Home About Services Jurisdictions Insights FAQs Contact Us Client Login Terms & Conditions Privacy Policy Get a Quote →
Home › Insights › AI-Assisted Examination of Patents in India: Accountability, Confidentiality and Justice after the 2026 CGPDTM Guidelines
11/17/2026 12:13 PM

AI-Assisted Examination of Patents in India: Accountability, Confidentiality and Justice after the 2026 CGPDTM Guidelines

Introduction : Discussions about artificial intelligence (AI) and patents typically involve whether a creation made by an AI can be patented. However, a more pressing question is related to the way in which the Indian Patent Office will make use of this technology in its decision-making procedures. The Office of the Controller General of Patents, Designs and Trade Marks (CGPDTM) has issued Guidelines for the Use of Artificial Intelligence in Patent Examination Procedures (Guidelines) on 7 August 2026, with the aim of maintaining the technical examination function of examiners preserving the independence of the decisions made by the Controller and quasi-judicial authorities.

According to the experts, AI can be effectively used for suggesting keywords and filtering the documents; but for examination, it ends in decisions that impact proprietary rights. This blog suggests that the main problem is an accountability gap. AI can have a role in an examination, but the officer remains responsible for it and the applicant cannot easily verify how far AI was involved. The Guidelines are good on principle, but weak on implementation. It should embed transparency within the examination process.

Legal Provisions

A. The Patents Act, 1970 and the Patents Rules, 2003

The Act differentiates between the Examiner and the Controller. An application is scrutinized in detail by an examiner after a request for examination has been made. An examiner checks on the statement of compliance, legality of objections to issuing patents, and conducts a search for prior art.  Referrals, however, come after publication as per the Rules. The Controller then reviews the report and takes action. The examiner’s report may not be open to public inspection or published.6 However, this is not the case with the first statement of objections issued by the Controller to the applicant together with any other prescribed documents following Rule 24B(3). As per section 13(4), a report does not equate to validation and no liability attaches to the Central Government or its officer.

If there are any objections in the report, section 14 mandates the Controller to share the gist of information with the applicant and offer a chance for a hearing if requested by the applicant in the stipulated time frame. Section 15 provides authority to reject the application. Sections 80 and Rule 129 require a hearing before any discretion is exercised adversely to the applicant. An appeal against appealable orders lies to the High Court under section 117A.

B. The 2026 Guidelines

The Guidelines are administrative in nature. The Guidelines state that AI will help but will not replace the Examiner or the Controller, and that any AI data influencing the examination must be verified manually, and that the officer is ultimately responsible for every official act. The guidelines prohibit unpublished applications, confidential information and materials of the deliberative process from processed using public AI applications, substituting AI for the officer’s application of mind on substantive matters, and citing AI-suggested references without independent verification. Recording of AI use, however, is something the competent authority “may prescribe,” and the officer declaration in Annexure-II is described as a “proposed” module.

AI Assited Patent

Legal Analysis

A. Can AI-assisted work influence examination?

The Guidelines’ own illustrations suggest that it will be possible due to prompt. For example, in novelty examination with one claim and one prior art reference it was established that if the prompt said that the parts of D1’s shaft should be treated as connected, there will be no novelty, and if the prompt said that it should focus on the detachable-attachment produced a finding of novelty.

It is possible for examiners to express a tentative view in the language of the question they pose. No less a danger to independent thought exists in this insidious anchoring of the mind than in outright delegation, which is not accepted. In a case in which an inventive step is refused, the Controller has the duty to perform an analysis of prior art, focus on the nature of the invention in question and how that invention would be obvious to those skilled in the art. A conclusion that is generated by AI and simply taken as one’s own would not meet such a requirement regardless of how well it is expressed.

B. Confidentiality of applicant material

One of the most powerful safeguards of the Guidelines is confidentiality. Section 144 protects the examiner’s report from public inspection and publication. That report must be distinguished from the first statement of objections issued to the applicant under Rule 24B(3).20 Applicants thus receive the grounds for seeking a response, but not necessarily the examiner’s own report, or the manner in which AI search and analysis was conducted. Because examination ordinarily follows publication, the specification is usually public by then; what remains protected rests on different provisions, namely officers’ duty of non-disclosure and secrecy directions, under which applications are not published while the direction operates. The prohibition on feeding such material to public tools tracks these duties. The weaker side is the “private” AI built into a subscribed search database, which the Guidelines label as more secure,  but doesn’t identify who provides these tools, whether the search includes certain types of queries, or whether vendors can re-use databases. They simply ask if there was any exclusion of confidential material; no questions regarding the exclusion of certain kinds of material are asked on the officer checklist.

C. Accountability for erroneous output

The Guidelines place full responsibility on the officer.26 Section 13(4) substantially restricts liability arising from the statutory examination process. The Guidelines leave no room for the applicant to seek relief from an adverse decision save for the statutory appeal and judicial review. However, the provision should not be interpreted as covering all claims that may be made as a result of unauthorised actions by the individual, for example, misuse of confidential information.

There are two types of error that need to be distinguished. The ones that are more common are fabricated legal citations. While the case was about judicial precedent and not patent examination, with the tribunal relying on non-existent precedents, the court in Pooja Ramesh Singh held that, fabricating paragraphs attributed to real precedents, and the Supreme Court’s holding that such a ruling was not a decision in law, even if the material had only an indirect relevance,  is relevant by analogy to patent examination. If the adverse patent decision is based on a fabricated citation, or on a source that was inaccurately represented, then the defect may affect the basis of the reasoning, the specific consequences of which can be evaluated under the Patents Act and the reasoning principles of the quasi-judicial decision. Even the Guidelines’ own annexure documents an AI system providing “verbatim” quotations from the order of a tribunal and only after multiple requests.

D. Procedural safeguards for applicants

Sections 14 and 80 and Rule 129 provides right of hearing for the applicant, and requests a section 14 hearing within prescribed time, but a section 117A appeal can be brought only against appealable orders.30 None of these provisions says whether AI contributed to an objection.Guidelines only stipulate that the machine translation will be documented in the reasoning for patentability;  other recordings are optional, disclosure to stakeholders is possible.

The gap is therefore threefold: AI can affect classification, search terms, claim extraction and initial comparison, leaving the officer responsible, and the applicant unable to easily gauge the extent of AI involvement. Prompt access to prompts, or internal deliberations, does not have to be the answer. The answer need not be unrestricted access to prompts or internal deliberations.

Recording the material AI assistance used for audit and provide to the applicant if it is a material basis of an adverse objection should suffice. This is reflected in the quasi-judicial authorities’ obligation to provide reasons which serves to promote transparency and the principle of justice appearing to be done.

E. Is human oversight enough?

The case for AI use is not trivial. While the Guidelines recognize that examination is document intensive, complex and time sensitive, oversight must be carried out and cannot be self-executing. The Guidelines themselves note that misguidance is hardest to detect when output is fluent, structured and persuasive. As held by the Delhi High Court, AI-driven tools cannot replace human intelligence in adjudication and is at best a tool for preliminary understanding. Oversight should be accompanied by recorded usage, verified sources and reasoned orders rather than being simply declared.

Relevant Case Laws

Pooja Ramesh Singh v. Jammu & Kashmir Bank Ltd. (Supreme Court, 2026): The Court set-aside orders passed by NCLT and NCLAT, relying on non-existent judgments/AI-generated paragraphs attributed to genuine judgments. It stated that it had a zero-tolerance policy with regard to the “creation of precedent” by AI tools that have not been verified and that AI cannot replace human adjudication, but did not preclude the legitimate use of AI. It was an analog application as it was concerned with judicial precedent.

Christian Louboutin SAS v. Shoe Boutique-Shutiq (Delhi High Court, 2023): In a passing-off suit, the plaintiffs relied on a ChatGPT query to establish their reputation. The Court declared that such an instrument may not be used to resolve legal or factual questions, as there might be “fictional cases” and “false responses”. The case does not provide a precedent for patent examination, but rather is a case for verification and human judgment.

Agriboard International LLC v. Deputy Controller of Patents & Designs (Delhi High Court, 2022): The Controller’s refusal under section 2(1)(ja) was a was a cursory order which merely reiterates the objection and fails to consider the objections raised by the applicant. It was set aside and remanded, the Court finding that the application of mind and the giving of a reasoned order are aspects of natural justice and the three elements of an inventive-step analysis should be discussed.

Kranti Associates (P) Ltd. v. Masood Ahmed Khan (Supreme Court, 2010): The Court summarised the law on reasons: quasi-judicial authorities are required to record reasons on their findings and all decisions in administrative cases, which affect anyone prejudicially ordinarily require them.

Practical Implications

The Guidelines provide vocabulary for patent agents and applicants for evaluating scrutinising First Statements of Objections. Consider an objection that refers to a claim feature in D1 when D1 does not mention it. In the Guidelines’ own mapping exercise, the handle and the detachable-attachment feature were not clearly disclosed in the cited passage. The applicant can state that in the exercise of the cited feature, the feature was not clearly described in the Guidelines, and ask for the feature to be re-described at the section 14 hearing at which D1 disclosures the feature; and, if the order is not well reasoned, challenge the order under section 117A.

For the Patent Office, some of the changes are administrative and already considered with respect to use of the AI tools: recording of the tool, nature of use and date of AI use, audit through the AI Governance Committee, and training of officers. Others require rule-making: a duty to disclose material AI assistance to applicants would require amendment of the Patents Rules through section 159. The Supreme Court has observed that AI governance ultimately requires public policy and enforceable rules and regulations, not judicial declarations alone.

Conclusion

The Guidelines rightly do not allow AI to replace the Examiner’s/Controller’s decision making process and the confidentiality rule is consistent with the Patents Act. The problem with them is their procedure. The recording is voluntary, the declaration is only proposed, applicants have no entitlement to know whether AI shaped an objection, and section 13(4) puts the accountability largely to appeals.

Four reforms follow. Administratively, the Patent Office may require the recording of material AI use in the examination file, and announce data-handling conditions for AI incorporated into databases subscribed by the Patent Office. It can mandate disclosures to the applicant regarding adverse objections based on AI’s material input, per the rule, and provide the Guidelines with legal support. This can be continued through the work of the AI Governance Committee, which has the power to review errors and revise the Guidelines, but so far, human oversight is an assurance, not a right that can be claimed by applicants.

Author :- Aastha Agarwal, in case of any query, contact us at Global Patent Filing or write back us via email at support@globalpatentfiling.com.

References

  1. Off. of the Controller Gen. of Patents, Designs & Trade Marks, Guidelines for the Use of Artificial Intelligence in Patent Examination Procedures § 1.1 (Aug. 7, 2026), https://ipindia.gov.in/dynamic/news-details/119 [hereinafter Guidelines].
  2. Rossitza Setchi et al., Artificial Intelligence for Patent Prior Art Searching, 64 World Pat. Info. 102021 (2021), https://doi.org/10.1016/j.wpi.2021.102021.
  3. The Patents Act, No. 39 of 1970, India Code (1970), §§ 11B(1), 12(1)–(2), 13(1)–(2).
  4. The Patents Rules, 2003, S.O. 493(E), Gazette of India, pt. II sec. 3(ii) (May 2, 2003) (as amended), r. 24B(2)(i).
  5. Patents Act, § 14.
  6. Patents Act, § 144.
  7. Patents Rules, r. 24B(3).
  8. Patents Act, § 13(4).
  9. Patents Act, § 14.
  10. Patents Act, § 15.
  11. Patents Act, § 80; Patents Rules, r. 129.
  12. Patents Act, § 117A.
  13. Guidelines, § 3.2.
  14. Agriboard Int’l LLC v. Deputy Controller of Patents & Designs, (2022) 290 DLT 315, ¶¶ 25–27 (Del.).
  15. Pooja Ramesh Singh, 2026 SCC OnLine SC 1258, ¶ 5.
← Back to All Insights