Wrestling with AI: plain-text version

[Written by a human:] Although the initial instructions, when this project began, included a request to track the tokens (the small chunks of text an AI model reads and writes, which can be a proxy for estimating the cost of the AI use), the tracking didn’t happen. When asked why it did not happen, Claude responded that the chat app doesn’t show a running token count, so tracking it would have taken a deliberate step at the end of each session and “nobody was assigned to do it.” Claude’s best estimate (completed retroactively) is [End of human-written text.] that this project took roughly 100 million tokens, almost all of them the AI re-reading the growing conversation at each step, including about 400,000 tokens of text the AI wrote. Only one of the three working sessions could be measured: it used about 44 million tokens, including about 155,000 written by the AI. The other two are estimated from their length.

What does that mean in dollars and energy?

Tokens don’t translate neatly into money or electricity, and Anthropic doesn’t publish energy figures for its models, so these are rough translations.

Dollars. If a developer had paid Anthropic’s published prices for this model ($4 per million tokens read for the first time, $0.20 per million tokens re-read, and $20 per million tokens written), this project would have cost roughly $40, and almost certainly between $25 and $60 (Anthropic, n.d.). (Unchecked by a person; AI research only.) That’s not what the person building this site paid: the AI was used through a subscription plan, not billed per token.

Energy and water. Google reports that a typical text request to its Gemini AI uses about 0.24 watt-hours of electricity and about 0.26 milliliters of water, roughly five drops (Vahdat & Dean, 2025). (Unchecked by a person; AI research only.) An independent estimate for ChatGPT puts a typical question at about 0.3 watt-hours, but close to 40 watt-hours when the AI has to read about 100,000 tokens of material at once (You, 2025). (Unchecked by a person; AI research only.) This project sent the AI several hundred requests, most of them very long. Our best guess is that it used somewhere between 0.1 and 3 kilowatt-hours of electricity, most likely between half a kilowatt-hour and 1 kilowatt-hour. One kilowatt-hour is about what a laptop uses running for a full day, or a microwave running for an hour. Using Google’s water figures, 1 kilowatt-hour would mean roughly 1 liter of water for cooling. Other researchers count more, for example the water used to generate the electricity.

These are estimates built on other companies’ numbers for other AI models, so treat them as a sense of scale, not a measurement.

Trade-offs we made

Building version 1 meant choosing between goals that pull against each other. We wanted the site to exist now, without a budget, so we used AI for work that people should do, such as drawing the brain scans and researching the science, and we labeled that work instead of hiding it. We used images from Plants vs. Zombies because we suspect students recognize them, even though that means part of the site can’t be openly licensed. We didn’t ask students about the value of this site before we built it, and we haven’t yet asked students to evaluate it. We simplified the science, showing two zombies, one view of the brain, and a few regions, which makes it easier to learn but less complete. We used a spooky font only for a few words, so the rest stays easy to read. And we chose to publish an unreviewed draft in public, with every unchecked claim labeled, because being open about our mistakes seemed better than waiting until we had none.

What AI did here, and what people should do

Who else should be at the table