AI Read 10 Years of Brazilian YouTube Climate Comments — Deniers Are Louder, Not More

A team of researchers wanted to know how ordinary people argue about climate change online — not politicians or scientists, just commenters. So they gathered more than 240,000 Portuguese-language comments posted on YouTube between 2014 and 2024, found through Brazil-focused climate searches. Then they used a language model (AI trained on huge amounts of text) from Meta, called Llama 3.1, to sort every comment into one of three buckets: believer, denier, or unclear. Doing that by hand would have taken years.

The trick was making the AI cheap and accurate enough. Instead of retraining the whole model — which costs serious money and computing power — the team used a technique that nudges the AI with a small add-on, then let it label easy examples on its own and feed those back in. The result was a sorting system that handled messy, sarcastic, and ambiguous comments far better than older methods, and didn't need an army of human annotators.

What they found flips a common assumption. Denialist comments were actually less common than pro-climate ones. But those denial comments attracted a much larger share of back-and-forth arguing across the divide. Picture a classroom where the climate-skeptical kid says less but starts more debates, while everyone else nods along in their own separate group. Pro-climate discussion, by contrast, mostly stayed inside like-minded threads — a classic echo chamber, where you hear your own view repeated back to you.

So what does this mean for you? It suggests the loudest voices online aren't necessarily the biggest group, and that your feed can make a debate look more evenly split — or more hostile — than it really is. The same AI shortcut could soon help researchers, journalists, and platforms track public opinion at a scale humans can't. The catch: this studied only Brazilian YouTube comments, not videos or other countries, and AI can misread jokes and sarcasm.