{"id":150632,"date":"2026-11-02T08:00:56","date_gmt":"2026-11-02T02:30:56","guid":{"rendered":"https:\/\/www.bananaip.com\/intellepedia\/?p=150632"},"modified":"2026-11-01T10:40:56","modified_gmt":"2026-11-01T05:10:56","slug":"ai-inventions-patents-sufficiency-of-disclosure","status":"publish","type":"post","link":"https:\/\/www.bananaip.com\/intellepedia\/ai-inventions-patents-sufficiency-of-disclosure\/","title":{"rendered":"AI Inventions and Patents: When Is the Disclosure Sufficient?"},"content":{"rendered":"

Note: <\/strong>This article is based on, and expands upon, parts of the author’s presentation at the Expert Exchange on AI, Patentability Criteria and Inventorship, organised as part of the thirty-eighth session of the Standing Committee on the Law of Patents (SCP) of the World Intellectual Property Organization (WIPO) in Geneva in November 2026.<\/p>\n

Introduction<\/h2>\n

Clear and complete disclosure of the invention for which a patent is desired is at the core of the quid pro quo<\/em> (give and take) that forms the basis of the patent system (WIPO, 2015). As a part of this give and take, the inventor discloses complete information about the invention to the public, and in return, takes exclusive rights over it in the form of a patent for a limited period of time (Masur and Ouellette, 2025). It is believed that this exchange between the inventor and the general public, implemented through the patent system, provides incentives to invent, invest, and design around while enriching the public domain and facilitating scientific, industrial, and economic progress (Aboy et al., 2024; Masur and Ouellette, 2025). Considering the importance of disclosure to the patent system, it is reasonable to expect that any problems or deficiencies relating to the law and rules governing disclosure of inventions will have a detrimental effect on the incentives and progress the patent system seeks to achieve.<\/p>\n

Over the years, whenever a new field, technology, or type of invention evolved, new laws, rules, and norms relating to disclosure have been created, or existing ones modified or fine-tuned, to ensure, enable, or facilitate clear and complete disclosure under the patent system. For example, as biological inventions evolved, a deposit framework was made available for disclosure of biological materials such as microorganisms (Budapest Treaty<\/em>, 1977, Article 3), and standards were created to facilitate disclosure of nucleotide and protein sequences (WIPO, 2024b). Similarly, in the field of computer software, specific rules were introduced to facilitate the submission of computer program listings (Code of Federal Regulations 2024, title 37, \u00a7 1.96), and in the field of chemistry and pharmaceuticals, standardised methods of representing chemical structures and formulae evolved to facilitate clear and precise disclosure of inventions (EPO, 2026a, Part A, Chapter IX, section 11.1). Today, the unique characteristics of AI inventions, which include the nature of AI models, their dependence on training data, and difficulties in reproducing their results, have raised questions about whether existing disclosure requirements are suitable for inventions in the field, and what changes in law, rules, or examination guidance may be appropriate (Upputuri and \u00d6zkan \u00dcner, 2023; Aboy et al., 2024; WIPO, 2026).<\/p>\n

In this context, this article seeks to provide an understanding of the problems relating to disclosure of AI inventions in patent applications, and the options available to address them while preserving the incentives and public benefits that the patent system seeks to achieve. The article starts with an introduction to AI and its constituent parts, and then explains the requirements relating to sufficiency of disclosure. Following that, it considers the disclosure challenges posed by AI, the guidance issued by different patent offices, and how that guidance might apply to an illustrative invention. It then discusses the policy options set out in the report of the World Intellectual Property Organization (WIPO) on patentability criteria and AI inventorship, before concluding with a recommendation for a progressive approach, under which disclosure requirements are heightened initially and relaxed as the technology and knowledge in the field develop, and with a view on why AI-generated inventions call for regulatory disclosure rather than a sui generis<\/em> system.<\/p>\n

Understanding AI Inventions: Model, Training Data, and Implementation<\/h2>\n

Before looking at the disclosure issues relating to AI inventions, it is useful to understand what AI is and what goes into making an AI system work, particularly because the term is used for a range of technologies and does not have one settled meaning (Widder, Whittaker and West, 2024). Broadly, an AI system receives inputs and processes them to produce outputs, which may take the form of predictions, recommendations, decisions, or generated content (Ramos and Pianese, 2026). In a machine learning system, the system learns patterns from data and uses what it has learnt to process new inputs, and in a generative AI system, such learning is used to produce new content such as text, images, or audio (Widder, Whittaker and West, 2024).<\/p>\n

The development of a large generative AI model generally involves a series of steps (Lin\u00e5ker et al., 2026). It starts with the collection, cleaning, and preparation of data for training. A base model is then trained using that data, a chosen model design, training code, development tools, and computing resources, in a process commonly referred to as pre-training. The model may then be refined through further training on selected examples, human feedback, and safety measures, before it is tested, documented, and released for use. It may also be adapted for a particular language, field, or purpose, or modified to run with fewer computing resources, depending on how it is intended to be used (Lin\u00e5ker et al., 2026).<\/p>\n

As can be seen, an AI system is made up of several parts that contribute to its development and operation. These parts may broadly be understood in three groups, namely, the model, training data, and documentation and implementation (White et al., 2024).<\/p>\n

(a) The Model:<\/strong> The first group includes the architecture and the weights of the model. Architecture refers to the design or structure of the model, and weights refer to the numerical values learnt during training, which influence how inputs are processed and outputs are produced (Ramos and Pianese, 2026; White et al., 2024). To understand a model, one therefore has to understand both its structure and the learnt values through which that structure operates.<\/p>\n

(b) Training Data:<\/strong> The second group includes the data used for training and testing, the code used to prepare that data and train the model, and the methods and resources used in the training process (Widder, Whittaker and West, 2024). It also includes the contribution of people who select or label data, evaluate outputs, and otherwise take part in developing the model, because what a model learns depends on these choices and processes as much as on its architecture (Widder, Whittaker and West, 2024).<\/p>\n

(c) Documentation and Implementation:<\/strong> The third group includes the materials surrounding the model, such as benchmarks, model cards, technical reports, and other documentation, together with the product or interface through which the model is made available for use (Lin\u00e5ker et al., 2026; Ramos and Pianese, 2026). These materials help others understand how the model was developed, tested, and implemented, and the conditions under which it is meant to operate (Lin\u00e5ker et al., 2026).<\/p>\n

Depending on the nature of the AI invention, an applicant for a patent may need to disclose all or some of the information relating to the model, the training process and data, or the details of implementation in a particular system or application (Aboy et al., 2024). The question, then, is which of these parts must be disclosed, and in how much detail, for a person skilled in the art to carry out the invention for which patent protection is sought.<\/p>\n

Patent Disclosure Requirements<\/h2>\n

With respect to disclosure in a patent specification, the Agreement on Trade-Related Aspects of Intellectual Property Rights (TRIPS Agreement<\/em>) provides only a broad requirement. It asks member states to require patent applicants to disclose the invention in a manner sufficiently clear and complete for it to be carried out by a person skilled in the art, and makes the requirement of best mode optional (TRIPS Agreement<\/em>, 1994, Article 29.1). With some variations, such as those relating to how claims have to be supported, whether best mode is required or not, and whether possession of the invention has to be proved, major jurisdictions follow the basic standards required under the TRIPS Agreement<\/em> (WIPO, 2015; Xiao and Zhang, 2026; WIPO, 2026).<\/p>\n

To acquire a patent, the patent applicant must disclose details of the invention by filing a patent specification. The specification has to disclose details of the invention in the form of a written description that is sufficient to enable a person skilled in the art to carry out the invention without undue experimentation (TRIPS Agreement<\/em>, 1994, Article 29.1; WIPO, 2015; Masur and Ouellette, 2025). Depending on the nature of the invention, the written description may include drawings, flow charts, computer program listings, chemical and mathematical formulae, sequence listings, deposits of biological material, and working examples (Patent Cooperation Treaty<\/em>, 1970, Article 7; PCT Regulations<\/em>, 1970, Rule 5.1(a)(v); Budapest Treaty<\/em>, 1977, Article 3; WIPO, 2024b; Upputuri and \u00d6zkan \u00dcner, 2023). The patent specification must also include claims that clearly and concisely define the boundaries of the invention, and these claims have to be fully supported by the content in the specification (WIPO, 2015; Ouellette, Fang and Ouellette, 2025). Some countries may also require patent applicants to disclose the best mode of carrying out the invention known to them (TRIPS Agreement<\/em>, 1994, Article 29.1; Patents Act 1970 (India), s. 10(4)(b); United States Code 2018, title 35, \u00a7 112(a)). Additionally, some countries may also require the applicant to show from the written description that the inventor was in possession of the claimed invention on the date of filing (United States Code 2018, title 35, \u00a7 112(a); WIPO, 2015).<\/p>\n

The sufficiency of disclosure in a patent specification is assessed from the perspective of a person skilled in the art, that is, a person skilled in the field to which the invention relates (TRIPS Agreement<\/em>, 1994, Article 29.1). This is a hypothetical person, who is assumed not only to possess average knowledge and ability in that field, but also to be aware of the common general knowledge in the field as on the relevant date, which is usually the date of filing or priority (WIPO, 2015; Fabris, 2020). A specification will be considered enabling only if this person is able to carry out the invention based on the disclosure in the specification and the common general knowledge relating to the invention (Aboy et al., 2024). Therefore, who the person skilled in the art is, and what constitutes the corpus of common general knowledge, play an important role in determining whether a given disclosure is sufficient or not.<\/p>\n

Although major jurisdictions follow the same basic standard, there may be variability among countries regarding the scope, nature, and extent of information required to satisfy the disclosure requirement (Xiao and Zhang, 2026; WIPO, 2026). Even where the legal standards are broadly the same, countries and patent offices may apply them differently (WIPO, 2015; Fabris, 2020). Consequently, the same specification may be found sufficient in one country and insufficient in another, even though both countries follow the same basic standard under the TRIPS Agreement<\/em>.<\/p>\n

Disclosure Challenges and Requirements of AI Inventions<\/h2>\n

The disclosure challenges relating to AI inventions can be understood through the three groups discussed earlier, namely, the model, training data, and documentation and implementation. For each group, the question is what information the skilled person needs to carry out the claimed invention, and what makes it difficult to provide that information in a patent application (Aboy et al., 2024; WIPO, 2026). The answer depends on where the contribution lies, because an invention relating to a new model architecture may call for different information from an invention that uses an existing model in a particular application (WIPO, 2026).<\/p>\n

(a) Challenges Relating to Model Disclosure<\/h3>\n

There are several patent disclosure challenges with respect to inventions claiming AI models. These primarily stem from the fact that AI models are non-deterministic and probabilistic, which makes their functioning unpredictable and uncertain (Xiao and Zhang, 2026; WIPO, 2026). Additionally, the specific functioning of AI models is not humanly explainable in most cases, and it is difficult, if not impossible, to disclose the relationship between inputs and outputs unambiguously (Nakayama, 2021; Robinson, 2026). As a result, the extent of disclosure required to enable a person skilled in the art to carry out, or reproduce, the model is much higher than what might be required for deterministic and rule-driven AI or software inventions (Widder, Whittaker and West, 2024; USPTO, 2020).<\/p>\n

To make the disclosure of an AI model clear, complete, and sufficient, the following may be required:<\/p>\n

(i) Design and architecture:<\/strong> The type of model, its architecture and structure, including the relevant layers, modules, and connections, and the mathematical functions through which it produces the claimed result. Merely naming the type of model, such as a neural network or a convolutional network, may not be sufficient where the invention depends on a particular arrangement or modification (Xiao and Zhang, 2026).<\/p>\n

(ii) Model weights and parameters:<\/strong> The parameters and user-specified settings, or hyperparameters, that determine how the model operates, and where necessary, the trained weights themselves (Aboy et al., 2024). Making the weights available may allow another person to run the model, even where its functioning cannot be explained.<\/p>\n

(iii) Evaluation:<\/strong> The methods used to validate the model, the performance measures applied, and the results showing that the model achieves the claimed result, together with the conditions under which those results were obtained (Xiao and Zhang, 2026; White et al., 2024).<\/p>\n

(iv) Fine-tuning:<\/strong> Where a pre-trained or base model has been adapted for a particular field or purpose, the details of that adaptation, including the data, methods, and settings used for further training, because the claimed result may depend on the adaptation rather than on the base model (Lin\u00e5ker et al., 2026).<\/p>\n

(v) Training data and process:<\/strong> An account of the data and process through which the weights were obtained, because the weights by themselves do not explain how the model was arrived at (Widder, Whittaker and West, 2024; Ramos and Pianese, 2026). This aspect is discussed in more detail under training data below.<\/p>\n

Unless there is substantial common general knowledge with respect to the model in question, details relating to different aspects of the invention ranging from design and architecture to model weights and training data may be required to make the disclosure sufficient. Where the invention uses a well-known model in a routine way, a general reference to the model may be enough, but where the claimed result depends on a new architecture, a particular arrangement, or a specialised training procedure, those details may be required to be disclosed (Aboy et al., 2024). The disclosure challenge is, therefore, twofold: identifying the information that is necessary, and communicating it in sufficient detail to enable the model to be reproduced despite its complexity (Aboy et al., 2024; WIPO, 2026).<\/p>\n

(b) Challenges Relating to Training Data Disclosure<\/h3>\n

The functioning and performance of an AI model depend, to a large extent, on the data on which it is trained, and on how that data is selected, prepared, and used during training (Aboy et al., 2024; WIPO, 2026). Disclosure of training data is, however, difficult, as training datasets may be very large, commercially confidential, or subject to restrictions relating to privacy, licences, and contractual obligations (WIPO, 2026; Xiao and Zhang, 2026). In many cases, permission to use data for training does not necessarily include permission to disclose that data to the public, and the applicant may therefore face restrictions with respect to what can be included in a patent application (WIPO, 2026). Further, even where the model design and data are available, variations between training runs may produce differences in behaviour or performance, making exact reproduction of the model difficult (WIPO, 2026).<\/p>\n

To make the disclosure relating to training data clear, complete, and sufficient, the following may be required:<\/p>\n

(i) Characteristics of the data:<\/strong> The type, source, quantity, and relevant characteristics of the data used for training and testing, particularly where the claimed result depends on them (WIPO, 2026).<\/p>\n

(ii) Selection and preparation:<\/strong> How the data is selected, labelled, and prepared for training. If a particular pre-processing step is material to the invention, a statement that the data was cleaned, normalised, or transformed may not tell the skilled person enough to perform that step (WIPO, 2026).<\/p>\n

(iii) Training process:<\/strong> The learning method, the training settings, and the steps through which the model is trained, as naming a learning algorithm may leave important training choices undisclosed, even though those choices decide whether a working model can be produced (Aboy et al., 2024).<\/p>\n

(iv) Example dataset or access:<\/strong> At least one dataset suitable for training the claimed model or, where that is not possible, an alternative means of making the data available, such as a deposit with a designated institution (Aboy et al., 2024).<\/p>\n

Here again, the extent of disclosure required depends on the common general knowledge in the field. Where the data and training methods are standard and well known, a reference to them may be sufficient, but where the claimed result depends on particular data or a particular training procedure, those details may have to be disclosed (Aboy et al., 2024). Similarly, where a correlation in the training data cannot be presumed from common general knowledge, the description has to support it, through actual experimental results or other evidence (Nakayama, 2021). In many cases, a description of the relevant characteristics of the data and its preparation may be sufficient without disclosure of the dataset itself, but where the invention cannot be reproduced without access to particular data, whether and how that data has to be made available remains an open question (WIPO, 2026).<\/p>\n

(c) Challenges Relating to Documentation and Implementation Disclosure<\/h3>\n

Even where the model and its training are adequately disclosed, the skilled person may still need to know how the model is put to work in a particular system or application, and how it produces the claimed technical result (WIPO, 2026). A statement that a model predicts, classifies, or controls something may not provide this information where the relationship between the inputs and the result is neither disclosed nor apparent from common general knowledge (Nakayama, 2021). Furthermore, enablement and reproducibility may also depend on the specific hardware, software libraries, and deployment environment, which may not be disclosed in patent specifications (Xiao and Zhang, 2026).<\/p>\n

To make the disclosure relating to documentation and implementation clear, complete, and sufficient, the following may be required:<\/p>\n

(i) Inputs and outputs:<\/strong> What inputs the model receives, how they are obtained and processed, and how the outputs of the model are used to produce the claimed technical result (Xiao and Zhang, 2026).<\/p>\n

(ii) System integration:<\/strong> How the model operates together with the other components of the system, such as sensors, processors, and other devices, with flow charts and illustrations where appropriate (Upputuri and \u00d6zkan \u00dcner, 2023).<\/p>\n

(iii) Computing environment:<\/strong> Details of the hardware, software libraries, and deployment environment, where the reproduction of the claimed result depends on them (Xiao and Zhang, 2026).<\/p>\n

(iv) Performance across the claimed scope:<\/strong> Test results, benchmarks, and other performance information showing that the system works under the conditions and in the applications covered by the claims (Aboy et al., 2024).<\/p>\n

(v) Documentation:<\/strong> Information of the kind found in technical reports and model cards, such as how the model was tested and the conditions under which it is intended to operate, to the extent that such information is necessary to enable the invention (Lin\u00e5ker et al., 2026; White et al., 2024).<\/p>\n

As in the case of the other two groups, the extent of disclosure required depends on the common general knowledge in the field. Where the implementation follows routine engineering practice, the skilled person may be able to supply much of it, but where the claimed result depends on a particular arrangement, a particular way of processing inputs, or a particular use of outputs, that arrangement has to be disclosed in the specification (Aboy et al., 2024; Nakayama, 2021).<\/p>\n

The Person Skilled in the Art and Common General Knowledge<\/h3>\n

What amounts to sufficient disclosure may vary depending on what is claimed, what is required to enable the claims, what the common general knowledge relating to the invention is, and who the person skilled in the art is (Aboy et al., 2024; Masur and Ouellette, 2025). As AI technology is interdisciplinary and the art is not yet mature, who the person skilled in the art is, and what should be taken as common general knowledge, may have to be decided on a case-by-case basis. For a given invention, the person skilled in the art may be an individual, such as a machine learning engineer or a data scientist, or a team of persons with expertise in machine learning, software engineering, statistics, and the field in which the invention is applied, such as medicine or chemistry (EPO, 2026c; Caleb Suresh Motupalli v. Controller of Patents<\/em>, 2025). Likewise, the extent of common general knowledge may vary. Where AI has become part of routine implementation in a field, the common general knowledge may be large and widespread, as opposed to a field where the use of AI is not very prevalent (Nakayama, 2021). Further, as AI technology is progressing rapidly, the common general knowledge in the field, and the level of knowledge expected of the person skilled in the art, may change significantly within a short period, and as sufficiency is assessed as on the date of filing, the same disclosure may be found insufficient in an application filed today, but sufficient in a similar application filed a few years later. Therefore, what information relating to the model, data, and implementation needs to be disclosed depends on several general and specific factors, ranging from the maturity and predictability of the field and how routine the necessary experimentation is, to the nature of the AI invention, the characteristics of the person skilled in the art, and the status of common general knowledge (In re Wands<\/em>, 1988; WIPO, 2015; Masur and Ouellette, 2025). Recognising this, most major patent offices have issued some examination guidance on what may be considered sufficient and insufficient disclosure for specific types of AI inventions.<\/p>\n

Patent Office Guidance on AI Inventions<\/h2>\n

Patent offices in different countries have addressed disclosure issues relating to AI inventions through examination guidelines, illustrative examples, and discussions of how existing disclosure requirements apply to the technology (WIPO, 2026). At a general level, most offices have given guidance relating to AI inventions within, or alongside, their frameworks for computer-related, computer-implemented, or computer software-related inventions, although the terminology and approaches differ (CGPDTM, 2025; EPO, 2026b; JPO, 2024). Their guidance covers the information that may have to be disclosed relating to the model, its training, and its application, while recognising the role of the skilled person’s knowledge in determining how much detail has to be disclosed (WIPO, 2026).<\/p>\n

(a) Indian Patent Office:<\/strong> The Indian Patent Office’s Guidelines for Examination of Computer Related Inventions, 2025 provide guidance on sufficiency of disclosure for AI inventions through non-exhaustive illustrative scenarios and hypothetical examples (CGPDTM, 2025). These scenarios relate to input-output transformations, already trained models, data pre-processing, reinforcement learning, improvements to a computer’s internal structure or operations, specific characteristics of training datasets, and new learning algorithms or new methods of creating or preparing data. As per the guidelines, the disclosure required depends on the nature of the invention that is claimed, and depending on the invention in question, the patent applicant may have to fully or partially disclose the model architecture with sufficient structural detail, the characteristics of the training data, the pre-processing steps and their correlation with the model, the correlation between input and output data, the agent interactions and actions in reinforcement learning, and\/or validation metrics and test results. Though the Indian Patent Office’s guidance acknowledges that common general knowledge can supplement disclosures in a patent specification, it nevertheless requires detailed disclosures of the model, data, and implementation as relevant to the invention to avoid sufficiency issues (CGPDTM, 2025).<\/p>\n

(b) China National Intellectual Property Administration:<\/strong> Under China’s Guidelines for Patent Examination, as amended with effect from January 2026, the patent applicant may have to disclose the necessary modules, layers, or connections of the model, the essential training steps and parameters, and the relationship between the inputs and outputs of the model (CNIPA, 2025).<\/p>\n

(c) Ministry of Intellectual Property, Republic of Korea:<\/strong> Under South Korea’s examination practice guidelines for the AI field, issued in March 2026, the patent applicant may be required to disclose the specific technical means for implementing the invention, such as the training data, data pre-processing methods, the structure of the learning model, and the loss function, together with the correlation between the input data and the output data of a trained model (Ministry of Intellectual Property, 2026). For inventions based on reinforcement learning, the applicant may have to disclose the correlation among the agent, the environment, the state, the action, and the reward. The guidelines, however, recognise that the absence of an explicit disclosure of a particular implementation means may not make the disclosure insufficient where a person skilled in the art would clearly understand it in view of the common general knowledge at the time of filing (Ministry of Intellectual Property, 2026).<\/p>\n

(d) Japan Patent Office:<\/strong> The Japan Patent Office has provided case examples in its Examination Handbook on how its enablement and support requirements apply to AI-related inventions, with the latest examples having been added in March 2024 (JPO, 2024). As per these examples, where training data contains multiple types of data, the patent applicant may be required to disclose information from which a correlation among them can be recognised, such as an explanation or statistical information, unless such a correlation can be presumed in view of the common general technical knowledge at the time of filing. Where a product is claimed to have a function estimated by AI, the applicant may have to disclose an evaluation of that function using a product that has actually been made, unless the accuracy of the AI estimation is verified, or the common general technical knowledge at the time of filing shows that such an estimation can substitute for an actual evaluation. The examples also indicate that a broad claim may not be supported where the correlation is disclosed only for a narrower set of data, and that an insufficient description cannot be cured by experimental results submitted after filing. Further, where an invention relates to the generation of training data, the applicant may have to recite in the claims both the AI to be trained and the content of the training data that solves the problem (JPO, 2024).<\/p>\n

(e) European Patent Office:<\/strong> The EPO’s 2026 Guidelines for Examination state that disclosure may be insufficient where the mathematical methods and training datasets are disclosed in insufficient detail for the skilled person to reproduce the technical effect of the invention without undue burden, using common general knowledge, over the whole scope of the claim (EPO, 2026b). Where the technical effect depends on particular characteristics of the training dataset, the patent applicant may be required to disclose the characteristics required to reproduce that effect, unless the skilled person can determine them without undue burden using common general knowledge. In general, however, the guidelines do not require the applicant to disclose the specific training dataset itself (EPO, 2026c).<\/p>\n

(f) United States Patent and Trademark Office:<\/strong> The USPTO has not issued disclosure guidance specific to AI inventions, and it examines the disclosure of AI inventions under the general requirements in its Manual of Patent Examining Procedure (MPEP), including its guidance on computer-implemented inventions (USPTO, 2019; USPTO, 2020; USPTO, n.d.). The USPTO may not have issued specific guidance on AI inventions because, when it sought public comments on AI and intellectual property, the majority of commenters were of the view that there are no unique disclosure considerations for AI inventions (USPTO, 2020). Section 2161.01 of the MPEP directs examiners to assess whether the patent applicant has disclosed the computer and the algorithm that perform the claimed function in sufficient detail to show that the inventor was in possession of the invention, and whether the specification teaches the skilled person how to make and use the invention without undue experimentation (USPTO, n.d.). The MPEP also makes it clear that merely restating the function recited in the claim is not sufficient, and that it is not enough that the skilled person could write a program to achieve the claimed function (USPTO, n.d.).<\/p>\n

Following the decision of the US Supreme Court in Amgen Inc. v. Sanofi<\/em> (2023), the USPTO issued a memorandum in January 2024 stating that examiners have to continue to use the Wands<\/em> factors in assessing whether the full scope of the claimed invention is enabled without undue experimentation (USPTO, 2024; In re Wands<\/em>, 1988). These factors are applicable to all inventions, including AI inventions, and they include the breadth of the claims, the nature of the invention, the state of the prior art, the level of ordinary skill, the level of predictability in the art, the amount of direction provided by the inventor, the existence of working examples, and the quantity of experimentation needed (USPTO, 2024).<\/p>\n

(g) Canadian Intellectual Property Office:<\/strong> CIPO has also not issued final disclosure guidance specific to AI inventions, but towards that end it revised a part of its Manual of Patent Office Practice (MOPOP), and published it for consultation in June 2025 (CIPO, 2025). The revised portion of MOPOP states that a patent application covering a machine learning algorithm would not meet the sufficiency of disclosure requirement if training it would require undue experimentation on the part of the skilled person. The draft treats the activities required to reduce a specific series of logic instructions to computer code as part of the common general knowledge of a skilled programmer, but to meet the sufficiency requirement, the patent applicant may have to disclose sufficient instructions to allow the skilled person to arrive at an operable embodiment (CIPO, 2025).<\/p>\n

At present, some patent offices have issued guidelines or guidance relating to AI inventions, some are in the process of issuing or improving their guidance, and others continue to rely on their general disclosure requirements. While the Indian Patent Office and the Korean Ministry of Intellectual Property have given detailed guidance on what the patent applicant may have to disclose for specific types of AI inventions, offices like the EPO have kept their guidance general. Further, patent offices such as the JPO and the Indian Patent Office have given guidance by providing case examples and hypothetical examples indicating when a disclosure would be considered to be sufficient, and when it would not be. Though there is some variability among the offices that have issued guidance, all of them recognise that, in applying existing disclosure requirements to AI inventions, attention has to be paid to aspects that are specific to AI, such as the model, the training data, and the correlation between inputs and outputs.<\/p>\n

Applying the Guidance: A Bicycle Assistance System<\/h2>\n

As explained in the earlier parts, the disclosure needed for sufficiency varies depending on the invention, the statutory requirements in each country, and the guidelines adopted by different patent offices. Though there are several commonalities among patent offices, there are also differences. This part uses an example of an invention that brings together a model, training data, and a system through which the model is used, to understand what disclosure might be sufficient, and how the guidance of each patent office may apply in practice.<\/p>\n

Illustrative Invention<\/h3>\n

Assume that a person has come up with an invention relating to a bicycle assistance system that provides live guidance to a person with blindness while the person is riding. Cameras and distance sensors scan the surroundings, while other sensors measure the bicycle’s movement. The AI model identifies paths, obstacles, and moving objects, and uses these inputs to select guidance that is communicated through audio messages and tactile signals on the handlebars. The rider uses this guidance to steer, slow down, or stop.<\/p>\n

The model is initially trained using labelled images, distance measurements, and bicycle movement data representing different riding conditions. Reinforcement learning is then used to adapt its guidance function to environments in different countries, taking into account road layouts, traffic rules, road surfaces, and other road users. In simulated environments, useful and timely guidance is rewarded, while missed hazards, delayed warnings, and unsafe directions are penalised.<\/p>\n

The implementation includes the placement and coordination of sensors, processing of live inputs, communication of guidance, and responses to uncertain inputs or equipment failure.<\/p>\n

What Each Office May Require<\/h3>\n

All the offices considered below require the disclosure in the specification to enable a person skilled in the art to carry out the claimed invention without undue experimentation or burden, taking into account the common general knowledge at the time of filing. None of them would accept a disclosure that merely states that the AI model identifies obstacles and selects appropriate guidance, as their guidelines make it clear that a statement of the function or result, without the technical means of achieving it, is not sufficient. The differences among the offices lie in how much detail may have to be disclosed for each group, the form in which it may be disclosed, and the extent to which common general knowledge may be relied upon. These are discussed below with respect to the model, the training data, and the documentation and implementation of the bicycle assistance system.<\/p>\n

(a) The Model:<\/strong> All the offices would require the patent applicant to disclose the model that recognises paths, obstacles, and moving objects, and selects the guidance to be given, in a manner that enables the skilled person to build and run it. A mere reference to a neural network or a machine learning model would not be sufficient before any of them. However, there are differences in what information and details would be required across offices. Under the Indian guidelines, the patent applicant may have to disclose the architecture of the recognition and guidance models of the bicycle system with structural detail, such as the type of network, the number and types of layers, and the activation functions, along with training details such as the loss function, the optimiser settings, and the number of training cycles. The Korean guidelines may similarly require disclosure of the structure of the model and its loss function, though the structure of a well-known recognition component may not have to be disclosed explicitly if the skilled person would clearly understand it from common general knowledge.<\/p>\n

The Chinese guidelines may require the applicant to disclose the necessary modules, layers, or connections of the models, and the relationship between the camera images and the distance and movement measurements on the one hand, and the guidance produced on the other. The JPO may focus on whether the relationship between the camera images and measurements and the guidance is disclosed in the specification, or can be presumed from common general technical knowledge. If it cannot be presumed, the patent applicant may have to disclose an explanation or statistical information showing that the images and measurements correlate with the paths, obstacles, and moving objects identified by the model, and with the guidance it selects, as a mere statement that the model has been trained to give timely guidance may not be sufficient. As the USPTO has not issued disclosure guidance specific to AI inventions, its existing guidance on computer-implemented inventions may require the patent applicant of the bicycle system to disclose, through formulae or flow charts, the algorithm through which the system converts its observations into a particular instruction. Similarly, the EPO may require disclosure of the model and the mathematical methods in sufficient detail to reproduce the claimed effect of timely and reliable guidance across the whole scope of the bicycle system claims in the patent application.<\/p>\n

(b) Training Data:<\/strong> All the offices may require the patent applicant of the bicycle system to disclose enough about the training data for the skilled person to train a working model, though none of them may require the entire set of labelled images and sensor measurements to be disclosed. Under the Indian and Korean guidelines, the applicant may have to disclose the type and source of the images, distance measurements, and movement data, the riding conditions they represent, how they were labelled, the pre-processing steps applied to them, and the correlation between the data and the guidance produced. The JPO may require the applicant to disclose an explanation or statistical information supporting the correlation between the measurements and the timing and type of guidance, unless that correlation can be presumed from common general technical knowledge, and may not allow a claim covering all riding conditions if the correlation is supported only for some of them. The EPO may require the applicant to disclose the characteristics of the data that are necessary to reproduce the claimed guidance, such as the range of riding conditions, obstacles, and sensor inputs represented, along with an indication of at least the input data suitable for training the model. The EPO may also require the training data to cover the range of conditions within the scope of the claims, and the data used for training to correspond to the information that is actually available to the bicycle system when it operates, which in this case would be the live images and measurements available while riding.<\/p>\n

(c) Documentation and Implementation:<\/strong> All the offices may require the patent applicant of the bicycle system to disclose how the system is put together and operates, so that the skilled person can trace how the detection of an obstacle leads to a particular audio or tactile instruction. This may include the types and placement of the cameras, distance sensors, and movement sensors, how their live inputs are combined and processed, and how the selected guidance is delivered through audio messages and handlebar signals. Under the Indian guidelines, the applicant may have to disclose the specific sensors used to observe the surroundings and the devices through which guidance is given. The Korean guidelines indicate that general functional block diagrams or flow charts may not be sufficient, and the applicant may have to disclose how the hardware and software are concretely configured. The JPO may require the applicant to disclose test results from a bicycle system that has actually been made, showing that it detects obstacles and delivers timely audio and handlebar guidance as claimed, unless the accuracy of the model’s estimations can be verified in some other way. Based on its existing guidance relating to computer-implemented inventions, the USPTO may require the applicant to disclose the algorithm through which the system processes the live inputs and selects and communicates each instruction, including how it identifies an unreliable sensor reading and responds to equipment failure if these form part of the claims. The EPO may similarly require the applicant to disclose how these functions are carried out, so that the claimed guidance can be reproduced across the conditions claimed.<\/p>\n

As can be seen, all the patent offices would require a patent applicant to disclose details relating to the model, the training data, and the implementation, but there are differences in the extent of disclosure that might be required. It must, however, be noted that this analysis is based on a generalisation of the guidance provided by the patent offices, and a clear understanding may not be possible without specific examples of how each patent office has assessed the disclosure in particular AI invention-related patent applications. Further, there may be significant variability even within a patent office, depending on the discretion of the examiner, and the guidance of the patent offices can therefore provide only a general level of understanding. Some decisions of patent offices are likely to be challenged before appellate bodies and courts, and the outcomes of those challenges may provide greater certainty about the disclosure expected for AI inventions. That, however, may take several years, as it has, and continues to, in other fields of technology.<\/p>\n

WIPO’s Policy Options on Sufficiency of Disclosure<\/h2>\n

WIPO has been working on the patent-related challenges posed by AI inventions for several years now (WIPO, 2024a). After several consultations, engagements with stakeholders, and expert exchanges, it has recently published a report on the expert exchange on legal and policy options in the areas of patentability criteria and AI inventorship, along with a summary of the policy options based on relevant WIPO documents, as a part of the proceedings of its Standing Committee on the Law of Patents (WIPO, 2026). Annex 2 of the document outlines the challenges posed by AI inventions in the context of patentability and inventorship, and proposes policy options that may be considered by Member States (WIPO, 2026). The policy proposals include two options relating to sufficiency of disclosure for AI inventions. The first proposes the continuation of the existing standards, with clarifications on how they may be applied to AI inventions, and the second proposes a stringent set of disclosure standards for AI patent applications (WIPO, 2026). In the context of AI-generated inventions, the document also refers to a sui generis<\/em> system, raised by some commentators, to evaluate and protect such inventions (WIPO, 2026).<\/p>\n

(a) Applying current standards with additional clarity:<\/strong> Under the first option, WIPO proposes that current disclosure standards be applied with additional clarity, through minimum, flexible and functional disclosure requirements that recognise the inherent opacity of AI systems (WIPO, 2026). Under this option, the patent applicant may not have to provide a complete technical specification, and establishing the plausibility of the claimed correlation through theoretical explanation, empirical data, or statistical results may be sufficient, provided that it is substantiated and not merely asserted. To implement this option, patent offices may require the applicant to disclose actual performance results, such as accuracy rates, precision and recall metrics, or benchmark comparisons, may treat the absence of performance data as a presumptive indication of insufficiency, and may adopt reproducibility standards similar to those applied to biotechnology inventions, which call for statistical evidence of reliable performance rather than identical results. WIPO also states that, where national patent laws permit, patent offices may implement these requirements through examination guidelines, without amending the patent law (WIPO, 2026).<\/p>\n

(b) Strict or enhanced disclosure requirements:<\/strong> Under the second option, WIPO proposes strict or enhanced disclosure requirements, under which the patent applicant may have to disclose all training data and datasets, the complete model architecture, including layers, activation functions, and hyperparameters, the training methodologies, and the pre-processing steps (WIPO, 2026). Under this option, the applicant may be required to submit or publicly deposit the training datasets, or to provide access to them on a case-by-case basis, wherever it is determined that the invention would otherwise be insufficiently disclosed. WIPO, however, notes in the context of this option that the confidentiality of proprietary datasets is a significant obstacle, and that the option may give rise to conflicts with trade secret protection, privacy, and licensing obligations (WIPO, 2026).<\/p>\n

(c) A sui generis<\/em> system for AI-generated inventions:<\/strong> In the context of AI-generated inventions, WIPO also refers to the option of a sui generis<\/em> IP regime, raised by some commentators, which could have its own rules on inventorship, ownership, inventive step, and disclosure, along with shorter terms of protection (WIPO, 2026). In this regard, WIPO records the view of critics that such a system may be considered only if empirical evidence demonstrates a need for additional incentives (WIPO, 2026).<\/p>\n

Closing Thoughts<\/h2>\n

AI inventions, particularly those that are non-deterministic and probabilistic, pose several challenges to meet the disclosure requirements of the patent system. These include a range of issues from what disclosure is sufficient to enable a person skilled in the art, and what disclosure is required to support the claims and show possession, to who the person skilled in the art is and what constitutes common general knowledge. Case law in this area has not yet developed sufficiently, and patent offices have, in recent years, started addressing AI inventions through existing disclosure requirements, with guidelines and case examples indicating how those requirements may apply to the model, the training data, and the implementation. In its attempt to help Member States find solutions to the patent-related challenges posed by AI inventions, WIPO has proposed two policy options relating to disclosure, one, to maintain current standards with additional clarity, and the other, to make disclosure requirements stringent (WIPO, 2026).<\/p>\n

Both the disclosure-related policy options proposed by WIPO are workable, and patent offices already lean towards one or the other, with offices such as the Indian Patent Office and the Korean Ministry of Intellectual Property giving more detailed guidance, and offices such as the EPO taking a more flexible approach. Instead of choosing between the two options, however, a progressive approach that draws on both over time may be better suited to AI inventions. Under such an approach, heightened requirements based on elements of the second option, short of mandatory full disclosure of datasets, would be adopted at present, and relaxed towards the first option as the art progresses.<\/p>\n

At this stage, the AI art is uncertain and unpredictable, the capabilities attributed to the person skilled in the art are limited, and common general knowledge is still developing, and as a result, the skilled person has little to rely upon beyond what is disclosed in the specification. In this scenario, heightened requirements, under which the patent applicant would have to disclose detailed information relating to the model, the training data and its preparation, the training process, and the basis for concluding that the claimed result can be achieved, would ensure that the public receives the knowledge necessary to carry out the invention at a time when much of that knowledge cannot be assumed. As the art matures and becomes more certain and predictable, as the person skilled in the art assumes more capabilities, and as common general knowledge grows, information that initially had to be disclosed in detail may become part of common general knowledge, and the requirements may progressively be relaxed.<\/p>\n

Such a progressive approach is particularly suited to AI, as a large number of applications have already been filed and are being examined, with the number of generative AI patent families alone having risen from 733 in 2014 to more than 14,000 in 2023 (WIPO, 2026). Heightened requirements at this stage would ensure that patents granted on these applications are supported by adequate disclosure, while the examination of these applications, and the decisions of appellate bodies and courts on them, would provide the experience and certainty needed to calibrate and relax the requirements over time. As the requirements of enablement and claim support are often used to narrow claims that go beyond what is disclosed, heightened scrutiny at this stage would also help keep the scope of AI patents commensurate with the contribution disclosed, and prevent a few entities from gaining control over large areas of AI technology through broad claims. In this manner, inventors and applicants would receive protection for their AI inventions, and the public would receive the information needed to put those inventions into practice, and the give and take of the patent system would be preserved.<\/p>\n

With respect to AI-generated inventions, there is no need for a sui generis<\/em> system. IP-led incentive systems, including the patent system, are meant to incentivise humans to invest their intellect, time, and resources in creating and disclosing inventions, and as WIPO notes in its document, the incentive rationale is difficult to sustain where an invention is generated by a machine that has no capacity to respond to incentives (WIPO, 2026). AI-generated inventions, therefore, do not justify an incentive framework, and those who develop and use AI systems will continue to generate such inventions without the need for exclusive rights. What may instead be required is a regulatory framework that ensures the disclosure of the inner workings and details of AI-generated inventions, including the AI systems, data, and processes through which they were generated, as such inventions raise issues relating to safety, privacy, liability, and several other legal concerns that call for transparency. AI regulations such as the EU AI Act already impose transparency obligations of this kind, which operate independently of patent disclosure (Holst et al., 2026; Regulation (EU) 2024\/1689, Article 53).<\/p>\n

Whether through progressively calibrated patent disclosure for AI inventions, or regulatory disclosure for AI-generated inventions, the objective remains the same, namely, that the knowledge underlying AI inventions reaches the public, as the give and take of the patent system has always intended.<\/p>\n

Comments and Inputs<\/h2>\n

This article is a work in progress, and the author welcomes critical comments, inputs, and feedback. Readers may write to the author at kalyan@bananaip.com<\/a> with their comments and suggestions for improving the article.<\/p>\n

Acknowledgement<\/h2>\n

The author thanks Mr. Sandeep Hegde, Legal Counsel, Oracle, for his valuable comments and inputs, which helped improve this article.<\/p>\n

References<\/h2>\n

Aboy, M., Price, W.N. II, Raker, S. and Liddell, K. (2024) The sufficiency of disclosure of medical artificial intelligence patents<\/a>, Nature Biotechnology, 42(6), pp. 839\u2013845. doi: 10.1038\/s41587-024-02270-8.<\/p>\n

Agreement on Trade-Related Aspects of Intellectual Property Rights<\/em> (TRIPS Agreement) (1994) Annex 1C to the Marrakesh Agreement Establishing the World Trade Organization, 15 April 1994. United Nations Treaty Series<\/em>, 1869, p. 299.<\/p>\n

Amgen Inc. v. Sanofi<\/em> (2023) 598 U.S. 594 (US Supreme Court).<\/p>\n

Ariad Pharmaceuticals, Inc. v. Eli Lilly & Co.<\/em> (2010) 598 F.3d 1336 (Fed. Cir.) (en banc).<\/p>\n

Budapest Treaty on the International Recognition of the Deposit of Microorganisms for the Purposes of Patent Procedure<\/em> (Budapest Treaty) (1977) Adopted at Budapest on 28 April 1977, as amended on 26 September 1980. Available at: https:\/\/www.wipo.int\/wipolex\/en\/text\/283784<\/a> (Accessed: 26 October 2026).<\/p>\n

Caleb Suresh Motupalli v. Controller of Patents<\/em> (2025) C.M.A. (PT) No. 2 of 2024, 2025:MHC:293 (High Court of Judicature at Madras).<\/p>\n

Canadian Intellectual Property Office (CIPO) (2025) Draft\u2014Revised MOPOP Chapter 22: Computer-Implemented Inventions<\/em>. Consultation draft, June 2025. Gatineau: CIPO. Available at: https:\/\/ised-isde.canada.ca\/site\/canadian-intellectual-property-office\/en\/draft-revised-mopop-chapter-22<\/a> (Accessed: 24 October 2026).<\/p>\n

China National Intellectual Property Administration (CNIPA) (2025) Explanation of the 2025 amendments to the Patent Examination Guidelines<\/em> [in Chinese]. 4 December 2025. Beijing: CNIPA. Available at: https:\/\/www.cnipa.gov.cn\/art\/2025\/12\/4\/art_66_202935.html<\/a> (Accessed: 24 October 2026).<\/p>\n

Code of Federal Regulations<\/em> 2024, Title 37, Chapter I, Part 1, \u00a7 1.96, \u2018Submission of computer program listings\u2019. Washington, DC: United States Government Publishing Office. Available at: https:\/\/www.govinfo.gov\/content\/pkg\/CFR-2024-title37-vol1\/pdf\/CFR-2024-title37-vol1.pdf<\/a> (Accessed: 26 October 2026).<\/p>\n

Controller General of Patents, Designs and Trade Marks (CGPDTM) (2014) Guidelines for Examination of Patent Applications in the Field of Pharmaceuticals<\/em>. October 2014. Section 4, \u2018Markush claims\u2019. Mumbai: Office of the CGPDTM, Government of India.<\/p>\n

Controller General of Patents, Designs and Trade Marks (CGPDTM) (2025) Guidelines for Examination of Computer Related Inventions (CRIs), 2025<\/em>. Mumbai: Office of the CGPDTM, Government of India. Available at: https:\/\/ipindia.gov.in\/storage\/uploads\/docs-operator\/7fdd559b-1ff8-4a87-8c36-033e7356b39b.pdf<\/a> (Accessed: 24 October 2026).<\/p>\n

European Patent Office (EPO) (2026a) Guidelines for Examination in the European Patent Office<\/em>. April 2026 edn. Part A, Chapter IX, section 11.1, \u2018Chemical and mathematical formulae\u2019. Munich: EPO. Available at: https:\/\/www.epo.org\/en\/legal\/guidelines-epc\/2026\/a_ix_11_1.html<\/a> (Accessed: 26 October 2026).<\/p>\n

European Patent Office (EPO) (2026b) Guidelines for Examination in the European Patent Office<\/em>. April 2026 edn. Part F, Chapter III, section 3, and Part G, Chapter II, section 3.3.1, and Index for Computer-Implemented Inventions. Munich: EPO. Available at: https:\/\/www.epo.org\/en\/legal\/guidelines-epc\/2026\/f_iii_3.html<\/a>and https:\/\/www.epo.org\/en\/legal\/guidelines-epc\/2026\/g_ii_3_3_1.html<\/a> (Accessed: 24 October 2026). Index for Computer-Implemented Inventions<\/a> (Accessed: 28 October 2026).<\/p>\n

European Patent Office (EPO) (2026c) Guidelines for Examination in the European Patent Office<\/em>. April 2026 edn. Part G, Chapter VII, section 3, \u2018Person skilled in the art\u2019. Munich: EPO. Available at: https:\/\/www.epo.org\/en\/legal\/guidelines-epc\/2026\/g_vii_3.html<\/a> (Accessed: 29 October 2026).<\/p>\n

Fabris, D. (2020) From the PHOSITA to the MOSITA: Will \u201csecondary considerations\u201d save pharmaceutical patents from artificial intelligence?, IIC \u2013 International Review of Intellectual Property and Competition Law, 51, pp. 685\u2013708. doi: 10.1007\/s40319-020-00953-8.<\/p>\n

Holst, L., Mayer, V., L\u00e4mmermann, L., Urbach, N. and Wendt, D. (2026) The impact of the EU AI Act\u2019s transparency requirements on AI innovation, in Flath, C.M., Gust, G., Thiesse, F. and Winkelmann, A. (eds) Artificial Intelligence, Data, and Decision-Making<\/em>. Lecture Notes in Information Systems and Organisation 82. Cham: Springer, pp. 135\u2013148.<\/p>\n

In re Wands<\/em> (1988) 858 F.2d 731 (Fed. Cir.).<\/p>\n

Japan Patent Office (JPO) (2024) Case Examples pertinent to AI-related technologies<\/em>. Examination Handbook for Patent and Utility Model, Annex A, provisional English translation, including case examples added in March 2024. Tokyo: JPO. Available at: https:\/\/www.jpo.go.jp\/e\/system\/laws\/rule\/guideline\/patent\/document\/ai_jirei_e\/jirei_e.pdf<\/a> (Accessed: 28 October 2026).<\/p>\n

Lin\u00e5ker, J., Osborne, C., Ding, J. and Burtenshaw, B. (2026) A cartography of open collaboration in open source AI: Mapping practices, motivations, and governance in 14 open large language model projects<\/a>. arXiv preprint, arXiv:2509.25397v2.<\/p>\n

Masur, J.S. and Ouellette, L.L. (2025) Disclosure puzzles in patent law, University of Chicago Law Review, 92, p. 1609.<\/p>\n

Ministry of Intellectual Property, Republic of Korea (2026) Examination Practice Guidelines by Technology Field<\/em>. English edn, March 2026. Chapter I, \u2018Artificial intelligence field\u2019. Daejeon: Ministry of Intellectual Property. Available at: https:\/\/www.kipo.go.kr\/upload\/en\/download\/EXAMINATION_PRACTICE_GUIDELINES_BY_TECHNOLOGY_FIELD.pdf<\/a> (Accessed: 24 October 2026).<\/p>\n

Nakayama, I. (2021) Patentability and PHOSITA in the AI era\u2014A Japanese perspective<\/a>, in Lee, J.-A., Hilty, R.M. and Liu, K.-C. (eds) Artificial Intelligence and Intellectual Property. Oxford: Oxford University Press, chapter 5. doi: 10.1093\/oso\/9780198870944.003.0006.<\/p>\n

Ouellette, L.L., Fang, V. and Ouellette, N.T. (2025) How will AI affect patent disclosures?<\/a>, Nature Biotechnology, 43, pp. 26\u201328. doi: 10.1038\/s41587-024-02515-6.<\/p>\n

Patent Cooperation Treaty<\/em> (1970) Done at Washington on 19 June 1970, as amended and modified. United Nations Treaty Series<\/em>, 1160, p. 231.<\/p>\n

Patents Act 1970<\/em> (India), Act No. 39 of 1970, as amended.<\/p>\n

Ramos, S. and Pianese, F. (2026) Regulating transparency: A critique of the \u201cfree and open source AI\u201d definition in the EU AI act<\/a>, AI and Ethics, 6, article 317. doi: 10.1007\/s43681-026-01177-1.<\/p>\n

Regulation (EU) 2024\/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act)<\/em> (2024) Official Journal of the European Union<\/em>, L series, 2024\/1689, 12 July 2024. Available at: https:\/\/eur-lex.europa.eu\/eli\/reg\/2024\/1689\/oj<\/a> (Accessed: 29 October 2026).<\/p>\n

Regulations under the Patent Cooperation Treaty<\/em> (PCT Regulations) (1970) Adopted on 19 June 1970, as amended. Geneva: WIPO.<\/p>\n

Robinson, W.K. (2026) Responsibility, transparency, and accountability in AI patents<\/a>, in Lim, D. and Yu, P.K. (eds) Inclusive Innovation in the Age of AI and Big Data. Oxford: Oxford University Press, chapter 13. doi: 10.1093\/9780197799444.003.0017.<\/p>\n

United States Code<\/em> 2018, Title 35 (Patents), \u00a7\u00a7 112 and 282. Washington, DC: Office of the Law Revision Counsel.<\/p>\n

United States Patent and Trademark Office (USPTO) (2019) Examining Computer-Implemented Functional Claims for Compliance With 35 U.S.C. 112<\/em>. Federal Register<\/em>, 84(4), 7 January 2019, pp. 57\u201363. Alexandria, VA: USPTO.<\/p>\n

United States Patent and Trademark Office (USPTO) (2020) Public Views on Artificial Intelligence and Intellectual Property Policy<\/em>. October 2020. Alexandria, VA: USPTO. Available at: https:\/\/www.uspto.gov\/sites\/default\/files\/documents\/USPTO_AI-Report_2020-10-07.pdf<\/a> (Accessed: 24 October 2026).<\/p>\n

United States Patent and Trademark Office (USPTO) (2024) Guidelines for Assessing Enablement in Utility Applications and Patents in View of the Supreme Court Decision in Amgen Inc., et al. v. Sanofi et al.<\/em> Memorandum to all patent examiners, 10 January 2024. Alexandria, VA: USPTO. Available at: https:\/\/www.uspto.gov\/sites\/default\/files\/documents\/112a-memo.pdf<\/a> (Accessed: 24 October 2026).<\/p>\n

United States Patent and Trademark Office (USPTO) (n.d.) Manual of Patent Examining Procedure<\/em>. 9th edn. Sections 2161, 2161.01, 2163, 2164 and 2165. Alexandria, VA: USPTO. Available at: https:\/\/www.uspto.gov\/web\/offices\/pac\/mpep\/s2161.html<\/a>and https:\/\/www.uspto.gov\/web\/offices\/pac\/mpep\/s2164.html<\/a> (Accessed: 24 October 2026).<\/p>\n

Upputuri, A.H. and \u00d6zkan \u00dcner, Z. (2023) Patent disclosure requirement for AI-assisted inventions: A comparative study of EU and India, Journal of Intellectual Property Rights, 28(2), pp. 114\u2013122.<\/p>\n

White, M., Haddad, I., Osborne, C., Liu, X.-Y.Y., Abdelmonsef, A., Varghese, S. and Le Hors, A. (2024) The model openness framework: Promoting completeness and openness for reproducibility, transparency, and usability in artificial intelligence<\/a>. arXiv preprint, arXiv:2403.13784v4.<\/p>\n

Widder, D.G., Whittaker, M. and West, S.M. (2024) Why \u201copen\u201d AI systems are actually closed, and why this matters<\/a>, Nature, 635, pp. 827\u2013833. doi: 10.1038\/s41586-024-08141-1.<\/p>\n

World Intellectual Property Organization (WIPO) (2015) Study on the Sufficiency of Disclosure: Summary<\/em>. Document SCP\/22\/4 Summary, 4 May 2015. Standing Committee on the Law of Patents, Twenty-Second Session. Geneva: WIPO. Available at: https:\/\/www.wipo.int\/edocs\/mdocs\/scp\/en\/scp_22\/scp_22_4_summary.pdf<\/a> (Accessed: 30 October 2026).<\/p>\n

World Intellectual Property Organization (WIPO) (2024a) Getting the Innovation Ecosystem Ready for AI: An IP Policy Toolkit<\/em>. Geneva: WIPO. doi: 10.34667\/tind.48978.<\/p>\n

World Intellectual Property Organization (WIPO) (2024b) WIPO Standard ST.26: Recommended Standard for the Presentation of Nucleotide and Amino Acid Sequence Listings Using XML<\/em>. Version 1.7, January 2024. Geneva: WIPO. Available at: https:\/\/www.wipo.int\/documents\/d\/standards\/docs-en-03-26-01.pdf<\/a> (Accessed: 26 October 2026).<\/p>\n

World Intellectual Property Organization (WIPO) (2026) Report on the Expert Exchange on Legal and Policy Options in the Areas of Patentability Criteria and AI Inventorship (SCP\/37), and Summary of the Policy Options Based on Relevant WIPO Documents<\/em>. Document SCP\/38\/6, 15 September 2026. Annex 2, sections I, III.F, and IV.B. Geneva: WIPO. Available at: https:\/\/www.wipo.int\/edocs\/mdocs\/scp\/en\/scp_38\/scp_38_6.pdf<\/a> (Accessed: 24 October 2026).<\/p>\n

Xiao, R. and Zhang, J. (2026) The sufficiency of disclosure in medical AI patents: Evidence from 865 granted patents across the US, China, and the EU, Journal of Law, Medicine & Ethics, pp. 1\u201314. doi: 10.1017\/jme.2026.10256.<\/p>\n

Disclaimer Regarding Use of AI<\/h2>\n

Artificial intelligence tools were used in preparing this article to assist with summarising the source material, refining the language, correcting grammatical errors, and checking references. The author has reviewed, revised, and approved the entire content, and takes full responsibility for the arguments, analysis, and conclusions in the article.<\/p>\n","protected":false},"excerpt":{"rendered":"

AI inventions bring together models, training data, and implementation, and as their functioning is probabilistic, not humanly explainable in most cases, and difficult to reproduce, disclosing them clearly and completely in a patent application is a challenge. This article examines what patent applicants may have to disclose for each of these, as per the guidance of major patent offices and WIPO’s policy options, and argues for a progressive approach under which disclosure requirements are heightened now and relaxed as the AI art matures.<\/p>\n","protected":false},"author":3,"featured_media":150635,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"iawp_total_views":0,"footnotes":""},"categories":[14,6],"tags":[3982,13296,13299,13163,7681,6591,7418,10352,5760,13295,13297,13298],"class_list":["post-150632","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-patents","category-intellectual-property","tag-ai-inventions","tag-ai-model","tag-ai-generated-inventions","tag-common-general-knowledge","tag-enablement","tag-patent-disclosure","tag-patent-office-guidelines","tag-person-skilled-in-the-art","tag-sufficiency-of-disclosure","tag-training-data","tag-trips-agreement-article-29-1","tag-wipo-scp"],"_links":{"self":[{"href":"https:\/\/www.bananaip.com\/intellepedia\/wp-json\/wp\/v2\/posts\/150632","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.bananaip.com\/intellepedia\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.bananaip.com\/intellepedia\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.bananaip.com\/intellepedia\/wp-json\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/www.bananaip.com\/intellepedia\/wp-json\/wp\/v2\/comments?post=150632"}],"version-history":[{"count":2,"href":"https:\/\/www.bananaip.com\/intellepedia\/wp-json\/wp\/v2\/posts\/150632\/revisions"}],"predecessor-version":[{"id":150636,"href":"https:\/\/www.bananaip.com\/intellepedia\/wp-json\/wp\/v2\/posts\/150632\/revisions\/150636"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.bananaip.com\/intellepedia\/wp-json\/wp\/v2\/media\/150635"}],"wp:attachment":[{"href":"https:\/\/www.bananaip.com\/intellepedia\/wp-json\/wp\/v2\/media?parent=150632"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.bananaip.com\/intellepedia\/wp-json\/wp\/v2\/categories?post=150632"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.bananaip.com\/intellepedia\/wp-json\/wp\/v2\/tags?post=150632"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}