The CGPDTM issued new guidelines on August 7, this time to regulate AI use in the patent examination process. In a 16-page document followed by annexures of illustrative examples and a checklist for the officers, the Guidelines for the Use of Artificial Intelligence in Patent Examination Procedures seek to maintain a balance between harnessing the benefits of using AI tools in patent examination while ensuring that their use does not compromise the technical, statutory, and quasi-judicial functions of the Office.
Credit where it’s due, the Guidelines are clearly aware of the risks of using AI for something as technical as patent examination and adopt an extremely cautionary approach when prescribing AI use by the officers. The document is replete with caveats about the need for human oversight over AI-generated outputs, while also identifying the tasks for which officers may and may not use AI. Some of these uses promise significant efficiency gains, particularly in identifying classifications and claim-feature extraction. My apprehension arises when the Guidelines contemplate the use of AI for tasks such as preliminary novelty and obviousness assessments, due to the cognitive influence that polished and seemingly authoritative AI-generated outputs may exert on human judgment.
In this two-part post, I’ll focus on the typical uses of AI in patent examination as identified by the Guidelines and whether the caveat of “human oversight” is cognitively sufficient in part I, and will ask what happens institutionally when the human oversight fails in part II.
Dear AI, Please Guide Me!
The Guidelines identify 12 typical uses of AI in examination, which I have categorized into 4 heads for analytical convenience.
- Patent Search and Prior-Art Identification
- Identifying candidate IPC/CPC classifications
- Generating search terms and concept clusters
- Prior-art search using officially subscribed AI tools
- Claim and Patentability Analysis
- Preliminary claim-feature extraction
- Preliminary novelty or inventive-step analysis
- Quick identification of clarity issues in claims
- Preliminary assessment of sufficiency of disclosure
- Drafting, Language and Legal Research Support
- Translation support
- Improving the structure or language of a draft Office communication
- Generating legal or technical citations
- Understanding case law and related legal or technical concepts
- Confidentiality and Risk
- Prohibition against the use of a public generative AI tool with unpublished patent application material
As mentioned above, all these uses come with a plethora of cautions and caveats, along with illustrations, clearly showcasing how unmonitored AI usage may eventually lead to an erroneous conclusion. The overall language across the document has been suggestive, and it carefully guides the officers as to how they can use AI in performing the above functions– ranging from search to examination to internal communication.
As can be seen, some of these uses are qualitatively different from some of the tasks grouped under “Claim and Patentability Analysis.” While the uses clubbed under “Patent Search and Prior-Art Identification” and “Drafting, Language and Legal Research Support” largely use AI to reduce the time spent on preparatory, organisational, and claim-feature extraction, tasks such as novelty and inventive-step analysis, and assessing sufficiency of disclosure go much closer to the substantive evaluative functions at the heart of patent examination. The efficiency gains here therefore come with a different kind of trade-off: once AI is used to frame or structure these inquiries, its output may cognitively influence the very human judgment which the officers are supposed to apply to come to their final decision.
Human in the Loop, AI in the Mind
Starting with the use of AI for Novelty and Inventive Step. The Guidelines state that using AI platforms may help with preliminary claim mapping and structured comparison. It cautions that each mapped feature must be independently verified against the actual prior-art document, but ultimately the final call on novelty or inventive step will be based on the officer’s own analysis and application of mind. This is because the Guidelines themselves identify some pressing concerns like misreading the cited documents and hindsight bias.
Despite flagging these concerns and explaining the use of AI through illustrations, the Guidelines state that such uses can be seen as an assistive input, and the examiner or controller should verify each mapped feature from the actual prior-art document. The problem, however, is not that such verification should somehow take place before the AI output is seen. Rather, it is whether verification conducted after exposure to an AI-generated mapping is sufficient to preserve an independent novelty assessment. Now, let’s take a step back and think. A claim-to-prior-art mapping is not a neutral exercise, and the way a claim is broken down and mapped against the prior art can influence how the officer conducts the novelty inquiry. The Guidelines seek to address this problem by requiring every mapped feature to be verified against the actual prior-art document and leaving the conclusion to the officer’s independent analysis. But verification takes place after the examiner has already been presented with an AI-generated analytical frame. So once the AI has proposed that a feature corresponds to prior art A, for instance, the examiner may approach the prior-art document by asking whether that proposed correspondence can be sustained, rather than independently asking whether the reference, read as a whole, discloses the claimed feature at all. In that sense, verification may confirm or reject individual mappings, but the examiner may still be influenced by the way the AI has flagged prior art A as relevant for the novelty assessment in the first place. A suitable alternative/ clarification to address this issue could have been the guidelines clearly specifying the sequence of reference, with making an initial independent review of the prior art mandatory before consulting the AI-generated mapping.
This tension goes a step further for the obviousness assessment because, unlike the novelty assessment (which ordinarily checks whether the invention is disclosed in one prior art), here the invention is seen as a whole, and its obviousness is tested by the person skilled in the art. The exercise is therefore not simply one of locating corresponding claim features, but of assessing the significance of the differences between the claimed invention and the prior art without reconstructing the invention with the benefit of hindsight. Over the years, different interpretations of the Patent Act have reiterated that while assessing obviousness, there should be no hindsight analysis during the obviousness inquiry. However, for an AI platform which is devised for similarity, pattern matching, and retrieval (as explained from the illustrations in the Guidelines), there is a risk that the output generated might be affected by this problem.
Another “black box” type problem stems from allowing the use of AI in novelty and obviousness assessment. If the examiner uses AI for preliminary claim mapping and the output then influences the subsequent examination, the applicant may not know how the mapping was done. And if the platform misses a crucial feature, the error may inadvertently become a part of the Examiner’s “reasoning”, making the guideline’s emphasis on “application of mind” incomplete as a safeguard.
As one can see here, the bigger problem, therefore, is that despite the safeguards stating that examiners and controllers need to cross-verify everything, allowing AI into novelty and obviousness assessment raises a larger tension of whether human verification can adequately neutralise the cognitive influence of AI-generated outputs. That too, after the Guidelines themselves acknowledging that use of AI may introduce issues such as hindsight bias and mosaicing of prior art stemming from that bias in the examination.
Moving to the sufficiency of disclosure. The Guidelines state that AI platforms can be used in identifying possible gaps in enablement or fair basis. This use is prone to risks such as over-reading, under-reading, and giving false confidence to the examiners in complex matters. To overcome this, the Guidelines suggest that the complete specification, drawings, examples, and common general knowledge be examined manually; the officer independently assess whether any alleged gap is real, material, and relevant; and any final sufficiency objection be framed only after independent technical and legal assessment. This poses a different kind of issue. AI may be useful in identifying gaps in the specification, but the identification of gaps does not make the disclosure insufficient. Whether an omission is legally material depends on the description, the scope of the claims, the common general knowledge of the person skilled in the art, and the extent of experimentation required. By allowing AI to identify potential “enablement gaps”, there is a risk that the AI platform may flag the missing information as insufficient, triggering the examiner to undertake the sufficiency inquiry from a ‘deficit’ point of view, rather than approaching the disclosure independently.
In part II of the post, I’ll touch on the institutional checks and balances that are needed in the aftermath of these guidelines.
[Thanks to Swaraj and Bharathwaj for the inputs on the post.]
