Generative artificial intelligence (AI) is changing what it means to learn and practice some of the skills traditionally associated with policy work. At the Bennett School, we are rethinking how to develop the knowledge, judgement, practical reasoning and human relationships that future policymakers need.
Generative artificial intelligence (AI) can draft a briefing, perform a basic statistical analysis and produce a plausible policy argument in seconds. These used to be core competences taught on most policy programmes around the world. As a result, public policy schools must now consider what these new AI capabilities mean, and how they can be deployed for their educational objectives.
Students are rightly asking what value a university education offers when AI can undertake many of the tasks they are being trained for. If a large language model (LLM) can explain concepts, answer questions and guide independent study, could it replace formal education altogether? Over the past few months, we’ve been reflecting on what knowledge, judgement and development our MPhil programmes in Public Policy and Digital Policy should provide in the context of this new technology.
Our starting point remains simple: students need a strong grasp of core policy theories, alongside key concepts in economics and statistics. An LLM can retrieve or explain many of these ideas in seconds, but looking something up is not the same as having internalised it. Students need enough knowledge of their own to recognise which concepts are relevant, connect them to one another, interrogate evidence and spot the limits of an argument. AI has therefore not changed what we regard as essential knowledge, but it has made us think more carefully about how students build and retain it.
The research by Bastani and colleagues illustrates this well. In a large field experiment with nearly 1,000 high-school students, the researchers compared different forms of GPT-4 support during mathematics practice. Students performed better while using AI, but those given unrestricted access did worse on a subsequent closed-book test once the tool was removed. That effect was largely avoided when the AI was designed to guide students with hints rather than simply provide answers. This finding is broadly in line with a much broader literature on “productive failure”. Experimental studies found that deliberately confronting learners with gaps in their understanding can improve constructive reasoning and help them address new problems.
In our teaching programmes we try to ensure that this kind of problem-solving is central as we put a range of different kinds of policy challenge in front of students. Think for example of a ‘wicked problem’ like urban congestion. Is the underlying driver urbanisation, inadequate public transport, e-commerce, too many roadworks, or a mixture of factors? We encourage students to figure out for themselves where they believe the main problem lies and what evidence would help them analyse the problem. Only then do we introduce them to theories and concepts that will help them conduct a more in-depth analysis into relevant processes and problems. The goal is not to make learning more difficult for the sake of learning, but to ensure that students do the intellectual work necessary through which analytical judgement is developed.
Capabilities that are repeatedly outsourced to AI are also capabilities students have fewer opportunities to build for themselves. Our conviction is that AI can undoubtedly support the development of skills relevant to policy analysis and process of policymaking, but the kinds of critical thinking and judgement involved in these activities is, and will remain, centred upon human capabilities.
Policy is made with and through other people. Preparing future policy leaders therefore requires more than just the command of basic concepts and technical skills. Students need opportunities to listen, persuade, negotiate, respond to challenge, and, sometimes indeed, to change their minds.
An AI system can, of course, simulate disagreement or negotiation. But it cannot fully reproduce what happens when another person challenges your assumptions in real time, or when you have to decide under pressure whether to defend your position, adapt or revise it. We therefore make deliberate space in the curriculum to practise these capabilities. Students work on oral and written communication in professional skills, confront the ethical dimensions of policy choices in a dedicated ethics course, and learn to navigate disagreement in a designated negotiation class.
A diverse, international student community is, we believe, particularly conducive to the development of these capabilities. Our cohorts include students from across the world, who bring with them experience of very different governments, institutions and societies. In engaging with each other they encounter assumptions, constraints and perspectives that may be unfamiliar to them and are challenged to explain and reconsider their own. That diversity is also a source of originality: different experiences can generate questions, interpretations and possible solutions that would be unlikely to emerge from any single shared perspective.
Beyond the dedicated modules, we also build opportunities to practice these capabilities into our case-based teaching. In a simulated cross-government negotiation, for example, students might represent different ministries or public bodies, each working from its own briefing, evidence and priorities. Their task in such an exercise is not simply to defend a position, but to navigate disagreement and reach an outcome that others can also accept. Afterwards, they step back from the result and examine the process itself, considering where the negotiation became difficult, which incentives and assumptions shaped behaviour, what shifted people’s positions and why some compromises proved possible while others did not.
Other universities are, we note, reaching similar conclusions. MIT’s August 2026 report on AI in teaching, learning and research training likewise emphasises the importance of people, community and the in-person educational experience.
Our approach to knowledge building and teaching also shapes how we think about assessment. Our relatively small class sizes allow us to experiment with formats such as oral exams and presentations, where students must apply concepts to unfamiliar problems, explain the evidence behind their conclusions and respond to challenge without relying on AI. We think this is particularly relevant preparation for policy work. Policymakers have always drawn on colleagues, advisers and other forms of support, and will increasingly draw on AI, but ultimately, they remain accountable for the recommendations they put forward. Assessment should therefore test whether students can understand and stand behind their work. Employers we speak to make a similar point: specific tools can be learned quickly on the job, while critical thinking and communication take much longer to develop.
Regular one-to-one supervisions also allow us to follow students’ intellectual development throughout the year. For dissertations, this means discussing in depth with students how a research question takes shape, how the literature informs it and why particular evidence is relevant. The final submission therefore emerges from a sustained, dialogic process of inquiry and feedback.
The approach we take is by no means driven by an outright hostility to this new technology. Our teaching has always drawn upon Cambridge’s strengths in science, technology and engineering. The new MPhil in Digital Policy, launching in early October, reflects our commitment to preparing policymakers for the challenges of AI and the digital world. Indeed, as AI makes some routine analytical and drafting tasks easier, we expect the capabilities emphasised here — deep knowledge, human judgement, communication and the ability to learn and work with others — to become more, rather than less, valuable.
When it comes to AI in our teaching and assessment, we’ll keep reviewing the evidence and listening to and learning from our students and peers.
The views and opinions expressed in this post are those of the author(s) and not necessarily those of the Bennett School of Public Policy.
