Are AI and Machine Learning Inventions Patentable in India? (2025 Guide)

Yes: AI and machine learning inventions are patentable in India, but only when the claims demonstrate a concrete technical effect, not merely an algorithmic advance. In India, the key barrier is Section 3(k) of the Patents Act, 1970, which excludes algorithms and computer programs “per se.” The CRI Guidelines 2025 (effective 29 July 2025) now provide a structured three-step test that determines whether a claim clears this bar. The short rule: if your invention produces a measurable technical improvement on a real-world system, it can be patented. If it produces only a better informational output, it cannot.

India has emerged as one of the fastest-growing jurisdictions globally for AI-related patent filings, with WIPO data indicating tens of thousands of AI-related applications filed over the past decade. Yet for every application that progresses to grant, many more stall, rejected on subject-matter grounds that applicants could have anticipated at the drafting stage. The fundamental question is not whether AI or machine learning (ML) technology is inventive, but whether the claims are framed around a concrete technical contribution rather than an abstract algorithm. This guide explains precisely how India evaluates AI and ML inventions under its 2025 examination framework, what the new CRI Guidelines require at each step, and, through worked examples, how claim drafting strategy determines the outcome. It also provides a concise comparison with USPTO and EPO standards for practitioners coordinating cross-border filings.

Section 5 covers PCT filing considerations for applicants seeking parallel protection in the US and Europe. India’s CRI Guidelines 2025, issued by the CGPDTM on 29 July 2025, represent the most significant shift in Indian AI patent examination practice in a decade. The guidelines are published on the Indian Patent Office portal (ipindia.gov.in). Read alongside the USPTO’s August 2025 AI Subject Matter Eligibility Memorandum and the EPO’s updated Guidelines effective 1 April 2025, a coherent multi-jurisdictional strategy is now achievable, provided the specification is drafted to support it.

1. The Core Legal Problem: Why AI Inventions Get Rejected

At every major patent office, the same structural issue arises. An AI or ML invention, at its core, involves a mathematical model, namely a neural network, a classification algorithm, or a gradient-descent optimisation, and mathematical methods, taken alone, are excluded from patent protection everywhere. The disagreement between jurisdictions is not whether to exclude pure algorithms, but how much technical context is needed to lift a claim out of the excluded category. Getting that balance right in the specification, before filing, is the single most impactful decision in AI patent prosecution.

Three practical consequences follow. First, a specification drafted purely around model architecture, training procedure, and accuracy metrics will face rejection in all three jurisdictions. Second, the same underlying invention, such as an ML model that reduces diagnostic error in medical imaging, can be claimed in a way that passes or fails subject-matter examination depending on how it is presented. Third, because specification quality at the time of filing determines the scope of protection available in India, investment in drafting precision before filing is more effective than attempting to broaden claims during prosecution.

2. India: Section 3(k) and the Technical Effect Doctrine

Under Section 3(k) of the Patents Act, 1970, mathematical methods, algorithms, business methods, and computer programs per se are excluded from patentability. The word “per se” is the operative qualifier. Courts and the Indian Patent Office (IPO) have consistently held that an invention is not excluded merely because it involves a computer program or algorithm; the exclusion applies where the claim is directed to nothing more than the algorithm itself.

2.1  The “Technical Effect” Test: How It Works

The judicial doctrine of technical effect , developed through a line of Delhi High Court decisions, has now been codified in the CRI Guidelines 2025 (effective 29 July 2025). The guidelines establish a structured three-step analysis for AI and ML claims:

1. Understand the invention as a whole. Do not assess individual components in isolation. Read the claim, specification, and drawings together.
2. Determine whether a technical effect is present. A technical effect is a concrete, reproducible improvement in a technical system, such as reduced computational load, improved signal fidelity, faster processing, measurably higher diagnostic accuracy, or lower memory consumption. Abstract improvements in information content alone do not qualify.
3. Rule on patentability accordingly. If a genuine technical effect is present, Section 3(k) does not apply, and examination proceeds on novelty and inventive step.

The CRI Guidelines 2025 explicitly state that AI-generated inventions, those created autonomously by AI with minimal human contribution, remain non-patentable because the inventorship requirement under Section 6 of the Patents Act cannot be satisfied. Only human inventors, or human inventors assisted by AI tools, can file.

2.2  Worked Example: Medical Imaging Diagnostic System

Practical scenario: An AI system uses a convolutional neural network (CNN) to detect early-stage pulmonary nodules in CT scans, reducing radiologist review time by 40% and false-negative rates by 18%.

Claim framed around the algorithm (excluded): “A method comprising: receiving a CT image dataset; applying a convolutional neural network trained on labelled nodule data; classifying regions of interest as nodule-present or nodule-absent.”

This claim describes the algorithm and its output classification. It produces an informational result, a diagnosis, and not a technical effect on a technical system. Under the CRI Guidelines 2025 three-step test, Step 2 fails. The IPO would issue a Section 3(k) rejection.

Claim reframed around technical contribution (patentable pathway): “A computer-implemented system for processing CT image data comprising: a processor configured to execute a CNN architecture optimised to reduce false-negative detection of pulmonary nodules having a diameter below 6 mm, wherein execution of the CNN reduces GPU memory allocation by at least 30% relative to a standard ResNet-50 baseline by applying a multi-scale feature pyramid with shared weight initialisation, the system producing a spatial heat map overlaid on the reconstructed CT volume for direct radiologist interface.”

This version anchors the claim in measurable technical effects: GPU memory reduction, specific architectural choices, and a technical output (spatial heat map integrated into the imaging system). In practice, quantitative benchmarking in the specification strengthens technical-effect arguments at examination and provides a defensible basis for distinguishing prior art during prosecution.

2.3  Key Judicial Developments

Several Delhi High Court decisions from 2023–2024 have shaped the current examination standard under CRI Guidelines 2025:

• Ferid Allani v. Union of India (2019, DHC) : Established that software-implemented inventions providing a technical effect are not excluded under Section 3(k). Remains the foundational authority.
• Microsoft Technology Licensing v. Assistant Controller (2023, DHC) : Reinforced that the IPO must evaluate claims as a whole and cannot reject on the sole basis that the claimed invention includes a computer program.
• Comviva Technologies v. Assistant Controller (2024, DHC) : Held that a business method authentication system incorporating technical security features was not automatically excluded; the court directed re-examination on technical merits. Signals the IPO must look beyond surface categorisation.
• Google v. Controller of Patents (2024, DHC, decided 2 April 2024) : The court set aside an IPO rejection and directed grant, finding that the invention demonstrated technical advancement over prior art. Reinforces that demonstrated technical advancement, not algorithmic novelty alone, is the operative test.

For the full CRI examination framework that underpins these judicial decisions, see CRI Guidelines for Computer Related Inventions in India. On the separate question of whether AI can be named as an inventor, see Can an AI be Recognised as an Inventor?.

3. How India Compares: USPTO and EPO at a Glance

India’s Section 3(k) framework is not unique in excluding bare algorithms, but it differs from the USPTO and EPO in how it measures the threshold for eligibility. At the USPTO, AI patent eligibility is governed by 35 U.S.C. Section 101 and the two-step Alice/Mayo test. The operative question is whether the claim integrates an abstract idea into a practical application, typically by improving a technological process or the functioning of a computer system. The USPTO’s August 2025 AI Subject Matter Eligibility Memorandum reinforces that claims referencing neural network training do not automatically recite an abstract idea, but claims that name specific algorithms such as backpropagation or gradient descent do, and require further analysis.

At the EPO, patentability is governed by Articles 52 to 57 of the European Patent Convention. The EPO’s updated Guidelines G-II, 3.3.1 (effective 1 April 2025) confirm that an AI or ML claim directed to a method using technical means, such as a computer, has technical character as a whole and is not excluded under Article 52(2) or (3). The real challenge at the EPO is inventive step: under the COMVIK approach, only technically characterised features count toward the inventive step assessment, which means an ML algorithm that merely classifies information contributes nothing inventive beyond running on a computer. The comparison table below summarises the key differences practitioners should account for when coordinating India filings with US or European prosecution.

For a deeper comparative analysis of Indian and European patent law, see Comparative Study of European and Indian Patent Laws.

4. Jurisdiction Comparison Table

The table below summarises the key patentability framework, exclusion basis, and claim-drafting priority for each jurisdiction covered in this guide.

Factor	India (IPO)	USA (USPTO)	EPO	Key Difference
Legal basis	Section 3(k), Patents Act 1970	35 U.S.C. § 101; Alice/Mayo	Art. 52(2)(3) EPC; COMVIK	India: statutory exclusion. US/EPO: case-law doctrine
What is excluded	Algorithms, math methods, computer programs per se	Abstract ideas, laws of nature, natural phenomena	Math methods, computer programs as such	Functionally similar but India uses “per se” qualifier
Core test	Technical effect / technical contribution (CRI Guidelines 2025)	Integration into practical application (Step 2A Prong Two)	Technical character + COMVIK inventive step	EPO has two hurdles; India/US effectively one decisive test
Key 2025 update	CRI Guidelines 2025 (29 Jul 2025): codifies technical effect; explicit AI/ML examples	Aug 2025 Memo: narrows “mental process”; clarifies neural-network claim analysis	Guidelines G-II 3.3.1 (1 Apr 2025): confirms AI/ML technical character if computer-implemented	India and USPTO updates are substantive; EPO update clarificatory
Claim drafting priority	Embed quantitative technical effects; compare to prior-art system baselines	Describe post-inference technical action on a system; avoid reciting specific algorithms by name	Disclose architecture, training data, and reproducible technical effect across claim scope	EPO requires broadest enablement; India rewards quantification
AI-generated inventions	Non-patentable; human inventorship required (Section 6)	Non-patentable; human inventorship required; AI-assisted ✓	Non-patentable; human inventorship required	Uniform across all three jurisdictions

5. Filing Strategy: PCT and Multi-Jurisdictional Considerations

For applicants seeking protection in India alongside the US and EPO, the PCT route provides a practical framework. The Indian Patent Office functions as a Receiving Office, ISA, and IPEA, giving applicants flexibility in ISA selection. Under the Patents (Amendment) Rules 2024, effective 15 March 2024, the Request for Examination (Form 18) deadline for PCT national phase applications in India coincides with the 31-month national phase entry deadline. Both must be met simultaneously. Restoration is generally unavailable under current practice, making deadline compliance non-negotiable.

From a specification standpoint, a PCT application that quantifies technical effects for India, describes post-inference technical action for the USPTO, and discloses training data and architecture in sufficient detail for the EPO will support prosecution at all three offices without requiring supplemental disclosure after filing. The claim drafting principles set out in Sections 2 and 3 above apply equally whether the route is direct national filing or PCT national phase entry.

6. Common Pitfalls and How to Avoid Them

The following patterns consistently cause rejection or narrowing across all three jurisdictions.

#	Pitfall	What Goes Wrong	Fix
1	Claiming the algorithm alone	Fails Section 3(k) in India; ineligible at Step 2A Prong One at USPTO; excluded “as such” at EPO	Add post-inference technical action on a system; embed quantitative technical-effect claim language
2	Novelty of domain as substitute for technical contribution	Applying ML to a new business domain does not create a patentable claim (Recentive, 2025)	Focus on how the invention improves a technological system, not what new domain it addresses
3	Vague specification: no architecture or training data details	EPO Article 83 sufficiency objection; USPTO §112 enablement rejection; India examiner finds no technical contribution	Disclose architecture, training dataset characteristics, and reproducible performance benchmarks in the specification
4	Missing the 31-month deadline for India national phase	Application treated as withdrawn; restoration is generally unavailable under current practice	Diarise 31-month deadline from earliest priority date; file Form 18 simultaneously (post March 2024 rules)
5	Claiming AI-generated inventions without human inventorship	Non-patentable in India (Section 6), US, and EPO; uniformly rejected at all three offices	Identify and document human inventive contributions; AI tools may assist but humans must conceive the invention

7. Conclusion

AI and ML inventions are patentable in India, the US, and Europe, but only when the claims are rooted in demonstrable technical contributions, not algorithmic novelty alone. The CRI Guidelines 2025 have materially clarified India’s position: the IPO now applies a structured three-step technical-effect analysis, drawing on the same judicial doctrine that the Delhi High Court has developed over the last decade. At the USPTO, the August 2025 Memorandum sharpens the mental-process boundary and reinforces that the operative question is whether ML output is integrated into a practical technical application. At the EPO, updated Guidelines G-II 3.3.1 confirm that AI/ML claims implemented on a computer have technical character; the inventive-step fight over COMVIK then determines grant.

For foreign applicants entering the Indian national phase on AI inventions, the most important decision is made before PCT filing: a specification that quantifies technical effects, discloses architecture and training data in reproducible detail, and claims post-inference technical action will survive examination across all three offices. Specification quality, ISA strategy, and deadline discipline, specifically the convergence of the national phase entry and Form 18 RFE at 31 months from priority under the 2024 amended Rules, are the three levers that determine outcome. Practitioners who embed the technical effect doctrine into drafting workflow before filing will encounter far fewer Section 3(k) objections in Indian prosecution.

Legal Disclaimer

This post is for informational purposes only and does not constitute legal advice. Statutory references reflect publicly available information as of the date noted above. For advice specific to your patent filing situation, consult a registered patent agent or attorney qualified in the relevant jurisdiction.

Frequently Asked Questions

Q1. Is a neural network itself patentable in India?

Not in isolation. A claim directed solely to a neural network architecture, without specifying a technical problem it solves or a measurable technical effect it achieves, will be excluded under Section 3(k) as a mathematical method or algorithm per se. To achieve grant, the neural network must be claimed as part of a technical system that produces a concrete, reproducible technical effect, such as reduced memory consumption, improved signal fidelity, or measurably higher diagnostic accuracy compared to a prior-art baseline.

Q2. Does using AI to generate an invention affect patentability in India?

Yes. Under Section 6 of the Patents Act, 1970, only a human inventor, or their legal representative, can file a patent application. Inventions created entirely by AI without meaningful human inventive contribution cannot be patented in India. Inventions where a human uses AI as a tool, while retaining inventive control over the conception, remain patentable, provided all other requirements are met.

Q3. How does the CRI Guidelines 2025 change examination practice compared to the 2017 guidelines?

The 2017 CRI Guidelines applied a relatively mechanical test that often led to rejections based on the presence of a computer program, without adequate analysis of technical contribution. The CRI Guidelines 2025, effective 29 July 2025, introduce a structured three-step test, expressly address AI, ML, blockchain, IoT, and quantum computing as potentially patentable domains, provide illustrative examples drawn from recent judicial decisions, and codify the technical-effect doctrine that courts had developed since Ferid Allani (2019). Examination is now required to assess the invention as a whole, not to identify algorithmic elements in isolation.

Q4. Can the same PCT specification support AI patent applications in India, the US, and the EPO?

Yes, if drafted to the highest common denominator. A specification that discloses architecture and training data in sufficient detail for EPO sufficiency purposes, quantifies technical effects for India’s technical-contribution test, and describes post-inference technical action for USPTO practical-application analysis will generally support examination at all three offices. Claim sets will typically need jurisdiction-specific adaptation during national phase prosecution, but a well-drafted PCT specification avoids having to supplement disclosure after filing.

Q5. What is the deadline risk for AI PCT applications entering the India national phase?

Under the Patents (Amendment) Rules 2024, effective 15 March 2024, the national phase entry deadline and the Request for Examination (Form 18) deadline both fall at 31 months from the PCT priority date, and both must be filed simultaneously. There is no option to enter the national phase and defer examination. Missing the 31-month deadline causes the application to be treated as withdrawn; restoration is generally unavailable under current practice. Practitioners should diarise the deadline immediately upon international filing and confirm entity classification before entry, as fees differ between large entity, startup, and small entity applicants.

Q6. Is the EPO’s AI/ML examination approach stricter than India’s after the 2025 guideline updates?

The EPO applies two sequential tests: technical character (first hurdle, usually cleared for computer-implemented AI claims) and inventive step using the COMVIK problem-solution approach (second hurdle, where most applications face objection). India’s CRI Guidelines 2025 apply a single structured test centred on technical effect. In practice, the EPO’s inventive-step analysis is the most technically demanding of the three jurisdictions because the COMVIK approach systematically strips non-technical features from the inventive step assessment. The USPTO’s Alice/Mayo test sits between the two: less demanding than the EPO on inventive step, but requiring clear integration of the abstract idea into a practical technical application.