Abhay Verma*
Introduction
Imagine a situation where the same AI gives two different answers to the same patent question. Let’s understand this with a simple example. A person filed a patent application for an invention of a surgical instrument with various features, including a shaft that can be detached. An AI system, after reviewing the claim and analysing the prior art, concluded that the invention was not new. Now, when the same AI was used again with prompts to focus specifically on the detachable part, it concluded that the invention was novel. A similar example was used by the Office of the Controller General of Patents, Designs and Trade Marks (CGPDTM) in its recent release of “Guidelines for the Use of Artificial Intelligence in Patent Examination Procedures” to show why AI cannot be independently relied upon to decide the patentability of an invention. This example begs a bigger question: If AI can influence what an Examiner notices, searches for and considers, is having a human “in the loop” really enough?
The recently released guidelines mentioned that AI is meant only to assist with tasks such as patent searching, classification, translation and claim analysis. Therefore, the CGPDTM, through its guidelines, proposed a model in which AI will only assist, while Examiners will verify the output and make the final decision. This model puts human judgment at the core of the whole examination process, but the real difficulty begins when we look at what happens before the Examiner reaches the final decision.
From Human Oversight To Algorithmic Influence
The distinction between checking an answer given by AI and checking the process that helped AI arrive at that answer cannot be ignored. Suppose an Examiner took the help of AI for classification and search terms for a patent claim. AI identified fifty prior-art documents related to the claim, and after reviewing all the documents, the Examiner concluded that the invention is novel. But what if the most relevant document was never included among those fifty documents given by AI? So, in this situation, even though the Examiner verified all the AI-identified documents, the most relevant document was actually never discovered by AI.
The guidelines recognised this risk and defined it as “Classification and search drift”, where the AI will divert the Examiner into an altogether different technical field by suggesting an improper classification or search. This demonstrates that the problem is not only about giving wrong answers by the AI, but it is about diverting the Examiner’s approach right from the start. The research on AI-assisted patent prior-art searching, likewise, shows that AI can assist in identifying and retrieving relevant prior art. While the Guidelines do highlight the problem and require human review and validation, they do not prescribe a specific method for testing whether the AI’s own classification or search pathway has caused relevant prior art to be omitted. This led to a very important question: Is the problem really solved through the verification of output when the actual problem lies in the search path?
The CGPDTM, through these Guidelines, stresses that the AI cannot independently decide whether a patent should be granted. Although these Guidelines place the ultimate responsibility on the Examiner to verify the result, they say less about how the Examiner should verify what the AI may have missed in the first place. This raises the question of how any error can be corrected if it remains invisible to the Examiner.
Sections 12 and 13 of the Patents Act, 1970 make it clear that prior-art investigation is not a merely post-searching checking exercise. Section 12(1)(c) requires the Examiner to take into account the results of the investigation carried out under Section 13, and Section 13 requires the Examiner to carry out an examination as to whether the invention claimed has been anticipated by any previous publication or prior claim. The Delhi High Court in Novartis AG v. Natco Pharma Ltd. & Anr. held that Section 12(1)(c), read together with Section 13, makes it mandatory on the part of the Examiner to conduct such an examination. Now, if AI influences the classification, search terms or references from which that investigation begins, the issue is not only whether the Examiner correctly verified AI’s results, but whether the Examiner independently assessed the scope and adequacy of the investigation.
The statutory framework also matters because novelty is not determined merely by finding similarities between an invention and isolated prior-art documents. According to Section 2(1)(l), an invention is not new if it has been disclosed through previous publications or prior use. Therefore, the Examiner must assess whether the claimed invention is already disclosed in the prior art being relied upon. The Delhi High Court in FMC Corporation & Ors. v. Natco Pharma Ltd. held that a prior-art document can only defeat the claim of novelty when all the features of the invention are found in that document and that separate prior-art documents cannot be mosaiced to establish novelty. In Wirtgen GmbH v. Controller General of Patents, Designs and Trademarks & Ors., the Court stressed the need for proper comparison between the claims and the cited prior art documents instead of merely reproducing parts of the prior-art document. This becomes important in AI-assisted examination because if AI fails to identify a relevant document or misses an important claim feature, the Examiner may be assessing novelty against an incomplete set of prior art.
When AI Shapes The Search, Who Controls The Examination?
The detachable shaft example presented earlier illustrates that AI does not rely exclusively on the information given to it, but rather the way in which the question is framed affects what AI considers relevant. Research by Setchi and others similarly examines how AI can assist with search terms, retrieval and ranking of relevant prior-art documents. Recently, new research has also been conducted by Ikoma & Mitamura to test whether large language models could be capable of conducting a patent novelty search by comparing claims with prior-art documents, and how the current models are limited in this respect. In terms of the procedure used for evaluation, if the initial inputs given to the AI are too restrictive, then the Examiner will actually be evaluating an entirely different set of documents. This shows that even if the Examiner cross-checks all the answers generated by AI, such an examination will be insignificant if the Examiner has not independently considered the search terms, classifications and other inputs that shaped those results.
Now, when we look at the international community, patent searching through the use of AI technology is no longer a matter of discussion. For instance, WIPO’S PATENTSCOPE now includes an AI-assisted platform that converts natural language descriptions into a patent search format. Further, the United States Patent and Trademark Office has also experimented with its Artificial Intelligence Search Automated Pilot Program for a limited time to evaluate AI-assisted prior art searching. All these developments show that the question is no longer whether AI should have any role in patent searching because it already does. The more important question is the extent of control an Examiner has when AI helps determine where and what to search.
The Guidelines raise a question about independence. An independent review of analysis and documents provided by AI does not mean that the Examiner was independent in terms of where to search, what terms to use, or which technical area to look into. Now, if these choices are shaped by AI, one can argue that even though the final decision is made by a human, the entire path that leads to such a result is not controlled by a human.
The Indian patent jurisprudence also requires more than a human officer formally making the final decision. In Agriboard International LLC v. Deputy Controller of Patents and Designs, the Delhi High Court opined that a decision to reject a patent application involves application of mind and reasoned decision, including careful consideration of prior art, the invention in question, and how the skilled person would move from the former to the latter. More recently, in Keranova v. The Assistant Controller of Patents and Designs, the Court found that an order could not be sustained where the Controller failed to properly engage with the relevant prior-art material and did not substantiate how the cited documents rendered the claims obvious. The same requirement is reflected in F. Hoffmann-La Roche Ltd. v. Cipla Ltd., where the Delhi High Court set out a detailed five-step test for inventive step, including identification of the skilled person, the inventive concept, the differences between the prior art and invention, and whether those differences would have been obvious while ruling out hindsight. The recent decision in Indigenous Energy Storage Technologies Pvt. Ltd. v. Deputy Controller of Patents and Designs reaffirmed the principles given by the court in F. Hoffmann and stated the sequential use of the inventive-step test and the negation of hindsight. Therefore, the guidelines recognise the need for the independent application of mind of the Examiner, but they fail to address how that independence should be demonstrated when AI has already influenced the process. Further, the application of AI technology raises a new question: if the system influences the classification of the documents, their keywords, and the choice of prior art references from which the examiner begins his analysis, can verification of AI-generated results alone demonstrate the independent application of mind required by this jurisprudence?
From Human In The Loop To Human In Control
The above-mentioned criticism does not mean that the suggestion is made for the exclusion of AI in patent examination. Instead, there is no doubt that the process of examination can be made quicker with the help of AI. The main issue is the necessity to determine the clear boundaries of using AI. The idea of human verification works well as long as AI does some routine tasks, but where AI is used in the examination process, especially in critical stages like classification, prior art search, claim mapping and novelty, the independent role of the Examiner becomes crucial. Therefore, there is a clear need for the guidelines to move from simply verifying AI output to exercising sufficient control over the examination process by the Examiner.
For this new suggested approach, the Examiner must take control over the search process itself. This approach is also consistent with established examination practice in the United States, where the examiner is expected to identify the field of search, select the appropriate search tools, and determine the search strategy. The classifications, keywords, and prior art results obtained with the help of AI should be treated as the starting point and not as the limits of the examination. It is important that the Examiners should not restrict themselves to the results provided by AI. Instead, they should analyse the suggestions given by AI, conduct searches using different terms and also check whether any claim limitations have been omitted. This approach can help to solve the problem of omission.
It is important that the use of AI can be tracked. The guidelines specify that all material uses of AI that have substantive importance are to be recorded along with the name of the AI tool used, the way it was used, and the time when it was used. The guidelines also mention audits and impact assessments. These measures improve traceability, but they do not necessarily show how AI influenced the search or how the Examiner responded to that influence. Therefore, a next step is needed that maintains sufficient documentation of how the AI contributed to the task of examination and what decision the Examiner made independently about it. This will lead to a significant increase in human oversight.
To further strengthen human oversight, it is necessary to add a specific challenge step where the Examiner will ask about the omission made by the AI instead of simply verifying the result provided by the AI. Under this step, the Examiner can try the use of alternative classifications or search terms, check whether there are any claim limitations that are overlooked and look for any possible prior art that is not mentioned in the results provided by AI. Recent research on human–AI collaboration in patent novelty assessment also highlights the continuing role of human feedback and intervention in evaluating AI-assisted results. There is no doubt that the inclusion of all these recommendations will shift human oversight from simply checking AI’s work to actively testing its blind spots.
The Blind Spot In The Loop
There is no doubt that humans should be responsible for the final decision, and the Guidelines convey the same principle. But the main problem is what happens before the decision is taken by a human. Suppose AI helps an Examiner decide the classifications, search terms and prior-art documents. Now the Examiner may end up checking only the material provided by the AI, and the Examiner will be restricted to that path without knowing what lies outside it. Thus, the blind spot is that the Examiner will never get the chance to examine the material that AI had not identified as relevant.
The new guidelines are important because they clarify that AI cannot replace the Examiner and that AI-generated outputs require human verification. But the bigger question is whether the involvement of human beings is meaningful when the important parts of the path that led to the decision of examination are identified by AI. Another real test is whether the Examiner can question the AI’s starting points, look beyond its results, and independently decide what matters. The human may be in the loop, but the main challenge is to ensure that the human is not simply following the loop.
* The author is an LLM student pursuing Intellectual Property Rights (IPR) at Hidayatullah National Law University (HNLU), Raipur. The author may be contacted at abhay.261054@hnlu.ac.in.
The author is greatful to Ms. Apurva Sharma, Assistant Professor at Hidayatullah National Law University, for encouraging the author to explore and write about recent developments in Intellectual Property Law.
This blog reflects the personal views of the author and does not necessarily represent the views of The Policy Chronicle.