The Turing Lectures: Making AI (truly) sustainable - from environmental costs to social impacts
What does it mean in real terms to make AI sustainable? What will it really take? It’s not news that AI has an environmental impact, and, as training models expand, so does the strain they place on the environment and energy grids around the world. AI models can transform society – improving efficiency, reducing wasted resources and accelerating innovation. Behind every model trained, every query processed and every dataset stored, though, lies a complex web of environmental, economic and social impacts. Dr Sasha Luccioni, climate and AI researcher, has featured on Business Insider's AI Power List for two years running, given two highly watched TED Talks on AI and, in 2024, was featured in both the BBC’s 100 inspiring and influential women from around the world and TIME Magazine's 100 most influential people in AI lists. 00:00 Welcome and Theme 01:25 Why Sustainable AI Matters 02:40 Meet Sasha Luccioni 04:33 What This Talk Covers 05:24 Origins of Artificial Intelligence 06:38 From Symbolic to Machine Learning 09:08 Transformers and the Scale Mindset 10:37 Consequences of Scale 10:54 Environmental Impacts and Transparency 15:10 Data Centres Energy and Emissions 17:40 Water Use and Cooling Tradeoffs 21:26 E-Waste and Hardware Turnover 22:45 Misleading Human vs AI Comparisons 24:56 Societal Impacts Beyond Carbon 25:50 Hidden AI in Everyday Tools 29:03 Surveillance Bias and Feedback Loops 30:59 AI for Good and Who Gets Replaced 32:30 Targeted Ads and Consumption Cycle 33:24 Why Ads Feel Like Eavesdropping 34:11 Cognitive Offloading and ChatGPT 36:32 Time Savings and Rebound Effects 37:43 Jevons Paradox Meets AI 39:40 Jobs Hype and Layoff Incentives 43:43 Technical Fixes Small Local Models 46:24 Measuring Emissions With CodeCarbon 47:14 AI That Helps the Climate 49:32 Regulation Standards and Energy Labels 53:36 Societal Push Democratisation Accountability 58:22 Organising at Work for Transparency 01:02:40 Q&A Start 01:05:12 Energy Score Meets Policy 01:07:08 Standards Are Hard 01:08:23 Org Guidelines For Fast AI 01:10:24 Coding Skills Vs Prompts 01:12:20 Estimating Copilot Emissions 01:13:36 VC Guardrails And Bubbles 01:16:08 Right Sizing Models With Routers 01:17:56 Holding Hyper-scalers Accountable 01:21:16 On Device Open Models Future 01:24:07 Shifting The AI Narrative 01:25:44 Incentivising AI For Good 01:27:58 Estimates Ranges And Trust 01:30:23 Closing Thanks And Wrap Up