The Promises and Pitfalls of Personalisation

How education theory keeps examples valid across contexts

I’m amused that my job as a (mostly) statistician centres on something I disliked at school and during my first degree. Real mathematics felt like solving cool logic puzzles; statistics felt like memorising tedious if–then rules for contrived scenarios about socks in drawers or the heights of children. At school, I didn’t care about socks or height - I loved music and politics. Imagine if I’d learnt statistics solely through examples from music and politics. Would I have loved it sooner? The question matters now because AI promises personalised tutors that connect dry subjects to students’ real interests.

Personalised examples have other potential benefits beyond agency and engagement. Anchoring to familiar contexts may aid initial comprehension and retrieval cues. The tech world is excited. If the education theorists of the last century were alive to see this, they might be rather less so. I hope to show you why.

In this post, I’ll show when personalisation helps, when it inadvertently breaks the very thing we are trying to learn, and how to tell the difference.

What to ask before you personalise

Teachers often develop a strong instinct for when a football story will help pupils grasp means in maths and when it will muddy migration in geography. That intuition is forged through disciplinary training and years of trial and error with examples in the classroom. If AI tutors and learning platforms are to choose suitable examples, we must turn those tacit judgements into an explicit, reusable framework. This is where education theory helps.

The call on whether a personalised example will help or hinder is nuanced, and many educational theorists have valuable things to say. In this post, I’ll focus on three who do most of the heavy lifting for these decisions: Marton, Maton, and Messick. They explain what must stay constant across examples, when context should travel or be constrained, and how to keep our judgments faithful to the thing we intend to teach.

Marton: What must stay the same?

Ference Marton’s Variation Theory claims that learners discern the essence of a concept when they encounter systematic variation in what is critical and relative constancy in what is not. In other words, understanding comes from seeing the critical feature come into relief through contrast, and then, once discerned, seeing it persist as non-critical features vary.

What does this mean for deciding whether examples are appropriate? First, identify the object of learning and its critical aspect(s). In maths, for finding the mean of five numbers, the critical relation is between total and count (e.g., same total/different count → different mean; same count/different total → different mean). Then vary the surface to support generalisation: goals in five matches, minutes of five songs, pages read on five evenings. In chemistry, suppose our concept is conservation of mass in a closed system. Juxtapose otherwise similar setups that differ on closedness (e.g., sealed vs unsealed vinegar with bicarbonate) so learners discern that condition as critical; then range over sealed examples—steel wool rusting in a closed jar, or a precipitation reaction in a stoppered tube—to show that, when closed, the total mass of reactants equals the total mass of products.

Marton’s point is not that context is irrelevant, but that variation should be used deliberately to throw the defining relation into relief and then to demonstrate its invariance across contexts.

In deciding whether a proposed personalised example is likely to help or hinder learning, Marton requires us to ask:

1. What is the critical feature or invariant I want students to notice?

2. Are my variations designed so that the invariant becomes unmistakable?

3. Does my personalisation keep the invariant intact, or does it fiddle with it?

4. Is this a concept where misconceptions or confusions are common, so examples must be crafted with clarity and care?

Maton: Should this travel?

Karl Maton’s Legitimation Code Theory (LCT) helps us understand the relationship between the abstract idea and the context it is set in. In this theory, semantic gravity describes how tightly meaning is tied to a specific context. When gravity is weak (abstract ideas), knowledge tends to travel well across different stories or settings; when gravity is strong (context does the work), changing the setting can quietly change the thing being learned. He sets out how teachers use a semantic wave, by deliberately moving down from abstraction to a concrete example before returning to the principle, thus ensuring students don’t get stuck in anecdotes.

He would ask us to first judge the gravity of our objective. Finding the mean of five numbers has weak gravity: whether the numbers are goals, pages, or song lengths, the rule sum ÷ count holds, so personalisation is safe. By contrast, judging the credibility of a historical source has strong gravity: the context (provenance, date, purpose, audience) is part of the competence. Here, you need to tightly specify what sorts of contextual examples are legitimate.

Maton’s ideas encourage us to ask:

• How strong is the “semantic gravity” here, i.e. does this idea genuinely travel, or is the context part of the concept itself?

• If I dip into a personalised story, do I ride back up and restate the principle, so learners don’t stay in the anecdote?

Messick: Will this add variance we don’t want?

Samuel Messick’s Validity Theory asks us where tasks genuinely reflect on the intended construct and whether unwanted influences - what he called construct-irrelevant variance - have been controlled. This way, the student (and teacher) has clarity about whether they fully understand how to complete the task, without distortions from other topics they may or may not have a knowledge of.

Before you personalise, ask: Will this story introduce variance we do not want? With our anchor, finding the mean of five numbers, the construct is a simple operation of sum ÷ count. A playlist of five track lengths is fine; but if one track is a medley counted as one, or a cup match counts double, you have smuggled in weighting and changed the construct. Likewise, framing the numbers as “your favourite team’s goals” risks motivational and attentional boosts unrelated to averaging - again, construct-irrelevant variance.

Messick would want us to ask:

• Does this personalised context add variance that is not part of the thing we intend to measure or teach?

• Could a pupil who loves this theme/example but cannot do the skill still look good - or vice versa?

• Can we cleanly and quickly judge whether the student knows the construct with a simple yes/no, without muddying it by a rubric-laden context?

Do Google’s personalised examples keep the knowledge intact?

Google’s Learn Your Way is a Labs experiment that takes a standard textbook chapter and lets learners set both the difficulty level and an interest theme (for example, sport). It then re-levels the prose, swaps in interest-aligned examples, and propagates the theme across multiple formats (audio, slides, mind maps, quizzes) so the personalised example appears consistently throughout. Google reports early results from a small, randomised study: students using Learn Your Way scored about 9% higher immediately and 11 percentage points higher on a delayed test than those using a standard digital reader (but note this is not a causal impact of personalised examples but rather of multiple adaptations to the textbook).

Google have certainly succeeded in building the technology to personalise examples, but the question for us is whether their technology has a framework that ensures the personalisations help learning rather than diluting it. Below, I apply the theoretical framework above to three Learn Your Way example cases on their website and ask, for each: does the personalisation keep the knowledge intact, does it aid understanding, and what guardrails would make it safer?

Teaching data structures & algorithms to a middle schooler who likes cooking

On the face of it, this is a good topic for personalisation. Many core ideas in data structures have crisp invariants (Marton) that travel well across contexts (Maton). If we keep the invariant front and centre and avoid kitchen “realism” that adds extra constraints, cooking is a safe and engaging skin. The judgement is binary (right/wrong) most of the time, which lowers the risk of construct drift (Messick).

Let’s see how Google’s actually applied cooking to data structures and algorithms:

In my view, based on the theorists above, three of Google’s cooking examples are solid as written; two need some adjustments to protect the invariant.

1. List → “recipe steps in order”. This example is partly right but needs a tweak. A list is an ordered, indexable structure. A recipe is an algorithm you execute. Blurring storage with execution risks pupils thinking the program is the data.

2. Set → “unique ingredients for the week”. This is a good example because a set keeps membership without order or duplicates. That said, it would be easy to drift into an incorrect example if quantities of the ingredients were added.

3. Map/dictionary → “recipe name ↦ ingredients and instructions”. This is a good example demonstrating look-up by key not by position.

4. Priority queue → “restaurant orders may be prioritised”. This is a fragile example that could easily introduce irrelevant variance because real kitchens imply multiple cooks and fuzzy policies (“might be prioritised”) which shifts the task towards scheduling and policy choice. That adds irrelevant variance to what should be a clean “always take the highest priority next” invariant. The invariant is only unmistakable with a single cook and a clear rule, e.g. “One chef has dishes each with a priority number; the chef always cooks the smallest number next.”

5. Graph → “food-pairing network”. This is a good example because nodes (foods) and edges (pairs that go well) express connections without extra baggage.

Teaching economic systems to a middle schooler who likes football

On the face of it, this is topic where personalised examples might work but need to be crafted with care. Personalisation can help only if the examples keep the defining features of each system in view (i.e. who owns resources, how decisions/prices are made (coordination by prices or by plan or by custom), and how goods and labour are allocated), and avoid football-specific rules that actually change the thing we’re teaching.

Of Google’s four football inserts, one is invalid, two are perhaps usable with a clarifying sentence, and one is solid with minimal context.

1. Youth academy presented as a command economy. This is not a legitimate example. An academy is a private organisation inside a market/mixed system. Internal top-down management ≠ society-wide central planning. This collapses organisational hierarchy into economic system, inviting the rule-of-thumb “any place with a boss = command”.

2. “The international transfer market” as a market. This example is partly legitimate, but fragile. There really is buying and selling, but the thing traded is contract/registration rights, not people; there are heavy regulations (windows, FFP) and it’s not a consumer market. These extras add unnecessary load and blur the criteria for “market”.

3. Superstar wages as “society values output, so income rises”. This is directionally OK, but needs a caveat. Why? It gestures at scarcity and demand, yet invites normative drift (“high pay = high social value”) and ignores market imperfections. It also nudges the task away from classifying systems and towards labour-market debates. A single clarifier would help: “Superstar pay is high because many fans pay for a very scarce skill; that doesn’t mean all valuable work is highly paid.”

4. Fan protests against the European Super League as pushback against globalisation. This is a legitimate example and useful with brief context. It squarely illustrates local/national institutions pushing back against a proposed trans-national structure.

Teaching early human evolution and migration to a middle schooler who likes photography

This topic is a red-light case for content personalisation. Early human evolution and migration is context-heavy and evidence-led. Understanding depends on dated finds, provenance, and integrating multiple representations (maps, timelines, artefact photos). It’s also high-confusability: pupils are easily pulled into appearance-based reasoning and “ladder of progress” myths. Personal interests like photography tend to change the mechanism – in this case from selection without foresight to human design - so personalisation risks distorting what is being learned.

Not surprisingly, Google only managed to create one personalised example for this entire topic. A photography “skin” that isn’t documentary (with scale and provenance) would push learners toward looks-based claims or linear “camera generations” analogies, both of which are invalid for the task. In that sense it’s a relief the system didn’t try to personalise most of the text: better no example than a misleading one.

The one example it did use is inappropriate. Google’s opener treats camera technology moving from plates → film → digital as an analogy for evolution. That swaps natural selection and inheritance for goal-directed design with human agency and deliberate improvement. It also compresses timescales and implies teleology (“things evolve to get better”). Cameras don’t reproduce; lineages do. As framed, the example is memorable but mechanistically wrong, so it primes exactly the misconceptions we’re trying to avoid.

Getting personalisation right

Personalisation has real benefits. Done well, it boosts motivation and helps pupils hook new ideas onto what they already know. A familiar setting lowers the barrier to entry and can make abstract concepts feel meaningful.

But it can also get in the way. Done badly, it invites misconceptions, drags in background knowledge that shouldn’t matter, and distracts from the thing we’re actually trying to teach. Compelling stories are not a substitute for conceptual clarity.

The good news is we’re not guessing. Education theory gives a usable rule-set to help us decide what examples are legitimate. We can personalise when invariants survive (Marton). We should constrain examples when context is the construct (Maton). We should ban examples that add construct-irrelevant variance (Messick).

But back to me and statistics. It’s too late to rescue my own school experience, but what about our children? Could they learn statistics through Harry Potter’s Hogwarts or Fantasy Football League? Often, yes. Much of school statistics has crisp, context-general structures and quick, binary checks, which makes it unusually well-suited to safe personalisation.

We can teach probability through goal-scoring in football or Quidditch. We can sample data on midfielders or on Hogwarts students and draw inferences. We can illustrate Simpson’s paradox by stratifying house-point totals, and Bayes-style updating when new evidence arrives about a player’s midweek training. So long as we keep randomness and independence plausible and avoid “magical determinism”, we’re likely to preserve the invariant while avoiding construct-irrelevant variance, doing so in examples pupils might actually enjoy.

Personalisation isn’t a magic wand and it isn’t a menace. It’s a design choice that either amplifies the signal or buries it. Used with care for the right types of knowledge, AI tutors won’t just feel engaging, they’ll help pupils learn the right thing, for the right reason, at the right time.