What the model is. Where your data stays.
What it will not do.
Four plain questions, answered without the two phrases that do most of the blurring in this market.
What this is
An AI transparency statement is a plain answer to four questions: what the model is, where your data stays, what the model can see, and what it will not do. This is ours. It is written for a teacher first and a data protection officer second, and it avoids the two phrases that do most of the blurring in this market.
Where your data stays
Where your data goes depends on how Aevum is running, and there are three configurations rather than one. In a school deployment, the software runs on the school's own hardware and pupils' work stays on the premises. During the trial, the model runs on a server Aevum operates in the United Kingdom, which is why the trial processes no pupil data at all. We set out all three, and which of our claims hold in each, on how Aevum runs.
What holds everywhere: your information is never pooled with other organisations, never used to train our models, and is not our business model.
How Aevum trains, precisely
“No training on your data” is not the same as “no training”, and we will not blur-sm the two. Aevum trains a shared base model on Dawn, the AIRR supercomputer at the University of Cambridge, on material we curated, before any customer ever sees it. In the building, that model can then adapt to how you use it, and that local learning stays local. Because a model that has learned from personal information is itself a kind of record, we treat it accordingly, with the access, retention and deletion obligations that follow.
What goes into the model
Aevum trains on licensed and openly-licensed material: UK curriculum and public-sector education content, open textbooks, and Aevum's own authored material. Content we do not have the rights to use is excluded, in code, before it can reach a corpus. We do not scrape the open web for other people's copyrighted work and quietly fold it in.
What the tool will not do
- We don't rely on US hyperscaler infrastructure to train our models, or run on a Silicon Valley account.
- We don't take your data to a cloud we control, pool it, or sell it.
- We don't carry advertising or monetise attention — you are the customer, not the product.
- We don't pretend “no training on your data” means “no training”.
- We don't pre-announce launch dates, run countdowns, or email you every week to keep you “engaged”. When it's ready, you'll get one email, with the link. That's the promise.
Inspecting the reasoning
A teacher can see why the tool suggested what it did, rather than trusting a black box. That is a design commitment we hold ourselves to: where we have measured it, we publish the measurement, and where the measurement is still running, the honest status stays “designed to, being tested” until the number exists.
What we publish before we claim
Every capability claim we make will be backed by a measurement we publish, and we publish the benchmark protocol before we run the experiment, so the test cannot quietly move to fit the result.
Disclosure
Aevum's atmospheric imagery is AI-generated on Aevum's own infrastructure, and we say so. A founder portrait is always a real photograph, never AI-generated.
Last reviewed .