You asked whether anyone has validated the way you're speaking right now — detailed, spoken-out reasoning that carries its own prerequisites, like explaining code to a rubber duck, except the speech itself is the work. The answer is yes for the core, and not yet for your exact version.
Explaining to yourself while you learn is one of the better-tested ideas in cognitive science. Your specific twist — every message self-encapsulating, pre-loaded so it can be read cold, out of context, and still explain the whole stack — has never been tested as a named practice anywhere the research could find. It is a plausible recombination of validated parts, not itself a validated technique.
The verdict, in plain English
Talking helps when the talking is explaining — generating reasons, patterns, and connections from memory. Just voicing thoughts without explaining them changes accuracy by essentially nothing. The benefit is in the effortful production, not in the performance of speaking.
And shared vocabulary really does make two minds faster — not as a metaphor, but in counted words and turns. What starts as a full sentence (“a person who's ice skating, except they're sticking two arms out in front”) collapses, once grounded, to two words: “the ice skater.”
Self-explanationg = 0.55
Prepare + teachg = 0.56
Generation effectd ≈ 0.40
Plain think-aloudr = −.03
Expressive writingr = .075
Your exact formatuntested
How to read this page
Every number carries its evidence tag
The research tagged each figure by how solidly it was sourced. Those tags are kept on this page, next to the number — they are part of the finding, not fine print.
[published] In a primary abstract or record the researcher read directly.[reported] Only in secondary summaries — agreeing across sources, but not re-checked against the primary text.[estimated] An arithmetic conversion derived from a published number.[unknown] No published figure was located. Said plainly, not padded.
Method: web-index search plus opening primary abstracts (PsycNet, ERIC, PubMed). Full texts were not systematically read — most effect sizes are from published abstracts or from secondary summaries that agree across multiple sources, tagged accordingly. Research date: 2026-11-19.
The evidence at a glance
Explaining moves the needle. Voicing alone doesn't.
Standardised effect sizes from the meta-analyses. Zero is “no difference from control.” Around 0.2 is small, 0.5 is medium. All bars start at zero.
Prepare + teach and self-explanation: [published] · Generation: [published] · *Expressive writing d ≈ 0.15 is [estimated] — converted from r = .075 [published], 146 randomised studies (Frattaroli 2006). Verbal overshadowing Zr = −0.12 [published] — the one negative bar: verbalising hurts recognition of hard-to-verbalise material. Plain think-aloud (r = −.03, 94 studies) is not drawn — it sits on the zero line.
Claim
Strength
Basis
Explaining (to self or other) improves learning / understanding
Moderate–strong
g = 0.55 [published] self-explanation, 64 reports; g = 0.35–0.56 [published] teaching; converging mechanisms
Generating (vs reading) improves memory
Moderate–strong
d ≈ 0.40 [published], 86 studies
Plain think-aloud is safe — doesn't hurt accuracy
Strong (as a null)
r = −.03 [published], 94 studies
Explaining exposes gaps in understanding
Moderate
Illusion of explanatory depth replicated; no pooled effect size [unknown]
Writing about thoughts improves processing / health
Weak
r = .075 [published], 146 studies; several nulls
The Feynman Technique by name works
Untested
No controlled test located [unknown]
Self-encapsulating messages work
Untested
No published work located [unknown]
The honest overall grade, in the report's words: the family is validated; the specific member is not. Dunlosky et al. (2013) rated self-explanation only “moderate utility” — real effects, lab-heavy evidence, thin classroom and long-term data.
Mechanism
Why explaining helps — five engines, not one
No single mechanism explains it. At least five converge, and different studies isolate different ones. The report's summary: generation is the engine, pattern-seeking is the steering, gap-detection is the dashboard, elaboration is the road-building, and the audience is the reason to drive carefully.
1
Generation / retrieval
Producing content instead of recognising it creates a stronger, more retrievable trace. Koh et al. (2018) suggest much of teaching's benefit is retrieval practice [reported].
2
Pattern-seeking
Asking “why?” constrains you to find a unifying rule that subsumes the cases — Lombrozo's subsumptive constraints. That's why explaining aids generalisation and transfer more than recall of details.
3
Gap detection + repair
Attempting an explanation exposes what you cannot say; good self-explainers then repair the gap. Explaining is a diagnostic instrument first, a learning instrument second.
4
Elaborative encoding
Generated explanations link new material to prior knowledge, creating more retrieval pathways back to it.
5
Monitoring + audience pressure
Explaining to a real or imagined other forces organisation and selection of main ideas. The protégé-effect expectancy finding — benefit accrues during preparation, before any teaching happens — shows the audience works partly as a mental frame, not only as a real listener.
The rubber duck, upgraded
The duck is not in the literature as a tested intervention. Its closest tested analogues are self-explanation and teachable agents — explaining to a non-judgmental listener who makes you organise, select, and notice gaps. An AI that can answer back is a more responsive duck; whether that helps or short-circuits the expectancy mechanism is untested [unknown].
Boundary conditions
When it hurts or fails
Hard-to-verbalise targets
Faces, colours, tastes, voices, spatial configurations: verbalising creates a biased verbal code that overshadows the original memory (Zr = −0.12 [published], small). Rule of thumb from the report: if the knowledge is perceptual and holistic, talk less; if it is conceptual and causal, talk more.
Insight problems — contested
Verbalisation can disrupt the non-reportable restructuring insight requires (Schooler et al. 1993) — but Fleck & Weisberg found no negative effect. Treat this as “possible harm, unresolved,” not settled.
Misleading patterns
Explaining drives pattern-seeking; if the salient pattern is wrong, explaining locks it in — the explanation impairment effect[reported]. Explaining is a pattern-amplifier, not a truth-detector.
Working-memory overload
Working memory holds roughly 4 chunks of novel information (Cowan 2001) [reported]. Long prerequisites and edge-case lists are exactly the kind of added load that could tip elaboration from schema-building into overload — for the reader as well as the writer. Where that crossover sits for this practice is not marked in the literature.
Nothing to elaborate with
Self-explanation and elaborative interrogation both need prior knowledge to connect to. Without it, learners generate plausible-sounding but wrong explanations (Woloshyn et al.) [reported].
No expectancy, no retrieval, no benefit
Teaching after studying without having expected to teach: g = −0.02 [reported]. Explaining from notes: no benefit over restudy. Explaining from memory: benefit (Koh et al.) [reported]. And all verbal reporting increases time to complete tasks [published] — it pays only if the deeper encoding is worth the time.
Your specific version
Self-encapsulating messages: what maps, what's novel
Your practice: every message carries its own prerequisites and edge cases, so any single message, read out of context, explains the stack. Here is how each element maps onto the literature.
Element of your practice
Closest literature
Fit
Explaining the reasoning in the message itself
Self-explanation (Chi; Bisra g = 0.55)
Direct — same act, written instead of spoken
Stating prerequisites before the point
Elaborated code (Bernstein) — principles made explicit, context-independent
Close analogy; Bernstein is sociological, not experimental
Stating edge cases / failure modes
Gap-detection (illusion of explanatory depth); self-explanation as repair
Partial — that literature is about discovering gaps by explaining, not pre-listing them
Any single message readable out of context
Grounding / common ground (Clark & Brennan) — but inverted
Clark & Brennan predict partners minimise joint effort by starting short and repairing as needed — the efficient equilibrium is terse messages plus cheap clarification. Self-encapsulating messages deliberately spend more speaker effort upfront. That is rational only when repair is expensive: asynchronous text, long gaps between messages, high cost of misunderstanding, or a partner whose memory is unreliable.
The literature supports this conditional reading: the principle is about total collaborative effort, and front-loading is the right trade when the acceptance phase is costly or unavailable. Your practice reads as an adaptation to a medium with poor grounding affordances — no copresence, no simultaneity, uncertain common ground — exactly the medium constraints Clark & Brennan list.
What is actually novel — stated as unvalidated inferences, not findings
1 · The unit of encapsulation
The literature studies explanation as a learning act or grounding as a dyadic act. Treating each message as a self-contained, re-readable artefact — built to survive being read out of context by a future self or a stateless model — is a documentation use of explanation no study found tests directly.
2 · Pre-loading prerequisites
Self-explanation research has learners explain given material. Actively anticipating what a reader would need and supplying it unprompted is audience design taken to an extreme — studied as a phenomenon, not as an intervention with measured effects [unknown].
3 · The dyad is human + AI
Every grounding and entrainment study here is human–human. Whether conceptual pacts and alignment work the same way with an AI partner — which does not forget, does not tire, but does not genuinely share common ground — is an open empirical question.
What is not novel: explaining to clarify your own thought (Chi, Vygotsky, Pennebaker); using a listener as a forcing function (Chase; rubber-duck practice in programming folklore); converging on shared terms (Brennan & Clark; Garrod & Anderson). In the report's words: you have independently re-derived a well-attested family of practices and combined them with unusual discipline. The combination and the artefact-orientation are the novel parts; the ingredients are not.
Shared vocabulary
Does a shared vocabulary actually make two minds faster?
Yes — robustly, in every paradigm that measures it — but in word counts and turn counts, not in a single pooled effect size. No meta-analysis of “shared vocabulary → communication speed” was located [unknown]; the claim rests on converging experiments with different measures.
“The next one looks like a person who's ice skating, except they're sticking two arms out in front” — trial 1 “The ice skater” — trial 6 [published]
Words fall
Clark & Wilkes-Gibbs (1986): within trial 1, word count fell 4.6 words per position from position 1 to 12; on trials 2 and 6 the declines were 1.0 and 0.4 [published]. Basic exchanges needing no refashioning rose from 6% on early trials to 84% on later trials [reported]. The same reference, once grounded, costs a fraction of the words and turns.
Terms converge — and stick
Brennan & Clark (1996): partners converge on the same terms and then appeal to that conceptual pact even when a simpler reference would do [published]. The pact is the efficiency device — it removes negotiation. Garrod & Anderson (1987): “entrainment is progressive, increasing as the dialogue proceeds” [published].
Alignment makes it cheaper to process
Pickering & Garrod (2004): alignment “greatly simplifies production and comprehension in dialogue” [published]. Shared vocabulary is not just shorter — the representations are already activated in both minds, so each message costs less to produce and less to understand. That is your latency point, in the literature's language.
The baseline is bad — and pacts don't transfer
Without entrainment, two people independently pick the same term for the same thing only ~10% of the time (Furnas et al.) [reported]. And pacts are partner-specific [published] — they do not transfer to a new partner. Your shared vocabulary with this AI is an asset tied to the dyad and to the message history that records it — itself an argument for self-encapsulating messages as a way of preserving the definitions. Caveat: partners can also align on a term while meaning different things by it — the illusion of common ground. Explicit definitions are a hedge against exactly that, though that hedge is itself untested.
Vernacular worth keeping — defined in place, as agreed
Grounding
The moment-by-moment process by which partners build common ground — mutual knowledge they both know they share — until both believe the other has understood “to a criterion sufficient for current purposes.”
Conceptual pact
A shared conceptualisation two partners settle on for a thing, then keep using. Once a pact exists, neither side re-negotiates the term.
Lexical entrainment
The convergence itself: partners coming to use the same words for the same things over the course of a conversation.
Subsumptive constraints
The pressure explaining puts on you to find a pattern broad enough to subsume — to sit above and cover — all the cases at once.
Elaborated code
Bernstein's term for speech that makes principles and operations linguistically explicit and context-independent — the closest named concept to self-encapsulating speech. Its opposite, restricted code, leaves meaning implicit and leans on shared background.
Protégé effect
Learning more because you are teaching — or preparing to teach — someone else. The expectancy finding: most of the gain happens in the preparation, before the teaching starts.
The evidence base
All 24 studies and sources
Grouped as the research grouped them: (a) core validation, (b) the communication layer, (c) boundary conditions. Each entry keeps what was tested, the measured result, the limits, and the primary source.
Showing 24 of 24
A · 01
Think-aloud protocols — Ericsson & Simon (1980); Fox, Ericsson & Best (2011)
Meta-analysis
Tested
Whether verbalising thoughts changes the process being reported. Ericsson & Simon proposed think-aloud only verbalises information already heeded in short-term memory, so it should not be reactive; only instructions demanding reasons or explanations should be.
Result
94 studies, nearly 3,500 participants [published]. Think-aloud effect on accuracy r = −.03, “indistinguishable from zero” [published]. Describing/explaining thoughts was significantly reactive — higher performance than silent controls [published]. All verbal reporting increased time to complete tasks [published].
Limits
Non-reactivity ≠ validity — Schooler (2011) argued non-reactivity “does not necessarily speak to its validity” [reported]. The meta-analysis aggregates accuracy, not learning or retention.
8 college students [published] studying physics worked examples aloud; spontaneous self-explanations counted, then a problem-solving posttest.
Result
High solvers averaged 82% correct on the posttest and 15.3 explanations per example; low solvers 46% and 2.8 per example [published] — as restated in Chi et al. 1994. Good explainers also monitored their understanding more accurately.
Limits
N = 8, correlational — good students may explain more because they are good students. That is why the 1994 eliciting study matters.
Source: Chi et al. 1989, Cognitive Science 13:145–182, doi:10.1207/s15516709cog1302_1 (DOI as cited in secondary sources; full text not opened)
A · 03
Eliciting self-explanations — Chi et al. (1994)
Experiment
Tested
24 eighth-graders [reported] reading a circulatory-system text: 14 prompted to self-explain after each line [published], 10 read the text twice.
Result
The prompted group improved more from pretest to posttest, with the largest gains on inference questions requiring knowledge not stated in the text [published]. Gain figures of 32% vs 22% overall and 22.6% vs 12.5% on the hardest questions [reported] — secondary summaries only, not verified against the primary text.
Limits
Small sample, single topic, short-term posttest; no delayed-retention figure verified [unknown].
Self-explanation meta-analysis — Bisra et al. (2018)
Meta-analysis · strongest leg
Tested
Meta-analysis of prompted self-explanation while studying or solving problems.
Result
69 effect sizes from 64 research reports [published]; overall weighted mean g = 0.55, random-effects [published]. ~5,917 learners [reported]. The effect held across domains, ages, and task types and after controlling time on task [reported]. Self-explanation beat instructor-provided explanation (g = 0.35) [reported]; choosing an explanation from a list gave g = 0.24, not significantly different from zero (k = 2) [reported].
Limits
Dunlosky et al. (2013) rated self-explanation only “moderate utility” — evidence in real educational contexts and for long-term retention is thin [published]. A maths-specific meta-analysis (Rittle-Johnson et al. 2017) found g = 0.39 for near transfer [reported].
Renkl et al.: 56 bank apprentices [published] learning compound/real interest, uniform vs multiple examples × spontaneous vs elicited self-explanations. Atkinson et al.: fading worked steps plus self-explanation prompts.
Result
Eliciting self-explanations supported transferable knowledge; low-prior-knowledge learners profited most on near transfer [published]. Fading + prompts gave medium-to-large gains on near and far transfer with no added study time [reported].
Limits
Most learners are passive, superficial self-explainers without prompting (Renkl 1997) [reported] — quality of explanation, not mere quantity, drives the effect. Worked examples lose effectiveness with expertise (expertise reversal) [reported].
3 category-learning experiments: explain why items belong to categories vs describe, think aloud, or free study. Tests the subsumptive-constraints account — explaining drives learners to find unifying patterns.
Result
Explaining promoted discovery of the broad (100%) rule governing category membership vs all three controls [published]. Description was better for learning item details [published]. Explaining anomalous observations may aid belief revision [published].
Limits
The same mechanism backfires: when a misleading pattern is present, explaining impaired learning [reported]. No standardised effect size located [unknown]. Lab categories, not real-world reasoning.
Source: Williams & Lombrozo 2010, Cognitive Science 34:776–806, doi:10.1111/j.1551-6709.2010.01113.x (DOI as cited in secondary sources)
A · 07
Explanation & learning by thinking — Lombrozo (2012; 2006)
Theory / review
Tested
A review chapter on the structure and function of explanations in learning and inference — a synthesis, not an experiment.
Result
Explaining recruits prior beliefs and explanatory preferences (simplicity, scope/unification) that jointly constrain subsequent processing [published]. “Learning by thinking”: new insight without new observations, via explanation.
Limits
Theoretical synthesis; no effect size. The virtues that help — simplicity, scope — are the same ones that mislead when the true pattern is complex or the simple pattern is spurious.
Chase et al.: Betty's Brain — students teach a computer agent by building concept maps, vs learning-for-self with identical software. Kobayashi: meta-analyses of learning by teaching.
Result
Teachable-agent students spent more time on learning activities and learned more, most pronounced for lower-achieving children [published]. Study 1: 62 eighth-graders; Study 2: 24 fifth-graders [reported]. Kobayashi (2019), 28 studies: preparing to teach g = 0.35; preparing + teaching g = 0.56 [published]. Face-to-face teaching after preparing g = 0.84 (k = 4 — fragile) vs indirect g = 0.48 [reported]. Kobayashi (2024), 39 studies: overall g = 0.27; with prior expectancy g = 0.48, without expectancy g = −0.02 [reported]. Koh et al. (2018): teaching without notes ≈ retrieval practice alone (difference d = 0.10) [reported].
Limits
Chase et al. is one system, schoolchildren, small samples. Expectancy is the active ingredient — explaining after studying without having expected to teach does ~nothing [reported].
Nothing, by that name, in a controlled trial. Feynman never published it as a method [reported].
Result
[unknown] as a named technique. Secondary reviews state: “No large study we could find has tested the Feynman Technique by name against other ways of studying” [reported]. Its components are validated separately: self-explanation (g = 0.55), retrieval practice, learning-by-teaching. A 2026 Grade 7 AI-integrated classroom study (mean 24.67 → 26.30, p = .014) [reported] is single-site, single-topic, and not a test of the plain method.
Limits
Do not cite the Feynman Technique as “proven.” Cite its mechanisms.
Slamecka & Graf: 5 experiments — generate target words from a rule + first letter vs read complete pairs; tested on recognition, free and cued recall. Bertsch et al.: meta-analysis.
Result
Generated words beat read words on every test [reported]. Bertsch et al.: 445 effect sizes, 86 studies [published]; overall effect 0.40, “almost half a standard deviation” [published]. N = 17,711 [reported].
Limits
The effect collapses for meaningless material: nonwords 0.05, anagrams −0.05 [reported]. Requires successful generation and enough prior knowledge. Difficulty barely moderated it (easy 0.41, hard 0.46) [reported] — effort alone is not the mechanism; semantic processing is.
Illusion of explanatory depth — Rozenblit & Keil (2002)
Experiment series
Tested
Rate your understanding of everyday devices (zipper, toilet, lock) on a 7-point scale, write a step-by-step explanation, re-rate, answer a diagnostic question, read an expert explanation, re-rate.
Result
Ratings fell at every stage, biggest drop right after the attempted explanation [published]. The effect is specific to explanatory (causal/mechanistic) knowledge; much smaller or absent for facts, procedures, narratives [published]. Original Experiment 1 N = 16 Yale undergraduates [reported].
Limits
No pooled effect size in the abstract [unknown]. The famous political extension (Fernbach et al. 2013) failed preregistered replication (Crawford & Ruscio 2021) [reported]; the core device effect replicated in 3 preregistered studies (Körner et al. 2024, N = 607) [reported]. This is the diagnostic half of thinking out loud: explaining exposes gaps rather than filling them.
Source: Rozenblit & Keil 2002, Cognitive Science 26:521–562, doi:10.1207/s15516709cog2605_1 (as cited) — Doczz full text
A · 12
Elaborative interrogation — Pressley et al. (1987, 1992)
Experiment
Tested
Undergraduates read sentences; the elaborative-interrogation group was asked “why?” and answered it, vs explanation-provided and reading-only controls; cued recall test.
Result
Elaborative interrogation 72% vs explanation-provided 37% vs reading 37% [reported] — multiple secondary sources agree; not verified against the primary paper. Mechanism: generating an explanation creates additional retrieval pathways and integrates the fact with prior knowledge.
Limits
Works best with sufficient prior knowledge (Woloshyn et al. 1994) [reported]; strongest for factual/conceptual material, not procedural skills [reported]. Dunlosky et al. (2013) rated it moderate utility [published]. A frequently quoted “0.59” (Hattie synthesis) [reported] could not be traced to a primary meta-analysis — treat as unverified.
Source: Pressley et al. 1987, J. Exp. Psychol.: Learning, Memory, and Cognition 13:291–300 (citation as given in secondary sources); Dunlosky et al. 2013 review — PDF
Write 15–20 min/day for 3–4 consecutive days about a stressful experience vs a trivial-topic control. Frattaroli meta-analysed experimental disclosure (writing or talking).
Result
Frattaroli (2006): 146 randomised studies [published]; average r = .075, positive and significant [published]; d ≈ 0.15 [estimated] conversion. Effects larger with recent stressor, sessions >15 min, ≥3 sessions, writing at home [reported].
Limits — strongest critique
Smyth's (1998) widely quoted d = 0.47 came from 13 studies under a fixed-effects model and has not been replicated [reported]. Meads & Nouwen (2005), 61 RCTs: “no clear improvement … in objectively measured physical health” [reported]. Zachariae & O'Toole (2015), cancer patients: g = 0.04 / 0.08 / 0.09, all non-significant [reported]. Ayers et al. (2018), n = 564: clean null [reported]. Pennebaker himself (2018): “Results have been difficult to replicate” [reported]. Relevance here is indirect — this literature is about emotional processing and health, not explaining ideas.
Source: Frattaroli 2006, Psychological Bulletin 132:823–865, doi:10.1037/0033-2909.132.6.823 (DOI as cited)
B · 14
Common ground & grounding — Clark & Brennan (1991)
Theory chapter
Tested
A theory chapter, not an experiment: grounding as the collective process by which partners accumulate common ground moment by moment.
Result
Grounding criterion: partners mutually believe the other has understood “to a criterion sufficient for current purposes” [published]. Principle of least collaborative effort: partners minimise joint work, trading speaker effort in the first presentation against refashioning effort later [published]. Eight medium constraints on grounding: copresence, visibility, audibility, cotemporality, simultaneity, sequentiality, reviewability, revisability [published].
Limits
Framework, no effect size. Direct tension for your practice: least collaborative effort predicts partners prefer short provisional expressions + repair over long perfect first messages — unless the medium lacks cheap repair.
Referring as a collaborative process — Clark & Wilkes-Gibbs (1986)
Experiment · cleanest speed evidence
Tested
Director–matcher tangram task, repeated over 6 trials; how referring expressions are proposed, refashioned, and mutually accepted.
Result
Words per figure and turns per figure declined across trials [published]. Within trial 1, word count fell 4.6 words per position from position 1 to 12; on trials 2 and 6 the declines were 1.0 and 0.4 words per position [published]. Basic exchanges rose from 6% on early trials to 84% on later trials [reported].
Limits
Lab task, 12 tangrams, one dyad type. Still the cleanest quantitative demonstration that shared history shortens communication.
Mechanistic theory (target article + peer commentary): in dialogue, interlocutors' linguistic representations become aligned at many levels via a largely automatic priming process.
Result
Alignment “greatly simplifies production and comprehension in dialogue,” enables local dialogue routines, and explains self-monitoring [published]. Humans are “designed for dialogue rather than monologue” [published].
Limits
Barr & Keysar (2004) and others: the automatic mechanisms are not dialogue-specific, and the model underplays explicit audience design and perspective-taking [reported]. No effect size — a model, not an intervention study.
Lexical entrainment / conceptual pacts — Brennan & Clark (1996)
Experiment ×3
Tested
3 experiments on repeated reference to the same objects in conversation: do partners converge on the same terms, and why?
Result
Partners come to use the same terms (lexical entrainment); once a shared conceptualisation — a conceptual pact — is established, partners appeal to it in later references “even when they could use simpler references” [published]. Pacts simplify over time, are abandoned when necessary, and are partner-specific [published].
Limits
Lab picture-description tasks. The baseline problem is severe: two people independently choose the same term only ~10% of the time (Furnas et al.) [reported] — entrainment is what rescues dyads from that.
Conceptual / semantic coordination — Garrod & Anderson (1987)
Corpus analysis
Tested
56 dialogues [published] from a cooperative maze game; analysis of 1,396 spatial descriptions [published].
Result
Pairs develop different local description schemes (path, coordinate, line, figural) and converge within a dialogue; “entrainment is progressive, increasing as the dialogue proceeds” [published]. Partners end up “playing the same language game” — an idiosyncratic local language.
Limits
One task domain (spatial locations); descriptive analysis, no causal effect size.
Sociological theory, not an experiment: elaborated code makes “principles and operations … linguistically explicit,” meanings relatively context-independent; restricted code leaves them implicit and context-dependent, relying on shared background [published]. Schools reward the elaborated code.
Result
No effect size — a framework. It is the closest named concept to self-encapsulating speech: deliberately elaborated, context-carrying messages.
Limits — strongest critique
Labov and others argued the distinction was read as a deficit account of working-class language — speakers of restricted codes reason elaborately in their own varieties, and researchers mistook unfamiliar vernacular for deficiency [reported]. Bernstein denied the restricted code was inferior [reported]. Applying it to human–AI communication is an analogy, not a tested prediction.
Source: Bernstein 1971, Class, Codes and Control Vol. 1 (Routledge); quoted text via New Learning Online
B · 20
Distributed cognition — Hutchins (1995)
Ethnography / framework
Tested
Ethnography of navigation aboard a US Navy ship (Cognition in the Wild): how a team determines position and course.
Result
No single person knows how to navigate the ship alone; the navigation team is a “cognitive and computational system” [published], with cognition distributed across people, artefacts (charts, alidades, logs), roles, and environment — a “social distribution of cognitive labor” [published].
Limits
Ethnography, no effect size, no controlled comparison. It licenses treating you + AI + message history as one cognitive system, but it does not predict that self-encapsulating messages improve that system — that is an application, not a finding.
Source: Hutchins 1995, MIT Press (discussed at PhilArchive)
6 experiments: verbalise a previously seen face (or colour) vs control, then recognition. A verbally biased representation interferes with the original visual memory.
Result
Verbalisers performed worse on recognition; visualisation did not impair recognition; verbalising a verbal stimulus marginally improved memory [published]. Original lineup difference ~25% worse [reported] (22% in Exp. 1 [reported]). Meissner & Brigham (2001): 29 comparisons, N = 2,018 [published]; Fisher's Zr = −0.12, “small, yet significant” [published]. Alogna et al. (2014) registered replication: immediate-description replications averaged a 4.01% difference; 20-min-delay replications 16% [reported].
Limits
Small, heterogeneous, delay-dependent; at delays ≥30 min the effect can reverse into verbal facilitation[reported]. It bites where the knowledge is hard to verbalise.
Verbalisation vs insight — Schooler, Ohlsson & Brooks (1993); Wilson & Schooler (1991)
Experiment ×4 · contested
Tested
4 experiments: verbalisation (retrospective or concurrent) while solving insight vs non-insight problems.
Result
Retrospective verbalisation made participants significantly less successful at insight problems; concurrent non-directive verbalisation impaired insight problems but had no effect on non-insight problems [published].
Limits / contest
“Solutions fall by 25%” [reported] — secondary, unverified. Fleck & Weisberg (2004, 2013) and Gilhooly et al. (2010) found no negative effect of think-aloud on insight problems [reported]. A 2025 study (Macchi et al.) replicated the dissociation: verbalisation hampered insight yet “dramatically aided” non-insight performance [reported]. The safest statement: the effect depends on whether the solution process is verbalisable. Wilson & Schooler (1991): analysing reasons impaired affective decision-making [reported], no numbers verified.
Source: Schooler et al. 1993, J. Exp. Psychol.: General 122:166–183 — Academia.edu abstract
C · 23
Cognitive load theory — Sweller (1988); Sweller, van Merriënboer & Paas (1998)
Framework
Tested
Framework from maths problem-solving: working memory is severely limited (~4 chunks for novel information, Cowan 2001 [reported]); long-term memory is effectively unlimited; learning = schema construction. Three loads: intrinsic, extraneous, germane.
Result
Worked examples beat unaided problem-solving for novices (worked-example effect) [reported]; support that helps novices can harm experts (expertise reversal) [reported].
Limits
No single effect size for “elaboration overload” [unknown] — a framework generating many specific effects. Germane load is contested: hard to measure independently of the other loads [reported]. Direct implication: self-explanation is germane load when it builds schemas, extraneous load when the elaboration itself consumes the budget needed for the actual problem.
Source: Sweller 1988, Cognitive Science 12:257–285, doi:10.1207/s15516709cog1202_4 (DOI as cited)
Vygotsky's theory: private speech (audible self-directed talk, ~ages 3–7) is social speech turned inward — a tool for planning, attention, monitoring — later internalised as silent inner speech. Winsler et al. review 15+ years of research.
Result
Evidence “largely supportive” of the Vygotskian view [reported]. Fernyhough & Fradley (2005): self-regulatory private speech most common at moderate difficulty and positively associated with concurrent task success [reported]. Private speech increases with task difficulty and obstacles [reported].
Limits
No meta-analytic effect size located [unknown]. The private-speech → success link is equivocal in some studies [reported]. Almost all evidence is developmental (children); your practice is an adult, written, dyadic version — outside this literature's core.
Source: Winsler, Fernyhough & Montero (eds.) 2009, Cambridge UP — front matter PDF
Open questions
What nobody has answered yet
Does the self-encapsulating format itself help, beyond the explaining it contains? No study isolates message format (encapsulated vs terse + repair) as an independent variable [unknown]. A clean test: same content, same dyad, encapsulated vs minimal messages, measuring re-reading comprehension after a delay.
Human–AI grounding. Do conceptual pacts and entrainment form with AI partners the way they do with humans, and does an AI's perfect recall change the optimal trade-off (front-load vs repair)? Almost entirely unstudied [unknown].
Optimal explicitness. Where is the crossover where added prerequisites and edge cases stop being germane load and become extraneous load — for writer and reader? Cognitive load theory predicts a crossover; its location for this practice is unknown [unknown].
Long-term retention. Most self-explanation evidence is short-term and lab-based. Does a year of thinking out loud compound, or does it plateau?
The insight / verbalisation boundary. Is the Schooler et al. (1993) insight impairment real, and does it apply to written, retrospective explaining (your mode) or only to concurrent think-aloud? Contested [reported].
Expressive-writing mechanism. Whether the small disclosure effect is driven by cognitive processing (narrative coherence) vs emotional habituation remains unsettled (Sloan & Marx 2004) [reported] — relevant only if you lean on the writing-as-processing claim.
Does explaining to an AI produce the protégé effect? Chase et al.'s agent could not answer back in natural language; Kobayashi's g = 0.84 face-to-face figure rests on k = 4 [reported]. An AI that can ask questions back might reach or exceed that — or might short-circuit the expectancy mechanism by doing the explaining itself. Untested [unknown].
Could not verify — figures the research flags, kept visible
Chi et al. (1994) percentages (32% vs 22%; 22.6% vs 12.5%) — [reported] from secondary summaries; the primary PDF was located but the numbers were not re-checked against it.
Pressley et al. (1987) 72% vs 37% vs 37% — [reported]; multiple secondary sources agree, primary paper not opened.
Elaborative interrogation “0.59” — [reported] as a Hattie-synthesis figure; could not be traced to a primary meta-analysis. Do not cite as a published effect size.
Williams & Lombrozo (2010) effect size — none located [unknown].
Schooler et al. (1993) “25% fall” in insight solutions — [reported] from one secondary source; not verified.
Rozenblit & Keil (2002) pooled effect size — none in the abstract [unknown].
Chase et al. (2009) sample sizes (62 eighth-graders; 24 fifth-graders) — [reported]; the abstract does not state N.
Kobayashi (2024) figures (g = 0.27; 0.48 with expectancy; −0.02 without) — [reported]; primary paper not opened.
Feynman Technique: any controlled test by name — none located [unknown].
Self-encapsulating messages as a named / tested practice — none located [unknown].
Shared-vocabulary → speed pooled effect size — none located [unknown].
Private-speech meta-analytic effect size — none located [unknown].
Cognitive-load “elaboration overload” effect size — none located; the theory is a framework, not a single effect [unknown].
Several DOI links are given as DOI identifiers as cited in secondary sources; the DOI resolver URLs were not all individually opened.
So — keep doing it?
On this evidence: yes, with two disciplines. Keep the speech explanatory — reasons, patterns, prerequisites, edge cases generated from memory — because that is the version with g = 0.55 behind it, not mere narration, which sits at r = −.03. And watch the load: the same prerequisites that make a message self-contained are the first thing that can tip it into overload. The format itself is your experiment. Nobody has run it. The literature says the ingredients work, the combination is coherent, and the test has never been done.