There Is No Universal Best Interface

Chapter 21 hands every reader the controls: declare depth, form, explanation order, evidence posture. The tempting next step writes itself — watch which controls each person sets, learn their type, and serve it automatically. Diagram people get diagrams. Detail people get detail. The interface learns your kind and stops asking.

This chapter destroys that shortcut and replaces it with something better. Not with “people are all the same” — they manifestly differ — but with an evidentiary standard for what kind of difference personalisation may act on. The attack has one exact target:

A person possesses a reasonably stable representational type, and matching instruction/interface modality to that type reliably improves task performance.

Everything else — preferences with agency value, strategies chosen per task, affordances of materials, expertise effects, accessibility constraints — survives. The type does not.

The weapon: the crossover interaction

Pashler, McDaniel, Rohrer and Bjork (Psychological Science in the Public Interest, 9(3), December 2008 — commissioned specifically to assess learning-styles evidence for educational use) set the criterion this chapter wields throughout. Correlational comfort (“visual learners prefer visuals”) proves nothing about benefit. The required pattern is a crossover interaction:

PERSON TYPE A:  representation A > representation B
PERSON TYPE B:  representation B > representation A

Each group must perform better in its matched condition — not merely like it, not merely differ overall. Pashler et al. found no adequate evidence base for deploying learning-style assessments, and the criterion has governed every serious test since, including the meta-analysis that most complicates the simple “myth” story. Clinton-Lisell and Litzinger (2024, content-read for this book in Chapter 2: 21 studies, 101 effect sizes, N=1,712) found a small aggregate matched-instruction effect (g ≈ 0.31) — but only ~26% of outcome measures showed the crossover the hypothesis actually requires, study quality was generally weak, and the authors themselves urged extreme caution and rejected adoption. Hattie and O’Leary (2025, Educational Psychology Review, DOI 10.1007/s10648-025-10002-w) sharpen the picture further: across meta-analyses specifically testing matching, they report overall effects around d = 0.04, with larger apparent effects coming from correlational work that collapses distinct concepts — styles, preferences, strategies — into one blurred category. The aggregate number is not the finding. The crossover frequency is the finding, and it is damning at exactly the strength honest science requires: not “zero everywhere” but “far too rare, weak, and low-quality to base a system on.”

Six things that are not the same thing

The conflation Hattie and O’Leary diagnose becomes this chapter’s central taxonomy — the instrument Part III uses from here on:

STYLE          "I am a visual learner."
               stable person classification; matching predicts performance
               → the attacked claim; evidence inadequate

PREFERENCE     "I like diagrams."
               legitimate user choice (Chapter 21's agency value)
               → no automatic performance claim

STRATEGY       "For this problem I draw a diagram."
               behaviour selected for a task
               → task-dependent, person-independent in principle

AFFORDANCE     "This topology is easier to perceive as a graph."
               property of representation × material (Chapter 5)
               → true regardless of who reads

EXPERTISE      "This explanation helps a novice but is redundant for an expert."
               person × domain knowledge × treatment
               → the licensed conditional (see below)

ACCESSIBILITY  "I require captions / screen-reader-compatible structure."
               constraint, not learning style
               → honoured unconditionally, never "matched"

Collapsing these into “learning style” manufactures apparent evidence while destroying the distinctions adaptation needs. Preference stays legitimate as control — Chapter 21 stands. Strategy and affordance explain most of what looks like person-fit without any person-type. Accessibility is non-negotiable and non-inferential. And expertise carries the chapter’s positive result.

The replacement: expertise reversal

The expertise-reversal effect demonstrates everything the type model promised — different people benefiting from different interfaces — with the conditioning variable being knowledge rather than identity. The 2025 meta-analysis (60 experimental studies, 5,924 participants, 176 effect sizes; Learning and Instruction, DOI 10.1016/j.learninstruc.2025.102142, verified in depth): lower-prior-knowledge learners benefit from higher instructional assistance (d = 0.505), higher-prior-knowledge learners do better with lower assistance (d = −0.428), moderated by assessment type, educational status, and domain — asymmetric (helping novices matters more than unburdening experts) and well-confirmed enough to guide instruction with stated caveats. No visual/auditory identities anywhere: the interaction is prior-knowledge × assistance, measurable per domain, changing as the person learns — today’s novice treatment becomes tomorrow’s redundancy. Genuine person × treatment interactions exist elsewhere too (e.g. memory-strategy skill and working-memory capacity interacting with multimedia conditions; Seufert et al., Learning and Instruction 19(1), 2009), so the chapter refuses the mirror-error that people never differ. Its principle is stronger than denial:

Require evidence for the particular person × treatment relationship. Do not infer it from a global type label.

The discriminating experiment

EXP-22 gives the stable-style model a fair fight rather than executing it by design. Declared presentation preferences recorded first; then four materially different tasks (topology/dependencies, exact numerical comparison, temporal sequence, qualified argument) each counterbalanced across prose/diagram/table/timeline; comprehension, transfer, decision accuracy, time, confidence, and preference measured separately. Four rival models compete: H1 stable-style (preference × representation predicts performance consistently); H2 task-fit (structure × representation wins regardless of person label); H3 expertise (prior knowledge × support level); H4 preference (declared liking predicts satisfaction/effort more than performance). Results recorded per person per task — never as person-label tables — so conditional evidence accumulates without manufacturing categories. The chapter embraces its most likely discovery in advance: preference stable while performance effects vary — diagrams liked throughout, prose winning qualification-heavy material, tables winning exact comparison. That vindicates Chapter 21 (control has agency value) while establishing that preference authorises no performance claim. No contradiction; two different objects, finally separated.

What this chapter earned

Stable modality-based types are an inadequate basis for adaptation — the crossover bar stands, the evidence falls short, and the conflation behind apparent revivals is named. In their place: preference as control, strategy and affordance as task properties, accessibility as constraint, and expertise as the demonstrated conditional — with the evidential rule that every person × treatment claim must be earned particularly, never inherited from a label. What remains open is learning: observations now exist (task, representation, preference, performance, correction, prior knowledge) with nowhere yet to accumulate.

Person → fixed type → fixed interface is dead. What can repeated interaction learn without resurrecting it as personality?

References

• Pashler, H. et al. (2008). Learning Styles: Concepts and Evidence. Psych. Sci. Public Interest, 9(3). Verified (venue/criterion): crossover-interaction standard; inadequate base for deployment.
• Clinton-Lisell & Litzinger (2024). Front. Psych. 15:1428732. Content-read (Ch 02): g≈0.31, ~26% crossover, weak quality, authors reject adoption.
• Hattie, J. & O’Leary, T. (2025). Learning Styles, Preferences, or Strategies? Ed. Psych. Rev. DOI 10.1007/s10648-025-10002-w. Verified citation-level: matching-specific d≈0.04; style/preference/strategy conflation thesis (numbers as cited, full-text deep read scheduled at Part III pass).
• Expertise-reversal meta-analysis (2025). Learning and Instruction, DOI 10.1016/j.learninstruc.2025.102142. Verified in depth: 60 studies/5,924 participants/176 effects; d=0.505/−0.428; asymmetry; moderators.
• Seufert et al. (2009). Memory characteristics and modality: ATI study. Learning and Instruction 19(1). Verified citation-level: memory-strategy/WM-capacity interactions exist — against over-denial.

Proposed experiment EXP-22: four rival models

Status: PROPOSED. Per the design above (declared preferences; four task types × counterbalanced representations; six separate measures; H1–H4 competition; per-person-per-task records, no type labels; fair-chance H1). Failure criteria: H1 wins consistently (type model vindicated — kept as live risk); no model separates (tasks insufficiently distinct); preference predicts performance better than task-fit (conditional thesis collapses). Artifacts: preference records, task instruments, per-cell outcomes, model-comparison analysis. What success would not justify: inferring types from H2/H3 wins — conditional effects are not identities, and Chapter 23 must learn them without categorising people.