Digital sobriety for AI chats: practical habits that lower impact

You do not need to quit AI assistants to care about footprint. Most impact per decision usually comes from how long your prompts and answers are, which model tier you pick, and whether you repeat the same context in every thread. Small habits add up — especially across a team.

1. Shorten the loop

Fewer input and output tokens generally mean less compute per exchange. Edit prompts to remove filler, ask for bullet points when a memo is overkill, and avoid pasting entire PDFs when a summary paragraph would do.

2. Right-size the model

Frontier models are not always required for drafting or classification. When quality allows, smaller or faster endpoints can materially change energy per task — your provider’s docs and benchmarks are the source of truth for that trade-off.

3. Reuse context instead of re-uploading

Starting a fresh chat and re-attaching the same 20-page brief trains the model on duplicated tokens. One thread, one canonical context reduces repeated work for you and for the infrastructure behind the API.

4. See your own pattern

Indicative CO₂e per reply — derived from usage metadata, not from reading your message text on our servers — helps build intuition. The carbon-llm Chrome extension surfaces estimates in the chat UI; the My footprint page aggregates trends when you use the extension.

5. Keep perspective

Individual chat emissions are usually small next to flights, heating, or diet — but at enterprise scale, and when AI is embedded everywhere, measurement and procurement still matter for CSRD-style narratives and for honest product marketing.