Patent Disputes Over AI-Assisted Product Features In Consumer Apps

Introduction

The technology of eliminating digital blurriness is called an anti-blur function. The function helps improve photo quality before images are saved through the camera application. For example, the online shopping application helps determine what a current customer is likely to buy in the future. Furthermore, the wellness application uses sensors to create personalized notifications for the user. The payments application is used to warn the user in case an illegal transaction is detected. Generally, all four of them share one key aspect. Each of them is regarded as an “AI feature” by their producers. The legal issues raised by the current situation include: what kind of data is collected, what steps of processing are performed, what kind of hardware is used, and what is the result.

The distinction is essential in India. The startup cannot escape the legal responsibility only due to the use of a different algorithm. Thus, the idea cannot be patented. According to section 3(k) of the Patents Act of 1970, it is stated that any patented abstract idea cannot be regarded as a patent potential.

According to the information presented in Indian jurisdiction reports, there is currently no substantial body of case law dealing with infringement claims in relation to the characteristics of AI in the consumer applications studied in this article. Nevertheless, some indications can be extracted from CRI cases and from patent disputes that are based mainly on software. The conclusion to be drawn is that the resolution of the issues will depend more on the actual technical implementation and wording of the claims rather than on "AI-powered" applications being involved.

The Legal Filter: Technical Application, Not AI Branding

The AI, machine learning, and deep learning technologies are specifically covered by the Indian Patent Office’s CRI Guidelines 2025. Even though the CRI Guidelines recognize that a practical and working technical solution can be eligible for patents, the CRI Guidelines view the creation of abstract models and learning algorithms as theoretical and mathematical. Furthermore, there should be enough disclosure such that the specification identifies relevant data features, inputs and output mappings, model and training methodology, as well as technical reproducibility and validation.² These Guidelines lay down the procedure for examining patent claims of AI technologies, but the guidelines are not an alternative to the Act.

This approach has been adopted from the judgment of the Delhi High Court in Ferid Allani v. Union of India³, which held that "technical effect or technical contribution" needs to be established in respect of an invention involving computers. Since cookies, authentications, and networking communications involved more than just an interface, but were related to computer networking itself, an innovation related to network authentication in Microsoft Technology Licensing v. Assistant Controller of Patents & Designs⁴ did not fall under Section 3(k).

In case of consumer applications, two BlackBerry cases determine the line between a computer program per se and an invention. The Delhi High Court, while considering allegations that involved assigning "confidence levels" to media items, selecting media items to make space on the device and updating a list using the cache manager in 2024, dismissed the Section 3(k) challenge since the alleged combination improved storage, retrieval, synchronizing, and functionality of the device.⁵

However, in April 2026, the same High Court upheld the rejection of cases made by people who received a message from a portable device whose messages were coloured. While there was no increase in the functionality of the hardware or the performance of the system, not getting the wrong person was an advantage to the user. The same is true for the case of Kroll Information Assurance v. Controller General of Patents, where the claim was taken as a whole and found to be fundamentally a computer program or an algorithm, despite mentioning a processor, memory, and network.

Thus, the "software versus hardware" distinction does not provide the answer. What matters is the choice between a sufficiently claimed technological solution (such as latency reduction, decreased memory usage, increased bandwidth, better computer security, optimized sensors, or direct control of a gadget) and an abstract criterion or convenience for the end-user.

Claim Construction Will Decide the Dispute

Claim development is the initial stage in patent infringement. As per the Delhi High Court's interpretation in F. Hoffmann-La Roche Ltd. v. Cipla Ltd.⁷, claim terms generally carry the meaning they have to a skilled person on the relevant date. Any ambiguity can be addressed through the specification, though the analysis has to be claim-based. Contrary to the usual practice of comparing the accused product with that manufactured by the patentee, the accused product will now be compared with the patent claim.

Since several pipelines can generate equivalent outputs, such a practice is essential when dealing with features of AI. A feature vector created based on data from a sensor, a certain confidence threshold, an on-device cache, and a post-inference stage can be indispensable to a claim. An application using server-side lookup to reach the same decision lacks an important limitation. However, replacing the original neural network with an equivalent one will not always preclude a claim that is formulated at the architectural level.

The same component-wise approach has to be applied to the evidence. Briefs on infringement or non-infringement have to link claims to the appropriate product or process according to the Delhi High Court Rules for Patent Suits.⁸ Examples of useful evidence for an application include versioned builds, architecture diagrams, API traces, model cards, SDK documentation, code snippets, test logs, and telemetry. While such documents may point out a feature, marketing claims hardly cover all technical steps.

 A further layer emerges due to cloud architecture, where data may be collected by a smartphone, inferred by a vendor's server, and presented or used by the startup. In Indian law, there is little development regarding the divided performance of software process claims. Consequently, it would be incorrect to assume that using an external API means eliminating liability, and instead it is necessary to match each claim element to its actor and location.

The Software-Hardware Overlap

The demand for new hardware has been strictly refused in Indian courts. In the case of Raytheon Co. v. Controller General of Patents & Designs, the Delhi High Court held that "the patent office should consider the technological contribution or effect" and there was no justification for new hardware. But incorporation of conventional hardware in decorative language cannot be achieved. There will be no way in which a general telephone, processing unit or memory would make any abstract rule technically specific, as is seen from the cases involving the BlackBerry and Kroll inventions.

A similar question of substance over form is shown in the important telecommunications decision in Lava International Ltd. v. Telefonaktiebolaget LM Ericsson¹⁰. There was a dispute over physical realization of an encoder but the claims were studied along with the whole specification and a speech encoding patent was declared invalid for its lack of originality. An appropriate patent position for AI app patent claims is somewhere between just claiming the mathematical logic of the model and connecting the mathematical logic to an anonymous "processor."

Defensive Drafting for AI-Enabled Features

Before releasing the concept publicly, defensive drafting should begin as the engineers can try alternatives and articulate the invention while still having the possibility of doing so. First, identify the technical problem and baseline. For example, the existing camera pipeline may be problematic in specific lighting conditions, utilize a certain memory space, or produce specific latency. Secondly, specify the specific process causing the improvement, not just the desired result.

Secondly, show the whole implementation process chain. It includes data collection, preprocessing, relevant architecture of the model, training/calibration, inference, post-processing, and consequent actions of the device/network. To prove the inventive effect, the 2025 Guidelines specifically focus on architecture, dataset characteristics, training method, and validation. Disclosures should enable the claimed invention without undue experimentation, though there is no requirement to disclose a sensitive dataset.

Third, consider deployments that are expected. The disclosure should consider interactions with sensors, memory, cache, processor, and networks; consider embodiments in devices, edge and cloud; and provide some fallbacks. It may be appropriate to file claims in the method, system, and device categories, though prohibited substances cannot be remedied by changing the claim category. Further, claims need to have sufficient technical detail so as to survive a 3(k) and sufficiency challenge without having an overly specific model name that can be substituted by a competitor.

Last, add important limitations to any filing. Under Section 59, only small modifications are possible; claims must remain within the scope and additional matter cannot be included. Narrowing or clarifying amendments that have been supported in the original disclosure have been permitted by the Delhi High Court; however, this provision does not mean that a deficient disclosure can be rebuilt during examination.

How Startups Should Manage Infringement Risk

"Freedom-to-operate," or FTO, assessment should be considered by the company as an aspect of the product rather than a legal review at the last minute. It is critical to consider the existing claims, prosecution history, ownership, time period, and status in India. The title and abstract of patents cannot be used as filters. Even pending Indian petitions need to be considered since rights can be extended based on post-publication activity under the Act.¹¹

Two queries need to be distinguished while reviewing. Firstly, is each and every element of a valid claim being performed by the said feature, whether directly or through its legally relevant equivalent? Secondly, can the claim be challenged under Section 3(k), originality, inventive step, sufficiency, or some other head? A bad patent may continue to incur legal expenses until it is narrowed, invalidated or expired, which means that it is different from having no patent at all. Statutory grounds for revocation can also be availed in an infringement suit under Section 107.¹²

The startup could change or skip a critical step, change the technology process, seek a license, have a reasoned non-infringement or invalidity opinion, or disable the function if the risk appears genuine. Design arounds should be documented relative to the claim chart. The contracts between vendors should be reviewed closely as well. As an illustration, an AI API or SDK licensing deal could provide for limited liability, no patent indemnification, damage cap, or coverage only for the AI model and not the model-app-device combination of the startup.

A Product-Feature Patent Review Process

Before design freeze and again before release, each material AI feature should pass the following gate:

1. Feature intake: Create a data-flow diagram and one-page description that includes the client, server, model, sensors, storage, outputs, vendors, and user/device activities.
2. Search and triage: Find pertinent patent families, owners, expiration dates, and legal status by searching granted and published Indian claims around the technological mechanism rather than just the product category.
3. Claim map: Determine the limitations of each significant independent claim. Connect each restriction to the intended execution, the accountable party, the performance site, and the accessible data.
4. Dual risk score: Evaluate company exposure, validity strength, and infringement likelihood independently. Note presumptions and unsolved technical details.
5. Determine whether to clear the feature, design around a limitation, obtain a license or opinion, or postpone the launch. Assign an owner and a deadline for each mitigation.
6. Release record and monitoring: Maintain the approved claim chart, architecture version, and build identifier; reopen the review whenever the model, vendor, data flow, deployment location, or feature logic change; and keep an eye on high-risk pending applications. 

Conclusion

"AI" is not viewed as a patent classification by the Indian courts. Instead, the question they ask is if the accused product performs the important steps of the claim and if the feature described in the claim shows any technical implementation as a whole. The distinction between recipient colour coding and media selection optimized for storage exemplifies why customer value alone is not enough.

It is, therefore, far more useful for startups to ask, "What does the feature make the computer, network, or device do—and which patent claims describe that mechanism?" than "Does this feature use AI?" It is much cheaper to answer the former question upfront in the development process than the latter during an infringement letter.

Endnotes

^{1 The Patents Act, No. 39 of 1970, § 3(k) (India).}

^{2 Office of the Controller Gen. of Patents, Designs & Trade Marks, Ministry of Com. & Indus., Gov’t of India, Revised Guidelines for Examination of Computer Related Inventions (CRIs), 2025 (2025).}

^{3 Ferid Allani v. Union of India, 2019 SCC OnLine Del 11867.}

^{4 Microsoft Technology Licensing, LLC v. Assistant Controller of Patents & Designs, 2023 SCC OnLine Del 2772.}

^{5 BlackBerry Ltd. v. Assistant Controller of Patents & Designs, 2024 SCC OnLine Del 6142.}

^{6 Kroll Information Assurance v. Controller General of Patents, 2025 SCC OnLine Del 4614.}

^{7 F. Hoffmann-La Roche Ltd. v. Cipla Ltd., 2015 SCC OnLine Del 13619.}

^{8 High Court of Delhi Rules Governing Patent Suits, 2022, Notif. No. 29/Rules/DHC (notified Feb. 24, 2022).}

^{9 Raytheon Co. v. Controller General of Patents & Designs, 2023 SCC OnLine Del 5770.}

^{10 Lava International Ltd. v. Telefonaktiebolaget LM Ericsson, 2024 SCC OnLine Del 2497.}

^{11 The Patents Act, 1970, S 10(4)(c), No. 39, Acts of Parliament, 1970 (India).}

^{12 The Patents Act, 1970, S 107(1), No. 39, Acts of Parliament, 1970 (India).}

https://chatgpt.com/s/m_6a7aea512564819198592f1f813c5095

The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.