Governing with AI: Future-ready policy decisions need adaptive systems

Artificial intelligence can help enable real-time decision-making, but its potential depends on how govts adapt

It is evident that artificial intelligence (AI) tools and their capacity to bring together immense knowledge and undertake complex analysis can contribute a lot to the world of governance and policymaking. But whether it does, or by when, depends upon how well the government can incorporate AI within its own processes. 

Governance is no doubt a complex set of activities characterised by limited information and resources to undertake them. The activities themselves require an aggregation of qualitative and quantitative information, beliefs and preferences, ethics and morality, democratic norms and the role of lobbies, history and path dependence, capacities and resources, and other factors. And the fact that geopolitics is making the policy environment so volatile makes it even more complex. In such a situation, we need to ask the simple questions: What might AI bring to the table, and is it worthwhile thinking about how to use it for governance? 

  Answering the first question — how AI can help government— is easier. Many ideas have been proposed, including those related to better targeting of welfare, improved monitoring, better policy design, superior initiatives focusing on behaviour change, improved analysis of scenarios and alternatives, automated audits and real-time monitoring, almost real-time analysis of citizen feedback, and perhaps even redress of grievances, among many others. 

  Each of these requires massive amounts of information and the ability to analyse it. What in the distant past took months to do, can be completed in a few weeks or days, with the use of IT-driven databases readily available to analysts. With AI, this can occur in a few minutes or even less. A query raised in a meeting in an AI-pervasive world can be answered during the conversation itself rather than requiring days of additional information gathering and analysis. This becomes even more relevant given that the intellectual difference between those in the senior levels of governance and those in the lower rungs is quite high, limiting the flow of information and analysis up the governance ladder. AI can empower lower-level functionaries and through them enhance the ability of those at higher levels. 

  Decisions can, therefore, be hastened, provided all the decision-makers are in the room. Benefiting from AI requires a different kind of decision-making process than is currently the norm. In other words, AI can help immensely, but only if the process of decision-making within the government changes character. 

  Moreover, India has a unique opportunity related to AI that many other countries, including developed ones, do not. Digital public infrastructure (DPI) rolled out initially for inclusion purposes, has spread through most parts of national, state and even local governments. Since AI works best when it has access to granular and deep information, the vast and growing DPI network, therefore, provides an additional opportunity to leverage this asset for decision-making in the government.  

  No doubt there are challenges with AI itself, and they will likely remain even when AI goes on steroids — with quantum computing driving it instead of conventional chips. Three main challenges are: Hallucination, inherent reasoning biases, and unevolved ethical and moral core. Hallucinations are when AI tools make up information, something they tend to do more often when less information is available. Since AI models require to be trained on pre-existing information, all our inherent biases — including those related to gender and religion — are absorbed by AI tools. And finally, the concept of right and wrong is absent in such tools. For all three of these concerns, I am sure many guardrails will emerge, and some already have; but they will be like Band-Aids, put up by people with diverse cultural ethos, professional experience, and thought processes from those who make decisions. 

  These are universal issues, and they may make the arguments against the use of AI quite compelling for some. Notwithstanding the fact that some global decision-makers are no different, we would want to eliminate these AI flaws from entering the policymaking space in India. AI should only be used to assist decision-making, not be assigned decision-making powers — at least for now. But even within that limited role, there are many opportunities. 

  There are many decisions that governments take that do not involve societal-scale impacts but have smaller, individual impact. And therefore, the human-AI interface needs to be different depending upon the scale of the potential impact. There are opportunities where quick small gains are possible with greater use of AI — improved website management, for example.  Consider, for instance, the non-revenue-collecting government websites, many of the older ones do not work well. The departments do not have the bandwidth to monitor every page and enforce performance standards on the website operators. Taken together, the improvement of the thousands of government websites could have a massive impact, even though each individual improvement may seem marginal. 

  Orthogonal to this is the issue of decision frequency. Some decisions, such as specific laws to change, are low frequency. Here the best role that AI can play is to better inform policymakers on different options available, and potential impacts under different scenarios. The role of AI in such a situation would be more like that of an information assistant or consultant. On the other hand, decisions such as how to allocate food among PDS shops are more frequent and mistakes can be corrected easily. Here AI can be assigned a greater allocative role under bureaucratic oversight.  Apart from frequency and impact, two other dimensions to assess AI could be costs and complexity.  The larger point is that by unpacking the problem, we can better identify the parameters under which AI should be used, which, in turn, can help hasten the benefits that we can derive from it. 

While it is evident that AI can be a great assistant for policymaking, it is not yet clear how. By unpacking decision-making within the government on the basis of frequency and potential impact, among other dimensions, a framework can be created that will enable both greater speeds, and improved quality of decisions. With the success of DPI and welfare delivery, many avenues have opened up where AI tools can use granular data to inform policymakers in real time. If, on top of that, the government were to tweak some of its internal processes, the potential impact can be significantly greater.  

The author is an economist. The views are personal