How do you stop automatic captions from changing product claims?

01

Which files should be frozen before caption QA begins?

Freeze the final video, final mixed audio, approved claim language, and caption file generated for that exact export. Do not review captions against an earlier script or a different cut.

The shooting script is not necessarily a transcript of the delivered performance. A creator may paraphrase, pause differently, omit a word, or record a replacement line. W3C advises that an existing script will probably need editing to match the final audio and that a transcription should not add, adapt, or “correct” what was actually said (W3C).

Use two references for different purposes. The final audible performance answers, “What should the captions transcribe?” The approved claim sheet or approved script answers, “Was the spoken claim authorized?” If the creator says “eliminates stains” but only “may help reduce stains” was approved, do not silently weaken the caption while leaving the stronger audio. Stop delivery and escalate the audio mismatch.

1. Identify the exact final video filename and version.
2. Confirm that no audio replacement or remix is pending.
3. Attach the approved transcript, script, or claim sheet.
4. Generate or export captions from that final version only.
5. Label the track as generated or draft until caption QA is complete.

02

Which words should receive claim-sensitive marking?

Mark any token whose removal, substitution, or movement could change what the product appears to do, for whom, under which conditions, or to what extent.

Start with brand, product, model, ingredient, and feature names. Then mark quantities and measurement language: numbers, percentages, prices, decimal points, dimensions, doses, volts, watts, milligrams, millilitres, hours, days, inches, and degrees. “15 mg” becoming “50 mg” is not an ordinary typo; it is a different quantity.

Also mark negatives such as “not,” “never,” “doesn’t,” and “without”; qualifiers such as “may,” “can,” “helps,” “designed to,” and “up to”; and conditions such as “when used as directed” or “with compatible devices.” Include comparisons, population limits, availability restrictions, sponsorship notices, safety language, and other required disclosures.

This priority reflects the FTC’s position that advertisers need a reasonable basis for objective express and implied claims before dissemination, with support matching the claim conveyed (FTC advertising substantiation statement). Caption QA does not substantiate a claim, but it can prevent transcription from creating a different one.

03

How should each marked claim be compared with the audio?

Review each marked phrase at reduced playback speed and compare it directly with the audio. Correct meaning before punctuation, capitalization, spelling style, or line appearance.

Use a simple issue log with four columns: audio, caption, risk, and action. If the audio says “may help reduce” and the caption says “helps reduce,” restore “may.” If “up to 12 hours” becomes “12 hours,” restore “up to.” If “does not contain” becomes “does contain,” stop delivery until the reversal is corrected and replayed in context.

Check character by character when reviewing model numbers, decimal points, percentages, prices, compatibility references, and units. Do not rely on the sentence looking plausible. W3C notes that automatic text generally requires editing and should not be treated as sufficient unless its accuracy has been confirmed (W3C caption guidance).

Do not add helpful-sounding product details that the creator did not say. If clarification is required, revise the underlying audio or obtain approval for intentional on-screen copy rather than disguising new copy as transcription.

1. Listen once without reading to establish what was said.
2. Replay while reading the caption track.
3. Verify every highlighted token against the audio and approved language.
4. Correct the caption or escalate the audio; do not hide an audio problem.
5. Replay the corrected section from before the caption begins until after it ends.

04

How do you test whether captions create a stronger net impression?

Watch the finished video muted and state the product promise conveyed by the captions alone. Then watch with sound and compare the two interpretations.

Ask whether the caption removes uncertainty, drops a limitation, broadens compatibility, changes a time period, or makes an individual experience look like a general promise. Check captions alongside headlines, demonstrations, pack shots, disclaimers, and other on-screen text rather than evaluating each phrase in isolation.

FTC health-product guidance explains that express and implied claims are evaluated through the ad’s overall net impression. It also says necessary qualifications should be noticeable, understandable, close to the claim, and consistent with the rest of the ad (FTC guidance). Although that guidance addresses health-product advertising, the review question is broadly useful: what would a reasonable viewer understand from the complete presentation?

If muted viewing produces a stronger promise than the audio, the caption track has failed even when most individual words are spelled correctly.

05

Can timing and line breaks change a claim?

Yes. A caption can temporarily present an unqualified claim if its limitation appears late, disappears too quickly, or is split into a separate caption.

Keep “may,” “up to,” “when used as directed,” and similar limitations in the same readable phrase as the language they modify. Avoid displaying “Eliminates stains” first and adding “in one laboratory test on one fabric type” only in the next caption. The first frame can be read and captured without its condition.

Break longer captions at logical phrase boundaries, keep them synchronized with the speech, and allow viewers enough time to read the complete thought. W3C recommends synchronized captions, generally using one or two lines and logical line breaks (W3C transcription guidance).

After editing timing, scrub frame by frame at caption boundaries. Look for isolated numbers, detached units, orphaned negatives, and qualifiers that vanish before the main claim.

06

Why should captions be approved as a separate deliverable?

Video approval should not automatically approve its caption track. Captions contain both text and timing, so they need their own identity, reviewer, status, and relationship to a specific video export.

Record the video filename, video version, caption language, file format, caption version, reviewer, review date, and approval status. A practical filename is campaign_product_cut03_en-US_captions_v04_APPROVED.srt. Store generated, corrected, and approved files separately so an old track cannot be mistaken for the current one.

YouTube allows creators to upload timed caption files, edit caption text or timing, download caption files, and upload revised versions (YouTube Help). Editing an automatic track creates a revised caption track, which makes clear version labels operationally important (YouTube Help).

ATIYO can preserve the brief, approved language, asset version, caption-QA decision, revision history, and reusable learning together. Media performance remains in the ad platform. ATIYO preserves creative context and learnings so teams can understand what changed, what was approved, and what should carry into the next version.

1. Approve the final video export.
2. Complete claim-sensitive caption QA against that export.
3. Save and label the approved caption file separately.
4. Link the caption approval to the matching video version.
5. Include both files and their status in the delivery manifest.

07

When must caption QA be repeated?

Repeat claim-sensitive QA whenever the video, audio, language, timing, caption styling, or publishing system changes.

Recheck cutdowns, reordered scenes, audio replacements, aspect-ratio variants, burned-in captions, translations, transcript auto-sync operations, and platform-generated tracks. A correct master caption file does not prove that a later derivative or viewer-facing platform track is correct.

Treat every translation as a new claim-bearing deliverable. YouTube distinguishes same-language captions from subtitles that translate spoken language (YouTube glossary). A translated product name, quantity, qualifier, or compatibility condition therefore needs review by someone who understands both the target language and the approved claim.

Finally, inspect the published video as a viewer would. Confirm that the intended caption track is active and that any platform-generated version does not replace or conflict with it.

1. Re-run the marked-token comparison after each transformation.
2. Check line breaks and timing in the actual aspect ratio.
3. Validate translations against approved source claims.
4. Inspect the platform’s viewer-facing caption track after upload.
5. Archive the approved derivative and record what changed.