If the behaviour changes when the role, prompts and design change, where exactly does the “critical friend” reside?
It is tempting to place it inside the AI.
The system asks difficult questions.
The system offers alternatives.
The system challenges.
Therefore the system is the critical friend.
But the evidence we have just followed makes that description too simple.
Tang and Putra did not open a generic chatbot and discover a dialogic partner waiting inside it.
They designed one.
Their Dialogic Science Teacher was deliberately configured around reverse prompting, Socratic questioning and a pedagogical account of dialogic learning. The study is explicit that educational chatbot design is not merely a technical problem. The theoretical and pedagogical position matters.
The resulting behaviour was therefore produced by an arrangement.
A model.
Instructions.
A task.
A curriculum context.
A learner.
A sequence of exchanges.
And a human-designed purpose for what those exchanges were supposed to do.
Change the arrangement and the relationship changes.
The same issue appears in the critical-friend literature.
Frambaugh-Kritzer and Stolle found that their AI critical friend could offer feedback and thoughtful questions, but the usefulness of those responses was constrained partly by how the interaction was prompted.
Dikilitaş and Farrell make the dependency more explicit: ChatGPT-mediated reflection is user-driven. The quality and specificity of the prompt help determine what becomes possible.
This creates a strange difference from the human critical friend.
A human colleague may negotiate the role with us.
They may refuse our framing.
They may say the question itself is wrong.
They bring expectations we did not write.
With AI, much of the early relationship can be configured before the exchange begins.
“Act as a Socratic opponent.”
“Do not give me the answer.”
“Challenge my assumptions.”
“Ask one question at a time.”
“Give me a counter-position.”
Each instruction changes the kind of interlocutor that appears.
That does not make the interaction fake.
It makes the design part of the evidence.
Tim Gander and David Parsons make this unusually visible by describing an AI-powered “critical friend” for student research proposals.
Their 2024 paper is a work-in-progress report, so it does not yet establish the outcomes it proposes to evaluate.
But the design problem is already there.
The AI critical friend is built into an assessment process to provide targeted feedback and guidance while the researchers examine critical AI literacy and transparent collaboration.
Again, the unit is not simply “AI”.
It is AI given a role inside a workflow.
This may explain why asking whether AI is a critical friend keeps becoming slippery.
The noun encourages us to imagine a stable entity.
The practice keeps showing us a configuration.
Model behaviour depends on instructions.
What the model can see.
What evidence it can access.
What the human reveals.
What task is being attempted.
Whether the AI is asked to answer, question, oppose, synthesise or remain silent.
What the human does with the response.
And what other people remain in the arrangement.
The critical work may therefore sit less neatly inside either participant than the name suggests.
It can emerge from the conditions of interaction.
That has consequences.
A well-designed critical prompt does not guarantee a good judgement.
A customized chatbot does not guarantee that the learner remains intellectually active.
A role instruction can be followed too obediently.
A system asked to challenge can manufacture opposition merely because opposition was requested.
Design makes some behaviours more likely.
It does not make their outputs true or their challenge worthwhile.
That is where the human remains consequential.
Somebody still has to decide whether the challenge matters.
Whether the alternative is credible.
Whether the evidence warrants movement.
Whether the AI has misunderstood the situation.
Whether the conversation should continue.
Or whether the role itself needs changing.
Perhaps the “critical friend” is therefore not a thing we have found inside AI.
It is a relationship we partly arrange around it.
Partly, because the system still produces responses we do not specify word by word.
Partly, because the human can be changed by what arrives.
Partly, because the surrounding institution, tool and task constrain what either side can do.
And partly, because other people may still be in the room.
Which raises a question that the replacement debate tends to hide:
If the AI is only one part of the arrangement, what happens when human interlocutors remain in it too?
References
Tang, K.-S. and Putra, G.B.S. (2026; first published online 2025) ‘Generative AI as a Dialogic Partner: Enhancing Multiple Perspectives, Reasoning, and Argumentation in Science Education with Customized Chatbots’, Journal of Science Education and Technology, 35, pp. 128–140. doi:10.1007/s10956-025-10240-1. Open-access article.
Frambaugh-Kritzer, C. and Stolle, E.P. (2025; first published online 2024) ‘Leveraging Artificial Intelligence (AI) As a Critical Friend: The Affordances and Limitations’, Studying Teacher Education, 21(2), pp. 188–211. doi:10.1080/17425964.2024.2335465. Publisher record.
Dikilitaş, K. and Farrell, T.S.C. (2026; first published online 2025) ‘Using ChatGPT as a “Friend” in reflective critical friendships’, ELT Journal, 80(1), pp. 117–125. doi:10.1093/elt/ccaf053. Open-access article.
Gander, T. and Parsons, D. (2024) ‘Evaluating the Impact of an AI Critical Friend on Student Research Proposals, Critical AI Literacy, and Transparent Collaborative Assessment Practices’, ASCILITE 2024 Conference Proceedings, pp. 459–464. doi:10.14742/apubs.2024.1201. Conference paper.