Dialogic education has promising potential for reducing polarization, widely seen as a threat to democracy (Wegerif, 2022; Parker, 2023). Engaging students in an internally persuasive discourse (IPD) (Bakhtin, 1981) means creating a space where students examine their vested truth in light of critique and alternatives presented by a different, conflicting Other (Matusov, 2009). Successful implementation of IPD increases students’ polyphony, manifested in legitimizing the right of other opinions (other voices) to exist and engaging in a dialogic relationship with this voice (Parker, 2023). It could bring democracy to life inside the school (Apple & Beane, 2007; Gilbert, 2020).

In previous work, we developed and successfully implemented a pedagogical model aimed at IPD. Our design relied on the replete evidence in the literature that a dyadic interaction — students with textual, inanimate representations of the Other, conflicting voice — is less likely to generate IPD because students’ reading is mediated by the mechanism of appropriation/resistance (Wertsch, 1998). Namely, students tend to unquestionably accept representations in line with their in-group voice and ignore or reject (with ostensive argumentative efforts) the Other voice (Brand et al., 2023).

We thus structured a triadic interaction— students from both sides of the conflict and text. The hypothesis was that the animated Other is flexible and attuned to one’s voice, thereby metaphorically “amplifying” the text. Nonetheless, meticulous scaffolding is required to (a) prevent the deterioration of hot discussions into mere disputes, (b) enable a safe space to argue and criticize, and (c) encourage reasoning and re-examination.  

In one successful implementation of this model, Israeli post-secondary students, Jewish and Arabs, e-investigated an event from the Israeli-Palestinian conflict. As expected, the discussions were disputatious. Nonetheless, they were fruitful. While students did not abandon their in-group narratives, their voices became polyphonic, that is, enriched by the Other voice. This was expressed, for example, in moving from a zero-sum viewpoint on historical events and employing moral judgment to a portrayal of an entangled relationship between the agents and assuming (some) accountability towards in-group historical agents (Ben-David Kolikant & Pollack, 2015).

Intuitively, chatbots based on large-language models (LLMs) (e.g., ChatGPT, Llama, Bard) can be used to scale dialogical education because, owing to their nature, they could enable, provoke, and facilitate a productive dyadic interaction—student and text. Specifically, the text that a chatbot provides is not inanimate, it “talks” and hence can dynamically attune the responses to the interlocutor. Moreover, it can introduce students to a myriad of voices and ideas attuned to the unfolding conversation.

The use of chatbots also lessens the need for careful structuring of the encounter, aimed at preventing “explosions”, students being offended or stressed by the Other, which may lead to the opposite result, a boost to polarization. Since chatbots are not human, there is no fear they would be offended by interlocutors. Additionally, students can feel safe to utter their voices and critique, ask for clarifications, experience uttering the Other’s voice, and admit that they changed their minds or realized there is merit in the other’s viewpoint without feeling that they betrayed themselves and/or their in-group.

To gain insights into the potential and limitation of LLMs to scale dialogic education, in particular the engagement of students in IPD, we fine-tuned an LLM with a corpus of discussions in which IPD was evident. Then, we conducted discussions on controversial topics with the chatbot and analyzed its discursive moves. Our focus was on how, if at all, the chatbot provokes and enables its interlocutors to revisit their ideas. 

We fine-tuned “Llama-2-7b-chat-hf” with a corpus of 1000 discussions taken from Reddit/Change My View (CMV). Llama 2 is a collection of pre-trained and fine-tuned generative text models, which (a) range in scale from 7 billion to 70 billion parameters; (b) are auto-regressive language models that use optimized transformer architecture; and, importantly, (c) can be optimized for dialogue use cases. We named the chatbot obtained “LlamaLo” (meaning ‘why not’, in Hebrew). CMV is self-described as “A place to post an opinion you accept may be flawed in an effort to understand other perspectives on the issue” (www.reddit.com/r/changemyview/). CMV is heavily moderated. To encourage users to respond to each other, whoever succeeds in shifting or expanding (i.e., changing) the view of the original poster can be rewarded with a Delta (∆). The idea was that LlamaLo would grasp the discursive “ground rules” embedded in discussions with Delta and use them in future conversations. Owing to the high quality of discussions in CMV, they are commonly used for natural language processing (NLP) and social science research, ranging from argument mining to the study of the effects of forum norms and moderation (Dutta et al., 2020; Na & DeDeo, 2022; Nguyen & Young, 2022). The Delta reward is perceived in those studies as an indicator of a productive discussion since it declares change or expansion in view. We then discussed with LlamaLo 10 controversial topics (e.g., religion and state; bi-national conflicts). We examined its responses to several discursive situations we had created (e.g., unreasoned disagreement, fake knowledge, complex argumentation, and critical questions). We analyzed the conversations, focusing on LlamaLo’s (a) quality of arguments presented, (b) extent of knowledge added, (c) transactivity, i.e., building on the interlocutor’s utterances, and (d) discursive acts that invite the interlocutor to expand and refine their voice.

The significance of this work is in the proof of concept of the possibility to scale dialogic education, employing a dyadic interaction between a chatbot and users. Specifically, our preliminary findings are encouraging. LlamaLo, for the most part, presented alternative ideas using well-grounded claims and added relevant knowledge. It mitigated the disagreement (i.e., softening) and provided to-the-point critique and alternative claims. It also made discursive moves, inviting the interlocutor to continue the conversation with probing questions, such as “What do you think?”. However, similarly to other LLM-based chatbots, it was not free of flaws, such as hallucinations. Also, sometimes it stuck to one point rather than enabling the conversation to expand. We are now in the process of further improving Llamalo by fine-tuning the base model and formulating effective prompts in order to scrutinize the potential and limits of such a tool. This phase lays the ground for future work, in which we will carefully and thoughtfully design a pedagogical model that leverages the learning potential of the dyadic interactions—student and chatbot. Then we will carry out design-based research to examine and improve the learning that takes place when the model is implemented in schools.

Apple, M. W., & Beane, J. A. (2007). Schooling for Democracy. Principal leadership, 8(2), 34-38. Bakhtin, M. (1981). The dialogical imagination: Four essays. (C. Emerson and M. Holquist, Trans.). Austin: University of Texas Press. Ben-David Kolikant, Y., & Pollack, S. (2015). The dynamics of non-convergent learning with a conflicting other: internally persuasive discourse as a framework for articulating successful collaborative learning. Cognition and Instruction, 33(4), 322-356. Brand, C. O., Brady, D., & Stafford, T. (2023, June 27). The Ideological Turing Test: a behavioural measure of open-mindedness and perspective-taking. https://doi.org/10.31234/osf.io/2e9wn Dutta, S., Das, D., & Chakraborty, T. (2020). Changing views: Persuasion modeling and argument extraction from online discussions. Information Processing & Management, 57(2), 102085. Gibson, M. (2020). From deliberation to counter-narration: Toward a critical pedagogy for democratic citizenship. Theory & Research in Social Education, 48(3), 431-454. Matusov, E. (2009). Journey into dialogic pedagogy. Nova Science Publishers. Na, R. W., & DeDeo, S. (2022). The Diversity of Argument-Making in the Wild: from Assumptions and Definitions to Causation and Anecdote in Reddit's" Change My View". In J. Culbertson, A. Perfors, H. Rabagliati & V. Ramenzoni (Eds.), Proceedings of the 44th Annual Conference of the Cognitive Science Society (pp. 969-975). Nguyen, H., & Young, W. (2022, March). Knowledge Construction and Uncertainty in Real World Argumentation: A Text Analysis Approach. In LAK22: 12th International Learning Analytics and Knowledge Conference (pp. 34-44). Parker, W. C. (2023). Education for Liberal Democracy: Using Classroom Discussion to Build Knowledge and Voice. Teachers College Press. Wegerif, R. (2022). Beyond democracy: Education as design for dialogue. In Liberal democratic education: A paradigm in crisis (pp. 157-179). Brill mentis. Wertsch, J. V. (1998). Mind as action. Oxford university press.