Ai Impact Risk Acronym Slide 1080

Ai Impact Risk Climate Title

🤷‍♂️ Disclaimers

• I assume basic familiarity with AI and its environmental impact
• This is a rapidly evolving field
• Data is scarce and often unreliable
• I'm not an environmental scientist

Let's focus on Scope 1 emissions

"Operational emissions dominate AI hardware lifetime at 70% to 90% of total emissions, with manufacturing emissions under 25% and data center construction emissions under 5%."

Other environmental effects

Ai Impact Risk Climate Ewaste

what_uses_more_quiz_slide_a

what_uses_more_quiz_slide_b

🚀 Champions v. 💀 Critics

open_ai_microsoft_sparks_synergetics_india

bender_emily_stochastic_parrot

💀 Misleading claim: "LLMs are parrots"

LLMs can discover things stochastically.

Stochastic processes can surface insights.

🚀 Misleading claim: "AI's impact is a known quantity"

Industry has no incentive to share data.

Industry opacity makes true environmental impact difficult to measure

💀 Misleading claim: "massive water per prompt"

A bottle of water per email assumes:

• 📧 10-50 medium-length ChatGPT queries per email
• 📍 Location with poor water usage efficiency
• ⚡ No improvement in data center efficiency since 2020

The "bottle of water per email" claim oversimplifies the issue

🚀 Misleading claim: "negligible water per prompt"

• ⚔️ Mistral reports 45mL per query (2025).
• 📊 If a "median" query includes code then it could be an order of magnitude lower than a typical text query.
• 🚰 Requires municipal water, which strains local infrastructure
• ☀️ 2/3rds of recent data centers built in water-stressed areas (Bloomberg 2025)

Water usage estimates vary significantly depending on query type

🚀 Misleading claim: "negligible electricity per prompt"

Results vary dramatically based on:

• 🌍 Local energy source (x2)
• 🌡️ Local data center climate (x3)
• 🔢 Number of prompts (x10)
• 🧩 Prompt complexity (x30)

Electricity usage varies widely based on multiple factors

🚀 Misleading claim: "A low per-query footprint = low impact"

• 📈 The rise of agents: "One user consumed 5.7 million tokens in 13 hours (while he slept)."
• 🔄 The replacement fallacy (Rebecca Yeager).
• 🌍 Local impacts can far exceed global averages.
• 🏗️ AI may be a trojan horse to build data centers for non-AI uses.

❌ Misleading claim: "focusing on AI distracts from serious impacts"

• We waited too long to examine the harms of social media; AI offers a chance to model better behavior around a new technology. This time we can get it right!

Learning from social media: an opportunity to act earlier

AI's impact on digital curation tasks

Ordinary archival tasks are relatively low in energy consumption compared to AI tasks.

🤷‍♂️ What can we do?

🛠️ Tools

• What Uses More
• What Uses More Sources
• Code Carbon
• Your Digital Life
• CollaborAITE

What Uses More Exercise

Can you find equivalences by tweaking the parameters, e.g. how many:

• 📝 AI paragraphs equals 30 minutes of 🐈 TikTok?
• 🎬 AI videos equals an hour of 🍿 Netflix?
• 🖥 AI lines of code equals 1 hour of 🎦 Zoom?
• etc.

Links for today's workshop:

• What Uses More
• Your Digital Life
• Code Carbon
• CollaborAITE lead Greg Nelson
• AI IMPACT RISK framework (website, video, infographics, and quiz to help remember AI downsides)
• How to defeat a data center in 5 steps
• Learning With AI toolkit with 500 crowdsourced resources, strategies, and speaker list