GenAI in Government Policy Drafting: From Production to Verification

Generative artificial intelligence has become one of the leading usages of artificial intelligence systems by a larger base of audience. While there are already enough instances to highlight the use of GenAI across the education or art sector, it is no longer a distant vision of seeing its usage in governance systems. There have already been documented instances of usage of popular platforms such as ChatGPT by Washington City officials and their use by private consultants such as Deloitte. The problem is not very far away from India either as studies have highlighted the use of AI generated text in government documents. While the IT ministry has already brought in rules for photorealistic synthetic media (Deep, 2026), we are hence not far away from the labeling of AI generated text, something that is discussed in the given scenario.

The use of the AI system to summarize research and to generate policy language is a classic example of the use of economics of cheap generation, which reduces the cost of generation, but doesn't increase the verification capacity, hence causing the governance problem of a verification bottleneck.

The economics of cheap generation is the primary economic mechanism over here. With the use of the Generative AI systems in the pilot project, it is seen that while there might be an initial capital cost of setting up such GenAI tools, after the initial cost, the marginal cost that is the cost of generating additional information, approaches a near-zero value. As there is little cost of usage of GenAI tools after their creation, this effectively means tasks such as summarizing academic research, generating draft policy briefs and recommending regulatory languages, all of which was done by government officials previously, can be done at virtually no extra cost, and at a significantly lower time because of the new tools. All of the aforementioned tasks which would require human input and potential cost of hiring such human labor is no longer relevant as the GenAI tools can do it at a much cheaper price. Hence, the cost of generating content itself is very cheap, and the mechanism is of cheap generation.

It must be highlighted that one could also see delegated execution as another economic mechanism over here. It must be highlighted that summarizing research and drafting are largely in the domain of generative content, hence, it cannot be the primary, but rather only a secondary economic mechanism, and beyond the ambit of this policy brief.

In the absence of cheap generation, the initial governance problem was that of producing information or producing the governance drafts, or having to go through academic research. Now, with the advent of GenAI and the economics of cheap generation, the essential governance problem shifts from producing information to verifying the information. Initially, the cost of production of information was typically high as it would require detailed research and writing skills, which require comprehensive training and would require a few trained professionals to do the same. With the advent of cheap generation, the cost of information production has gone significantly down, enabling organizations to produce information at the humungous scale that was not envisioned before. While the cost as well as the scale of generation has changed, the capacity of the organization to verify the generated content has not necessarily changed. Initially, there was some congruence between the generated output and the verification capacity. Now, as the generated output has exponentially increased, the verification has become scarce, and the capacity is insufficient, hence causing a verification bottleneck. 

Hence, the question now is, who verifies these automated outputs? How can the verification capacity be increased? Could AI even be trusted with the verification of content that it itself has generated, with flaws in it as seen in the given case study. There is also a philosophical problem of epistemic externalities, particularly as cabinet notes and other decision-making instruments, if they are victim to the GenAI hallucination, then it can significantly erode the quality of decision environments, and subsequently, the decision itself.

For such complex problems, the use of Hood's NATO framework provides guidance on designing policy instruments. The use of nodality or information based tools is particularly suitable in the given case. The proposed tool is to ensure that the GenAI tool in itself is built on government documents and government data and only provides content that is developed from this data. Only a government approved database of research papers can be a part of this training data set. This ensures that the risk of hallucination is lower in the tool. Complimentary to this, there must be a proper audit requirement, which must not be restricted to the algorithm, but to the usage as a whole, requiring labeling of content as AI-generated.

The basic mechanism is that of having auditing requirements. By restricting the database of the GenAI tool, the organization ensures that there would be lesser chances of hallucination, and reduces any potential information asymmetry. It brings about the need for verification, or usage of discretion before processing the given information. To illustrate, the problems of ghost citations or references to laws that don't exist will be infrequent occurrence as the data on which the GenAI tool primarily works would be an accurate data set, hence moving the verification requirement to ex ante instead of post generation. The auditing requirement of flagging the content highlights that for data of lower priority, such as internal notes, the requirement for verification could be lower, while high priority sensitive content, such as the cabinet notes would have a higher requirement for verifications. Hence, the limited verification capacity is matched as per the priority of verification. verification requirement. Essentially, the tool ensures that limited verification capacity is distributed as per priority of need.