Maike is designed to give a writing conversation somewhere concrete to go: back into the student’s essay. It diagnoses an argument, selects a question about a weakness, explores that question with the student, then helps them plan a revision. The student writes the change and can resubmit the essay for another cycle.
That is the architecture described by Lucile Favero, Tanja Käser, Juan Antonio Pérez-Ortiz and Nuria Oliver in Thinking Is Not Writing: Designing AI Chatbots to Connect Reasoning and Revision, an arXiv preprint submitted on 5 November 2026. Peer-review status was not verified as of 7 November. This account examines the authors’ intended design, without independent testing of the tutor.
The authors identify a practical problem: asking useful questions about an argument does not by itself ensure that the discussion reaches the page. Their proposal makes planning the student’s own edits an explicit part of tutoring. Whether that design produces lasting learning is a separate question.
Essay diagnosis supplies a pool of eight questions
Maike is designed for short argumentative essays by secondary-school students. In the architecture description, sections 3.1–3.4, diagnosis operates at two levels.
First, a model identifies parts of the argument: the main position, claims supporting it, premises supporting claims or opposing arguments, and alternative viewpoints. It labels components Strong or Weak, explains its judgment, identifies missing component types and selects two components it judges weakest.
Separate model calls examine the essay as a whole:
- Logical cogency: whether the reasoning is acceptable, relevant and sufficient.
- Rhetorical effectiveness: how clearly and persuasively the essay presents its case.
- Dialectical reasonableness: how fairly and coherently it engages with alternatives.
These calls produce Weak, Fair or Strong labels, explanations and confidence outputs. They are internal model judgments used to guide the dialogue. The diagnoses are not shown to the student as scores or corrective feedback, and the model’s confidence should not be read as a calibrated measure of accuracy. The writing task excludes external sources, so the prompts instruct the model not to penalize missing citations.
Four question generators use these diagnoses to produce eight critical questions. One generator targets weak components and quotes the relevant passage; the other three address the whole-essay traits without quoting particular passages.
A separate model call, called the Judge, chooses a question. Its priorities start with relevance to the diagnosed weakness and whether the student can answer without external research, followed by manageable scope, avoiding repetition, Socratic form and clarity. The diagnosis therefore supplies both a topic and a reason for asking about it.
Think explores the issue; Write turns it into a revision plan
The selected question enters Think mode, where the tutor is intended to help the student clarify an idea, justify it and examine its reasoning. Write mode then asks how that thinking could change the essay: what needs clarifying, qualifying, connecting or developing. The student chooses and writes the actual change.
According to section 3.5.1 and the workflow in section 3.2, students can explicitly advance between stages. Otherwise, the system advances after five conversational turns in each mode, excluding requests to elaborate on a question. After Write, it explores another question in Think. Once the question pool is exhausted, it prompts the student to revise and resubmit.
Editing is available throughout. A student can change and resubmit the essay in either mode, without waiting for all eight questions. Resubmission triggers revision review, followed by fresh diagnosis and question generation.
An elaboration option is intended to rephrase or simplify the latest question while preserving its substance. Its instructions prohibit supplying the answer. That is a prompt constraint; the architecture description alone cannot establish how consistently the tutor follows it.
Each reply uses an evolving account of the conversation
Maike’s adaptation relies on what the authors call a student model. For each typed contribution, the system updates its interpretation of the student’s immediate need, readiness to progress, apparent understanding and possible misconceptions. It also summarizes addressed and open issues.
These are model-generated inferences, rather than validated measurements of a student’s knowledge or readiness.
In the response-generation process, sections 3.5.2–3.6, that updated model produces a one-line instruction about the teaching purpose of the next reply. The system generates three candidate responses using this instruction, the student model, selected question, current mode and preceding dialogue. Another call selects a response, prioritizing responsiveness to the student, progress on the question and fit with the intended strategy.
The candidate-response instructions ask for plain language, one question and attention to the student’s latest contribution. They also prohibit substantive answers, sample arguments and essay sentences. In Think mode, questions can explore clarification, assumptions, reasons and evidence, viewpoints, implications, or the question itself.
When the essay is resubmitted, a review call compares the old and new versions using the retained student model. This is how the design is intended to connect earlier discussion with later changes. The documented continuity is within a session; it does not establish cross-session profiles or a data-retention policy.
Figure 5 illustrates how a claim might change
The paper’s Figure 5 is an expressly constructed illustration, not an exchange taken from the study. Its caption and accompanying section 3.7 text show how the stages are meant to connect.
The example begins with a passage treating a correct answer as evidence of learning. A question challenges that assumption. During Think, the example student distinguishes copying an answer from understanding it. During Write, the student considers how to incorporate that distinction, and the revised passage qualifies the original claim.
The useful detail is the connection between the question and a specific edit. The example shows the workflow the authors intend; it supplies no evidence about how often students follow that path.
Maike excludes fact-checking; learning claims need further evidence
The authors define a limited role for Maike: it does not retrieve external information, verify factual accuracy or correct grammar. A question about the structure of an argument therefore cannot be assumed to establish whether its underlying facts are true.
There are also limits to this account. The source text accessible for this article ended partway through section 3.7. Implementation details, methods, results and supplements were inaccessible. We could not verify the exact deployed model, full prompts, failure rates or privacy and retention arrangements, and do not assess the classroom outcomes here.
Establishing retained learning would require later independent writing assessments. Transfer would require evidence on new tasks, topics or contexts without the same assistance.
