We use only the storage this website needs to work, such as remembering light or dark mode. We do not use analytics or advertising cookies. If we ever add them, we will ask you first, and nothing optional runs unless you accept.
Why learning data matters now in UK education and training, published studies of AI tutoring, and the research behind each teaching method the tutor uses. None of these studies tested Nested Learning.
Across UK education and training, providers are asked to show what learners gain, while skills gaps widen and funders want proof. Each fact below links to its official or published source.
Skills gaps
Employers and local areas report skills gaps, with digital skills among the most discussed.
An AI-assisted analysis of the 38 local skills improvement plans in England found digital skills development and apprenticeship programmes among the ten most common themes.
Why it matters for learning data: A tutor taught from an organisation's own materials can give every learner structured practice on the skills that organisation needs.
Apprenticeships: growth, and the evidence funders ask for
Apprenticeships are growing, and providers must evidence every hour of off-the-job training or risk losing funding.
Apprenticeship starts rose 8.7% to 308,770 in August 2025 to April 2026, and participation rose 4.6% to 735,500, compared with the same period a year earlier (provisional figures).
Providers must hold quantifiable evidence of off-the-job training that meets the official definition, which includes online learning and simulation exercises (paragraphs 78 and 91).
“The evidence must be quantifiable and must meet the OTJT definition”
In 2019 the National Audit Office reported that the funding agency had limited assurance that apprentices received their off-the-job training (then set at 20% of their time).
“The ESFA has limited assurance that apprentices are spending at least 20% of their time on off-the-job-training.”
Ofsted does not review providers' internal data, but inspectors may discuss with leaders how they use assessment data and what it tells them about apprentices' achievement.
“Inspectors do not review providers' internal data. However, they may discuss with leaders how they use internal assessment data”
Why it matters for learning data: Every tutor session is recorded with its date, the scores for each turn and a summary. Where the training plan includes it and it meets the off-the-job definition, that record can support a provider's own evidence of learning. It does not replace the provider's evidence pack or its judgement.
Higher education: showing learning gain, under recruitment pressure
Universities are asked to show what students gain, while their finances depend on uncertain recruitment.
In TEF 2023, the OfS noted there is no national measure of educational gain; a review of 51 Gold-rated providers' submissions found measuring gains and students' distance travelled to be complex.
“We recognise that there is currently no national measure of educational gain”
The revised TEF will rate student experience and student outcomes, and the OfS is consulting on what evidence the assessments will use; providers rated Requires improvement may face student recruitment limits.
“including what evidence would be used in the assessments”
UK student entrants rose 3.5% in 2024-25 but were 8.6% below the sector's forecast, and 42.7% of institutions could report deficits in 2025-26 without mitigating action.
Why it matters for learning data: Turn-by-turn analytics give staff a record of how students' understanding develops during a module: one form of evidence of learning alongside the measures the OfS already uses.
Workplace learning: justifying the spend
Learning teams are asked to show the value of training, and many find it hard.
16% of learning and development professionals named struggling to show return on investment as a barrier, and 12% said they had no way of knowing which efforts are most effective (1,108 respondents).
Why it matters for learning data: Analytics show what each learner understood, not just that they completed a course, so managers can set evidence of learning beside what training costs.
AI tutoring: policy and the market
Government and the market both expect AI tutoring to grow, with safety at the centre.
The government invited edtech and AI companies to build safe AI tutoring tools that could benefit up to 450,000 disadvantaged pupils a year in Years 9 and 10, with tools available to schools from 2027. Nested Learning is not part of this programme.
“All tools must meet rigorous UK safety standards, align with the national curriculum, and be co-designed with educators.”
A commercial estimate puts the UK AI tutors market at US$123.8 million in 2025, forecast to reach US$924.7 million by 2033 (28.6% a year from 2026). It is a market-research estimate, not official data.
Why it matters for learning data: Safety, alignment with the curriculum and teacher control are what buyers will be asked to show; they are built into how the Nested Learning tutor works.
Published research
Peer-reviewed studies of AI in education
Six peer-reviewed studies and reviews of AI tutors and of how teachers and learners work with AI. They are studies of AI in education in general, not of Nested Learning.
Randomised controlled trial University students
AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting
Students learned more, in less time, from an AI tutor built on research-based teaching practice than from an in-class active-learning session, and reported higher engagement and motivation.
What it does not show: It studied a different AI tutor, not Nested Learning.
Systematic review 28 studies; 4,597 school students
A systematic review of AI-driven intelligent tutoring systems (ITS) in K-12 education
Létourneau, Deslandes Martineau, Charland, Karran, Boasen and Léger (2025). npj Science of Learning 10, 29
Intelligent tutoring systems generally improved school students' learning and performance, though the gains were smaller when they were compared with non-intelligent tutoring systems.
What it does not show: The studies were quasi-experimental and mostly short; the authors call for longer, larger studies. Not a study of Nested Learning.
Digital tools improved learning (g = 0.65). Intelligent tutors and simulations did better than other tools, and effects were larger when tools supported teaching rather than replaced it.
What it does not show: It covers digital tools in secondary maths and science, not Nested Learning.
Multiple case study 4 real-world educational use cases
Different lenses on human-AI collaboration: towards a comprehensive view on hybrid intelligence in education
Horvers, Molenaar, Snoeijen, Costache, van der Graaf, Kramer, de Mooij, Topali, Vermeire and Sleegers (2026). Behaviour & Information Technology (published online 7 October 2026)
AI took over bounded analytical tasks, while teachers interpreted, checked and acted on what it produced. The authors conclude AI may redistribute tasks between teachers and AI “without displacing teachers' pedagogical judgement”.
What it does not show: Four cases, analysed qualitatively; it does not measure learning outcomes or study Nested Learning.
Longitudinal multiple-case study over seven weeks 6 university English learners (3 focal cases)
Understanding development of learner agency in GenAI-assisted oral argumentation through the critical GenAI literacy-based dialogic teacher feedback
Li, Chen and Sun (2026). Studies in Higher Education (published online 12 October 2026)
Learners got better at prompting generative AI, but their ability to judge the content and logic of its output developed less consistently. Learner agency developed “in differentiated and nonlinear ways”, and sustained, responsive teacher feedback mattered.
What it does not show: Six learners in one English-language setting; qualitative, so it shows patterns rather than effects. Not a study of Nested Learning.
Each learning mode draws on research about a teaching method, mostly with human teachers. These studies did not test AI or Nested Learning.
Why tutoring
Bloom (1984), The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One-to-One Tutoring. Educational Researcher 13(6), 4–16.
Summary of two doctoral experiments School pupils in two studies (Anania; Burke)
Pupils tutored one-to-one, with regular checks and corrective feedback, scored about two standard deviations above pupils taught in a conventional class of about 30. The average tutored pupil did better than 98% of the class.
What it does not show: Small, short studies with school pupils. Later reviews found a smaller effect for human tutoring (VanLehn, 2011). Not a study of AI or of Nested Learning.
VanLehn (2011), The relative effectiveness of human tutoring, intelligent tutoring systems, and other tutoring systems. Educational Psychologist 46(4), 197–221.
Review of experiments Published experiments comparing tutoring with no tutoring
Human tutoring raised scores by about 0.79 standard deviations over no tutoring, well below two. Computer tutors that give feedback on each step of a problem came close, at about 0.76.
What it does not show: Mostly science, maths and technology, and older than today's AI. Not a study of Nested Learning.
Chi and Wylie (2014), The ICAP framework: linking cognitive engagement to active learning outcomes. Educational Psychologist 49(4), 219–243.
Review and framework Laboratory and classroom studies of learning activities
Learning rises as learners move from listening, to doing, to building their own explanations, to dialogue in which each answer builds on the other's reasoning.
What it does not show: A framework built from earlier studies; some later experiments have not found the same order.
Rosenshine (2012), Principles of Instruction: Research-based strategies that all teachers should know. American Educator 36(1), 12–19.
Research summary for teachers Cognitive science, studies of effective teachers and of learning supports
Ten principles, including a short review at the start, new material in small steps, many questions, worked examples, guided practice and frequent checks for understanding.
What it does not show: A summary rather than a new experiment; some principles are broadly stated.
Laboratory experiments 120 university students (Experiment 1)
Students who were tested on a passage remembered more of it two days and a week later than students who spent the same time restudying it, though restudying did better on a test five minutes later.
What it does not show: Short prose passages in a laboratory, with no feedback; not a study of oral examinations or AI.
Gibbs (1988), Learning by Doing: A guide to teaching and learning methods. Further Education Unit, Oxford Polytechnic.
Guide for teachers Not an experiment
Sets out a six-stage reflective cycle: description, feelings, evaluation, analysis, conclusion and action plan. It is widely used in health, social care and teacher education.
What it does not show: A practical model, not a tested intervention: it structures reflection rather than proving gains.
Cook, Hatala, Brydges and colleagues (2011), Technology-enhanced simulation for health professions education: a systematic review and meta-analysis. JAMA 306(9), 978–988.
Systematic review and meta-analysis 609 studies; more than 35,000 trainees
Compared with no training, simulation training was linked to large gains in knowledge, skills and behaviour, and moderate gains in patient outcomes.
What it does not show: Health professions only, and compared with no training rather than other teaching; not AI role-play.
Hattie, Biggs and Purdie (1996), Effects of learning skills interventions on student learning: a meta-analysis. Review of Educational Research 66(2), 99–136.
Meta-analysis 51 studies
Study-skills support worked best when it was taught within the subject and helped students plan, monitor and reflect on their own learning, rather than as general skills on their own.
What it does not show: Covers study-skills programmes in general; not a study of coaching by AI.
Wood, Bruner and Ross (1976), The role of tutoring in problem solving. Journal of Child Psychology and Psychiatry 17(2), 89–100.
Observational study of one-to-one tutoring 30 children aged 3 to 5
Introduced 'scaffolding': the tutor keeps the learner on task, simplifies it, points out what matters and demonstrates, then withdraws support as the learner succeeds.
What it does not show: Young children on a block-building task; it describes tutoring rather than measuring gains.
Published research is encouraging but still young. A 2025 randomised trial found university students learned more, in less time, from a research-based AI tutor than from an in-class active-learning session, and a 2025 review of 28 school studies found intelligent tutors generally improved learning. None of these studies tested Nested Learning.
Which research is each learning mode based on?
Each mode draws on research about a teaching method: Socratic Partner on Chi and Wylie's ICAP framework, Rosenshine on Rosenshine's principles of instruction, Oral Examiner on retrieval practice, Gibbs Reflection on Gibbs' reflective cycle, Simulator on simulation-based training, Academic Coach on study-skills research, Academic Feedback on Evans' review of assessment feedback, and Adaptive Learning on scaffolding and tutoring research.
What evidence of off-the-job training do apprenticeship providers need?
Providers must hold quantifiable evidence that each hour of off-the-job training meets the official definition, according to the apprenticeship funding rules for 2025 to 2026 (paragraph 91). If the minimum is not met, or evidence is not available, funding is at risk of recovery (paragraphs 81.4 and 310).
Does Ofsted look at providers' learning data?
Not directly. Ofsted's further education and skills inspection toolkit (June 2026) says inspectors do not review providers' internal data, but they may discuss with leaders how they use internal assessment data and what it tells them about learners' and apprentices' achievement.
Is there a national measure of learning gain in UK higher education?
No. For TEF 2023 the Office for Students said there is currently no national measure of educational gain, and an independent review of 51 providers' submissions found measuring gains complex. The OfS is now consulting on what evidence the revised TEF will use.
How were these studies chosen?
We cite peer-reviewed or well-established work, link to each source, state its design and sample size beside its finding, and say what it does not show. Every study is labelled as research on AI tutoring or teaching methods in general, not on Nested Learning.
See it working with your own course
A pilot runs on your own course materials, with your learners. Choose a time for a call in our booking calendar.
