Three students replay a Science lesson after school near Clementi. The teacher’s voice is clear on one laptop and difficult to hear on another, so they switch on captions. A sentence about evaporation appears beneath an animation. One student notices that the caption says condensation. Another is already copying the words into notes. The picture continues. The lesson has not stopped, yet the knowledge travelling from teacher to student has quietly changed.
Can automatic captions be wrong in educational videos? Why do speech-to-text mistakes matter in Science, Mathematics and English learning? What should parents and teachers do when a transcript contradicts the audio? Those questions are more than technology trivia. Many pupils rely on captions to hear, process, review or search educational material. Captions can create real access. But a caption can also turn a spoken distinction into a different claim, especially when technical words, negative terms, numbers and speaker changes carry the idea.
The important discovery is that a caption is another representation of the lesson, not a guaranteed copy of the lesson. Speech recognition produces text from sound. The result can be edited, checked and made accurate; it can also be wrong or incomplete. The best learning method is not to discourage captions. It is to keep access while teaching a reliable route from suspicious line to verified meaning.
The 50-second answer
Automatic captions are generated by software that estimates what someone said. They can be helpful as an immediate aid, especially where no human-prepared captions exist, but they may mishear technical vocabulary, numbers, units, names, negatives or sentence boundaries. An error can change the concept, not merely the spelling. W3C’s Web Accessibility Initiative warns that automatically generated captions generally need checking and editing before they can meet accessibility needs.
For learners, a strange caption is a cue to pause and verify against the audio, a reliable transcript, the visual demonstration or a teacher’s clarification. Students who cannot access the audio must be offered an equivalent reliable route; they should not be told that listening harder is the solution. For teachers, the priority is to proof the captions that carry essential concepts, remove meaning-changing errors and provide a correction path. Parents can help children identify contradictions without quietly rewriting an unfamiliar scientific explanation themselves.
The learner’s real problem is a mismatch between channels
An educational video often contains four informational channels. Spoken language explains what is happening. Visual language presents diagrams, gestures, demonstrations and on-screen words. Captions display a time-aligned text representation of speech and significant sound where appropriate. A separate transcript, if provided, may support reading, searching or reviewing the whole recording. These channels should help one another. When they conflict, the learner faces an extra decision: which channel represents the intended idea?
Experts sometimes resolve such a mismatch almost automatically. A Science teacher knows that a pool of water becoming vapour is evaporation, so the incorrect caption condensation looks suspicious. A novice may not have that safety net. The same word error that an expert spots in one second can enter the novice’s notes as a new definition. The learner may later answer a question confidently using the captioned term. Confidence does not protect against an error in the source representation.
The right response is a small verification habit. Ask whether the word fits the observed change, the surrounding explanation and the standard scientific meaning. If one channel disagrees with the others, mark the uncertainty and check the source. Do not choose the written version merely because written text looks more official. Nor should a student be asked to trust an inaccessible audio channel they cannot hear.
What speech recognition does and does not know
Automatic speech recognition maps an audio signal to a sequence of words using statistical patterns and learned language structures. It must decide where words begin and end, which similar-sounding expression is most probable, where punctuation belongs and sometimes which speaker is talking. Technical vocabulary can be underrepresented in ordinary speech data. Background noise, unfamiliar accents, overlap between speakers, microphone distance and rapid delivery can make the mapping harder.
This is not the same task as understanding the scientific experiment. A captioning system may output a plausible English sentence while making a technical mistake. An educational model that explains evaporation asks about liquid particles escaping into the gas phase. A speech recogniser hearing a noisy syllable asks which word sequence best matches the sound and language pattern. The recogniser can choose a common word that fits ordinary speech but fails the Science context.
Students do not need a lecture on neural-network architecture to appreciate this distinction. They need one sentence: the computer is estimating words from sound; it is not certifying the scientific truth of those words. That frame turns caption checking into a meaningful reading activity rather than an arbitrary warning that “AI sometimes makes mistakes”.
The most dangerous errors are small
A caption that turns photosynthesis into a nonsense phrase may be easy to notice. A caption that changes increase to decrease can be harder and more damaging because both are valid scientific words. One omitted not can reverse a statement. One missing decimal point can change an example’s quantity. Misplacing the word only can alter the condition under which an explanation is true. These edits are small at the character level and large at the meaning level.
Consider the spoken sentence, “The mass does not disappear when the ice melts.” If the caption drops not, the apparent claim becomes the opposite of the intended conservation principle. A student copying the text has not simply made a spelling error; the scientific relation between change of state and mass is inverted. The repair is not to memorise the correct sentence in isolation. It is to ask which observation and principle determine the statement’s truth.
The same applies in Mathematics: “Multiply the denominator, not the numerator” or “The intercept is negative” could be changed by a missing word or sound. In English, a misplaced speaker label can assign a claim to the wrong character. The category of risk is semantic, not merely typographic.
A Clementi table: three lines that cannot all be right
Imagine an entirely fictional clip on dissolving. The animation shows sugar dispersing in water. The audio says, “The sugar has not vanished; its particles are dispersed through the water.” The captions read, “The sugar has vanished; its particles are dispersed through the water.” The first half of the caption contradicts the second. A pupil who knows little about dissolving might notice the contradiction only after the teacher asks, “Does the mass of the sugar simply cease to exist?”
A useful classroom comparison places three columns on the board: what the speaker said, what the caption displayed, and what the science supports. The last column is not automatically identical to the first. Teachers, too, can make mistakes. The student must use the experiment and reliable scientific knowledge to decide. This keeps the lesson from becoming blind obedience to either human voice or machine text.
The repair task can be small. Ask learners to underline the word whose change reverses meaning. Then require a new explanatory sentence in their own words. A student might write, “Dissolving spreads the solute through the solvent; it does not mean the solute’s matter is destroyed.” This demonstrates concept recovery rather than mere caption editing.
When captions are an access route, not a preference
Captions can be essential for students who are deaf or hard of hearing. They can also assist learners processing speech in a noisy setting, following an unfamiliar accent, reviewing a technical term or learning through an additional language. The response to caption error must preserve these benefits. Saying “turn captions off and listen” excludes some of the very learners captions are meant to serve.
W3C distinguishes captions from transcripts and explains why accurate, synchronised captions matter. A reliable accessibility plan may include reviewed captions, a downloadable transcript, on-screen terminology, clear diagrams and opportunities to ask questions. Different media needs require different accommodations. A transcript alone may not substitute for captions when timing and speaker identification matter; captions alone may not provide the easy navigation needed for close study.
CAST’s Universal Design for Learning guidelines support offering information in multiple usable ways while preserving meaning. Multiple formats are helpful only when students can actually access equivalent important information. If audio says one thing and text says another, the presence of two modalities is not by itself a success. Equivalence needs checking.
Captions and transcripts are not interchangeable
Captions appear in coordination with video. They follow when words are spoken and can identify important non-speech audio such as a bell or reaction sound. A transcript usually provides a more continuous record that can be read, searched or printed. A learner watching a slow-motion diagram may need timed captions to connect explanation with visual movement. Another learner preparing revision cards may prefer the transcript to locate the definition of resistance without replaying ten minutes of footage.
Each format introduces its own failure modes. The caption may appear too early or too late. The transcript may omit a diagram reference such as “this arrow”. A transcript copied into a worksheet may lose mathematical formatting or speaker labels. Captions can cover important on-screen data if poorly positioned. A teacher should judge the formats by the learning job: watching an experiment, reconstructing a procedure, checking a term, or studying an argument.
A good lesson package might therefore include the video, reviewed captions, a concise list of key terms and a linked written summary. This is not an instruction to produce four completely separate lessons. It is an invitation to ensure that the essential idea remains recoverable when one access path is difficult for a learner.
Science case one: the difference between melting and dissolving
Teachers can make the caption mechanism visible with two short invented clips. In one, a solid cube of ice changes into liquid water. In the other, sugar is stirred into water. Suppose the automatic captions mistakenly use the word melt for both. The students can still see two changes that may look superficially similar: something solid stops being visibly solid. Yet the scientific processes differ. Melting is a change of state; dissolving involves a solute dispersing in a solvent.
Ask learners to build a diagnostic question: “What happened to the material, and what would happen if the water were removed?” The answer should not depend on the caption alone. Students can draw before-and-after particle representations at an age-appropriate level, name the process and explain one important distinction. If the caption is wrong, the comparison becomes a safe learning opportunity after the accurate information has been provided.
But be careful with the teacher’s role. A caption error should not be left in an assigned video as a deliberate surprise when students do not know it is there. The point is to teach verification, not to create hidden barriers. In real coursework, correct the resource and clearly identify any earlier version that was defective.
Science case two: a missing negative word in a fair test
An instructor says, “We must not change the amount of water between groups, because we want to test the effect of temperature.” The caption reads, “We must change the amount of water between groups.” A learner then designs an experiment in which both temperature and volume differ. When the outcomes change, the learner cannot tell which factor contributed to the difference. A small transcript omission has damaged the logic of the comparison.
This case exposes why context checking matters. The word because announces a reason. The intended reason, isolating the effect of temperature, conflicts with the incorrect captioned instruction to vary volume too. Students can use the structure of a fair test to detect the mismatch. They should identify the manipulated variable, what is measured and what should remain reasonably controlled. The correction becomes more robust when they can explain why the alternative procedure would make the conclusion weaker.
The teacher can test whether the repair worked with a changed example. Instead of water volume and temperature, use light exposure and plant growth in a hypothetical investigation. Ask which conditions must remain comparable. If students transfer the control logic, they learned more than a corrected line of text.
Mathematics case: numbers are especially unforgiving
A presenter says, “Zero point zero five.” The automatic caption prints “zero point five.” Both phrases are grammatically ordinary and mathematically different. The first represents five hundredths; the second represents five tenths. A student who copies the caption may be off by a factor of ten. The problem is not that Mathematics is uniquely vulnerable to technology. It is that compact notation can encode large quantity differences in small visual marks.
Other risks include negative signs, powers, fraction language and units. Spoken “x squared” might appear as “x to”, or a unit name might be omitted. A caption can state “multiply by four” when the screen shows division by four. If the video displays the original equation clearly, students can cross-check. If the screen contains only a talking head, the risk is greater. Teachers can reduce dependence on perfect speech recognition by showing important equations and units as carefully proofed on-screen text.
A useful practice is an equation witness: after hearing a calculation, the learner writes the relation independently from the displayed notation and checks that the audio, captions and mathematical structure agree. If they do not, ask for clarification instead of selecting whichever seems most convenient.
English case: who said what?
A literature video may involve a narrator reading dialogue, then explaining the character’s motive. Automatic captions can omit quotation marks and misidentify when one speaker stops and another begins. The resulting transcript may make a character sound as though they expressed the teacher’s interpretation, or make the teacher appear to endorse a character’s questionable statement. The words may be individually correct while the attribution is wrong.
In comprehension, attribution is not decoration. Whether a sentence is spoken by a character, thought by a narrator or supplied by a commentator determines what it can support. Students can use quotation marks, speaker labels, surrounding text and the original passage to reconstruct the intended source. Ask: “Which voice made this claim, and what evidence identifies the voice?” The goal is not to become an audio engineer. It is to preserve the chain from speaker to meaning.
This skill also matters when students quote captions in assignments. If the teacher requires a quotation from a video, pupils should use a reliable transcript or verify the wording with the original media and give the location or timestamp as instructed. A machine-generated caption is a provisional record unless reviewed.
Timing errors can change meaning without changing a word
Imagine a video in which a teacher points first to a red line and then to a blue line. The caption reads, “This one rises more quickly.” If the text appears after the pointer moves, the student may attach “this one” to the wrong line. Every word of the caption may be correct, yet the time alignment produces a false association. This is a different error from speech recognition. The word sequence is intact but the reference is misplaced.
Timing becomes especially important in demonstrations, gestures, maps and animated diagrams. Deictic words such as this, that, here and now borrow their meaning from the speaker’s position and the visual moment. A transcript stripped of timing can lose that anchor too. Teachers should minimise unsupported “this one” references and name the object when possible: “the blue line representing temperature” is more durable than “that line”.
When reviewing captions, inspect not only spelling but synchronisation at moments where a new visual state appears. A caption that arrives a second late may be harmless during general introduction and consequential during a rapidly changing experiment. Quality checking should follow conceptual risk.
The caption risk map
An efficient caption audit is not simply a spelling check from beginning to end. Start with words that carry the lesson’s logic. Scan for negation: not, never, except, unless. Then quantity: decimal points, fractions, ranges and rates. Check technical terms that resemble ordinary words. Check causal language: because, therefore, if, provided that. Check comparison language: greater, smaller, before, after. Check speaker identity and visual references.
This is a risk-based list, not a claim that every other word is unimportant. An incorrect person’s name can matter in history, and a location can matter in geography. The priority is to locate the places where one small textual error could force the learner to build a different model. A teacher with twenty minutes to review a lesson should protect those conceptual hinges first, then complete the wider accessibility review as resources allow.
Where a resource is distributed to many students, a formal caption review should follow relevant accessibility requirements and organisational processes. A teacher’s quick classroom check can catch obvious problems, but it is not a substitute for required accessibility compliance or specialist testing.
Why an error may survive into notes and homework
Once a learner copies a wrong caption into a notebook, the error can travel. It may be repeated during revision, included in a flashcard, used in a homework explanation and eventually appear in an assessment. The student then looks as though they have a misconception when the starting point was a corrupted media representation. This chain is especially hard to detect if the incorrect sentence sounds plausible.
One repair is to teach source-marked notes. Instead of copying the entire subtitle stream, the learner records a concise claim and its source location. If a line seems uncertain, mark it with a question rather than converting it into a fact. After verification, write the corrected relationship in the student’s own words. For Science, the note should include the mechanism and boundary; for Mathematics, the relation and units; for English, the speaker and textual context.
This method preserves independent learning. Students should not become dependent on teachers correcting every transcription typo. They learn to recognise the high-stakes categories of error, verify when possible and ask focused questions when necessary. Accuracy becomes an active skill rather than a hidden promise made by the interface.
Teacher route: before uploading a recorded lesson
First, determine who needs captions, a transcript or another access format. Do not assume one format serves everyone. Next, check the audio recording quality and the presence of on-screen representations for critical formulas and terminology. Generate captions if useful, but review and correct them before releasing them as the authoritative accessible version. Listen at meaningful transitions and check the difficult vocabulary, negatives, numbers and speaker boundaries.
Then test the final delivery. Turn the sound off and follow the captions while watching. Do they allow the scientific idea to be reconstructed without hearing the voice? Turn the captions on with the sound and inspect whether text obscures important labels. If assistive technology or institution-specific accessibility testing is available, use it on representative material. Provide a clear route for students to report an error without publicly exposing their disability status or private circumstances.
If an error is found after release, correct the caption file or resource where possible, note the substantive correction and alert students who may have used the old version. A corrected transcript is not enough when the underlying video continues displaying the wrong caption to later viewers.
Learner route: from suspicion to usable knowledge
During ordinary study, pause when a caption contradicts a diagram, accepted definition or nearby sentence. Do not assume the contradiction is your own ignorance. Rewind within the permitted platform controls, compare the audio or authoritative transcript if accessible, and identify the exact word or time alignment causing the mismatch. Then write a corrected one-sentence explanation and try a new example without looking at the captions.
If a learner cannot hear the audio, the verification route must remain accessible. They can compare proofed on-screen material, a teacher-approved transcript, a textbook reference or a teacher’s written clarification. The burden should not be shifted to an inaccessible modality. If the caption is part of formal assessment content, report the discrepancy through the authorised channel before making unsupported assumptions about the expected answer.
For speed, use a two-column note: “Caption says” and “Verified idea”. This keeps the error visible long enough to prevent accidental recopying but does not turn the notebook into an archive of machine failures. The final learning artefact should contain the accurate principle, the reason it holds and one changed application.
Parent route: help the child verify, not guess
Parents often encounter the issue when a child insists that “the teacher said this” while the textbook says something else. Instead of deciding instantly who must be wrong, ask the child to show the precise timestamp and explain what each source states. This converts a vague disagreement into a checkable one. If the source is a school-assigned video, send the discrepancy to the teacher with the exact line and timestamp rather than publicly attacking the recording.
For young children, the adult can model noticing a difference between can and cannot or more and less. For older students, invite a more demanding comparison: What mechanism makes one claim plausible? Does a diagram support the spoken or written version? Could both be true under different conditions? These questions teach evidence-based reasoning while preserving the teacher’s role in confirming the curricular explanation.
Do not make the child listen repeatedly to sound they cannot reliably access. If captions are essential, request a corrected accessible text or other suitable support. The goal is equal access to the concept, not endurance of a defective interface.
A five-minute class challenge
Show a teacher-made sentence and a deliberately altered text version labelled as a practice example, not as an actual machine transcript. Spoken or printed source: “The circuit is not complete, so current does not flow through the bulb.” Altered caption: “The circuit is complete, so current flows through the bulb.” Ask students to identify exactly which words changed, draw the circuit condition, and explain what observations would distinguish the two statements.
Then change the task. Provide two accurate statements and ask students not to label one wrong merely because the wording differs. “Current flows when a closed conducting path is available in the simple circuit” and “A complete circuit allows current to flow” can express related ideas at different levels. Verification must avoid the opposite error of treating every paraphrase as corruption. The target is preservation of scientific meaning, not identical strings.
The class challenge ends with a new principle: check the semantic difference that matters. A comma may be harmless. A negative word may not be. A synonym can preserve meaning in one context but change it in another. Students should learn to explain which kind of change has occurred.
When captions are accurate but comprehension is still weak
A learner may have perfectly proofed captions and still be unable to explain the concept. Captions address an information-access route, not all the learning operations that follow. Students still need prior knowledge, practice, retrieval, feedback and opportunities to transfer. A transcript can make a lesson searchable without making the student’s memory permanent. A polished subtitle file cannot replace a missing explanation of photosynthesis.
This caveat protects a useful distinction. Access fidelity asks whether the learner can receive the intended content. Learning success asks whether the learner can understand, retain and use it. Both matter. If the captions are wrong, repair the access route. If they are right but the concept is unfamiliar, teach the concept. If the student knows the concept but struggles to write an answer, teach expression. Conflating these problems wastes time and can unfairly label students.
The Education Endowment Foundation’s guidance on cognitive load and metacognitive strategies is relevant to instruction after content becomes accessible. However, it does not establish a universal effect size for correcting Science captions. That specific classroom recommendation is an application of accessibility and learning principles, not a promised numerical gain.
Common failure modes and quick repairs
Failure one: treating unreviewed automatic captions as official notes. Repair by labelling them provisional and providing reviewed material. Failure two: correcting spelling but ignoring lost negation. Repair with a meaning-risk audit. Failure three: assuming a transcript substitutes for time-aligned captions. Repair by matching format to learner need. Failure four: telling a deaf pupil to listen instead. Repair by providing an accessible verification route. Failure five: placing captions over key diagram labels. Repair by testing the final video layout.
Failure six: taking every paraphrase as a mistake. Repair by checking whether the scientific relation changed. Failure seven: keeping the old incorrect captions live after a correction note. Repair the actual delivery. Failure eight: relying on sound alone when units and notation matter. Display proofed equations and terms. Failure nine: blaming student understanding for corrupted inputs. Check the received representation before diagnosing learning. Failure ten: stopping at a corrected caption without assessing knowledge. Ask the student for an independent explanation or new application.
Each failure demonstrates the same lesson. The resource is a pipeline from speaker to audio, from audio to text, from text and visuals to the learner’s model. The pipeline can fail at more than one point. Good teaching identifies the point rather than assigning all problems to the student.
A three-pass correction method for a difficult sentence
A caption that seems odd deserves a method, not an instant guess. First, locate the claim that matters for the lesson: the part that changes a cause, direction, quantity, condition or scientific relationship. Second, replay a short span with the visual context and, where possible, compare it with the teacher’s slides or source text. Third, rewrite the idea in a fresh sentence and ask whether it produces the same prediction. This final pass is important: repeating a corrected line without understanding it can still leave the learner unable to use the concept.
Imagine a lesson about evaporation. The spoken explanation is “A higher temperature generally increases the rate of evaporation, if other relevant conditions stay the same.” An automatic transcript may drop increases, lose generally, or fail to capture the condition. Each change has a different consequence. Missing increases destroys the direction. Missing generally converts a useful tendency into an absolute claim. Missing the condition leaves the student unaware that surface area or air movement can also affect the observation. To check meaning, the pupil can draw two comparable setups, name what stays constant, and predict which one evaporates faster. A caption is verified not merely when its letters look plausible, but when the learner can connect the sentence to a sensible model.
Now consider an English comprehension recording. The teacher says, “The narrator did not believe the witness was lying.” A caption that omits not reverses the narrator’s attitude, and an otherwise accurate essay could then be built on a false premise. Replay alone might not resolve an ambiguous recording. The student should then consult the full passage or ask the teacher. Saying “the caption is wrong” is less useful than identifying the exact disputed word and how the two interpretations affect the answer.
A good revision notebook records the resolved idea in the student’s own language, with an optional timestamp linking back to the lesson. It should not become a repository of unverified machine output. For younger students, a simple note can use three boxes: I heard or saw, I checked against, and What the idea means. Older students can add why the initial version was tempting. Teachers can model this briefly using a deliberately edited sentence, then return to the Science or English objective so the lesson remains about understanding rather than caption technology.
What uncertainty should remain after checking?
Sometimes a recording is genuinely ambiguous. A teacher may have misspoken, a diagram may be missing, or the class may be using simplified terminology appropriate to an early level. In those cases, a correction should not pretend to possess certainty. Mark the claim as unresolved and identify the next authoritative point of comparison: the syllabus-level explanation, the referenced diagram, the teacher’s clarification or an established textbook treatment. There is no educational advantage in replacing a doubtful transcript with a confident but untested reconstruction.
This distinction is useful beyond recordings. Science ideas travel through speech, captions, slides, notebooks, summaries and examination answers. Every transfer can alter them. A reliable learner notices when a representation has become easier to read but less faithful to the original reasoning. The transfer is successful only when the underlying relationship survives and the learner can explain it independently.
FAQs
Are automatic subtitles reliable enough for school lessons?
They can be useful as a provisional aid, but accuracy varies with audio quality, technical vocabulary and context. W3C warns that automatic captions generally require checking and editing to meet accessibility needs. Schools should not assume unreviewed output is fully equivalent to the spoken content.
What if the teacher’s audio is wrong and the caption is right?
That can happen too. The goal is not to crown audio automatically. Compare the claim with reliable subject knowledge, the demonstration and the teacher’s correction. A verified lesson should repair whichever channel contains the error.
Can one wrong caption really change a Science answer?
Yes. Losing not, swapping increase with decrease, changing a technical term or moving a decimal point can reverse the intended relationship. Other mistakes are merely cosmetic. Judge errors by their effect on meaning, not by their number of letters.
Is a transcript better than captions for revision?
A transcript can be easier to search and annotate; captions better support alignment between speech and moving visuals. Many students benefit from both. Neither format guarantees learning without active explanation and practice.
Should a student use caption text as a quotation?
Use a reviewed transcript or verify the exact wording against the source under the teacher’s citation rules. An unreviewed automatic line may contain errors. Provide the timestamp where helpful and distinguish the speaker from the captioning system.
What about captions in another language?
Translation adds another layer where meaning can change. Technical terms, negatives and units need special care. A translation should be checked for conceptual accuracy as well as grammatical fluency, with appropriate language support for the learner.
Can students fix captions themselves?
They can notice and report suspect lines, and may annotate personal study notes. Editing the official caption track usually requires the owner or authorised editor. Never modify a school assessment resource and represent it as the official original.
Do captions reduce the need for a teacher?
No. Captions make information available through an additional route. Teaching also involves selecting examples, explaining mechanisms, checking understanding, repairing misconceptions and adapting to learner needs. Accurate access enables these activities rather than replacing them.
Are caption mistakes proof that AI is unsuitable for education?
No. They show that automated tools have specific strengths and limitations. The sensible response is an accountable workflow: use automation where helpful, review important outputs, correct errors and retain a human route for questions that affect learning or access.
Sources and research notes
W3C Web Accessibility Initiative, Captions/Subtitles, explains why automatic captions need verification, why speech-text mistakes can change meaning and how captions differ from other media resources: https://www.w3.org/WAI/media/av/captions/ . W3C, Making Audio and Video Media Accessible, offers broader guidance on captions and transcripts: https://www.w3.org/WAI/media/av/ . CAST Universal Design for Learning Guidelines 3.0 (2024) supports accessible representation of essential information: https://udlguidelines.cast.org/representation/perception/ . Education Endowment Foundation, Clarity is King—reducing extraneous load (2024) discusses how presentation can introduce additional working-memory demands: https://educationendowmentfoundation.org.uk/news/eef-blog-clarity-is-king-reducing-extraneous-load . IES What Works Clearinghouse, Providing Reading Interventions for Students in Grades 4–9, provides relevant background on vocabulary and contextual meaning, not a claim that the programme specifically improves caption quality: https://ies.ed.gov/ncee/wwc/practiceguide/29 . Accessibility recommendations and classroom verification examples above are reasoned applications; they are not reported results from a caption-correction trial.
For the distinct problem of a voice explanation disappearing before the learner can process it, see How Transient Information Works. For a different medium where visual input loses information, see Why a Homework Photo Can Be Harder to Read Than the Page. This page owns the automatic-caption-to-scientific-meaning boundary and does not replace those neighbours.
Takeaway: preserve the idea across every version
Back at the Clementi table, three students are watching the same lesson and receiving different versions of one sentence. The problem is not that one prefers captions. It is that the teacher’s idea must survive its journey through audio, text and visual timing. A good learner learns to notice a doubtful line. A good teacher makes essential captions accurate and accessible. A good parent asks for the exact discrepancy rather than choosing a side from memory. When the representation changes, the concept should remain inspectable. That is the small, practical promise behind a reliable educational video.