AI summary by TwinMind
why I took down my climate science video
Sabine Hossenfelder
12m 29s
Initial Video Publication and Takedown
- Published strongly worded video about climate science study on LA wildfires
- Video disappeared day later, also deleted X/Twitter post
- Took down video after 12 hours when someone claimed misreading of statistical results
- Set video private to avoid spreading misinformation on policy-relevant topic affecting millions
World Weather Attribution Study Critique
- Study examined LA wildfires from January with press release claiming:
- "Human induced warming from burning fossil fuels made Peak January fire weather index 30% more probable"
- "High confidence that human induced climate change primarily driven by burning fossil fuels increased likelihood of devastating LA wildfires"
- Author's analysis showed results were statistically insignificant
- Results compatible with climate change having no effect on LA wildfires
- Study online for month, no other climate scientists noticed statistical insignificance
Climate Science Community Criticism
- Many climate scientists know World Weather Attribution Center studies are "sketchy" but stay silent
- Silent because speaking up would be "politically inconvenient"
- Pattern similar to foundations of physics destroying research area reputation
- Scientists look away from "crap going around in media" thinking "all attention better than no attention"
- Reason for author's distrust of scientists despite appearances
Extreme Event Attribution Methodology Problems
- Run climate models twice: once without global warming, once with current level
- Count similar events to target (LA wildfires), compare probabilities
- Major issues:
- Many climate models don't correctly predict extreme events initially
- Throw out most models (poor procedure)
- Results depend on exact definition of extreme event
- Too many details = never shows up in simulation (comparing zero to zero)
- Calculate probability ratio: >1 = more likely, <1 = less likely
Statistical Analysis Confusion
- Results table showed mean value 35% higher probability (1.35 ratio)
- 95% confidence interval included 1.0 (no effect)
- Author interpreted as not statistically significant
- Paper doesn't explain what they mean by statistical significance
- Standard: 95% confidence interval for significance (some areas require 99% or 99.999%)
Color Coding and Table Legend Problems
- Table caption claims all changes are increases
- Says decreases marked orange, increases marked blue
- Author questions if colors were mixed up
- Caption distinguishes light orange (non-significant) vs dark orange (significant)
- Multiple orange shades used inconsistently across tables
- Completely messed up table legends and color codes
Verification Process
- Double-checked with climate scientist who agreed (not statistically significant)
- Fed entire report to ChatGPT and Grok - both agreed not statistically significant
- Colleague emailed lead author after 3 days
- Received out-of-office reply, then response from co-author confirming:
- "Changes in intensity and likelihood are unsurprisingly not statistically significant"
- Light orange designation confirmed
Study Quality Assessment
- Rapid attribution study (not peer-reviewed)
- Not stated in press release
- Author considers study "so bad" with "sincere doubts even authors read it"
- Worse quality than initially thought after closer examination
Broader Context and Clarifications
- Climate change may have made wildfires more likely in some world regions
- LA wildfires specifically - this analysis tells nothing about climate influence
- Multiple research groups work on extreme event attribution
- Various methods available, some better/worse than others
- Not all extreme event attribution research is unreliable
- Some research "isn't crap" but author's opinion of climate scientists damaged
YouTube Algorithm Impact
- Taking video down temporarily "nukes YouTube recommendations"
- Algorithm treats private video as underperforming
- Stops recommending video after republishing
- Requests viewers click and let video run to help algorithm recovery
