Ai statement

We use Ai. We are also wary of it. Both are true at once, and this page says plainly what we're doing, for the councils and SENCOs we work with.

What we're doing

Over the last year we have been experimenting with AI, carefully, to find out what it can genuinely do for SEND. Anything on this site that uses AI carries our AI mark.

We have our reservations. So far the results are interesting, not perfect, and enough to make us feel that AI used responsibly has real advantages.

We are members of LGAi, the Local Government Association's peer network of over 250 council officers looking at AI critically. Our principle of collaboration means we learn alongside councils. 

Our concerns

AI makes things up

AI models hallucinate, fluently and confidently. In SEND that is a serious risk, not a quirk. 

We build in stages and processes to make sure that AI does nothing from memory.

For example, in Ask Ai, every answer comes from real documents, every quote is checked against its source by code, and the citation is the product. You should read the summary critically. The citations are real, just as if you were gathering your highlighted notes from across many documents. 

It uses a lot of energy

This is the one we can't really test very well, until we start using a difference provider. At the moment we're using Anthropic's Claude, be we're still experimenting, and they don't tell us about energy use. Once we run an AI on and API (hosted in the EU) we'll have a better idea. 

But what we do know is that not all data centres are the same, some run on green energy, and some are just data farms for social media and advertising companies. 

We do not run AI where it is not needed.

For example, Dawn, our plain English checker, uses no AI at all, by design. It's just rules and word lists, running in your browser. 

Where your data goes

On principle, we believe in data sovereignty. We work hard to try to use open AI models based in the UK or EU. 

You should be able to answer three questions about any AI supplier

• which company, in which country, controls the system
• what leaves your estate, where it is kept, for how long, and whether it trains anything
• and what you take with you if you leave. We answer all three, per tool, in plain English, and our privacy notice has the detail.

We don't know who how laws will change in different jurisdictions, so we make sure we're flexible on the models we use. 

Our privacy and terms pages clearly explain how this works for each product. 

Whose writing it is

Sovereignty cuts both ways. Most of what our tools read is government guidance published under the Open Government Licence. Where it is someone's own writing, we ask the copyright owner, we record the licence, and if they say no it comes out.

It makes you stupid

People say this half as a joke. It is not. I know from personal experience that it's true.

If you ask an AI to do everything then the thinking, the part of the job that matters, is lost. So we design for your judgement, not around it. Citations first, prose second, answers kept short, and nothing that decides anything about a child.

So where we use AI, and make AI tools, we design them to keep your critical thinking ahead of everything else. 

For example, Ask Ai answers with citations first and prose second. You can see the sources (you've probably already read them at some point already). The quote is shown explicitly, not just a vague reference. The summary is there because it makes the search better, but don't copy it verbatim into a report.

AI tends to say way too much too. We're using what we know from writing well for the web (built into Dawn) to make sure that the text output is simple. 

AI vendor lock-in

We are making sure that we can swap out of one AI to another very quickly. We don't want anyone to be stuck with a long AI contract on the wrong model.

The AI model is a service, not the feature. What we own, and a council owns by working with us, is the content, the search, the checking. The model and its host are choices, made on results, and they can change. 

For examples, we have a standard set of questions for Ask Ai to evaluate new models and see how they compare. Once we start getting better results with a new model, we can easily make the change. 

What we will not do

• Use AI to write about a named child.
• Put anyone's writing into a tool without asking them.
• Publish an AI answer without a source you can open.
• Hide that AI was involved in something.
• Build a feature because it is possible rather than because it makes things better for you.

How we build, and what's next

Every tool that uses AI has a page saying what it is made of: the interface, the harness, where it looks things up, which model, who hosts it, and how we test it. Start with how Ask Ai is built (opens in a new tab). Dawn, our plain English checker, uses no Ai at all.

---

 

If you work in a council or a school and want to try any of it, or tell me where it is wrong, I hold office hours on Tuesdays at 10am. Bring a hard question.

This page will change as we learn.

cheers,
Jonny Holden