The Dangers

This will be the last weeknote of this series (series 21!) which started last summer. I’ve finished up at MHCLG and Sport England, now enjoying some time off between contracts.

I’m starting a piece of digital assurance work for a new client, but other than that I’m just reading, tinkering with a side project, and building up to the marathon on 19 April.

Wrapping up the Data & AI Lab

Last week we wrapped up the Data & AI Lab at Sport England. This meant documenting the work, the principles which guided us, clearly describing the problems we worked on, and packaging things so that other people could use them.

The lab is being paused until it can be resumed, so I wanted the people who come next to know how we worked, what worked and what we’d change. I wrote up our guiding principles, our methodology (plus what we’d change), and other docs that try to pass on our mindset.

That’s always been really important to me. Heck, it’s one of the main reasons I work in the open. So it didn’t really feel like work, it was fun.

We got some really good feedback from the teams we worked with:

‘If you’d have asked us to spend 4 hours in a workshop at the start of this, we’d have said we didn’t have the time. But seeing what you’ve done now, we’d find 2–3 days at least to spend with you.’

Always good to leave people wanting more!

Symbols to think and talk with

Half of what I did was making symbols to think and talk with through AI prototyping. Not making a thing we’d go and build, but rather making a thing to provoke conversation and define more edges.

There is a danger with prototyping in general – let alone prototyping with AI – that people may believe that what you’ve created is a finished, functioning product. So I caveated every demonstration with what became a mantra: ‘We might have got more wrong than we’ve got right, and this is all about talking to each other: we’re not going to build anything you see.’

In one instance I didn’t get to control that message, and the prototype was demo’d without caveats, which set hares running. Word of warning to anyone trying to poke holes in reality: consider making caveats more pronounced. Create friction to ensure people read them.

Reality drift

People are placing outsized trust in AI products, but you can’t blame them: these things are marketed as omnisolutions, tools capable of meeting any need you might have. This is particularly pronounced when people turn to AI products for help with analysis, not realising the risks inherent with LLMs as a technology.

It’s really hard to explain that to non-technical people though. The risks stem from how the technology is built. Common practices in analysis can help prevent errors, but if you’re outsourcing your thinking to an algorithm, you won’t know about those. And loads of people just switch off when you start explaining technical concepts.

So I came up with a term to help describe the problem: reality drift. If you’re using Copilot or ChatGPT or Claude for user research analysis, data analysis, or generating insights to inform your strategy, please do read why reality drift is a danger, and how to avoid it, on the Boring Magic blog.

Oh yeah, please do follow the Boring Magic RSS feed.

Running

My ankle continued to hurt during training runs, so two weeks ago I stopped running altogether. This means I’ve missed the last two weeks of my training programme and tapered early. That’s not the best prospect when you’ve an upcoming marathon staring you in the face, but I’ve decided not to be worried about it.

My thinking is that I’ve got enough base fitness to pull it off. I’ve run a marathon before – one with 1,300 metres of elevation! – and the other week I ran 27km on a hangover. So I can run over halfway while feeling shit and being dehydrated, and I’ve got the mental fortitude to press on regardless.

I’ve been cycling to keep the fitness up. The pain is basically non-existent now, so the swelling has clearly gone right down. My osteopath recommended a few 10km runs on grass to test it out, which I can use as a guide on whether to proceed or not.

A few years ago I’d never consider running a 10km without perfectly completing a training programme. But now I’ve got just enough experience to make a confident-ish bet I could finish 42.2km on almost-complete training.

Fuck it, we’ll find out.

Bookmarks

· Weeknotes

0 replies, 0 reposts, 2 likes