About CPWB

Our use of AI

Last updated: October 17, 2026

We’ll be honest: there’s a real tension in an environmental conservation group using artificial intelligence. AI’s data centers consume electricity and water, and we don’t think an organization devoted to protecting Waquoit Bay should pretend otherwise.

We use these tools deliberately and sparingly — to do more with the small, volunteer-driven team we have, so that limited resources go toward protecting the bay rather than administrative overhead. Most of the summaries in our Research Library, newsletter archive and news section were drafted with AI from the linked source, then published by us; each one says so with a “Drafted with AI assistance” note that links here. Two automated jobs on this website also use AI: gathering partner organizations’ events for our calendar, and suggesting related reading at the end of articles. We’re mindful that this choice carries an environmental cost.

Rather than claim to cancel that cost out, we’d rather just be transparent about it. Based on published estimates, a typical AI query uses roughly a quarter to a third of a watt-hour of electricity, and between a few drops and a few milliliters of water, depending on whether you count only data-center cooling or also the water used to generate the electricity.

These are order-of-magnitude figures, not a precise ledger. The real numbers depend on factors the providers don’t fully publish, and we’d rather show you an honest estimate than a false precision.

Our commitment is restraint: we use these tools only where they genuinely free up volunteer time for the bay, we don’t spin up wasteful uses, and we keep watching a fast-moving field as more efficient and transparent options emerge. We welcome your questions and your accountability on this — send us a message.

What we send to the AI service

Our AI tools run under CPWB’s own account with Anthropic, the company that makes Claude. They read public web pages and the articles on this site. We don’t send them information about our visitors, members or donors.

How we estimate this

• Every AI request this website’s automated jobs make is logged with its size in tokens, the units AI models read and write. The figure above adds up the past 12 months and updates daily.
• We convert each request into “typical queries”: longer requests count as more than one, and every request counts as at least one. Larger models count for more than smaller ones. These weightings are our own judgment, because the providers don’t publish them.
• Drafting our archive is included as an estimate. We drafted about 210 summaries with Claude in chat sessions that keep no usage log, so we count each one as 20 typical queries — enough to cover reading the source, writing the summary, and the model’s reasoning along the way. It’s based on a typical post rather than measured, and could be off by a factor of several in either direction.
• We multiply by published per-query estimates: 0.24–0.34 Wh of electricity, 0.26–3 mL of water and 0.03–0.3 g of CO₂-equivalent. The upper carbon figure is our own conservative allowance for electricity from carbon-intensive grids.
• Not included: the AI used to build this website and its tools, AI-powered web searches, and the energy used to train the models.

Sources:

• Google Cloud, “Measuring the environmental impact of AI inference” (median Gemini text prompt: 0.24 Wh, 0.03 g CO₂e, 0.26 mL of water)
• Towards Data Science, “Let’s Analyze OpenAI’s Claims About ChatGPT Energy Use” (about 0.34 Wh per average ChatGPT query)
• Jegham et al. (2025), “How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference” (method, and water used in generating electricity)