Artificial intelligence (AI), machine learning (ML), and deep learning (DL) have rapidly evolved from buzzwords to the backbone of India’s technological transformation. From AI-powered medical diagnostics revolutionizing healthcare to smart agriculture systems optimizing crop yields, these technologies are reshaping every sector of the Indian economy. This surge has positioned India as the fifth-largest jurisdiction globally for AI patent filings, with over 86,000[1] AI-related patents filed between 2010 and 2025, representing more than 25% of all technology patents in the country. However, securing patent protection for AI/ML/DL innovations presents a unique set of challenges that go far beyond traditional patent categories.
Patents grant exclusive rights only when an invention satisfies the statutory requirements of patentability being novelty, inventive step, and industrial application. For AI/ML/DL inventions, there is another critical consideration that has historically created significant barriers for innovators. Under Section 3(k) of the Indian Patents Act, 1970 (hereinafter referred to as “the Act”), the mathematical methods, algorithms, business methods, and computer programs “per se” are excluded from patentability. This exclusion represents the single most formidable challenge for AI/ML/DL patent applications in India.
Because AI/ML/DL technologies are fundamentally grounded in mathematical models, learning algorithms, statistical models, and computational methods, patent applications must transcend abstract algorithmic descriptions and demonstrate concrete technical effects through specific implementations that interact with hardware, systems, or real-world processes. The challenge has always been determining where the line exists between excluded abstract algorithms and patentable technical innovations.
The Office of the Controller General of Patents, Designs & Trade Marks (CGPDTM) published the Guidelines for Examination of Computer Related Inventions (CRIs), 2025 (hereinafter referred to as the CRI Guidelines 2025) on July 29, 2025, marking a transformative moment for AI innovation in India. These guidelines serve as the new examination standard for patent applications involving artificial intelligence, machine learning, deep learning, and related emerging technologies, fundamentally changing how these inventions are assessed for patentability. The guidelines codify the judicial focus on “technical effect,” “technical advancement,” or “technical contribution” as the primary criteria for overcoming Section 3(k) exclusions. For AI/ML/DL inventions, this means the focus shifts from whether an invention involves algorithms to whether it produces concrete implementable technical improvements or solutions.
The CRI Guidelines 2025 explicitly state that the mere presence of a computer program in a claim does not automatically render it ineligible if the subject matter demonstrates genuine technical advancement. This represents a monumental shift from the previous mechanical approach that often rejected AI inventions based on their algorithmic nature alone. The guidelines establish that abstract ideas including mathematical formulas and theoretical AI/ML/DL constructs remain unpatentable due to their lack of practical application. Patentability emerges when AI/ML/DL innovations successfully transform these abstract principles into tangible, real-world applications that demonstrate concrete technical effects. The CRI Guidelines 2025 also establish clear distinctions between different categories of AI-related inventions. AI-generated inventions, created autonomously with minimal human intervention, are explicitly non-patentable since artificial intelligence cannot satisfy the requirement of being the “true and first inventor” under Section 6 of the Act. However, AI-assisted inventions, where AI tools support human inventors, are not categorically excluded and can achieve patentability if they meet standard criteria and demonstrate technical effects.
Comprehensive Disclosure Framework
The guidelines provide clear disclosure requirements specifically designed for AI/ML/DL inventions i.e. the technical detail should enable reproduction of the claimed invention without undue experimentation by a person skilled in the art. Patent specifications must now provide comprehensive descriptions of training data nature and characteristics, neural network architectures and topologies, algorithm types and optimization techniques, and performance metrics with concrete validation evidence.
For training data, the applicant must provide explicit identification of data nature, clear correlation between data characteristics and the technical problem solved, and quantified evidence of data quality. For model architecture, comprehensive descriptions of neural network topology, detailed algorithm explanations, and mathematical foundations underlying novel approaches are required. For preprocessing requirements, explicit documentation of all data pipeline steps, clear correlation between raw input and processed learning data, and sufficient detail enabling reproduction of preprocessing algorithms must be provided. For reinforcement learning implementations, detailed specifications of agent-environment interactions, comprehensive descriptions of states and action spaces, explicit reward function definitions, and policy learning mechanisms are mandatory. For novel algorithm disclosure, complete mathematical formulations, computational step documentation, and algorithmic innovation details with benchmarking against established baselines are required.
The guidelines offer various examples to demonstrate how AI/ML/DL inventions can achieve patentability through proper technical disclosure and demonstration of technical effects. Some of them are as follows:
Healthcare Applications: Facial Recognition Methodology
Consider an invention for improving facial recognition accuracy for elderly individuals in the healthcare and safety applications. In this scenario, a facial recognition model is trained on a diverse dataset of individual aged 65-90, with various aging features such as wrinkles and loose skin, as well as common occlusions. The dataset also accounts for variations in lighting, poses and accessories such as glasses.
For technical implementation, the invention utilizes a modified FaceNet architecture, a pre-trained deep convolutional neural network model enhanced with an age aware embedding loss. The invention demonstrates patentability by achieving X% recognition accuracy for elderly individually compared to Y% (X>Y) for a baseline model trained on a general dataset. The invention thus provides reliable person identification in elderly care environments, reduces false negative in safety monitoring and enables automated fall detection and health event tracking.
For sufficiency, the applicant must disclose dataset traits, neural network architecture (FaceNet Variant) and quantitative comparison with baseline models showing material improvement. The applicant must also provide explanation why generic face datasets would not suffice the requirement and how specific dataset traits impact the performance of the model.
Industrial Applications: Predictive Maintenance
Consider an invention for predicting wind turbine failures through multivariate time-series sensor data. In this scenario, LSTM (Long Short-Term Memory) neural networks process time series sensor data. The data including measured vibration, temperature and speed collected at a certain time intervals. The input provided is the sensor logs from the turbines which includes data such as wind speed, direction and power output.
For technical implementation, the invention employs LSTM architectures for time-series sensor data, preprocessing algorithms that filter environmental noise while preserving failure-indicative signal patterns, learning approaches that incorporate failure patterns from multiple turbines. The invention generates failure predictions for industrial machines.
For sufficiency, the applicant must disclose the LSTM architecture specifications, sensor preprocessing methodologies including data labelling methodologies, and quantitative performance metrics comparing the multi-variate approach against traditional threshold-based systems. The applicant can also provide explanation of why conventional single-sensor monitoring would not suffice and how the integration of multivariate time-series sensor data impacts failure prediction accuracy.
Agricultural Applications: Precision Farming Revolution
Consider an invention for classifying crop types using satellite imagery and deep learning. In this scenario, raw satellite data undergoes multiple preprocessing stages prior to classification. The preprocessing pipeline applies atmospheric correction algorithms using Sen2Cor to normalize environmental variations, calculates Normalized Difference Vegetation Index (NDVI) to enhance vegetation signatures, and performs image segmentation based on vegetation indices and texture filters. The processed images are then classified using a CNN (EfficientNet) trained on labeled plots with crop type annotations, where the correlation between raw data and processed learning data directly impacts classification performance.
For technical implementation, the invention employs a CNN architecture (EfficientNet) with specific implementational parameters and their inter-relationships disclosed, preprocessing pipelines that include atmospheric correction (Sen2Cor algorithm), NDVI calculation methodologies, and segmentation logic based on vegetation indices and texture analysis. The preprocessing stages correlate to the end model by transforming raw satellite data into processed learning data optimized for crop type classification. The training dataset encompasses labelled plots with defined characteristics relevant to the preprocessing challenges, with specific labelling methodology and volume disclosed. The invention demonstrates patentability by achieving X% classification accuracy, with preprocessing contributing to a Y% performance boost compared to using raw imagery. For sufficiency, the applicant must disclose the step-by-step preprocessing pipeline including atmospheric correction (naming the Sen2Cor algorithm), NDVI calculation, and segmentation logic. The applicant must also disclose the different implementational parameters and their inter-relationship used in the CNN model structure (EfficientNet), the training dataset characteristics that are relevant to the preprocessing challenges including labelling methodology and volume. The applicant can also provide a clear explanation of how pre-processed data improves classification performance over raw imagery, along with a comparative performance benchmark which shows the material effect of preprocessing on classification accuracy.
The CRI Guidelines 2025 mark a transformative moment for AI/ML/DL patent protection in India. By establishing clear technical effect frameworks, comprehensive disclosure requirements, and systematic examination methodologies, these guidelines provide the roadmap the Indian patent community has long needed. The emphasis on substance over form, technical contribution over algorithmic classification, and real-world impact over theoretical advancement aligns India’s patent examination practices with AI innovation realities.
Success in this new landscape requires embracing the guidelines’ fundamental message: AI/ML/DL inventions achieve patentability not by avoiding their algorithmic nature, but by demonstrating genuine technical contributions that solve real-world problems. The guidelines provide the framework, but the challenge now lies in using it effectively to secure meaningful protection for AI innovations that will shape India’s technological future.
