Where we stand on Ai

Last updated 16 September 2026

We use Ai. We are also wary of it. Both of those are true at once, and this page is our attempt to be clear about what we're doing, for the councils and SENCOs we work with.

Why we're doing this at all

Over the last year we have been experimenting with Ai, carefully, to find out what it can genuinely do for SEND.

So far that has produced Pop Pop Library, a searchable collection of over 1,100 SEND documents, and Ask Ai, a research assistant that sits on top of it and answers questions with a citation you can check. We have also built Survey, which puts one question to 40 councils' draft SEND reform plans at once, and an 18 step walkthrough of what an Ai search box on a SEND Local Offer could become.

We are members of LGAi, the Local Government Association's peer network of over 250 council officers looking at Ai critically. We would rather learn alongside councils than try to support them from outside.

Below are the 4 worries people raise most, and what we actually do about each one.

It makes things up

Ai models hallucinate. They write fluent, confident sentences that are simply not true. In SEND, that is not a quirk. It is the whole risk.

So Ask Ai is not allowed to answer from memory. It only ever sees real pages from real documents in the Library, found by a search first. It writes from those pages and nothing else. Then every quote in the answer is checked, by code, against the exact page it claims to come from.

In testing it caught the model paraphrasing a sentence from the SEND Review and presenting it as a direct quotation. The words were plausible. They existed nowhere. We also found and fixed a subtler bug where a genuine quote was credited to the wrong source, and had been passing as verified.

In a blind test on 16 real SEND questions in September 2026, Claude, the model we use, produced 95 quoted citations, and 7 per cent of them could not be verified. An open model we tested alongside it produced 8 quotes, and half of those could not be verified. Both made the same mistake on one question, calling real sources "the SEND Code of Practice" when they were not. So as we refine the model, and how it works, take everything you see in Ask Ai as

The rule for you is simpler. Read the source. Ask Ai exists to get you to the right page faster, not to be the page.

It uses a lot of energy

This is the one we can be least precise about, and we would rather say so than guess.

Every question on Ask Ai makes 3 calls to a model, and the main one reads up to 15 pages of source text before it writes a word. Anthropic does not publish an energy figure per question, and we have not found an independent one. The nearest public number is Google's, for its own Gemini service. A typical text prompt there uses 0.24 watt hours of electricity and about 5 drops of water. That is small per question, it comes with caveats, and it is not our model.

What we can control, we do. Our own website has been audited against the W3C Web Sustainability Guidelines. It ships almost no JavaScript, hosts its own fonts, and its analytics set no cookies. It also still serves 31MB of oversized images, which is on the list and not fixed yet.

We also do not run Ai where it is not needed. Dawn, our plain English checker, uses no Ai at all, by design. It is rules and word lists, running in your browser. Ask Ai is capped at 3 free questions, then 50 a day once you sign in. Partly that is cost. Partly it is that nobody needs to ask a machine 200 questions about EHCPs before lunch.

Where your data goes

Fundamentally we belive in data sovereignty.

When you ask a question on Ask Ai, your question and the source pages are sent to Anthropic, the company that makes Claude, and processed in the United States. Anthropic does not use it to train their models and deletes it from their systems within 30 days. The transfer is covered by the UK International Data Transfer Addendum. We keep a copy of your question and its answer for 12 months, so we can see how the tool is used and improve it. All of that is set out in our privacy notice.

The Library itself, every document and the search index, lives on our own server with OVHcloud, a French hosting company. Nothing about it is rented from an Ai company.

We would prefer the model to be in Europe too. We have been testing models hosted in Paris by Scaleway, and an open model on OVHcloud, precisely so that sending your question to the United States stops being a given. So far the honest result is that the US model answers better and makes fewer things up. We would rather tell you that than pretend, and we will keep testing.

Two more things. Nothing you type into Dawn leaves your browser. And nothing you enter into Spoke, our local offer checklist, is ever sent to an Ai. It stays in our database, in your council's workspace, and nowhere else.

Whose writing is in the Library

Data sovereignty cuts the other way too. Most of the Library is government guidance, tribunal decisions and inspection reports, published under the Open Government Licence. Some of it is writing by charities, parent carer forums and individual authors. For those, we ask. Each author gets an email and a one minute form to say yes, or title and link only, or no. If they say no, their documents come out. We record a licence against every document, and where we do not know it, the record says so.

It makes you stupid

People say this half as a joke. The research says it is not entirely one.

In 2025 MIT Media Lab wired up 54 people with EEG and had them write essays with an Ai, with a search engine, or with nothing. The Ai group showed the weakest brain connectivity, felt the least ownership of what they had written, and often could not quote their own essay back a few minutes later. Microsoft and Carnegie Mellon surveyed 319 knowledge workers the same year. The more people trusted the Ai, the less critical thinking they reported doing.

That is exactly the failure a SENCO cannot afford. The judgement about a child is the job. So we design against it.

Ask Ai answers with citations first and prose second. The sources open. The quote is on the page. The summary is a way in, not a way out. The Library is the main thing and Ask Ai is a door into it, which is why the Library came first and gets most of our time. Dawn is deliberately rules based, because a writer who is shown the rule learns it, and a writer who is shown a rewrite learns to press the button.

What we will not do

Use Ai to write about a named child.
Put anyone's writing into the Library 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 a council has asked for it.

How we build, and what's next

We use Ai to help write code as well. A person reads it, tests it and takes responsibility for it. Most of Ask Ai's quote checker exists because a person noticed the model getting something wrong.

What we are trying to learn next is how any of this supports the people doing the actual work. Can a SEND team see across 40 reform plans in an afternoon instead of a month. Can a SENCO find the paragraph of the Code they half remember without reading nearly 300 pages. Can a parent get a plain answer about EHCP timelines with the source underneath it. Those are the questions we are testing, in public, with real documents.

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. The date at the top tells you when we last looked at it.

cheers,
Jonny Holden