The Teacher-Architect: Using Historical Logs to Preserve

The Teacher-Architect: Using Historical Logs to Preserve describes a classroom approach. Teachers design AI tasks so learners show clear evidence of their thinking. In AI education, a historical log is a time-ordered record of a learner's prompts, AI responses, source checks, revisions and reasons for accepting or rejecting suggestions. Teachers use the log as formative evidence, not proof of authorship, so they can question the learning process when AI outputs are fluent but opaque (Bearman & Ajjawi, 2023).

For example, in a Year 9 history lesson on the causes of the First World War, a learner records the first AI summary, flags an unsupported claim, checks it against two sources and rewrites the prompt. The teacher can then ask why that claim was rejected, where the evidence improved and what the learner would do differently next time.

Understanding AI in Education

AI worries echo earlier debates about calculators. Educators feared that easy access to machines would weaken mathematical fluency, or learners' quick and accurate use of basic maths. Yet the research record is mixed: Hembree and Dessart's meta-analysis found that calculator use alongside normal instruction generally did not damage paper-and-pencil skills. It could also improve problem solving and attitudes to mathematics.

The classroom lesson is not that tools automatically improve learning. It is that teachers have to redesign the task. Historical work on educational technology shows that new machines change learning only when curriculum, assessment and classroom routines change with them.

The same principle could apply to generative AI. A documented historical log can make prompt choices, revisions, source checks and rejected AI suggestions visible, so the assessment conversation moves from "did the machine write this?" to "what evidence of thinking can the learner show?"

‍

‍

‍

‍

From Black Box to Audit Trail

A safer way to describe the classroom model is as a documented human-AI inquiry cycle. The learner records what they asked, what the model returned, what they doubted, what they checked and how their next question changed.

This process-focused approach values learning, not just the final piece of writing. Black (1998) argued that assessment supports learning when teachers use evidence to adapt their teaching. The same formative principle applies here. The teacher can see partial understanding, ask follow-up questions and give feedback while the learner's thinking is still developing.

For teachers, the design question is clear: how can the task stop learners from offloading their thinking to AI and require them to use AI as a tool for checking, explaining and improving their reasoning?

‍

‍

The Audit Trail of Thought

It is best to treat documentation as indirect support for self-regulated learning. It gives learners a record they can use to plan, monitor their work and evaluate after the task. This fits with metacognition research. It does not prove that any single AI log format improves outcomes.

Cognitive offloading research explains both the opportunity and the risk. External tools can reduce the processing load of a task, but learners still need to monitor whether the tool is helping or replacing their thinking (Risko & Gilbert, 2016; Flavell, 1979). A log is useful only when it prompts that monitoring.

Without that record, AI use can become invisible. The teacher sees an answer, but not the rejected ideas, source checks or moments where the learner changed direction. The value of the historical log is therefore evidential and pedagogical: it creates something concrete to question.

‍

‍

‍

‍

Research findings on documentation

Ataş and Yildirim's design-based research on shared metacognition is useful, but it is narrower than a direct study of AI history logs. Shared metacognition means learners thinking together about how they are learning. The study supports orientation-planning, monitoring and evaluation-reflection in online group learning. These ideas can guide the design of AI tasks.

This matters especially in the context of generative AI. Bjork (1994) argued that learners can misjudge what they know when fluent processing feels like mastery. The machine's "fluency" can create the same "fluency illusion", where the learner believes they understand a topic simply because the AI has summarised it clearly and confidently (Bjork et al., 2013). The Audit Trail disrupts this illusion by requiring learners to "show the work" of their logic, much like traditional formative assessment practices.

In practice, the learner should remain in the loop: checking claims, asking for alternatives, rejecting weak answers and documenting why the final judgement is theirs. That is a classroom routine, not a guarantee built into the technology.

‍

‍

‍

‍

Shifting the assessment focus

When teachers use the Historical Log, attention shifts from the final product to the documented development of thought. This helps the Teacher-Architect see the scaffolding of the learner's thinking.

Assessment that asks for the process can also make feedback more useful. Instead of marking only a polished final answer, the teacher can respond to the point where the learner misunderstood a concept, accepted a weak source or failed to challenge an AI suggestion.

Decision records are not forensic proof on their own, but they do raise the quality of the evidence conversation. A learner who can explain why a prompt changed, why a source was rejected and why a claim survived checking is showing more than output completion.

‍

‍

Detecting Authentic Inquiry

One concern for teachers is that learners can ask AI to simulate a historical log after they have finished a paper. There is no technical guarantee that a log is authentic, so schools should avoid treating it as automatic proof.

AI-generated logs can look plausible. AI-text detectors are not reliable enough to settle authorship questions on their own. Detection studies show that generated text can be hard to identify, especially after paraphrasing. They also show that human writing can be misclassified.

The practical safeguard is not a branded construct. It is a pattern of evidence: timestamps where available, drafts that match the lesson sequence, source notes, teacher questioning and visible changes of direction that the learner can explain.

‍

‍

‍

‍

The fingerprints of real learning

Look for visible challenge rather than smooth compliance. A stronger log shows the learner questioning an answer, asking for evidence, checking a source and changing course when the evidence does not support the claim. That is useful classroom evidence, but it is not a validated marker of human agency.

Teachers can use timing and task sequence as context, but they should still ask the learner to explain the choices. A timestamp may support a conversation; it should not replace professional judgement.

Sadasivan and colleagues argue that reliable AI-text detection is difficult in real classrooms and other practical settings. Liang and colleagues found that GPT detectors can wrongly label non-native English writing as AI-generated. For this reason, historical logs should be part of a wider assessment conversation. They should not become a detector-led compliance routine.

‍

‍

Reclaiming Teacher Agency

Historical logs may reduce some uncertainty about AI use, but they also add a new routine for teachers to manage. The workload argument is credible only when the log is short, structured and directly connected to feedback.

The strongest case is formative, not administrative. A log can help the teacher identify where feedback is needed: the first weak prompt, the unsupported claim, the copied explanation or the missed opportunity to compare sources.

AI detectors are unreliable enough that schools should be cautious about using them as sole evidence (Weber-Wulff et al., 2023; Liang et al., 2023). A documented log gives teachers a richer basis for questioning, but it does not remove the need for judgement.

‍

‍

‍

‍

Targeted feedback at the point of need

When the log is structured by stages, feedback can be more precise. Teachers can ask about the moment a learner accepted a hallucinated source, ignored a contradiction or changed the question after new evidence appeared.

Teachers do not rely only on plagiarism checks after the work is done. They design inquiry-based learning pathways instead. They also look beyond the final product. They review learners' thinking processes and use them to give feedback.

‍

‍

Future Actionable Knowledge

Generative AI does mean schools need to rethink what counts as achievement. Producing text is no longer enough; learners need to show how they selected, questioned, verified and improved the information they used.

Bearman and Ajjawi (2023) argue that education needs pedagogy for working with AI systems whose outputs are not transparent. In other words, teachers need ways to teach when they cannot fully see how an AI answer was made. The World Economic Forum's Future of Jobs Report 2023 also lists analytical thinking, creative thinking, AI and big data as prominent workforce skills.

When information is easy to generate, learners need to direct, organise and check it well. Academic rigour means careful, disciplined thinking. It is shown through selecting evidence, checking sources and revising work responsibly. A polished paragraph alone is not enough.

For classroom purposes, the useful test is simple: can the learner explain how today's AI-supported work will help them make a better decision, solve a similar problem or ask a stronger question next time?

‍

‍

Using Historical Logs in Your Classroom

This research has practical implications for how you teach with AI. Consider implementing these approaches: Use it as a starting point for professional discussion: identify the learner's current need, record evidence from more than one lesson, and agree the next classroom adjustment with the SENCO or family.

1. Require documented inquiry: Ask learners to maintain a short log of their prompts, AI responses and judgements about whether each response is useful or flawed. This turns AI from a shortcut into a thinking tool. Ask: "Show me your conversation with the AI. Where did you disagree with it? Why?"
2. Audit the process, not just the product: Shift your assessment focus from the final essay or assignment to the documented process of creating it. Use questioning strategies that explore learner decisions: "Why did you ask the AI this specific question? What did you learn when it gave you this answer?"
3. Teach with cognitive load theory in mind: Use AI to offload mechanical tasks, such as generating first ideas or examples, so learners can focus on evaluating arguments, comparing perspectives and building schema in the subject.
4. Model visible challenge and source checking: Show learners how you use AI critically. Think aloud about hallucinations you catch, questions you ask and ways you challenge the AI's assumptions. This shows that working with AI requires challenge and checking, not passive consumption.

‍

‍

Frequently Asked Questions

‍

Next Step for AI-Supported Assessment

Learners gain agency when AI use is visible enough to question. A historical log does not prove authorship, but it can make the learning process clearer by showing prompts, revisions, checks and changes of direction.

Teachers gain more useful evidence when the task asks learners to justify logic, compare sources and explain why they trusted or rejected an AI suggestion.

Used carefully, historical logs move the classroom conversation away from detector scores. They bring attention back to assessment design. They help teachers ask clearer questions about process, evidence and accountability.

‍

‍

References

Akgun, M., & Toker, S. (2024). Evaluating the effect of pretesting with conversational AI on retention of needed information. arXiv. View arXiv record.

Ataş, A. H., & Yildirim, Z. (2024). A shared metacognition-focused instructional design model for online collaborative learning environments. Educational Technology Research and Development. View DOI record.

Bearman, M., & Ajjawi, R. (2023). Learning to work with the black box: Pedagogy for a world with artificial intelligence. British Journal of Educational Technology, 54(5), 1160-1173. View Deakin record.

Bjork, R. A., Dunlosky, J., & Kornell, N. (2013). Self-regulated learning: Beliefs, techniques, and illusions. Annual Review of Psychology, 64, 417-444. View UCLA-hosted PDF.

Black, P., & Wiliam, D. (1998). Assessment and classroom learning. Assessment in Education: Principles, Policy and Practice, 5(1), 7-74. View university-hosted PDF.

Department for Education. Generative artificial intelligence (AI) in education. View GOV.UK guidance.

Flavell, J. H. (1979). Metacognition and cognitive monitoring: A new area of cognitive-developmental inquiry. American Psychologist, 34(10), 906-911. View DOI record.

Hembree, R., & Dessart, D. J. (1986). Effects of hand-held calculators in precollege mathematics education: A meta-analysis. Journal for Research in Mathematics Education, 17(2), 83-99. View ERIC record.

Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7), 100779. View Stanford SCALE record.

Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676-688. View PubMed record.

Sadasivan, V. S., Kumar, A., Balasubramanian, S., Wang, W., & Feizi, S. (2023). Can AI-generated text be reliably detected? arXiv. View arXiv record.

Weber-Wulff, D., Anohina-Naumeca, A., Bjelobaba, S., Foltýnek, T., Guerrero-Dib, J., Popoola, O., Šigut, P., & Waddington, L. (2023). Testing of detection tools for AI-generated text. International Journal for Educational Integrity, 19, 26. View open-access article.

World Economic Forum. (2023). The future of jobs report 2023. View report page.