Tony Blair and the local authorities

They’ve noticed us

The Tony Blair Institute has added to its recent series of reports extolling the opportunities of AI in government with a report looking at local government. Governing in the Age of AI: Reimagining Local Government promises that there are huge savings waiting to be realised by councils that embrace AI. It also provides some specific advice for councillors. Which is exciting. Because I actually am a councillor.

AI eh?

The report is frustratingly vague about what it means by AI. Based on some of the case studies it references the report is taking a very expansive definition. It includes machine learning, predictive modelling, remote sensors even data analytics and digital service design.

Is it helpful to badge these things as AI? I guess it’s sexier than digital, data and technology. At its heart that’s really what this report seems to be calling for. Fixing the plumbing, improving case management systems, better data standards. I’m all for this. We’re all all for this I think. Some of us have been all for this for ten or twenty years.

So allow me a moment to be a little frustrated that the bigwigs in the TBI seem to have turned up, looked into a local authority and said “gosh there are some issues here, had you ever thought about using technology to transform your services and maybe getting IT suppliers to provide you with modern tools?”.

They have fed some data about an unnamed local authority into a model which claims AI (or, more prosaically, better tools) could save us 25% of staff costs. Which sounds pretty exciting. If a bit difficult to believe. They don’t talk about the cost of the tools but we are allowed to understand that these will be much less than the cost of the staff.

The future

Let’s look at the solutions the TBI folks are proposing.

A Devolved AI Service (DAIS) funded by central government but owned by local government. Anything we don’t have to pay for is superficially attractive but if you don’t pay the piper you don’t, in reality, get to call the tune. Still taht’s just details. DAIS would do a number of interesting things including encouraging the development of shared data standards, prising open monopolistic markets and bolstering technical procurement capacity.

These are actually good ideas which people in the sector have been talking about for quite a while (Local GDS anyone?). I’m inclined to think that there may be some benefits in this cross-sector approach. Some.

The local government sector is large and complex though which multiple stakeholders locally and across the UK Government. A single technical consultancy for local government (stick AI in the name if you feel you must) is going to struggle to have the sort of impact we need at scale. Ultimately we are talking about changing in a complex system and this Local GDS, sorry: DAIS, is a solution for a complicated frame.

Carry on governing

There are some sections which I think could be charitably classified as “those things you’re already doing… do more of those”.

Transcription technologies have the potential to save social workers time in writing up notes the report tells us. If we had better case management systems and shared data standards they could potentially save even more time. And, an under-appreciated potential strength of the technology it seems to me, they could make it easier for social workers (or other professionals) to focus on the relationship with their clients without being distracted by forms or notes. This is no doubt why local authorities are already experimenting with these tools.

In a section about helping citizens get the support they need (local navigation assistant) the report highlights issues with data linking, data sharing and content and service design. All of which are, in my experience, fair. None of which I would particularly badge as AI. Still if we have to call it AI to get people to look at it. So be it.

I didn’t get the section on planning at all. Does the TBI think AI can write our local plans? Will we also have AI planning inspectors? Will AI barristers provide LLM generated evidence at planning inquiries? Once again they list some case studies of interesting things that councils are doing in this space “do more of the good things”.

Call to councillors

Then we get to the bit just for me. Five things they say councillors should be doing

1. Champion AI innovation
    If I don’t have to call it AI I shall be very happy to continue to champion innovation around the use of data and technology.

2. Adopt tools that have been proven to work
    Yes. I mean I feel like that is what local government does. But I’m very much for this and against, for example, adopting tools that have been proven not to work.

3. Use convening power to de-risk adoption
    This seems to be a fancy way of saying “work with others” which I’m also in favour of.

4. Adopt existing standards and platforms
    Yes. And I would like to see more councillors talking about this.

5. Adapt the workforce and attract the right talent
    I mean, again, sort of the day job. It is not unhelpful to have a reminder to make sure that data and technology is factored into workforce planning.

This is fine

Overall this is fine. It’s OK. I recognise many of the problems that the report identifies. I think some sort of sector-wide consultancy might be helpful. I think the sector has been working on these issues and innovating in this area but we can always do more.

This is not reimagining local government. Which is not a problem other than that’s the title of the report. It’s imagining local government making better use of technology. Which is, frankly, fine.

I just don’t think this is really about AI unless you define AI so broadly that it becomes indistinguishable from computers.

Maybe it has to be packaged this way to get government to take it seriously. I don’t know.

Back to the coal face where we’re wrestling with ever increasing demand and diminishing funding and providing services which are, in a very large part, fundamentally about effective relationships between human beings not technology.

C19th landuse patterns (at a hack day)

Distribution of tithe map point data in Ceredigion

Hacking

I recently attended a Hacathon at the National Library of Wales. We were given access to a range of datasets and services from the library’s collections and given a few hours to see what we might build with them.

I went in a fairly conventional direction and this is a brief write up of what I attempted, how I did it and what I learned.

Tithe maps

I was interested in the tithe maps. Tithes were payments charged on land users. Tithe maps were produced between 1838 and 1850 to ensure that all tithes were paid with money rather than produce. The Library has digitised and georeferenced the maps. You can view the digital tithe maps online. The library has also transcribed the attributes associated with each field and these are available as points (each point geolocated to the appropriate field).

The land use of the field was recorded in these datasets.

For the hackday the Libaray gave us temporary internal API to access the data.

I wondered whether it would be feasible to use the data to create a land use map.

And basically it is.

Crunching the numbers

This is what I did.

I could have done this in QGIS but I really wanted to learn how to do it in R. I typically use R for data analysis unless it is spatial data. It would be smoother if I could do everything in R…

I downloaded the point data for Ceredigion (being big enough to have different land uses but small enough that it wouldn’t cause me problems in processing large datasets).

And I sourced a shapefile containing 1km grid squares from the Met Office (thanks MetOffice). This is for the UK but I trimmed it just to Ceredigion.

The landuse descriptions in the dataset are a bit messy as you would expect from transcription of 19th century handwriting. I generated a list of all the unique values and quickly grouped them by hand in a spreadsheet.

I made some fairly arbitrary decisions.

• Meadow and pasture are grouped together into “Pasture”

• Gardens, cottages and roads were all grouped as “Buildings”

• Anything that looked like a crop (including hay which you might argue is actually what you grow on a meadow) is “Arable”

• Lots of things including waste, gorse, moor etc were grouped into “Waste”

• “Water” includes, well, water

• “Woodland” also includes orchards.

In many cases multiple land uses were give for a single record. I just used the first one. Sorry but it was a hacathon so I was hacking.

Then it was a simple case of counting the number of points of each type in each polygon (which is lovely and easy to do with the sf package in R).

This enables some nice looking visualisations. For example if we display the most frequently occuring landuse type in each square:

Or we can show what percentage of points in each square were from a particular landuse.

Such as arable:

Or waste:

Or pasture:

My thinking is that we could compare this data to more modern landuse data and examine changes in landuse over time.

What surprised me is how low woodland came out in this analysis.

I couldn’t find an easy to work with dataset of landuse to do the comparison but I imagine that woodland would be a much heavier landuse especially in the south eastern band these days.

Limits

There’s one obvious limitation with the source data: fields are of different shapes and sizes but the landuse data is held against a point. So a big field counts the same as a small field.

I think this is somewhat compensated for by the fact that lots of smaller fields will have a lot more points so if we are looking at counts and proportions of points in a specific geography points in a larger fields will have a larger impact because there will be fewer points.

Tracing the geometries of the fields would solve that problem of course. But It would be a considerable labour for humans and I’m not sure how well ML would handle it. Nice project for a future hackday perhaps?

1km squares are quite large and it might be nice to use smaller squares to get more granular data. The process would be exactly the same. For a hackday proof of concept I think this is fine.

What did I learn?

The data was nice to work with (thanks National Library) but as with almost all data projects data cleaning was the most time consuming aspects.

I think looking at this approach with someone who had a bit more knowledge of 19th century land-use descriptors and handling multiple land-uses in a more sophisticated way would be what I’d do next.

I DID learn how to handle this sort of analysis in R and that should prompt me to reach for R next time I have to process some spatial data.

Thanks to Jason Evans and the other National Library staff who made this possible.

Code and data

The (very hacky) R Script and the relevant data is on GitHub tithe-maps-2025. Thank you to the National Library for allowing me to share the Tithe Map data.