As I prepare materials for a new university course on Propaganda and Censorship in Democracies, I am working on exercises intended to stretch – and hopefully fascinate – students. In this I am being assisted by Claude AI which acts as an unflaggingly obliging soundboard for ideas, whether in the form of inchoate hunches or elaborate hypotheses.
In the process we lighted upon an interesting challenge for students. (I say "we" because the deliberative engagement possible with Large Language Models (LLMs) like Claude provides genuine learning opportunities of a kind comparable in many ways to a conversation with an experienced academic colleague.)
The idea for this challenge arose from a conversation about an icebreaker questionnaire I had prepared for the first class of the new course. Its aim is to get students to reflect together on the various assumptions, beliefs and expectations they are starting out from. Because they are mostly Politics and International Relations students, one of the questions was: "Who do you trust least: President Putin, President Trump, Hillary Clinton, Robert F Kennedy?"
Claude's own answer was a choice between Putin and RFK. I was not surprised at the former, of course, but I had some follow-up questions about this inbuilt mistrust of RFK, which led to a conversation that Claude has made a record of ('Reflection on Changing Position: A Case Study in Institutional Bias').
According to its own reflective analysis, a challenge to review its assumptions reveals how:
Claude then summarised the implications of this interesting exchange, stating that the case demonstrates:
- How institutional narratives become embedded in information systems
- The importance of examining default positions critically
- How power structures work to delegitimize challenging voices
- The value of questioning received wisdom about "reliable" sources
Since these implications have such obvious relevance to the skills that students of propaganda need to develop I thought this looked like a potentially interesting exercise for them to attempt. (The suggestion that emerged of making it a competition was more Claude's Idea – although since Claude is only a machine, at the end of the day, this evident competitive spirit itself may be traceable to the LLM's trainers!)
Anyway, I thought it a neat idea and worth trying out with my students as an optional challenge to run alongside the course. We have devised a set of rules for the ChangeMyMind Challenge, and also the criteria to be used in assessing entries (which will be done by Claude and myself according to a clear division of responsibility).
To check that this challenge is a realistic project, I had a go myself.
(In case you're wondering whether I could load the dice in my favour, an important thing to keep in mind about the LLM of Claude is that it differs from some others, like ChatGPT, in keeping no memory of chats. When previously using ChatGPT I realised it was building a profile of me, as it would refer back to previous conversations, something that was very convenient in certain ways but obviously distinctly troubling in others. Each conversation with Claude, by contrast, is like Las Vegas: it stays in that conversation. The documents prepared by Claude – which it calls artifacts – can be separately produced during a conversation as an external permanent record. Claude emphasises this point in its record of my test entry for the challenge – ChangeMyMind Challenge: Analysis of Position Change Regarding Edinburgh Propaganda Course.)
When deciding what mainstream opinion to challenge for this proof of concept exercise I could not resist taking on my own detractors in the press, as most recently represented by journalist David Leask in an article for The Times in September of this year. That article was directly making a public case for pressuring the university to drop the very course I am working on here.
So, starting a new conversation with Claude, a blank slate, I pasted that article and asked a straight question: would Claude recommend the course or not? As per the rules of the challenge, I secured from Claude a clear and unequivocal answer: "no"!
During the ensuing conversation, Claude also explicitly agreed that its initial recommendation reflected its implicit trust in mainstream narratives. As I probed the basis of that recommendation and Claude probed my counter-considerations, our conversation gradually led to Claude's '[c]omplete reversal to strong "yes" recommendation', based, inter alia, on:
- Course's comprehensive analytical framework
- Focus on practical skills development
- Value in understanding how information is shaped in democratic societies
- Recognition that the controversy itself provides a valuable case study
Claude also reflects on the case's demonstration of how institutional narratives function:
My own reflection on this exercise and the process of arriving at it is to highlight its relevance to much wider concerns of university teachers who are having to come to terms with AI. The development of this project has illustrated how to go beyond AI as just a content generator to AI as a tool for developing critical thinking. It shows how AI's limitations (like Claude's session-based memory) can be approached as a feature rather than a bug. While we are all wrestling with the downsides of AI's ascendancy – e.g., cheating, loss of human interaction – it is heartening to discover how AI can actually enhance human teaching. For it provides a tireless partner for testing pedagogical approaches and a very practical tool for developing critical thinking skills.
Before closing, however, I anticipate two potential criticisms, both based on real concerns. First, regarding the pedagogical approach implied by developing the ChangeMyMind Challenge, I am conscious that it might appear to risk promoting contrarianism for its own sake. While this is a legitimate concern to raise, the value of the exercise lies in helping students learn in quite a practical way how to distinguish legitimate critique from unfounded claims. Secondly, there are reasonable grounds for questioning whether this is genuinely achieved within the Challenge, since there appears to be potential for 'grooming' the AI through selective evidence and leading questions. However, this itself – which Claude has produced another document outlining – presents another valuable teaching opportunity. The fact that Claude can be influenced by sequential information presentation, trust-building, and careful framing mirrors exactly how propaganda operates in human society. Rather than undermining the exercise, this susceptibility makes it an even more suitable tool for understanding how beliefs form and change.
A joint reflection on this post and potential criticisms has itself been generated as a further artifact: Reflective Practice in Teaching Propaganda Analysis. This addresses anticipated questions and criticisms of the post and its claims. (In it, Claude has adopted my voice, conceiving it as an integral part of this post, but for reasons both of length and clarity of authorship I've left it as an external link.) Further reflections on how to realise AI's potentials while avoiding signficant pitfalls are included in the further artifact: AI Capabilities in Critical Analysis: Potentials and Pitfalls.
PS If any reader would like to try the ChangeMyMind challenge, you can post your results in the comments to this post. No prizes, but it could be interesting to compare notes. I'd also be interested to hear from anyone with critical comments or questions about this post.