#5850.0ScienceT1 — official sourceCooling

Generative AI uses more energy on non-English text

Spotted 2026-11-15 02:08 (UTC+8) · original event 2026-11-05 · sources: Hacker News, University of Michigan, Futurity · spreading on 3 platforms

Heat score

Rank score: 6040% — where it sits on the charts
Discussion score: 8040% — how much people are talking about it
Cross-platform score: 7020% — how many platforms it spread to
Composite heat score:50.0rank × 0.4 + discussion × 0.4 + cross-platform × 0.2 — includes a -20-point cooling decay (peak passed, attention fading).

Why it matters

AI is priced in English: if your first language is not English, the same answer costs up to 15 times the energy and comes back worse, so the people who can least afford the waste pay the most for it.

The angle

Speaking Hindi to AI costs 8x the energy of English: the Michigan study that measured the hidden language tax, and why your tokenizer decides your billSuggested format: Short video · for AI users worldwide, non-English speakers, AI energy and fairness followers

Evidence chain

• HN #7, 104 pts/70 cmt (2026-11-15)• University of Michigan Engineering study 'The Language-Energy Divide' (Naihao Deng et al., EMNLP 2026, DOI: 10.18653/2026.emnlp-main.2113): generative AI uses 1.4 to 180 times more energy on non-English text than equivalent English• Hindi 7.9x, Bengali 15.3x• about 5.2 billion people need more than twice the energy to use AI in their first language• measured on Qwen 3, Gemma 3 and Llama 3.1 across 122 languages (Belebele dataset) with the ML.Energy framework• more tokens per word is the driver, and non-English output also tends to be less accurate• U-M News 2026-11-05, Futurity 2026-11-09, TechXplore and ScienMag covered it• 2026-11-15 03:07 update: HN #10, 125 pts/77 cmt• score unchanged 76.0 (rank 85, discussion 70, cross 70)• 2026-11-15 04:07 update: HN #14, 136 pts/84 cmt• rank 85 -> 75• 2026-11-15 05:07 update: HN #13, 145 pts/87 cmt• score unchanged 72.0 (rank 75, discussion 70, cross 70)• 2026-11-15 06:07 update: HN #13, 159 pts/89 cmt• score unchanged 72.0 (rank 75, discussion 70, cross 70)• 2026-11-15 07:07 update: HN #15, 165 pts/90 cmt• score unchanged 72.0 (rank 75, discussion 70, cross 70)• 2026-11-15 08:07 update: HN #15, 170 pts/95 cmt• score unchanged 72.0 (rank 75, discussion 70, cross 70)• 2026-11-15 09:07 update: HN #16, 174 pts/98 cmt• score unchanged 72.0 (rank 75, discussion 70, cross 70)• 2026-11-15 10:07 update: HN #15, 180 pts/98 cmt• score unchanged 72.0 (rank 75, discussion 70, cross 70)• 2026-11-15 11:07 update: HN #17, 187 pts/99 cmt• score unchanged 72.0 (rank 75, discussion 70, cross 70)• 2026-11-15 12:07 update: HN #19, 188 pts/102 cmt• score unchanged 72.0 (rank 75, discussion 70, cross 70)• 2026-11-15 13:07 update: HN #24, 191 pts/102 cmt• rank 75 -> 60• 2026-11-15 14:07 update: HN #24, 196 pts/103 cmt• score unchanged 66.0 (rank 60, discussion 70, cross 70)• 2026-11-15 15:07 update: HN #23, 200 pts/104 cmt• score 66.0->70.0 (rank 60, discussion 70->80, cross 70)• 2026-11-15 16:07 update: HN #26, 203 pts/105 cmt• score unchanged 70.0 (rank 60, discussion 80, cross 70)• 2026-11-15 17:07 update: HN #27, 206 pts/107 cmt• score unchanged 70.0 (rank 60, discussion 80, cross 70)• 2026-11-15 18:07 update: HN #28, 210 pts/111 cmt• score unchanged 70.0 (rank 60, discussion 80, cross 70)• 2026-11-15 19:07 update: HN #29, 214 pts/111 cmt• score unchanged 70.0 (rank 60, discussion 80, cross 70)• 2026-11-15 20:07 update: HN #26, 215 pts/112 cmt• score unchanged 70.0 (rank 60, discussion 80, cross 70)• 2026-11-15 21:07 update: HN #25, 218 pts/113 cmt• score unchanged 70.0 (rank 60, discussion 80, cross 70)• 2026-11-15 22:07 update: HN #29, 224 pts/119 cmt• score unchanged 70.0 (rank 60, discussion 80, cross 70)• 2026-11-15 23:07 update: dropped out of HN top 30• cooling -20 (70.0 -> 50.0)