TL;DR
An empirical analysis of Hacker News posts finds roughly 65% register as negative and those posts average 35.6 points versus a 28-point site average, a ~27% edge. The result holds across six different sentiment models, though causation between negativity and attention remains unresolved.
What happened
A researcher studying attention dynamics on Hacker News analyzed 32,000 posts and 340,000 comments to measure sentiment and its relationship to engagement. Using six classifiers — three transformer-based models (DistilBERT, BERT Multi, RoBERTa) and three large language models (Llama 3.1 8B, Mistral 3.1 24B, Gemma 3 12B) — the analysis found a persistent negative skew: about 65% of posts are labeled negative. Posts classified as negative averaged 35.6 points on Hacker News compared with an overall average of 28 points, a roughly 27% performance premium. The negative bias appears across all tested models, though individual distributions vary and some models required inverted scales. The author notes that much of the negative content is technical critique rather than personal attacks, and raises the open question of whether negative framing drives engagement or if controversial topics attract both negativity and attention. A preprint is available on SSRN and the author plans to publish the code, dataset, and a dashboard soon.
Why it matters
- Signals about what content draws attention can inform how tech discussions and product announcements are framed on developer-focused platforms.
- A persistent negative sentiment skew complicates automated moderation and community health metrics because critique and toxicity are not the same.
- Researchers and platform operators should consider model selection and possible classifier bias when measuring sentiment at scale.
- For journalists and analysts, the finding highlights that engagement metrics may reflect tone as well as substance, affecting how stories spread.
Key facts
- Dataset: 32,000 Hacker News posts and 340,000 comments analyzed.
- Sentiment split: roughly 65% of posts labeled as negative by the study's classifiers.
- Engagement: negative posts averaged 35.6 points vs. a 28-point overall average (≈27% higher).
- Models tested: DistilBERT, BERT Multi, RoBERTa, Llama 3.1 8B, Mistral 3.1 24B, Gemma 3 12B.
- Negative skew persisted across all six models, despite variation in distributions and some inverted scoring scales.
- DistilBERT was used for the author's dashboard because it runs efficiently in a Cloudflare-based pipeline.
- Definition of 'negative' in the analysis includes technical criticism, skepticism of announcements, complaints about practices, and API frustration.
- The author distinguishes substantive technical critique from personal attacks and notes most negativity observed is not toxic.
- A preprint reporting these results is available on SSRN; the author intends to release code, the dataset, and a dashboard.
What to watch next
- Release of the public code, dataset, and dashboard that the researcher has pledged to publish (confirmed in the source).
- Further analyses or updates from the author addressing whether negativity causally drives engagement or both are driven by controversial topics (confirmed as an open question in the source).
- Whether the sentiment classifiers will be recalibrated or replaced in future iterations of the dashboard (not confirmed in the source).
Quick glossary
- Sentiment analysis: Automated techniques that classify text according to emotional tone or stance, commonly labeled as positive, negative, or neutral.
- Transformer (model): A neural network architecture widely used in natural language processing that relies on attention mechanisms to process text.
- LLM (Large Language Model): A class of machine learning models trained on large text corpora to generate or analyze human-like text.
- Classifier calibration: The process of adjusting a model so its predicted probabilities or labels accurately reflect real-world distributions and meanings.
- Hacker News: A technology-focused discussion forum and social news site where users submit links and comment on tech and startup topics.
Reader FAQ
Does the study prove negative posts cause higher engagement?
Not confirmed in the source; the author notes causation is unresolved and both directions are possible.
What data and scale did the analysis use?
The analysis covers 32,000 posts and 340,000 comments from Hacker News.
Which models were used to label sentiment?
Three transformer models (DistilBERT, BERT Multi, RoBERTa) and three LLMs (Llama 3.1 8B, Mistral 3.1 24B, Gemma 3 12B).
Will the code and dataset be available?
The author says they will publish the full code, dataset, and a dashboard soon (confirmed in the source).
Is the negativity observed mainly toxic or personal attacks?
According to the author, most negative posts are substantive technical critique rather than toxic personal attacks.
65% of Hacker News Posts Have Negative Sentiment, and They Outperform January 6, 2026 • 300 words • 2 min read • ∞ Posts with negative sentiment average 35.6 points…
Sources
- 65% of Hacker News Posts Have Negative Sentiment, and They Outperform
- The value of downvoting, or, how Hacker News gets it …
- The Value of Downvoting, or, How Hacker News Gets It …
- Crowdsourced Censorship: The Irony of Hacker News
Related posts
- AI server chip rush set to push DRAM and HBM prices up as much as 70% by mid-2026
- Pairing NotebookLM with Gemini removes notebook isolation and connects notes
- System lets you control your Mac remotely using natural language