This post is a little different. Instead of a written round-up or essay, I’ve used Google’s NotebookLM to generate an AI-simulated podcast conversation about my area of academic interest: AI safety in medical education. NotebookLM can take a set of documents (in this case, my own recent publications) and simulate a dialogue between two speakers discussing their contents. The result is a kind of “AI-mediated” conversation that sits somewhere between a podcast and an automated journal club.
In this post I’ll explain to you how I generated this podcast episode, and discuss potential use cases in our work as educators.
NotebookLM
NotebookLM is an AI research tool from Google DeepMind and Labs that creates personalised “notebooks” powered by a large language model fine-tuned for retrieval and synthesis. It allows users to upload source materials (papers, notes, transcripts, PDFs, etc.), which the system then indexes into an internal semantic map. This forms a representation of key ideas, entities and relationships. When prompted, the model can generate summaries, explanations, or simulated conversations that explicitly reference those uploaded documents rather than general internet data.
In practice, this means the dialogue you’ll hear isn’t drawn from random web text: it’s generated by the model grounded in the source material. NotebookLM effectively acts as both a retrieval-augmented generation (RAG) system and a knowledge synthesiser, designed to help researchers explore, annotate and cross-connect their own materials in new ways.
In addition to audio summaries, NotebookLM can also create video/slide presentations, mind maps and conversations that feel more like traditional chatbots. NotebookLM featured in my ‘Best AI Tools for Educators and Reserachers’ list last week:
Sources
I uploaded a bunch of my own publications, which all explore the role of generative AI in medical education, including work on AI-authored exam questions, bias in AI-generated medical imagery, and the emerging need for safety frameworks and disclosure practices. Together, these papers provide a foundation for understanding both the opportunities and the risks that AI introduces to teaching, assessment and professional identity formation.
Discussion
So, is it any good? In terms of the style, the conversation is coherent, if occasionally over-confident, and it highlights some interesting tensions around automation, authorship and oversight. It is incredibly natural, but I occasionally found myself encountering the uncanny valley although I’m not sure I could put my finger on exactly what shattered the illusion.
As an educational artefact, it’s surprisingly effective. It feels like overhearing two well-briefed colleagues work through a topic. For learners, this kind of AI-generated discussion might serve as a bridge between reading and reflection: a way to hear ideas being explored rather than simply reading them. There is perhaps a very exciting tool here for learners who struggle with reading, including people with dyslexia or SLDs.
For educators, it could point toward new ways of producing dynamic learning materials from research outputs. I’ve had a go at brainstorming some potential use cases, but I’d love to hear your ideas in the comments:
Automated journal clubs: Upload a set of research papers and generate an AI-moderated conversation summarising key findings and controversies. Of course, this rather defeats the educational value of an actual journal club, which most benefits learners who fully participate in the discussion.
Case-based learning prep: Combine clinical guidelines, case notes and textbook excerpts to create briefing summaries or simulated tutor–student discussions.
Curriculum mapping: Ingest module handbooks and learning outcomes to identify overlaps, gaps and misalignments across years or themes.
Assessment design support: Upload question banks and learning objectives to help flag duplication, alignment issues or generate example stems for discussion (with human review).
Faculty development: Generate explanatory dialogues about new pedagogical or regulatory frameworks, helping busy clinicians grasp educational theory quickly.
Student revision aids: Turn lecture slides or readings into self-quizzing dialogues or Socratic tutorials that help learners test understanding.
Research synthesis: Import qualitative data or literature to produce structured thematic summaries or conceptual frameworks for early analysis.
AI safety exploration: Use transcripts and policy documents to simulate stakeholder conversations about governance, bias and ethical deployment in teaching.
