# AI Patents: What Can Be Protected and How to Prepare

This guide explains the issues to consider and the materials to prepare when seeking patent protection for AI technology.

Source: https://www.iplexlaw.co.kr/en/blog/1538326

HOME / NEWS & INSIGHTS NEWS & INSIGHTS AI Patents: What Can Be Protected and How to Prepare This guide explains the issues to consider and the materials to prepare when seeking patent protection for AI technology. AI & Software 2026.09.17 published IPLEX 22 min read AI patent applications are not limited to new model architectures. Potentially protectable features can arise throughout a system, including data preprocessing, training-data generation and selection, training methods, retrieval and generation, inference resource management, distributed training and device control. Where a system uses an established model or public library, the assessment can focus on any additional technical improvements. A patent application starts with a specific solution, rather than a description of a function. Explain the inputs, processing steps, interactions between components and limitations of existing approaches that the solution addresses. Then assess novelty and inventive step against the prior art and whether the specification adequately supports the proposed claims. This guide is for researchers, developers and technical managers familiar with model training and inference, fine-tuning, embeddings and retrieval-augmented generation (RAG). It focuses on Korean patent applications. The implementation examples are hypothetical and illustrate potential areas of protection. Patentability requires an individual assessment of the prior art, implementation and claims at the time of filing. Core requirements for AI patent protection Statutory invention requirements and concrete information processing In Korea, AI inventions are assessed under the examination criteria for computer-related inventions. Beyond identifying a mathematical principle or abstract idea, the analysis considers whether software-based information processing is concretely implemented using hardware. Merely adding a computer or server to a claim is not sufficient. For example, stating the aim of recommending content to suit a user's preferences differs from describing a processing architecture that converts input histories into a defined representation, generates candidates and selects among them under latency or resource constraints. Even when the implementation is specified, novelty and inventive step require separate assessment. Novelty and inventive step Novelty is assessed against prior art disclosed before filing. Relevant material may include patent documents, papers, conference presentations, technical documentation, publicly available code and product disclosures. A technology that is new to a company is not necessarily new for patent purposes. Inventive step concerns whether a person skilled in the art could readily derive the claimed differences from the prior art. Choosing a known model, adding established modules or automating an existing task with AI may not provide a sufficient distinction. Explain the combination of features, the technical difficulty addressed and the effects attributable to those differences. An improvement in performance does not automatically establish inventive step. Assess whether the result comes from additional data or computation, or from the claimed technical features. Nor is accuracy the only relevant measure: latency, memory use, communication volume, error recovery and device-control stability may also matter where they relate to the technical problem. Sufficiency of disclosure and support for the claims The specification must enable a person skilled in the art to implement the invention. It should disclose the necessary inputs and outputs, data characteristics, preprocessing, model or training method and inference process. The claims must be supported by the description and clearly define the subject matter. Not every application requires disclosure of the entire source code or trained model weights. However, omitting details needed to implement the core invention can create problems. Distinguish established techniques from the independently developed improvements when deciding how much detail is required. Types of technologies for which protection may be considered 1 Data collection and temporal alignment Differences in sensor sampling rates or communication delays can cause data from different times to be combined in a single input. For example, a system analyzing vibration and current in rotating equipment could detect reference events marking changes in operating state, align the signals in time and mask uncertain intervals. Features to assess include reference-event detection, time-offset estimation, alignment-confidence calculation, mask generation and use by downstream models. Explain the conditions and criteria applied, rather than merely stating that the data is aligned. Compare the approach with existing interpolation or synchronization methods and assess its effects when sensor data is missing or operating tasks change. Filing materials may include raw-signal types, sampling intervals, delay ranges, handling of alignment failures and consistent application of the approach during training and inference. For a preprocessing invention, explain how raw data becomes an input to the training or inference process. 2 Image preprocessing and region of interest selection Small-defect detection often involves a trade-off between image resolution and processing speed. One approach identifies candidate regions using texture changes and image-acquisition conditions, reanalyzes only those regions at high resolution and combines global and local features. The relevant features extend beyond using regions of interest. They include selection criteria, region expansion, overlap removal and the way global and local features are combined. Compare processing requirements and missed-defect rates at a given detection performance, accounting for defect size. Keep image resolution and hardware conditions consistent when comparing results. 3 Controlling training sample selection and labeling Active learning selects uncertain data for labeling, but the specific selection and update mechanisms still require assessment. For example, a system could score prediction uncertainty, rarity within a product group, overlap with existing samples and sensor status, then use the results to adjust the sample budget for the next cycle. Document score calculation, exclusion criteria, sample-budget allocation and the order in which selected samples enter training. Compare performance or reductions in duplicate selections under the same labeling budget. The assessment concerns differences from known selection methods, rather than the label 'active learning' itself. 4 Generation and verification of synthetic data When generative models supplement limited data, validating the generated samples may be central to the invention. A hypothetical system uses process conditions and material information to guide generation, then admits samples to training only if the generated defect's shape and location fall within physically plausible limits. Potentially protectable features extend beyond the generated image. They may include construction of generation conditions, validation rules, rejection or regeneration of unsuitable samples and mixing with real data. Comparing results on real test data with and without validation helps explain the role of each step. 5 Multimodal feature combination and model structure When a system combines images, audio or vibration signals, input quality can vary. A separate quality-assessment path could determine feature weights or select a processing path. The design may also provide alternative paths when a sensor is unavailable. Describe each input representation, the quality signals, where features are combined, how gates operate and how the components connect during training. Compare the approach with simple concatenation, averaging and conventional attention. Relevant measures include performance with degraded inputs and the additional computational cost. Describing how components interact is more useful than assigning a new name to the architecture. For claim preparation, distinguish interchangeable modules from connections essential to the solution's operating principle. 6 Loss functions and training schedules Changes to training objectives may also warrant patent assessment. For example, a system could analyze recent training errors to update loss-term weights, while limiting the update magnitude when instability is detected. This could address changes in the relative importance of errors across product lines or process states. Explain every variable and how the equations operate within the training loop: the source of the statistics, update intervals and how updates continue through training. Compare the approach with selecting a known loss function or routine parameter tuning. 7 Fine-tuning and adapter operation Fine-tuning an openly available foundation model may involve additional inventions. Applying LoRA to a particular model is not, by itself, a sufficient technical distinction. Examine the specific mechanisms for selecting, sharing, loading and switching task-specific adapters. For example, a system might classify incoming tasks, load only the required adapters, reuse shared components and validate caches when switching adapters. Describe adapter training separately from runtime selection, since the technical features and responsible actors may differ. 8 Knowledge distillation and model compression Rather than transferring a large model's full output to a smaller model, a hypothetical approach selects task-relevant features and changes the teacher guidance used at stages where the smaller model performs poorly. Assess feature-selection criteria, the correspondence between teacher and student models and the combination of loss terms. To explain compression effects, compare latency, memory use and task-specific performance loss on the same device, alongside model file size. Data showing which errors increase at higher compression ratios and which features mitigate them can clarify the technical contribution. 9 RAG document chunking and retrieval In enterprise document retrieval, chunking can separate a provision from its exceptions. A hypothetical approach links related sections using heading hierarchy, reference identifiers and product-applicability information, then retrieves groups of sections relevant to the question type. The potential invention lies in chunk-boundary criteria, stored relationships, retrieval-expansion rules and context-exclusion conditions, rather than the use of RAG itself. Compare these features with fixed-length chunking and known hierarchical retrieval. Measures such as evidence recall, context length and latency should reflect the practical problem. 10 Grounding verification and follow-up retrieval A highly similar document is not necessarily valid evidence. One approach filters documents by product version and period of applicability, matches statements in the generated answer to supporting passages, then retrieves further evidence or withholds unsupported output. Assess applicability extraction, document-validity checks, statement segmentation, matching criteria, follow-up retrieval scope and stopping conditions. The goal of accurate answers or eliminating hallucinations is not enough on its own. Explain verification costs and how the system avoids withholding output unnecessarily. 11 Automated prompt construction and output-format control Distinguish an attempt to monopolize particular instructions or wording from a computer-implemented method for constructing prompts. A hypothetical approach classifies input errors, selects tool specifications and examples for each type, arranges them within a token budget and uses output-structure checks to construct the next input. Describe the relationship between input analysis and example selection, context priority, format-violation checks and changes made on retry. Assess whether the approach amounts to changed wording or a routine combination of known prompting techniques. Keep the product's purpose distinct from its technical implementation. 12 AI agent execution state and recovery Agents calling external tools can encounter duplicate execution on retry or inconsistent state after partial failure. A possible architecture associates execution identifiers and state records with tool calls, checks completion after an interrupted response and determines which steps to repeat or how to restore a defined state after failure. Assess this approach against prior art in distributed processing and transaction handling. Connecting an LLM does not, by itself, make an existing recovery technique inventive. Explain how planning, tool execution and state checks interact to address a specific problem. 13 Quantization and selective recomputation When reduced precision increases errors for particular layers or inputs, precision allocation may account for both layer sensitivity and device operating conditions. For example, the system could recompute selected stages at higher precision when an error signal is detected during inference. Examine sensitivity estimation, precision allocation, switching conditions and the scope of recomputation. Compare these features with known mixed-precision techniques and measure conversion costs. A different bit width alone does not establish a sufficient technical distinction. 14 Context cache and memory management Repeated use of the same context makes caching important to processing cost. To prevent stale representations after a document changes, a system could track context-fragment versions and dependencies, then invalidate only affected cache entries. Describe dependencies, change detection, invalidation scope, recomputation order and handling of model or adapter changes. Identify the difference from conventional caching. Relevant effects may include reduced incorrect reuse, memory consumption and latency. 15 Splitting inference between a device and a server Inference placement may respond to device temperature, battery state and network latency. For example, a system could estimate computation and transmission requirements at candidate split points, select a partition for current conditions and use an on-device fallback during a communication failure. Inference partitioning and general resource optimization may already be known. Examine the model's intermediate representations, device memory structure and preservation of state during transitions. Describe latency immediately after a transition and behavior when connectivity returns. 16 Robot perception and control A hypothetical robot uses cameras and tactile sensors to predict slippage and adjust grip force within limits that prevent damage to the object. Assess sensor-feature fusion, prediction confidence, permissible-force calculation and control updates. Specify how prediction scores affect physical actions. For reinforcement learning, describe the relationships among the agent, environment, states, actions and rewards as they relate to the invention. Explain how the device's permitted operating range constrains exploration during training and how the system selects a path under uncertainty. 17 Process quality prediction and closed-loop control Predicting process conditions and controlling production in response may involve distinct technical features. Where control inputs have delayed effects, a system could include control history and delays in its state representation, then limit changes according to prediction uncertainty. Describe how model outputs, control commands, device responses and subsequent inputs connect. Examine control stability beyond the display of predictions. Useful measures may include quality variation, control-signal oscillation and recovery time after abnormal conditions. 18 Medical image analysis and diagnostic assistance Medical-imaging inventions may concern acquisition-condition correction, region segmentation, registration of different image modalities or uncertainty estimates. For example, an information-processing device could select a correction model based on acquisition conditions and provide analysis results with confidence information. In Korea, surgical, therapeutic and diagnostic methods performed on humans require separate assessment of industrial applicability. Distinguish devices and information-processing technologies from methods of medical treatment by examining the actual claims; labeling a claim as a device does not automatically change the result. Patentability is also separate from medical-device approval and clinical effectiveness. 19 Federated learning and distributed updating Distributed updates may be aggregated across institutions with different data distributions and communication conditions. A hypothetical architecture uses update deviation and validation reliability to determine aggregation weights, then adjusts the parameters transmitted to participants with unstable connections. Describe aggregation criteria, update selection and exclusion, delayed-update handling and retransmission conditions. Identify which steps are performed by the central server and which by participating institutions. Keeping raw data local does not guarantee privacy: identify the potential information leakage and the specific measures addressing it. 20 Performance monitoring, retraining and deployment Operational performance can deteriorate because of sensor failures, temporary environmental changes or new data types. A system could combine input statistics, sensor status and error patterns from delayed ground-truth labels to identify the cause, then select inspection, preprocessing updates or partial retraining. The potential contribution lies in linking cause assessment to the chosen response, beyond a monitoring dashboard. A new model might be deployed only after meeting validation criteria and rolled back if operating conditions are no longer satisfied. Identify AI-specific state assessments or data-processing relationships that differ from general deployment automation. Defining the subject matter for protection Algorithms and mathematical principles Distinguish mathematical formulas or abstract algorithms from inventions that use them in concrete information processing. Whether the subject matter qualifies as an invention depends on the claim as a whole; including a formula is not decisive either way. Explain how it operates on actual inputs through processing, storage, computation or control. Datasets and trained model weights A large dataset or different learned numerical values do not, by themselves, confer patent protection over an entire dataset or weight file. Assessment may focus on data generation, transformation, selection and use, or the model's specific structure and function. Claims directed to a trained model require separate consideration of how the subject matter is defined and whether the invention and disclosure requirements are met. Contractual rights to use data, copyright-related rights, trade secret management and patents are different issues. Filing a patent application does not secure the right to use training data or external models. Service concepts and business rules Descriptions such as AI consultation, AI pricing or AI recommendations do not identify a technical distinction. Examine input processing, model interactions and use of outputs. Even for a business-oriented invention, the assessment concerns the actual information-processing architecture; commercial performance does not replace a technical contribution. Materials and structures proposed by AI Where AI helps identify a new material, compound or mechanical structure and the resulting product is claimed, the patentability and disclosure requirements of that field apply. Assess whether predictions alone sufficiently support manufacturability or the stated effects. Using AI does not replace the assessment of novelty, inventive step or whether the result can be implemented. Materials to prepare for a patent assessment Technical issues and existing approaches Document where existing approaches fail. Instead of saying accuracy is low, identify the data and operating conditions associated with the errors. Relevant papers, publicly available code and existing product workflows help establish a basis for comparison. Then identify the changes: preprocessing criteria, model connections, inference paths or other features. Distinguish the system's overall function from the core features of each potential invention. Relationship between training data and model Record data types, collection conditions, label definitions, preprocessing and data-splitting criteria. Explain how data characteristics relate to the intended prediction or control. A list of inputs alone does not establish the relationship between input and output. If the contribution lies in model architecture, describe module connections, inputs and outputs, and differences between training and inference. If it lies in training, document the loss function, sample selection, update order and stopping criteria as implemented. Where the application of a model is central, input construction and use of outputs may need more detail than the established model itself. Comparative experiments and technical effects Choose metrics that test the claimed effect. Defect detection may call for missed-detection and false-positive rates; RAG may require measures of supporting evidence and latency; distributed processing may involve memory use and communication volume. A single accuracy measure need not explain every effect. Record the comparison conditions. Differences in model size, training time, hardware, input length, batch size or data splits can complicate interpretation. Removing or replacing a new feature can help show its contribution. Clearly distinguish measured results from expected effects. Alternative implementations and applications Consider whether the same solution works with other models, sensors or deployment environments. Separating essential features from optional implementations helps avoid unnecessarily narrow descriptions. Adding variants without enough detail to implement them does not, however, support broader protection. If a numerical value is central, explain why it was selected and the range over which it applies. Distinguish a value incidental to one product from a condition that matters to solving the technical problem. Aligning claim strategy with the business model Patent protection is defined by the claims. Check that the claimed features reflect the benefits described for the technology. Omitting essential features or adding unnecessary steps can weaken the protection sought for the business. Distinguish features used only on training servers from those present during inference on customer devices. If different entities train the model, provide it and operate the service, the appropriate claim subjects may differ. Consider method, device, system and recording-medium claims in light of the actual implementation. In distributed systems, determine whether key steps are divided among operators or countries. Whether a claim combining all those steps is suitable for enforcement is a separate question. Server-side and device-side processes with independent technical significance may warrant separate assessment. One service may contain several potential inventions, but they need not all appear in one application. Assess grouping and division in light of technical relationships, unity of invention, release timing and development plans. The aim is to capture the core technical features, rather than simply increase the number of filings. Coordinating disclosure and filing Paper submissions and presentations, technical blogs, public repositories and product demonstrations can affect patentability, depending on what is disclosed and who can access it. Identify potential inventions and filing dates before disclosure. Korea's exceptions for certain prior disclosures have conditions and procedural requirements; equivalent treatment cannot be assumed abroad. If disclosure has occurred, preserve the material, date, disclosing party, audience and any access restrictions. Distinguish confidential collaboration from public disclosure, and separate the disclosed technology from improvements that remain confidential. Choosing between patent protection and trade secret protection depends on whether the technology can be observed externally, its commercial importance, the consequences of disclosure and the likelihood of independent development. Broad patent protection cannot be expected if the specification omits information necessary to implement the invention. Frequently asked questions about AI patent applications Can technology built on an open model be assessed for patent protection? Yes. Using an openly available model does not rule out patent assessment. Look for additional technical improvements in input processing, training, model combinations or inference. This does not mean the model's known architecture can be claimed as your own invention. Licensing terms and patent issues require separate review. Do good fine-tuning results make an invention patentable? Good results alone are not decisive. Identify specific differences in data preparation, training procedures or adapter selection. Assess whether the approach routinely applies a known fine-tuning method to new data or introduces a distinct mechanism addressing a technical difficulty. Can prompts be protected? Distinguish prompt wording from the processing methods used to construct and apply it. A specific combination of input analysis, context selection, tool specifications, response validation and retry control may warrant assessment. Particular wording alone is unlikely to support broad protection. Is building a RAG system enough to obtain a patent? Filing an application and meeting the requirements for grant are different matters. Identify changes to a conventional RAG architecture, such as retrieval based on document structure, version-validity checks, grounding verification or conditional follow-up retrieval. Are full source code and training datasets required? Full code and complete datasets are not required for every invention. The disclosure must nevertheless enable a person skilled in the art to implement it. The necessary detail depends on the data, labels, preprocessing, relationship to the model and processing conditions. Can I apply before the product is finished? A finished product is not a universal filing requirement. The technical concept must, however, be sufficiently developed to describe an implementable solution. An idea limited to goals and expected results may need further design or validation. The evidence required also depends on the field's predictability and the claimed effect. What do I do if the paper or code has already been made public? Promptly document what was disclosed, when, by whom and the intended filing countries. Assess whether Korea's exception for prior disclosure may apply and whether the required procedures can be met. Identify any improvements that remain undisclosed. Do not assume an application can be filed at any time after publication. If AI is used as a development tool, who is the inventor? Current Korean practice treats natural persons as inventors. Using an AI tool neither establishes nor rules out inventorship. Identify the individuals who contributed to the inventive work through problem definition, conception of specific solutions, revision and verification. Affiliation, seniority or merely giving instructions does not determine inventorship. Can I do business freely if I get a patent? Patentability and freedom to operate are separate questions. An improvement may be patentable while remaining subject to earlier rights. Before launch, assess freedom to operate in key markets and review the relevant agreements and licenses. Can multiple patents come from one AI service? Yes. A single AI service may contain distinct improvements in data processing, training, retrieval, inference optimization and device control. Assess each feature's independent technical significance and its relationship to the others, rather than dividing applications mechanically by function. Filing strategy should also account for later improvements and disclosure plans. How to Start an Invention Review You do not need to begin with legal drafting. Prepare technical information showing where existing methods fail, what you changed, input and output flows, comparative results and alternative implementations. Include relevant papers, public code, planned disclosure dates and the participants and roles in any collaboration. Potential AI inventions emerge from specific solutions developed in practice, rather than a service name. Examples include inputs reflecting data reliability, adaptive training, evidence-grounded generation, inference under device constraints and prediction linked to physical control. Comparing these features with prior art and developing enabling descriptions and claims are central to filing preparation. Read the Korean source This article reflects the information available when it was published. Contact us to discuss your circumstances. Discuss this topic ↗ All articles TALK TO IPLEX Discuss your IP questions We consider your technology and business needs together. ↗ Contact us Newer Australian Trademark Filing and Registration: Procedure, Timing and TM Headstart ↗ Older Non-Use Trademark Cancellation: Responses Illustrated by Cases ↗ Related insights AI & Software 2026.10.01 FiX: fine-grained forgetting in softmax attention Yongduck Kim examines FiX’s feature-wise gates, numerical implementation and paged cache, distinguishing reported gains from unresolved limitations. ↗ Read article AI & Software 2026.09.30 MHAR: Reading earlier layers through different feature subspaces Yongduck Kim examines Multi-Head Attention Residuals: depth routing, reported training results, implementation costs and the relationship between technical features and effects. ↗ Read article AI & Software 2026.09.28 Column: Claude Computer Use and the Data That Trains AI Agents Writing for AI Times, IPLEX Managing Partner Yongduck Kim examines the training data behind computer-operating AI agents through U.S. Patent No. 12,585,862. ↗ Read article

- https://www.iplexlaw.co.kr/en
- https://www.iplexlaw.co.kr/en/blog/category/ai
- https://www.iplexlaw.co.kr/forum/view/1538326
- https://www.iplexlaw.co.kr/en/contact
- https://www.iplexlaw.co.kr/en/blog
- https://www.iplexlaw.co.kr/en/contact
- https://www.iplexlaw.co.kr/en/blog/1538373
- https://www.iplexlaw.co.kr/en/blog/1538124
- https://www.iplexlaw.co.kr/en/blog/fix-fine-grained-forgetting-attention
- https://www.iplexlaw.co.kr/en/blog/fix-fine-grained-forgetting-attention
- https://www.iplexlaw.co.kr/en/blog/mhar-multi-head-attention-residuals
- https://www.iplexlaw.co.kr/en/blog/mhar-multi-head-attention-residuals
- https://www.iplexlaw.co.kr/en/blog/claude-computer-use-agent-patent
- https://www.iplexlaw.co.kr/en/blog/claude-computer-use-agent-patent