AI in explained: this guide covers what it means, who it applies to, the step-by-step process, documents required, fees, due dates and penalties in India — so you can stay compliant with confidence and avoid costly mistakes.
The Patent Office has published guidelines on how its own examiners and Controllers may use artificial intelligence tools in examination. They are not about whether an invention that uses AI can be patented; that question is dealt with in our CRI Guidelines series. This article explains the first part of the guidelines: their scope, the terms they define, the fourteen risks they list, and the rule of human oversight. It also explains what an applicant may reasonably expect, and question, when an office action shows signs of AI assistance.
The Guidelines, as published by the Patent Office, are the Office's guidance and do not have the force of law; the Patents Act, 1970 and the Patents Rules, 2003 as now in force prevail. The Office revises its guidelines, so check the current version on ipindia.gov.in.
The Guidelines allow AI as an aid in tasks such as screening, classification, search, translation support, drafting support and technical comparison. They list fourteen risks, from hallucinated citations to bias and confidentiality, and say the examiner or Controller stays fully responsible for every official act. For an applicant, the practical result is that an office action must rest on verified prior art and the officer's own reasoning, never on an unverified AI output.
Objective and scope (paragraphs 1.2 and 1.3)
The Guidelines say they are meant to guide and regulate AI use in a way that supports efficiency and quality while preserving confidentiality, accountability, consistency and the independent application of mind by the examiner or Controller (paragraph 1.2). Their scope covers screening, classification, search, translation support, drafting support, technical comparison and knowledge retrieval (paragraph 1.3). The introduction explains why: patent examination is document-heavy, technically complex and time-sensitive, and AI tools can help with search terms and ranking documents, while the statutory and quasi-judicial functions of the examiner and Controller must not be compromised. It adds that the examples in the annexure were prepared with the help of public and private tools and that, because such tools are proprietary, none is named.
The terms the Guidelines define (paragraph 2)
| Term | What the Guidelines mean |
|---|---|
| AI and generative AI | Systems that produce predictions, recommendations, classifications, summaries or new content from input data and prompts |
| AI prompt | The text, command or query given to the model |
| AI tool | Software that uses such models to perform tasks, classed as public or private |
| Public AI tool | A tool open to the public, whether the version without charge or a paid one; training on broad public data makes it versatile but may limit depth and reliability, and data security may differ between versions |
| Private AI tool | A closed system with controlled access and data, suited to sensitive and proprietary information |
| AI-assisted output | Output produced with a human in the loop who gives direction and oversight, as opposed to fully automated output |
| Prompt engineering | Framing prompts to give the model context and instructions |
The distinction between public and private tools matters because confidentiality of an unpublished application is the first concern. Our article on permitted uses and prohibited uses explains how the Office applies it.
The fourteen risks (paragraph 3.1)
The Guidelines list the common ways AI can go wrong, state that the list is not exhaustive, and say these risks are why thorough human oversight is needed. In TaxClue's grouping:
- Input-driven risks. Profile sensitivity (output shaped by earlier use), prompt discipline (output depends on how the prompt is framed, so prompts must be clear, precise and limited to the purpose).
- Reliability risks. Hallucination or fabrication, including non-existent prior art or judgments; omission of critical detail such as a limiting feature, amendment or disclaimer; false pattern matching on surface similarity; loss of technical context in summaries or translations.
- Drift risks. Classification and search drift; over-breadth or over-narrowness of concepts.
- Transparency and trust risks. Black-box opacity; misguidance by output that is fluent but unsound; confidentiality risk from sharing unpublished material with an external tool.
- Data risks. Incomplete or outdated data, with no real-time access to all patent and non-patent literature; bias in training data and results.
- Language risk. Difficulty with complex patent language, drawings, flowcharts and nuances such as claim construction, enablement and inventive step.
On hallucination the Guidelines go further than naming the risk. Any resource an AI tool cites must be cross-checked from an authorised source, and where AI is used on cited material, its output must be compared with the original, especially for judgments, since AI may misread or overstate a principle.
If you are answering an office action and want its citations and reasoning checked point by point, our patent objection reply service can do that review.
Human oversight (paragraph 3.2)
The Guidelines say the Indian policy material on AI "repeatedly emphasises the need for human oversight, verification and accountability", and state that AI is "intended to assist, and not replace, the functions or the works of the Examiner or Controller" (paragraph 3.2). Any use of AI that may influence search or examination is subject to manual review and validation. The examiner or Controller remains fully responsible for every official act involving AI in any way, and that responsibility is neither diluted nor transferred. An AI-assisted output may be used only after the officer is personally satisfied of its correctness, relevance and appropriateness.
What an applicant can expect, and question
The Guidelines regulate the Office's conduct; they do not create a right of action or a new ground of reply. Used with care, they help an applicant frame a response.
- Check each cited document. If an objection cites a document you cannot find, or describes it in a way that does not match its text, say so in the reply and ask for a copy. The Guidelines themselves treat fabricated citations as a known risk.
- Read the mapping against your claim. If the objection relies on keyword similarity, point out the technical or legal difference; the Guidelines name false pattern matching as a known risk.
- Look for omitted features. If the objection overlooks a limitation, an amendment or a disclaimer, quote it.
- Ask for the officer's reasoning. The Guidelines say the officer must be personally satisfied; a hearing is the place to ask the Controller to explain the reasoning in his own terms.
- Do not assume AI was used. An office action need not say whether AI was used, and a reply should rest on the substance, not on speculation.
For the process these points fit into, see the examiner's report and first examination report and the search for anticipation and novelty.
Checklist table
| Risk in the Guidelines | What it can look like in an office action | What the applicant can do |
|---|---|---|
| Hallucination | A cited document that does not exist or says something else | Ask for a copy; show the mismatch |
| Omission of critical detail | An objection that ignores an amendment or limiting feature | Quote the omitted feature |
| False pattern matching | Prior art chosen on keywords alone | Set out the technical difference |
| Misguidance | Fluent but unsound reasoning | Answer each step; ask for a hearing |
| Loss of context in translation | A foreign document read inaccurately | Provide a certified translation |
A worked example
Corvid Analytics Pvt Ltd receives an objection that cites a foreign document as disclosing a "feedback valve" in its claim 1. Corvid's agent pulls the document and finds that it describes a pressure sensor, not a valve, and says nothing about the temperature limit in the claim. The reply sets out the difference line by line, quotes the limiting feature the objection left out, and asks for a hearing. The Controller considers the document afresh and withdraws the objection.
Common lapses
- Accepting a citation without reading the document.
- Replying with general denial instead of specific differences.
- Treating the Guidelines as a ground of appeal in themselves.
- Assuming an office action is AI-generated because it is fluent.
- Sharing unpublished material with an outside AI tool while preparing a response.
Need help with an office action?
An objection is only as good as its citations and its reasoning. If you want an office action tested against the documents it cites, speak to our patent objection reply team.
Key takeaways
- The Guidelines cover the Office's own use of AI in examination tasks.
- Fourteen risks are listed, from hallucination to bias and confidentiality.
- The examiner or Controller stays fully responsible for every official act.
- Applicants should verify citations, mappings and omitted features.
- For patenting AI inventions, see the CRI Guidelines articles.
Read next
- AI in patent examination: permitted uses, safeguards and prohibited uses
- AI in patent examination: Annexure I examples
- Chapter 9: examiner's report, FER, reply and hearing
- Chapter 9: search for anticipation and novelty
Disclaimer: Based on the manuals and guidelines published by the Office of the Controller General of Patents, Designs and Trade Marks that are named in the article, as consulted on 4 October 2026. They are guidance and do not have the force of law; the Patents Act, 1970 and the Patents Rules, 2003 as amended (including the 2024 amendment rules) prevail, and the current versions on ipindia.gov.in should be checked. This article is general information, not legal advice; check the official text before acting.