Self-Care or Tone-Deaf? Beyond Sentiment in 5,832 YouTube Comments

A computational social science analysis of 5,832 YouTube comments on Maudy Ayunda’s advice about managing distressing news. Two language models reviewed a date-blinded, stratified sample of 300 comments. Both-model critical labels rose across the selected cutoff, while strict source engagement was nearly unchanged; weak model agreement and no human validation keep context-loss findings exploratory. A separate screen covers 13 candidate videos whose repost status is not fully verified. The study does not establish viral onset, individual exposure, or causal effects.

Proof points

  • 5,832 Source Comments
  • 300 Dual-LLM Sample
  • Exploratory, Non-causal

Technologies

Python, Pandas, Computational Social Science, LLM Annotation, Statistical Analysis, YouTube Comment Analysis

Links

  • Source: private or not public

Problem

A burst of comments around a public video can look like a simple positive/negative sentiment shift, but polarity alone cannot show whether people responded to the source, changed its meaning, made evidence-based arguments, or attacked the person.

Approach

Analyzed a snapshot of 5,832 source-video comments, with two language models independently reviewing a date-blinded, stratified sample of 300 using the transcript and sourced Indonesia context. Compared model-coded stance, source engagement, and claim fidelity across a selected cutoff; screened 13 candidate videos separately because repost status was not confirmed.

Results

Post-stratified both-model critical labels rose from 8.3% before to 75.2% after the selected cutoff, but the strict share with source-engagement scores of at least 2 was nearly flat (8.3% versus 8.4%), with overlapping intervals on mean scores. Claim-fidelity and qualifier results are exploratory, with weak model agreement, small denominators, and no human validation. The cutoff is not a verified viral onset, and the observational data cannot establish exposure or causal effects.

Self-Care or Tone-Deaf? Beyond Sentiment in 5,832 YouTube Comments

Research Question

When audience attention to advice about managing distressing news surges during a national crisis, do comments still engage with the video itself? Sentiment alone cannot separate source-grounded criticism from claims that omit the video's qualifications, personal attacks, or reactions to an outside frame.

Corpus and Method

  • Source video: 5,832 timestamped YouTube comments on Maudy Ayunda’s Indonesian video about managing news exposure during disaster coverage, collected in a September 14, 2026 snapshot.
  • Annotation sample: Two language models independently reviewed a stratified, date-blinded sample of 300 comments. They received one comment at a time, the auto-caption transcript, and a short packet of sourced Indonesia context; dates, identities, periods, and neighboring comments were withheld.
  • Separate candidate-video screen: 72 comments from 13 candidate videos were analyzed separately because not all were verified as reposts.
  • Time comparison: September 11, 2026 at 00:00 WIB was selected as a deliberately early comparison cutoff, not discovered as the viral onset. Comment timestamps have mixed precision, and the original X post timestamp could not be verified.

Descriptive Findings

An interval-based changepoint analysis places a sharp rise in observed comment arrivals around September 14. The high-volume interval spans September 14 at midnight to September 15 at midnight WIB; the first hour-level comment is at 2:00 a.m. on September 14. This describes the YouTube snapshot, not when the video first went viral.

Post-stratified, both-model critical labels rose from 8.3% before the selected cutoff to 75.2% after it. This is a contrast in model-coded comments, not evidence that viewers changed their minds. The stricter source-engagement result was nearly unchanged: the share for which both models scored engagement at least 2 was 8.3% before and 8.4% after. Mean source-engagement scores moved from 0.72 to 0.62, with overlapping 95% bootstrap intervals.

Claim-fidelity scores raise a possible context-loss signal but remain especially uncertain: among sampled comments where both models agreed a checkable source claim was present, mean scores were 1.65 before the cutoff (10 comments) and 0.85 after (45). In a narrower qualifier test, both models marked the video's qualifier omitted in 1/1 pre-cutoff comments and 16/16 post-cutoff comments; those tiny denominators make this diagnostic only. A 16-comment screen for resemblance to a secondary-reporting frame does not establish copying or diffusion.

Limits and Next Step

The models disagreed on source engagement for about four in ten comments, and weighted κ for claim fidelity was 0.17. No human validation has been completed, and the auto-generated transcript has not been checked against the audio. The analysis cannot establish individual exposure, viewer-wide opinions, a viral-onset time, or causal effects. Human review of a date-blinded sample is the next step. Read the companion article, When a Self-Care Video Meets a National Crisis, What Happens to Context?.