Nous construisons l'IA dont l'enseignement supérieur a réellement besoin

Why AI in education is a sustainability question

Education's environmental footprint has always been dominated by physical things: travel to campus, heating and electricity for buildings, printed materials. Digitally delivered student support avoids some of that. But we are careful with that argument. We do not publish per-session carbon comparisons, because honest ones require assumptions we cannot verify for every institution.

What we can do is show you the choices we make to keep our own footprint small, and give you the data to assess it yourself.

First: almost no digital tool in education today is without AI features somewhere in it. Fully analog delivery is not realistic for most institutions, given budgets, student numbers, and expectations. The real question is not whether to use digital tools, but whether they are used mindfully, and whether suppliers are honest about their footprint.

Second: digital footprints deserve context, including ours.

Current estimates put a typical AI text query at roughly 0.24 to 0.34 Wh, figures shared by Google and OpenAI in 2025 and consistent with independent analysis by Epoch AI.

We share these numbers as digital literacy, not an excuse. They describe typical text queries. Longer reasoning tasks and agentic workflows, including parts of our own platform, use meaningfully more energy per interaction. And small per-query figures multiplied across billions of daily queries still add up: usage, not training, now accounts for an estimated 80 to 90% of AI's energy consumption, a figure also cited in the UNU-INWEH 2026 report and discussed further by Jisc's National Centre for AI.

The honest conclusion is not that AI is free. It is that mindful use of the right tool is what keeps digital education a net environmental good.

Our position on AI in education

Our position is the same one we give partner institutions: AI should be used where it is a genuine, efficient solution to a real problem, and nowhere else. This is not just an environmental principle, it is an educational one. AI does not replace educators, and not every task needs it.

That is why LearnWise is modular. Institutions adopt the components that solve their actual problems, rather than paying for, and powering, capabilities they do not need. Our customer success team works with every institution on using the products well, which includes not using them where a simpler tool does the job. LearnWise exists because universities accumulated too many overlapping edtech tools in a short space of time. Our goal is not repeating that pattern with AI.

We are not the authority on institutional AI policy, and better guidance than ours exists. For institutions developing their approach, we recommend the EAUC and Jisc guide, AI and Environmental Sustainability in Post-16 Education (June 2026), which includes a supplier transparency checklist and adaptable policy templates.

Their guide's own summary of the risks and opportunities is worth seeing in full, so we've pulled it directly below.

Our infrastructure and technical choices

Sustainability at LearnWise is built into how the product works, not bolted on afterward.

Measurable AWS infrastructure. LearnWise runs on Amazon Web Services. Our infrastructure emissions are reported through the AWS Sustainability console, covering Scope 1, 2, and 3 emissions attributed to our AWS usage, broken down by region and service, using both market-based and location-based accounting. The underlying calculations follow the AWS Customer Carbon Footprint Methodology v3.0 (October 2025), which is independently verified. Our infrastructure footprint is not an internal estimate. It is measured by a third-party-assured methodology any institution can review.

We make deliberate engineering decisions that reduce energy use per session. We do not train any AI models, so we do not carry that concentrated one-off energy cost of training. The larger and more recurring cost is inference: the energy used every time a model responds to a query. That is why we focus our engineering effort there and here is how our effort shows up in our product decisions:

Honesty about our current limits. Our primary model provider does not currently publish per-query energy data. Where we estimate inference energy, we use conservative third-party benchmarks and label them as estimates. We will update our figures as published data becomes available, including as more granular reporting emerges across the industry.

Sustainability is a work in progress

We think what a company refuses to say about sustainability tells you as much as what it says. So, plainly:

We do not claim to be carbon neutral or net zero. We have not completed a full Scope 1 to 3 emissions baseline, and we will not use those terms until we have.

We do not publish precise per-conversation carbon figures. The data to do this credibly does not yet exist across our stack, although provider disclosures in 2025–26 have begun to close this gap, and we update our estimates as they do.

We do not treat our cloud provider's commitments as our own. AWS has committed to being water positive by 2030 and reports a global data center water use efficiency of 0.12 liters per kWh (2024). Our workloads run within that infrastructure. Our own water footprint sits almost entirely inside our cloud provider's operations, and we will incorporate water into our reporting as our baseline work matures.