In Part I of the post, I discussed the contents of the Guidelines and how using AI may cognitively affect the officers while examining a patent application.
From Personal Satisfaction to Institutional Accountability: Where does the Buck Stop?
Now, I understand that some of the arguments in Part I are targeted towards the idea of using AI in patent examination and not the Guidelines per se. However, despite all the caveats for human oversight, the Guidelines state that the concerned officer needs to be “personally satisfied”/ “personally check” the output. It does not mention the corrective steps that the Office would undertake if there were lapses on the part of one of the officers while using these platforms. It is understandable that with the AI boom, many professionals and officers would turn to AI tools to ease their workload. This is particularly relevant in the context of the Patent Office, where concerns about overburdened examiners and Controllers are well known, with targets being set for them to meet within a specific timeframe (see this fantastic set of posts by Bharathwaj here, here, here, and here). With the seeming permission to use AI assistance in performing crucial tasks, it is only natural to drift towards a tendency for using them to quickly meet these targets. It is here that proper mechanisms and checks should be in place within the system.
Presently, the prescribed administrative measures in the Guidelines state that the officers may be required to record the use of AI in specified functions, stating the name and nature of the tool, date of use, and any other information necessary for supervision, audit, or quality review, without stating the rectifications to which the Controller General resort to, to fix any serious instances of… let’s say hallucinations. In those cases, would the applicant be supposed to file an appeal before the High Court to get the necessary remedy? And what if it’s the other way round and an applicant gets a patent based on hallucinated reasoning? Then the question is – who is supposed to file an appeal in that case? This is the same set of issues that we have seen in the context of orders being unreasoned (see here, here, here, and here). However, with the possibility of having AI platforms contribute to the reasoning (despite all the Guidelines’ caveats against it) makes this situation even more challenging because the source of the error may itself be difficult to identify. An applicant may be able to point out that a conclusion in an order is unsupported by the prior art or the specification, but may have no way of knowing whether that error originated in the officer’s own analysis, in an AI-generated mapping or summary, or in the officer’s reliance on an inaccurate AI output. This becomes particularly important where the Guidelines do not require the use of AI to be disclosed in the FER or the final order.
In other words, recording that AI was used is useful for audit purposes, but an internal audit trail without a corresponding remedial framework only tells how the error may have occurred, but it does not tell us what the institution will do about it. Perhaps the administrative-measures portion of the Guidelines should therefore go one step further and expressly provide for mechanisms to detect, review and rectify material AI-related errors before the ordinary appellate or revocation mechanism has to be invoked.
Private AI, Public Office, Little Transparency
The concerns around accountability and transparency get amplified when one realises the silence on the names of the platforms which are being used by the Office. The Guidelines have disclosed that the examples in the Annexure were prepared using public Generative AI tools, and certain private tools are already accessible to the Patent Office; however, the Guidelines state that “Since such tools are proprietary in nature, they have not been specifically identified in these Guidelines.” This does not make sense. A platform can be proprietary while its name is publicly known, and the Guidelines seem to be conflating the disclosure of the tool’s identity and the disclosure of proprietary information about the tool.
More importantly, it is necessary to understand how these AI platforms were procured and what safeguards govern their use. On a quick look at the CGPDTM website, I wasn’t able to find any call for bids for AI platforms specifically. But working on the assumption that instead of issuing a specific tender for AI platforms, the Office could have bundled it with a tender for patent-database subscriptions, I was able to find a Request for Proposals (RFP) for URL/IP based Patent Literature Database and Search Engine (pdf). Furthermore, in an RTI response shared by a friend of the blog (pdf), it was stated that subscriptions for Clarivate’s Derwent World Patents Index (DWPI) & DesignVision Database were taken by the Patent Office for 1 year from April 4, 2026 to April 5, 2027. It is worth pointing out that Clarivate is a British-American analytics company.
This raises a larger data-governance question, i.e., where does the information supplied to these tools actually go? The Guidelines rightly prohibit officers from uploading confidential and unpublished patent material to unapproved public Generative AI tools. But there are other tech sovereignty-focused concerns here. Merely saying that certain “private” tools may be used does not tell whether prompts and outputs are stored, where they are processed, how long they are retained, whether third-party providers or sub-processors have access to them, whether they may be used for improving or training models, or what contractual restrictions the Patent Office has imposed on such use. (On why these concerns are important, interested folks can check out this SpicyIP TV podcast episode with Dr. Kailash Nadh.)
A Hallucination in the Definition of Hallucination?
Another thing worth pointing out here is that the definition of hallucination seems to have a typo. Explaining Hallucination or Fabrication, the Guidelines state:-
”The AI tools may present a confident but unsupported statement, citation, explanation, or mapping. Such hallucinations may be extrinsic or intrinsic. An intrinsic hallucination occurs when the AI tool generates information that is not present in the provided context, thereby fabricating facts to fill a knowledge gap. An intrinsic hallucination occurs where the AI tool is given specific source documents or context in the prompt, but its output contradicts, distorts or misrepresents that provided information.”
If read closely, the explanation seems to mix up the concepts of intrinsic and extrinsic hallucinations with the statements contradicting each other by using the same term for different phenomena. As explained by Elyes Hajji et al, hallucination is extrinsic when the information is not present in the context, thereby nudging the AI model to fabricate information. Whereas hallucination is intrinsic when the context has the information, but the AI model misrepresents or contradicts what the source says.
Conclusion
The Guidelines in themselves are a welcome step attempting to bring structure and caution in the Patent Office’s use of AI in the patent examination process, but the “human oversight” caveats cannot be sufficient to respond to the different risks that follow from introducing AI into the examination process. The vigilance towards AI in patent law is rightfully justified considering that the integration of AI is not only limited to being used as a tool but it can also unsettle the fundamental understanding of crucial concepts such as Person skilled in the Art (See here for Roberto Dini’s argument on this and also check out Krishna Jani and Tanya Aithani’s post in response to Dini’s Dilemma), [Sidenote: This is also something that was discussed by Murali Neelakantan, Swaraj, Adarsh, Shivam, and me during the Summer School 2026 and we are looking to examine this in more detail beyond these Guidelines.] If AI is to become part of this regime (which it seems like it already has), the next step must be to build stronger institutional safeguards around its use, particularly on disclosure, auditability, corrective mechanisms, and accountability, when that oversight fails.
Thanks to Swaraj and Bharathwaj for the inputs on the post.
