Responsible AI Usage

How we use AI

Comprehenzo uses AI in two narrow places:

• Content generation — Anthropic's Claude models generate the listening passages, comprehension questions, and vocabulary for each scenario at a given level.
• Text-to-speech — ElevenLabs and Google Cloud Text-to-Speech convert the generated text into audio in the target language.

We do not use AI to grade your answers, recommend your next lesson, or personalize what you see. Those are deterministic — built from your own session history, the SM-2 spaced-repetition algorithm, and a curated curriculum graph. Keeping AI out of those paths means we don't pay an environmental cost on every screen tap.

Minimal energy and water footprint

Training and running large AI models consumes electricity and water (used for cooling). Inference — the act of generating a single response — is far cheaper than training, but it adds up fast at scale. Our architecture is designed to keep that footprint as low as we can make it:

• Generate once, serve many. Each scenario at each CEFR level is generated once and stored in our database. When you start a session, you receive pre-generated content. Two thousand learners doing the same scenario do not trigger two thousand AI calls; they trigger zero new ones.
• Audio is cached. Once the text-to-speech engine has spoken a phrase, the resulting audio file is stored and replayed for every subsequent learner who encounters it. We never regenerate audio that already exists.
• The smallest model that fits. We choose model sizes by job. Quick classification and routing tasks use small, low-energy models. We escalate to larger models only when the task genuinely requires it.
• Background batching. New session generation runs in a background queue with a small concurrency cap, so we're not spinning up bursts of inference compute that strand idle capacity.
• No model training on your data. Your sessions, answers, and progress are never used to train any AI model — ours or a third party's. Training is the most resource-intensive AI workload by a wide margin, and we don't do it.

We will not publish a fabricated number for grams of CO₂ or milliliters of water per session. The honest answer is that the per-session footprint of a cached, pre-generated lesson is small enough to be dominated by the network round-trips to deliver it. We're focused on keeping it that way.

Where we won't cut corners

Some uses of AI that we've deliberately ruled out, even though they'd be easy to add:

• No always-on AI tutor chat. The compute cost of an open-ended chat dwarfs a structured lesson, and the learning gain is unclear.
• No on-device large model. Running an LLM on your phone burns your battery and CPU; we keep generation server-side where it can be batched and cached.
• No AI-generated novelty for novelty's sake. New content is generated when the curriculum requires it, not on a continuous loop.

Where we want to do better

Honesty over greenwashing. Things we'd like to improve:

• We don't yet measure or publish per-session energy use directly. We rely on architectural choices to keep the number low rather than instrumenting it.
• We don't currently choose our hosting providers or AI vendors based on whether they run on renewable energy or in low-water-stress regions. That's on the list.
• The spaced-repetition algorithm could surface “review existing content” more aggressively over “generate new content,” which would push the cache-hit rate higher.

Giving back

Even with disciplined design, AI usage has a cost — to the grid, to water, and to communities that host the data centers. We commit to giving 10% of the founder's annual compensation to non-profits each year, split evenly between environmental groups and organizations supporting English Language Learners and social justice.

See this year's recipients →

Contact

Questions or suggestions about how we use AI? Email hello@comprehenzo.com.