the question “Do(es) the AI and ML method (steps) contribute to the technical character of the invention?”.
The nature of the data was also considered, with a distinction made between functional and cognitive data.
How to show a technical effect?
In reply to a question from the audience regarding the inclusion of tests to verify a technical effect, it was recommended to include in the specification as much information about the technical effect as possible.
The applicant/attorney dialogue indicated that frequent interaction with the inventor to identify the technical features and the corresponding technical effect was sometimes necessary.
Clarity
Examiners might raise objections of lack of clarity because of buzzwords and marketing terms in the claims. They were aware of the importance of carefully considering the scope of protection.
The patent attorneys recommended defining artificial neural network features such as input/output data and data network architecture.
The skilled person under the European Patent Convention (EPC)
The concept of the skilled person appears in Articles 56 and 83 EPC (inventive step and sufficiency of disclosure respectively). According to the Guidelines for Examination in the EPO, the skilled person might also be a team. The skilled person was aware of concepts and terminology used in the field of the application and had the means for routine work and experimentation, which could include AI tools if their use was common in the field in question.
Reference material: Guidelines for Examination and the case law
The 2018 update of the Guidelines for Examination (available as of October 2018 and entering into force in November 2018) would feature examples relating to AI as well as detailed information on the technicality of CII based on decisions by the EPO’s boards of appeal.
The patent attorneys emphasised that up-to-date knowledge of the case law was crucial, as were considerations relating to the different types of claim format available, as set out in the Guidelines for Examination (F-IV, 3.9).
IV. Current and future challenges in drafting patent applications for AIrelated inventions – some bold suggestions put forward
Identifying three types of AI patenting suggested
For the purposes of the discussion, the panellists identified three possible types of AI patenting:
– “Core AI”, where the challenge was that it often related to algorithms as such, which as mathematical methods were not patentable;
– Trained models/machine learning, where claiming variations and ranges might be an issue;
– AI as a tool in an applied field, defined via technical effects.
Patentability requirements in Europe and the US
Under the EPC, AI as a mathematical method was excluded from patentability when claimed as such. However, if a claim was directed to a method involving technical means (e.g. a computer) or to a device, its subject matter had a technical character as a whole and for that reason was not excluded from patentability. In the US eligibility presented a challenge because abstract ideas were not patentable. However there was currently little guidance on how to define “abstract”. Furthermore, the mere use of a computer to implement an abstract idea was not sufficient to pass the eligibility hurdle in the US. Eligibility was therefore more of an issue for AI applications in the US than in Europe.
Patenting AI algorithms as such
The panel was divided regarding the proposal to allow patents for AI algorithms as such. Currently they were not patentable, yet algorithms represented the core of AI.
One solution (generally favoured by the US interlocutors due to the challenges faced there) would be to legislate to make AI algorithms as such patentable.
Claiming trained models/machine learning
Comparative examples and parameter ranges might be needed and inventive step practices from other areas such as industrial chemistry might be relevant. It was suggested that the EPO could be more lenient regarding the technicality conferred by specific datasets and allow the “second use of a model” by analogy to second medical use claims in pharmaceutics. Uses should not be considered equivalent if arrived at by different means.
AI in an applied field, defined via technical effects
In fields such as autonomous vehicles and healthcare AI might be claimed as a tool embedded in a larger claim, defined as a functional feature providing a pasolution. In such cases, the claim as a whole should be assessed for inventive step, and not just the functionally defined AI tool.
AI start-ups
AI innovation increasingly came from start-ups with limited resources. There was a need for much faster searching and patenting and a more “accessible” patenting system so that companies could monetise their IP.
V. Post-grant aspects of AI-related patents – still little case law but some guidance exists
Japan
Claims for CII/AI inventions had to be drafted in a certain format to meet the patentability criteria. When it came to enforcement, it was a challenge to find and provide sufficient evidence, since it was not necessarily visible how the infringing product or method worked. For that purpose courts could issue “document production orders” to reveal the underlying systems. There were already some decided cases.
US
Uncertainty about the implementation of Section 101 of the US Patent Act, and thus the criteria for eligibility for patent protection posed a major problem for all applicants in the US. The new Commissioner for Patents at the US Patent and Trademark Office had announced that he would focus on finding a sustainable solution. The second issue was the patent validity decisions issued by the Patent Trial and Appeal Board. However, it was said that its decisions had become more patentee-friendly in the past three months.
Europe
Algorithms played a major role in AI. Since they were akin to mathematical methods, they were excluded from patentability where claimed “as such”. However, if a claim was directed to a method involving technical means (e.g. a computer) or to a device, its subject matter had a technical character as a whole and for that reason was not excluded from patentability – i.e. the eligibility hurdle was passed. When it came to enforcement, evidence of infringement was a challenge. Since it was often not possible to see how exactly the AI worked, it was very difficult to show that the infringing product used the same method. There were a lot of similarities to situations in chemistry.
Conclusion
Proving infringement was an additional difficulty because sometimes it was difficult to provide evidence. The availability of evidence and the ability to react suitably quickly were important when it came to enforcement, as was the predictability of the legal system. The audience suggested patent pools and reversing the burden of proof.
Each article is copyrighted to their original authors. The news is for informational purposes only and does not provide legal advice.