What are the possible effects that might emerge as artificial intelligence (AI) is introduced into the practice of contemporary public administration and policy analysis? The adoption of AI in the public sphere presents a number of opportunities, yet simultaneously raises serious challenges that must be confronted. AI can be used to improve government operational efficiency, but at the potential expense of displacing scores of public servants. The use of AI can improve decision quality and consistency, though might make obsolete the empathy of the ‘street-level bureaucrat’ and reveal the inability of AI to sensibly explain its reasons for its decision. AI might potentially revolutionize the relationship amongst citizens and stakeholders, the public service, and the political class, though at the expense of the thoughtful exchange of ideas and perspectives envisioned in the origins of democracy. And there are potentially profound challenges in regulating the use of AI (both within government, and in the private sector) where such uses impinge upon public values.
My aim here is not to resolve these dilemmas, but rather to provide readers with context for thinking about how AI might affect the public sector — specifically the public policy analysis function — over the longer term. In considering what AI might mean for how policy analysts carry out their duties, I take as a metaphor ‘Lasswell’s Robots’, imagining the possibilities and pitfalls where Harold Lasswell’s vision for the policy sciences intersects with the trajectory of artificial intelligence — or, what happens when an artificially intelligent robot achieves the capabilities of our best human policy analysts? Lasswell was clearly aware of the early advances made in the field of artificial intelligence in the 1950s, noting in his 1956 presidential address to the American Political Science Association “Great strides have been taken in brain design. Experimental models of robots have been built who solve problems of a rather complex order in a given environment”. He went on to ask: “when do we wisely extend all or part of the Universal Declaration of Human Rights to these forms? When do we accept the humanoids … as at least partial participants in the body politic?”. While it has taken much longer than he or many others of that era anticipated it would take for today’s AI and robots to reach even their current potential (let alone the status of sentient humanoids), the question is no less pressing. As much as the 1956 address asked broader questions about the appropriate role of the policy analysts in a democratic society, the idea of ‘Lasswell’s Robots’ is used here to ask what increasingly capable AI might mean for the future of the public service and its democratic foundations.
Artificial Intelligence in the Public Service
As AI research and development have accelerated in the past decade, a renewed enthusiasm for technology to ‘transform’ governments and societies has taken hold. (In this enthusiasm it is useful to recall how the popularization of “the information superhighway” in the mid-1990s gave rise to a “fairy godmother” period when progressive politicians sought “to associate themselves with the magical effects of her wand”, a period that also promised to ‘transform’ government and fix all manner of administrative inefficiencies, and ask whether we are putting a similar faith in the magicless of AI today). From a public policy perspective, governments around the world are seeking to maintain and sharpen their country’s competitive edge by funding further research and development, while also considering the appropriate governance of AI both within and outside of government. Governments are also working to respond to the regulatory challenges that AI raises, and are collaborating with civil society organizations, academics, and private firms to develop appropriate governance frameworks for AI both within and outside of government. To date, more than 30 countries and regions have published AI strategy documents, and principles to guide the adoption of innovative, trustworthy, and responsible AI have been developed by the OECD, the European Commission, and many countries.
Within the walls of government, public sector use cases for AI include business process automation, asset management, knowledge sharing, compliance and risk management, fraud and corruption detection, citizen service delivery, virtual service agents, and analytics for decision-making and policy design. The OECD predicts advances in AI for administrative efficiency, public decision making, healthcare, transportation, security, citizen and stakeholder relationships, regulations, and achieving the Sustainable Development Goals. AI can be used, for example, to analyze 311 call center data to improve internal management, service delivery, and strategic decision making.
Policy Analysis is What Policy Analysts Do
While opportunities for improving government operations and administration, enhancing personalized service delivery experiences, improving back-end process efficiencies, and ensuring consistency and timeliness in administrative decision making are important for the public service of the future, one key application focused on here is in using natural language processing (NLP) for undertaking policy analysis and supporting public policy decision making. That is, AI is now showing itself as a plausible tool to be used to support or replace the work of the policy analyst.
Of particular note is Lasswell’s introductory chapter in The Policy Sciences, where he advanced “policy analysis” as a term of art, seeking to differentiate it conceptually from the social sciences generally and political science specifically.
Consider the writing of a briefing note in support of policy making, where a policy analyst consults a variety of sources — both internal and external — and works to synthesize that body of evidence into useful decision support material. Policy analysis and decision support is the most tangible artefact of Harold Lasswell’s vision, where an apolitical assessment of the public and its problems, founded upon methodologically rigorous social science analysis, leads to the articulation of optimal strategies for dealing with the problem. Policy analysis provides support for decision making with the aim of contributing to a better decision than would have been made in the absence of such analysis. The formalization of policy analysis was articulated through the publication of Lerner and Lasswell’s edited volume inaugurating the policy sciences, where an integrated, multidisciplinary approach to the study of public problems and methods for developing recommended solutions first took shape. Harold Lasswell is widely considered to be the founder of the policy sciences, and his postwar writings provide the field with its earliest concepts. Of particular note is Lasswell’s introductory chapter in The Policy Sciences, where he advanced “policy analysis” as a term of art, seeking to differentiate it conceptually from the social sciences generally and political science specifically. And near the end of his career, Lasswell provided bookends to his career-long advance of the policy sciences perspective.
There is a rich literature on what policy analysts should do. But in terms of how policy analysts today actually practice their “art and craft” — beyond the facetious tautology of “policy analysis is what policy analysts do” — the empirical evidence on what policy analysts actually do in practice is less developed. The term policy analyst represents a wide-ranging category involving all public servants with a connection to policy analysis and formulation processes in government. Public servants in these positions may spend a lot of their time dealing with communications issues, planning and reporting, attending to operational concerns, and engaging in stakeholder management.
Alternatively, what do policy analysts think they do? Over two decades ago, Morçöl found that there was considerable support for quantitative positivism among policy professionals, especially among practitioners and professionals with educational backgrounds in economics, mathematics, and science — the veritable embodiment of Lasswell’s hope for the policy sciences. Under that model, policy analysis (with the emphasis on analysis) focuses on the clarification of the problem and the consideration of possible solutions. The artefact of the briefing note is more of a by-product of the analytical process, meant to summarize for the executive decision maker the cognitive and analytical processes the analyst has gone through. However, a more recent survey of practicing policy analysts in government found that, when asked to rank five policy analyst archetypes (“connector,” “entrepreneur,” “listener,” “synthesizer,” “technician”) in the order of how they understood and practiced policy analysis, the “synthesizer” archetype (defined in part as “consulting various sources to understand how a problem is conceptualized … [and] develop recommended ways to deal with the problem”) was overwhelmingly identified with, and the “technician” archetype (defined in part as “locating of primary raw data sources in order to undertake statistical policy research”, and more aligned with what Morçöl found) was consistently ranked lowest. Simultaneously over the past quarter century there appears to have been a profound shift away from traditional analytical activities undertaken by policy analysts towards the providing of support for the political agendas of ruling parties — the very status Harold Lasswell sought to rescue political economy from. As much as policy analysis is usually considered distinct from politics, the post-positivist policy perspective acknowledges the normative basis of policy analysis and the crucial role that politics plays in the process.
Contemporary policy processes are understood to emerge from complex actor constellations (both inside formal government and between government actors and external actors), and decisions are often made and implemented in a highly decentralized and informal manner. Part of the challenge of understanding the world of the modern policy analyst is tied up in the complexity of policy making. Complex public policy challenges are more than just ‘really complicated’ problems; they exhibit conditions such as uncertainty, rapid emergence, network and system effects, and non-linearity, and are subject to interpretation. The complexity of the modern public policy environment requires a response from the policy analysis system that is agile, open, accepting of mistakes and failure, and is able to learn and adapt.
Lasswell’s Robots … and the End of the Policy Analyst?
Is this a situation in which AI might excel? Advances in AI generally have led to dramatic gains in natural language processing (NLP) capabilities specifically. NLP tools (e.g., the GPT-3 tool from OpenAI) can also now quasi-independently write text—such as news stories or short essays—with very little human input or guidance, based on a short prompt, gathering information from online resources, filtering biased data, synthesizing a large volume of text, and even mimicking a style of writing. Recent experiments have demonstrated the ability of AI to generate text on different topics with a high level of accuracy.
NLP algorithms have the ability to generate text using a process similar to that used by policy analysts, raising the possibility that parts of the policy analyst’s skill set could be supplemented or even replaced by AI. A recent experiment tested whether current NLP technology is capable of producing policy analysis briefing notes that expert evaluators judge to be useful, finding that currently available NLP tools are not quite yet able to do so on their own but might serve as a useful supplement to the work of human policy analysts.
Could one of Lasswell’s Robots serve as a policy analyst today? Try this simple experiment: using your voice to communicate with your nearest smart device (Siri, on your phone, or a nearby Alexa device), ask it (her? them?) “What is the capital of Mali?” So long as you speak clearly with little background noise (so that the AI-powered NLP model can work its magic), your device will likely give you the correct answer. Next try this: “Siri, what is the best strategy to ensure high rates of vaccine uptake and compliance with public health orders in the country over the next year?” This question is likely to provide the equivalent of a web search (punctuated by the annoyingly perky “here’s what I found — check it out!”) with limited salience.
While the use of AI being demonstrated here is the ability of your device to understand the words you speak and take action based on that, the results of this wizardry are rather mundane. Where the answer to a query is clear and unambiguous, the correct answer is provided.
While the use of AI being demonstrated here is the ability of your device to understand the words you speak and take action based on that, the results of this wizardry are rather mundane. Where the answer to a query is clear and unambiguous, the correct answer is provided. Where the question is more nuanced and complex, the best your phone can do is provide you with quantity over quality — something that would get a junior policy analyst quickly reprimanded.
Yet assuming linear advances in the capabilities of NLP tools like GPT-3, could AI — say, within the next five years — independently write plausible, persuasive, and useful briefing notes? For policy analysis questions of the category “tell me about X”, we already have working models that can undertake this type of work.
But for questions of the type “what is the best strategy to choose in this complex setting”, the act of synthesizing a large volume of material into a two-page briefing note will fall short of meaningful decision support, and only serve to illustrate what policy analysts add to the consideration of policy problems, Advances in NLP text summarization are impressive and go beyond even today’s most powerful search engines. You will have likely already read an NLP-generated news story without realizing it. Yes policy analysis involves accessing and synthesizing a range of information into summary briefing notes. But the type of policy analysis required to support questions about inter alia strategic responses to complex settings requires much more than just synthesis. Policy analysts also critically assess what the wider research literature has to say. They consider what strategies and interventions are legal and within the authority of the relevant government. They survey public values, and evaluate citizen and stakeholder perspectives, when contemplating government action. They gauge how the Minister’s perspective and mandate informs how she might see the issue. These types of questions illustrate what AI can do for us (e.g., synthesize relevant information, draft background text, etc.) and what it can’t do (e.g., understand the wider political climate, know the mind of the minister, engage in constructive dialogue with stakeholders, etc.). Given all the things, noted above, that policy analysts do and the complex environments in which they operate, it is more science fiction than science to suggest that a robot policy analyst can (for the foreseeable future) do more than synthesize the textual material it has access to.
Nonetheless, AI-generated briefing notes that synthesize and summarize existing textual material are indeed within reach. To be clear, there is still much work to be done to develop an AI that can independently produce plausible, persuasive, and useful briefing notes. But given the speed with which NLP capabilities are developing, we should anticipate a not-too-distant future in which requesting a briefing note on a topic would result in the instantaneous production of one — either for the immediate decision-support needs of a public sector executive, or as the preliminary draft to be amended by the policy analyst. In short, we are at the cusp of creating one version of Lasswell’s Robots: AI models that can undertake some of the work of policy analysts in synthesizing large volumes of material into decision support documents. While the wholesale displacement of policy analysts in the public service by an army of Lasswell’s Robots is unlikely, this should not cause us to disregard the potential impacts of advancing technology on the future of the public service. To borrow a phrase from the impact of AI on radiologists: AI won’t replace policy analysts; but policy analysts who know how to effectively use AI will replace those who don’t.
The Democratic Implications of Lasswell’s Robots
But if policy analysts do come to embrace this functionality and take to creating instantaneous draft briefing notes, to let their personal robot do the dirty work of writing the first draft, what might be the implications for the future of policy analysis? Several concerns are worth noting.
What are the data origins of that first draft? Have bad actors found ways of ‘poisoning’ the databases that the robotic policy analyst has consulted, resulting in a misleading take on, say, public opinion? Or has Lasswell’s Robot been led down a particularly idiosyncratic rabbit hole that becomes a path of inquiry that is difficult to steer out of? Are future ways of considering an issue irresistibly colored by past framings?
What is lost by taking the easy route to a first draft? Getting the two-page summary of the four hundred articles, reports, blog posts, and position papers is certainly time saving. But what does the policy analyst lose by not immersing himself in the minutiae of the underlying documentation? What nuance of perspective is not available to the policy analyst when all they see is the summary?
Who bears responsibility for the briefing note? The emerging law around driverless cars is trending towards placing responsibility for collisions in the hands of the human at the wheel. Why not the manufacturer of the car? Or the programmer? Or the AI itself? Similarly, for public sector use of AI, public administrators and politicians cannot shield themselves behind a Kafkaesque veil claiming the denial of an application is because ‘the computer said no’.Traditions of public service responsibility and the unique position of policy analysts as advisors within a democratic system should not allow either policy analysts or decision makers to claim that the reason for their decision was because ‘the computer said so’.
 Harold Dwight Lasswell, “The Policy Orientation.” The Policy Sciences, edited by Daniel Lerner and Harold Dwight Lasswell, (Stanford, CA: Stanford University Press 1951): p. 3–15.
 Harold Dwight Lasswell, “The Political Science of Science: An Inquiry into the Possible Reconciliation of Mastery and Freedom.” The American Political Science Review, Vol. 50, No. 4 (1956): p. 961–79.
 Helen Margetts, Information Technology in Government: Britain and America (London: Routledge 1998).
 Kai-Fu Lee, AI Superpowers: China, Silicon Valley, and the New World Order (Houghton: Mifflin Harcourt, 2018).
 OECD,A Artificial Intelligence in Society (Paris: OECD Publishing, 2019); Ryan Calo, “Artificial Intelligence Policy: A Primer and Roadmap,” UCDL Rev. Vol. 51 (2017): p. 399-432; Tara Qian Sun and Rony Medaglia, “Mapping the Challenges of Artificial Intelligence in the Public Sector: Evidence from Public Healthcare,” Government Information Quarterly, Vol. 36, No. 2 (2019): p. 368–83.
 Miriam C. Buiten, “Towards Intelligent Regulation of Artificial Intelligence,” European Journal of Risk Regulation, Vol. 10, No. 1 (2019): p. 41–59.
 OECD, (2019); Corinne Cath, “Governing Artificial Intelligence: Ethical, Legal and Technical Opportunities and Challenges,” Philosophical Transactions. Series A, Mathematical, Physical, and Engineering Sciences, (2018). royalsocietypublishing.org; https://doi.org/10.1098/rsta.2018.0080; Virginia Dignum, “Ethics in Artificial Intelligence: Introduction to the Special Issue,” Ethics and Information Technology, Vol. 20, No. 1 (2018): p. 1–3.
 Daniel Zhang, Saurabh Mishra, Erik Brynjolfsson, John Etchemendy, Deep Ganguli, Barbara Grosz, Terah Lyons, et al. “The AI Index 2021 Annual Report,” arXiv [cs.AI] (Human-Centered AI Institute, Stanford University, 2021), arXiv. http://arxiv.org/abs/2103.06312
 Jamie Berryhill, Kévin Kok Heang, Rob Clogher, and Keegan McBride, “Hello, World: Artificial Intelligence and Its Use in the Public Sector,” OECD Working Papers on Public Governance 36 (Paris: OECD, 2019 ) https://doi.org/10.1787/726fd39d-en
 Weslei Gomes de Sousa, Elis Regina Pereira de Melo, Paulo Henrique De Souza Bermejo, Rafael Araújo Sousa Farias and Adalmir Oliveira Gomes, “How and Where Is Artificial Intelligence in the Public Sector Going? A Literature Review and Research Agenda,” Government Information Quarterly, Vol. 36, No. 4 (2019): p. 101392.
 Jamie Berryhill, Kévin Kok Heang, Rob Clogher, and Keegan McBride, (2019).
 Daniel J. Caron, Eric Lyall Nelson and Sara Bernardi, “Gouvernance Municipale et Information: étude Exploratoire Sur L’utilisation Optimale Des Données 311,” Canadian Public Administration: Administration Publique Du Canada, Vol. 63, No. 4 (2020): p. 602–19.
 Edward S. Quade, Analysis for Public Decisions (New York: Elsevier, 1975).
 Daniel J. Caron, (2020).
 Harold Dwight Lasswell, (1951), p. 3-15.
 Harold Dwight Lasswell, “The Emerging Conception of the Policy Sciences,” Policy Sciences, Vol. 1, No. 1 (1970): p. 3–14; Harold Dwight Lasswell, A Pre-View of Policy Sciences (Elsevier publishing company, 1971).
 Hank C. Jenkins-Smith, “Professional Roles for Policy Analysts: A Critical Assessment,” Journal of Policy Analysis and Management [the Journal of the Association for Public Policy Analysis and Management], Vol. 2, No. 1 (1982): p. 88; Douglas Toregerson, “Between Knowledge and Politics: Three Faces of Policy Analysis,” Policy Sciences, Vol. 19, No. 1 (1986): p. 33–59.
 Aaron Wildavsky, The Art and Craft of Policy Analysis (Palgrave Macmillan, 1979).
 Arnold J. Meltsner, Policy Analysts in the Bureaucracy (Berkeley and Los Angeles: University of California Press, 1976).
 Laurent Dobuzinskis, David H. Laycock and Michael Howlett, Policy Analysis in Canada: The State of the Art (University of Toronto Press, 2007).
 Göktuğ Morçöl, “Positivist Beliefs among Policy Professionals: An Empirical Investigation,” Policy Sciences, Vol. 34, No. 3-4 (2001): p. 381–401.
 Ola G. El-Taliawi, Nihit Goyal and Michael Howlett, “Holding out the Promise of Lasswell’s Dream: Big Data Analytics in Public Policy Research and Teaching,” The Review of Policy Research, Vol. 38, No. 6 (2021): p. 640–60.
 Göktuğ Morçöl, “Positivist Beliefs among Policy Professionals: An Empirical Investigation,” Policy Sciences, Vol. 34, No. 3-4 (2001): p. 381–401.
 Justin Longo, “Towards Policy Analysis 2.0: Platforms for Knowledge Sharing and Collaboration Among Policy Analysis,” In Policy Work in Canada: Professional Practices and Analytical Capacities, edited by Michael Howlett, Adam Wellstead, and Jonathan Craft (Toronto: University of Toronto Press, 2017): p. 302-321.
 John Forester, “Response: Toward a Critical Sociology of Policy Analysis,” Policy Sciences, Vol. 28, No. 4 (1995): p. 385–96.
 Harold Dwight Lasswell, (1951), p. 3-15.
 Frank Fischer, Reframing Public Policy: Discursive Politics and Deliberative Practices (New York: Oxford University Press, 2003).
 See https://openai.com/blog/gpt-3-apps/
 Laith Abualigah, Mohammad Qassem Bashabsheh, Hamzeh Alabool and Mohammad Shehab, “Text Summarization: A Brief Review,” In Recent Advances in NLP: The Case of Arabic Language, edited by Mohamed Abd Elaziz, Mohammed A. A. Al-qaness, Ahmed A. Ewees and Abdelghani Dahou (Cham: Springer International Publishing, 2020): p. 1-15; Kaiz Merchant and Yash Pande, “NLP Based Latent Semantic Analysis for Legal Text Summarization,” In 2018 International Conference on Advances in Computing, Communications and Informatics (ICACCI), (2018): p. 1803–7.ieeexplore.ieee.org
 Mahak Gambhir and Vishal Gupta, “Recent Automatic Text Summarization Techniques: A Survey,” Artificial Intelligence Review, Vol. 47, No. 1(2017): p. 1–66.
 Ben Armstrong, Megan Beretta, Evan Crothers, Michael Karlin, Dongwoo Kim, Justin Longo, Lorne Powell and Trooper Sanders, “Siri Humphrey: Design Principles for an AI Policy Analyst,” In Artificial Intelligence’s Societal Impacts, Governance, and Ethics: 2019 Summer Institute On AI and Society and Its Rapid Outputs, edited by Edward A. Parson, Alona Fyshe, and Dan Lizotte. AMII, CIFAR, UCLA Law / PULSE, (2019), https://aipulse.org/siri-humphrey-design-principles-for-an-ai-policy-analyst/
 Mehrdad Safaei and Justin Longo, “The End of the Policy Analyst? Testing the Capability of Artificial Intelligence to Generate Plausible, Persuasive, and Useful Policy Analysis,” In Facing the Challenges to Public Administration Research and Teaching (Canadian Association of Programs in Public Administration, 2021): p. 6-7.
 Alexander Hevelke, and Julian Nida-Rümelin, “Responsibility for Crashes of Autonomous Vehicles: An Ethical Analysis,” Science and Engineering Ethics, Vol. 21, No. 3 (2015): p. 619–30.
 E. Wihlborg, H. Larsson and K. Hedström, “The Computer Says No!’ -- A Case Study on Automated Decision-Making in Public Authorities,” 2016 49th Hawaii International Conference on System Sciences (HICSS), (2016): p. 2903–12.
 James Farr, Jacob S. Hacker and Nicole Kazee, “The Policy Scientist of Democracy: The Discipline of Harold D. Lasswell,” The American Political Science Review, Vol. 100, No. 4 (2006): p. 579–87.