Updated on
September 23, 2026
AI and Metacognition: What Teachers Need to Know
Harness ai metacognition teaching students to think critically. Learn how to use artificial intelligence as a Socratic partner rather than an answer engine.
AI metacognition means using a chatbot to question and test a learner's thinking, never to write the finished piece. Support becomes offloading when the AI does the last step, so learners mark a flawed draft or work backwards from an output to its prompt.
AI metacognition is the deliberate use of AI to help learners plan, monitor and evaluate their thinking rather than simply generate answers. In practice, a teacher can ask learners to compare an AI response with their own explanation, spot what is missing, and revise the prompt so the tool becomes a mirror for reasoning. This matters because AI can either strengthen self-regulation or quietly replace it. The teaching move is to keep learners explaining choices, checking evidence and improving subject knowledge.
Key Takeaways
- AI metacognition means using a chatbot to make learners plan, monitor and check their own thinking, not to produce the answer for them.
- The line between support and offloading is who does the last step: if the AI hands over a finished product the learner has offloaded, and if it hands over a part the learner must improve they are still thinking.
- Four routines carry the work: a prompt iteration journal, marking a flawed AI draft, a Socratic persona that only asks questions, and working backwards from an output to the prompt that made it.
- Good prompts need subject knowledge, so learners struggle to prompt well on a topic they have not yet studied.
- Assess the process, not the product: the journal, the critique or the chat log shows the thinking, and a finished essay does not.
What Is AI Metacognition?
Metacognition is knowledge and monitoring of one's own cognition, as described by Flavell (1979). Classroom tools and routines can help learners plan, monitor and evaluate their learning. The Education Endowment Foundation Teaching and Learning Toolkit reports on metacognition and self-regulation approaches.
Their average impact is eight months of additional progress. Teachers use large language models as cognitive mirrors, not just answer engines. The output reflects the quality of the prompt.
So learners can see the gaps in what they understand. Vague questions get generic answers. So learners must analyse and refine how they communicate.
AI can help make thinking visible, but without scaffolding it can also encourage cognitive offloading rather than learning. Focus on using AI to encourage deeper thought, instead of only preventing cheating. Learners need subject knowledge to create useful AI prompts. Prompting AI can help structure a learner's reflection and knowledge.
History teachers can have learners define the Industrial Revolution using classroom chat. If a learner types a vague request, they get a broad summary.
The teacher asks them to find words lacking historical detail. The learner rewrites their request for dates, locations, and inventions. They then monitor communication and build vocabulary (Holmes et al., 2022). The teacher then assesses the better question.
The Research Behind It
Flavell (1979) called metacognition thinking about thinking. Flavell (1979) described metacognitive knowledge, metacognitive experiences, goals or tasks, and actions or strategies. The familiar split into knowledge and regulation came later.
It is a simpler, classroom-friendly version of that research. Learners now use this thinking with digital tools. They must control how they interact with databases. The machine needs the learner to monitor it.
There is a tension between cognitive offloading and cognitive support when using generative AI. Cognitive offloading occurs when learners let the machine do the thinking, reducing their cognitive effort and harming memory retention. Support happens when the machine helps the learner reach a higher level of understanding than they could achieve alone.
The challenge is designing tasks that provide support while preventing offloading. If the AI provides the final product, the learner offloads. If the AI provides a component that the learner must then analyse or improve, the AI supports the learning process.
Molenaar (2022) says learners need guidance to use AI tools well. Learners often seek quick answers and lack self-regulation skills. Hattie (2012) notes the value of feedback. AI offers instant feedback, but learners must learn how to use it. AI provides a safe space to experiment, as it does not judge learners.
For example, a teacher provides a complex physics problem. A learner asks the AI for the answer. The teacher requires the learner to ask the AI to explain only the first step of the equation.
The learner reads the first step, attempts the second step on paper, and then asks the AI to check their working. This transforms the interaction from cognitive offloading into structured cognitive support. The teacher reviews the learner's working and the AI's feedback.
AI in the Classroom
The Prompt Iteration Journal
This strategy makes the metacognitive process of refining language visible and gradable. The teacher provides a complex task and requires learners to track their interactions with an AI tool in a dedicated journal. The focus shifts from the final generated text to the process of generation.
Learners record their initial prompt, copy the resulting AI output, and write a critical reflection on why the prompt succeeded or failed. They then write a second, improved prompt based on their reflection. This forces them to analyse their instructions and recognise how ambiguity leads to poor results.
A learner attempting to generate a summary of the water cycle logs their first attempt as too broad. The AI produces a university-level explanation. The learner writes a reflection noting that they failed to specify the target audience. They refine the prompt to ask for a summary suitable for a 10-year-old using specific scientific terminology. The journal serves as the assessed piece of work.
Critique the Machine
This approach builds critical evaluation skills. It treats AI outputs as flawed drafts, not authoritative sources. The teacher uses AI to generate an essay that is deliberately flawed, biased or incomplete. The teacher then hands this text to the class with a strict marking rubric.
Learners can evaluate AI output by checking it against textbooks and correcting the errors. That is the evaluation level of Bloom's taxonomy (Bloom et al., 1956). They mark the machine's work with rubrics and back each grade with evidence from the text (Wiggins, 1998). This flips the roles. The learner becomes the expert evaluator (Sadler, 1989).
Learners receive an AI biography of Churchill. It has timeline errors and political bias. They check the essay against primary sources. This applies the historical-thinking concepts of evidence and perspective (Seixas, 2004; Wineburg, 2001).
Learners rewrite the conclusion and fix the dates and the bias. That sharpens their critical reading. The teacher then assesses the corrected work and the reasoning behind it.
The Socratic AI Persona
Teachers can use system prompts to change how the AI interacts with learners, turning it into a dialogue partner. The teacher provides a custom prompt that instructs the AI to act as a Socratic tutor. The prompt explicitly forbids the AI from giving direct answers, instructing it to only ask guiding questions.
Learners must converse with the AI to solve a complex problem. Because the AI will not do the work for them, learners must articulate their hypotheses, test their ideas, and respond to the machine's probing questions. This makes their internal thinking visible in the chat log.
A learner struggling with fractions interacts with the Socratic AI. The AI asks what the learner thinks happens when the denominators are different. The learner types their guess.
The AI points out a logical flaw in the guess and asks another question. The teacher reviews the chat log to pinpoint exactly where the learner's understanding broke down. The teacher provides targeted support based on the chat log.
Reverse Engineering
Learners break down good examples to grasp how they work. The teacher shows an AI answer on the board but hides its source.
Learners work backwards to deduce the exact prompt that produced that output. They must identify the tone, the structural constraints, the specific vocabulary and the persona requested. Then they test their guess. They put their deduced prompt into the AI and compare the result with the teacher's original text.
The class reads a structured haiku about photosynthesis that includes specific scientific vocabulary. Learners must write the prompt that they believe created it. They quickly realise that simply typing 'write a poem about plants' does not work. They must refine their prompt to specify the poetic form, the exact biological process, and the required terminology, demonstrating comprehension of the subject matter. The teacher assesses the deduced prompts.
Practice table · Build It · AI metacognition
What it looks like in practice
Nine routines from teachers on social media.
| Routine, as teachers describe it | What you would notice | What the evidence says |
|---|---|---|
| Normalise not using it Name the rule in high-school physics: they already know how learning works, so opening a model does not fit that model.
|
No chatbot in the lesson. The thinking stays on paper and in talk. | This is one teacher's account, not trial evidence. NEU State of Education: AI (2 Apr 2026): 66 per cent of secondary teachers in English state schools agreed critical thinking had declined because of AI, against 28 per cent in primary. The EEF Toolkit +8 months is a mean for plan, monitor and evaluate in the subject, not a finding about a ban. |
| Teacher-mediated only Any AI the class sees is on the teacher’s screen. Learners talk about the output. Nobody prompts alone.
|
A projected window. Hands go up about the answer, not into a private chat. | Wiliam’s reply to Boxer, 28 Aug 2026. OECD PISA 2025 (association, not cause): higher-performing learners tend to be those who do not use AI; AI use is associated with better outcomes when learners are explicitly taught to interpret and assess AI-generated information. Those are two findings, not one comparison. Not a trial of mediation. |
| Teachers yes, learners no Staff may plan with a model. Learners produce in the room. Unsupervised writing is not the product.
@daisychristo |
The split is named in the staffroom. Take-home essays are treated as untrustworthy. | ResearchEd, 5 Sep 2026. Teachers split on whether to keep AI away from learner work or teach its use while changing homework. @paulhaigh (20 Jan 2025): teach them to use it and change homework, because workplaces will. @EnserMark (14 Jan 2025): the goal of homework is thinking, not a finished file. |
| Blank-column planning A human summary of a research book, with a blank column the teacher fills. Teacher-side, not a learner chatbot.
|
Two columns. The right one is handwritten. The photograph in the post is redrawn elsewhere, not republished. | Cogsci Book Summaries. “It's the interaction in making the summary that is important.” “No AI was employed at any stage.” EEF Teacher Choices on ChatGPT for KS3 science planning saved 25.3 minutes a week (56.2 vs 81.5). NFER: some teachers used the time for other teaching tasks; others reduced overall workload. |
| Judge the tool by thinking Before adopting a tool, name the thinking it forces. How human the voice sounds is not the test.
|
A staff conversation about the task, not about fluency. | Bakhtiari, Suriano and Mario Grasso (2026), the Frontiers paper the post links, found a statistically significant advantage for a GPT-4o avatar on knowledge acquisition (M=10.63 vs 8.92, p=.036). The tweet’s “Little or no impact” line is the post’s claim and is contradicted by that paper. The staff question still stands: judge the tool by the thinking it forces. Not a trial of this staff question. |
| Talk, not a prompt lesson Oral defence in the room. Learners explain, challenge and rebuild the idea out loud.
|
Talk, not a chatbot window. The product is a spoken account the class can hear. | Hebbes, Tes, 7 Apr 2026: talk is how humans explore; AI is the exploit engine. No study has tested it. Teachers almost never posted a lesson where learners judged an AI answer and it worked. |
| Explain-back, bot closed A fluent homework answer is not accepted until the learner can say it in their own words with the bot closed. A US post; it does not name a country or role beyond
@MariaPeiro305 |
A learner who wrote a finished file going silent when asked why. Handwriting alone is not the product. | A US post. @TheMayeflowers is the closest first-person account. No study has tested it. |
| Restrict the newest model The newest agentic model is not allowed for learner work. US high-school history, Cutler’s own classroom rule.
|
The restriction is named. This is one teacher’s rule, not a school-wide policy. | Medium, 7 Sep 2026: “as far as my classroom teaching goes, Astra isn't allowed.” Edutopia, 10 Jul 2026: at his school it is mostly up to the individual teacher; his own class prohibits AI unless he notes otherwise. Not a school policy. |
| Literacy crowding Gatsby objection The objection is that AI literacy is an extra requirement with nothing cut. US high school English.
|
A module that replaces nothing. Full books lose a week. | NYC Public Schools guidance: two 45-minute modules, not a course. The teacher’s post said “coursework”; the city page says modules. Name the gap. Not a trial of literacy modules. |
The EEF +8 months figure is the Toolkit mean. These routines are teachers' own accounts, not trial evidence.
A ten-minute routine: judge the answer before you read it
Use this routine when you want the thinking to stay with the learner rather than the chatbot. It takes ten minutes, needs one projected screen and a mini whiteboard each, and it works in any subject where an explanation can be right or wrong.
Minutes one to three: give the class a question you have already taught. Before anyone sees an AI answer, every learner writes what a good answer must contain: two or three things it has to say, and one mistake it might make. That prediction is the metacognitive move. It makes learners draw on what they know before the machine speaks.
Minutes four to six: show the chatbot's answer on the board. Learners score it against their own list. Did it include the points they predicted? Did it make the mistake they expected, or a different one? Take three answers from hands, not shouts.
Minutes seven to nine: pick one flaw the class found and improve the prompt together. Type the revised prompt and run it. Learners compare the two outputs and say in one sentence what changed and why.
Minute ten: each learner writes one line in their book. "The AI missed X, so next time I will ask for Y." That line is your assessment evidence. It tells you more than the AI text does.
Why the order matters. A chatbot writes fluent text whether or not it is right, and it will not flag its own errors. If learners read the answer first, they judge it against the answer, and the answer looks confident, so it wins. If they predict first, they judge it against their own knowledge, and the gaps become visible. This is the same offloading test the research section describes: the AI output is a part to improve, never the finished product.
Variations. In Year 7 science, the prediction is a checklist of key terms the answer must use. In GCSE history, it is the two perspectives the answer must include and the date it must get right. In a primary class, keep the whole routine on the board and do the prediction out loud together, which is the same move as the primary answer in the questions below.
Common Misconceptions
The main misconception is that AI is just an automated answer engine. That view assumes the technology only exists to fetch facts. In reality, AI is a cognitive mirror. It reflects how clearly the user is thinking.
Treat it only as a source of facts and you get cognitive offloading and shallow learning. Teachers must reframe the tool as a conversational partner. It needs careful direction and constant monitoring.
Schools wrongly focus on stopping cheating. Too much emphasis on plagiarism detection ignores the job of integrating AI into teaching. Assess the prompt, the critique or the chat alongside the finished work.
Change the assessments instead. Assess the process: the prompt, the chat log, or the learner's critique of the AI output. That fits the wider case for assessing learning processes rather than just final products (Sadler, 1989).
Prompt engineering needs more than just typing. Good prompts require subject knowledge. Learners should set constraints and verify AI responses when using generative AI. The learner’s subject vocabulary also matters, and learners predictably struggle to write good prompts on topics they have not yet studied, because they lack the schema needed to interrogate the AI output (Willingham, 2009).
AI gives fast text, but it lacks teaching insight (Miao and Holmes, 2023). Teachers build strong learner relationships, not machines (Luckin et al., 2016). Guide learners to verify AI outputs, identify bias and correct errors (O'Neil, 2016).
Consider, for instance, learners using AI for geography work. Instead of punishing them, change the task. Now, learners use AI to create counterarguments (Holmes, 2024). Learners must write essays, addressing and refuting these AI points. Assess the essays and AI argument refutation skills.
Worked Examples by Subject
Maths: Explaining the Steps
Maths teaching benefits when learners focus on the process, not just the answers. AI can act like an interactive textbook that explains methods. Learners can ask the AI about the rules of the quadratic formula, rather than just solving equations.
Learners read the AI explanation and then solve equations. They annotate each written step, stating which rule they applied. Teachers review annotations alongside the AI to check learner comprehension (Chi et al., 1981; Ericsson & Simon, 1980; Sweller, 1988).
English: Structural Analysis
Learners find identifying persuasive techniques tough. Teachers can use AI to create texts for analysis. Learners instruct the AI to write a speech arguing against uniforms, using the rule of three and rhetorical questions.
The teacher provides a highlighter code. Learners highlight the persuasive devices used by the AI within the generated text. They then write a short evaluation assessing whether the machine applied these devices effectively or clumsily. This forces the learner to move from reading for content to reading for structural mechanics. The teacher assesses the highlighted text and the evaluation.
Science: Hypothesis Testing
Learners design plant growth experiments with varied light. They list the equipment and the method, then check for flaws. In inquiry science, learners should spot variables, confounds and missing controls (Hofstein & Lunetta, 2004).
The AI can point out that the prompt did not specify the temperature or the volume of water. The teacher monitors as learners adjust their physical experiment and rewrite their method based on this critique. This interaction models the peer-review process essential to scientific research. The teacher assesses the revised method and the learners' explanations of the changes.
History: Bias Detection
Learners should grasp history's subjective nature. They ask AI to explain the First World War's causes. AI offers British and German viewpoints (Wineburg, 2001; Seixas, 2004). This helps learners compare national narratives (Lee & Ashby, 2000).
Learners compare AI outputs with primary sources, as prompted. They summarise how perspective changes the history. This builds on the historical-thinking work on perspective and evidence (Lee & Ashby, 2000; Seixas, 2004). Learners note the AI's errors, which strengthens critical reading of every text. The teacher assesses the summaries and how well learners spot bias.
Coalface · Build It · AI metacognition
What teachers did
Five classroom examples show learners planning, mapping ideas, restarting, comparing drafts and checking AI output against the text.
1 · Process
1 · The summary
2 · Blank column
3 · You fill it on Monday
4 · No AI was employed
A physics teacher posts a two-column planning sheet. The thinking is the filling-in, not the fluent left-hand text.
The page says the interaction in making the summary is what matters, and that no AI was used: Cogsci Book Summaries.
2 · Mind map
ChatGPT
2 · PE progression
3 · word banks
4 · model texts
5 · information texts
6 · reading level
7 · maths fluency
A Suffolk deputy lists teacher-side uses of ChatGPT. None of them is a learner prompting a tutor.
3 · Cause and effect
The stall · A finished file not an argument
2 · AI-able homework
What unstuck it · The teacher generates the miss
4 · Critique on the teacher's terms
Homework that asks for an argument is still thinking. If there is value in critiquing an AI answer, the teacher should generate the exemplar so it contains the miss they want named.
The page is Why children shouldn’t be using AI to do their homework.
4 · Before and after
Before · The model does the thinking
After · Astra isn't allowed his classroom rule
A high-school history teacher restricts the newest agentic model because it now does more of the thinking. That is his classroom rule, not a school-wide policy.
The teacher's classroom rule is: “as far as my classroom teaching goes, Astra isn't allowed.”
5 · Process
1 · ChatGPT said my answer was right
2 · Check the text page number, not a chat
3 · never using that garbage again
Half points back if they can explain the actual answer. The judgement happens against the source, with the bot closed.
Links to Other Theories
Universal Thinking Framework
The Universal Thinking Framework organises learning with clear actions. AI can help with the "create" stage. For example, AI can produce brainstorming ideas. Learners then evaluate those ideas (Bloom, 1956), refining the output. This supports learning without a blank page (Anderson & Krathwohl, 2001).
Webb's Depth of Knowledge
Webb's Depth of Knowledge sorts academic tasks by how cognitively demanding they are. Using AI to generate a list of facts is a Level 1 task: basic recall. Critiquing an AI essay against a marking rubric pushes the task to Level 3, Strategic Thinking, or Level 4, Extended Thinking.
So AI becomes a tool for reaching higher cognitive demands. The technology provides the baseline information instantly. That frees up lesson time for analysis.
Self-Regulated Learning
Self-regulated learning requires learners to plan their approach, monitor their progress, and reflect on their outcomes. Engaging with an AI chatbot demands all three phases. Learners must plan their prompt carefully, monitor the AI output as it generates, and reflect on how to adjust their strategy if the answer is inadequate. The AI acts as a feedback mechanism that only responds well to self-regulation.
For example, a teacher explicitly maps an AI task to Webb's Depth of Knowledge on the whiteboard. They tell the class that generating a summary of a novel is a Level 2 task. They explain that finding logical flaws in that summary and rewriting it to improve the academic tone is a Level 3 task. Learners actively aim for Level 3 by annotating the AI text with corrections, visibly tracking their cognitive depth. The teacher assesses the annotated text and the learners' explanations of their corrections.
Teachers on social media pushed back on this idea, and the objections below are worth reading before you plan anything. Each one is a real post, quoted as written, followed by what the replies and the evidence say. Use them to anticipate the staffroom conversation.
Putting it into practice · Build It · AI metacognition
The questions teachers raise
Use the questions below to decide which thinking learners must complete before AI can support the task.
question
“Students using AI for educational purposes is neither desirable nor inevitable.”
What the replies and the evidence say. Dylan Wiliam replied on the same thread: no learner-facing AI that is not mediated by the teacher. The replies do not all agree. Paul Haigh, answering Enser in January 2025, asked why not teach them to use it and change homework, because workplaces will. The live UK split is learner-facing versus teacher-facing, not a consensus ban.
question
“Kids will use LLMs to write the coursework, outsource their cognition, and you will accelerate the reverse Flynn effect”
What the replies and the evidence say. @AP_Teaching, in the same thread: “any AI-able coursework isn't worth the hard drive it is etched onto.” The same post says GCSE Computer Science dropped the controlled assessment because answers leaked online. The page does not prove Lucy Powell used the phrase “bring back coursework”; that wording is Daisy’s. The direction, more coursework, is what the replies are fighting.
question
“There is zero evidence that this is reducing workload. If anything it's increasing it because you have to check it.”
What the replies and the evidence say. The next line in the same post: if you do not check it, you have not planned it. YouGov for Bett, reported 27 Aug 2026: teacher AI use doubled, and only about a third of users said they were working fewer hours. The EEF ChatGPT lesson-prep trial saved 25.3 minutes a week. NFER: some teachers used the time for other teaching tasks; others reduced overall workload. Time saved is not hours back for everyone.
question
“Oh no. AI tutors for pupils with free school meals.”
What the replies and the evidence say. GOV.UK, 26 Jan 2026: tools for up to 450,000 disadvantaged learners by the end of 2027. NEU State of Education: AI: 49 per cent of teachers opposed the plan and 14 per cent supported it. Sal Khan told Chalkbeat that Khanmigo was a non-event for a lot of learners; the Tennessee middle-school trial found they messaged it on about a third of Khan Academy days. Do not sell an AI tutor as closing the gap.
question
“@NYCSchools' new AI policy says “all high school students will complete AI literacy coursework” and I'm asking what we are supposed to stop doing”
What the replies and the evidence say. NYC Public Schools guidance: two 45-minute modules, not a course, plus a moratorium on learner-facing generative AI in 2-K to 8th grade. The teacher’s post said coursework; the city page says modules. A colleague in the same thread asked what the policy even means. A US article that only says “teach AI literacy” will bounce unless it names what comes out.
In ten minutes you will see who is checking and who is copying, which no chat log will tell you.
Frequently asked questions
How do we prevent cognitive offloading?
Design tasks where the AI output is the starting point of the lesson, not the final product. Ask learners to critique, improve, format, or apply the generated text rather than simply submitting it. If the assignment requires learners to evaluate the machine's logic, they cannot offload their thinking; they must engage with the material to complete the task.
Can primary school children use AI metacognitively?
Yes, provided there is teacher support. In primary settings, the teacher should act as the sole driver of the AI, projecting the interface on the board. The class collectively decides what prompt to type, and they evaluate the response together. This models the metacognitive process out loud, teaching young learners how to question information before they interact with technology independently.
What if the AI gives incorrect information?
Inaccuracies offer learning chances. Teachers, ask learners to check AI facts using textbooks. This builds confidence and shows the need for human checks. Learners find errors in AI, using reliable sources.
Does prompt engineering take time away from subject content?
Prompting well needs subject knowledge. Learners need biology knowledge before they can specify useful details. Precise language expresses that subject knowledge. This is curriculum communication, not a separate technology skill.
How should we assess work that involves AI?
Assess the process of learning, not just the final product. Grade the Prompt Iteration Journal, the learner's written critique of the AI output, or the chat logs showing their line of questioning. This shifts the focus from grading a potentially machine-written essay to grading the learner's documented metacognitive process.
For example, a teacher addresses the fear of incorrect information by creating a weekly challenge called Spot the Inaccuracy. The teacher projects an AI text containing one deliberate historical or scientific error. Learners race to find and correct the error using their textbooks, turning a software flaw into an active learning game that demands close reading and factual verification. The teacher awards points for the first correct answer.
Project an AI chatbot on your whiteboard for the next lesson. Use a vague prompt about your topic. Ask learners to explain why the chatbot's answer is poor. This active learning strategy, suggested by Miao and Holmes (2023), develops critical thinking and is consistent with the wider literature on metacognition and evaluative expertise (Sadler, 1989).
Related reading: A Diamond 9 CPD lesson for ranking classroom AI uses
Authoritative guidance on AI in education: DfE guidance on generative AI in education, DfE support materials for using AI in education settings, Ofsted findings from AI early adopters in schools and FE, EEF Teacher Choices trial on ChatGPT in lesson preparation.
For the wider picture, explore our AI and EdTech tools hub, our home for evidence-based AI guidance across policy, lesson planning, and classroom practice.
See also our guide on AI chatbots and self-regulated learning.
References
Anderson, L. W. and Krathwohl, D. R. (eds) (2001). A Taxonomy for Learning, Teaching, and Assessing: A Revision of Bloom's Taxonomy of Educational Objectives. Longman.
Bakhtiari, E., Suriano, R. and Mario Grasso, G. (2026). AI avatars in education: effects on knowledge acquisition, academic motivation, and cognitive load. Frontiers in Psychology, 17, 1860755.
Bloom, B. S. (ed.), Engelhart, M. D., Furst, E. J., Hill, W. H. and Krathwohl, D. R. (1956). Taxonomy of Educational Objectives: The Classification of Educational Goals. Handbook I: Cognitive Domain. David McKay.
Chi, M. T. H., Feltovich, P. J. and Glaser, R. (1981). Categorization and representation of physics problems by experts and novices. Cognitive Science, 5(2), 121 to 152.
Christodoulou, D. (2014). Seven Myths About Education. Routledge.
Education Endowment Foundation (2024). Choices in EdTech: Using generative AI (ChatGPT) for KS3 science lesson preparation. Teacher Choices trial. Education Endowment Foundation.
Education Endowment Foundation (2026). Metacognition and self-regulation. Teaching and Learning Toolkit. Education Endowment Foundation.
Ericsson, K. A. and Simon, H. A. (1980). Verbal reports as data. Psychological Review, 87(3), 215 to 251.
Flavell, J. H. (1979). Metacognition and cognitive monitoring: A new area of cognitive-developmental inquiry. American Psychologist, 34(10), 906 to 911.
Hattie, J. (2008). Visible Learning: A Synthesis of Over 800 Meta-Analyses Relating to Achievement. Routledge.
Hattie, J. (2012). Visible Learning for Teachers: Maximizing Impact on Learning. Routledge.
Hofstein, A. and Lunetta, V. N. (2004). The laboratory in science education: Foundations for the twenty-first century. Science Education, 88(1), 28 to 54.
Holmes, W. (2024). AIED, coming of age? International Journal of Artificial Intelligence in Education, 34(1), 1 to 11.
Holmes, W., Porayska-Pomsta, K., Holstein, K., Sutherland, E., Baker, T., Buckingham Shum, S., Santos, O. C., Rodrigo, M. T., Cukurova, M., Bittencourt, I. I. and Koedinger, K. R. (2022). Ethics of AI in education: Towards a community-wide framework. International Journal of Artificial Intelligence in Education, 32(3), 504 to 526.
Lee, P. and Ashby, R. (2000). Progression in historical understanding among students ages 7 to 14. In P. N. Stearns, P. Seixas and S. Wineburg (eds), Knowing, Teaching, and Learning History: National and International Perspectives (pp. 199 to 222). New York University Press.
Luckin, R., Holmes, W., Griffiths, M. and Forcier, L. B. (2016). Intelligence Unleashed: An Argument for AI in Education. Pearson.
Miao, F. and Holmes, W. (2023). Guidance for Generative AI in Education and Research. UNESCO.
Molenaar, I. (2022). Towards hybrid human-AI learning technologies. European Journal of Education, 57(4), 632 to 645.
O'Neil, C. (2016). Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy. Crown.
Sadler, D. R. (1989). Formative assessment and the design of instructional systems. Instructional Science, 18(2), 119 to 144.
Seixas, P. (ed.) (2004). Theorizing Historical Consciousness. University of Toronto Press.
Sweller, J. (1988). Cognitive load during problem solving: Effects on learning. Cognitive Science, 12(2), 257 to 285.
Wiggins, G. (1998). Educative Assessment: Designing Assessments to Inform and Improve Student Performance. Jossey-Bass.
Willingham, D. T. (2009). Why Don't Students Like School? A Cognitive Scientist Answers Questions About How the Mind Works and What It Means for the Classroom. Jossey-Bass.
Wineburg, S. (2001). Historical Thinking and Other Unnatural Acts: Charting the Future of Teaching the Past. Temple University Press.
