Our next event!
🚨 6PM, WED 29th April 🚨
At the MoJ or online.
Lisa Trickey and Rebekah Wilson, LGA : LG2040: The State of Things to Come
Jonathan Slater (ex DfE Perm Sec), Visiting Professor at King’s College London and QMUL
luma.com/7nu2j853
Please share! Thanks 🙏
For me, it was watching the character arc of Woody Allen's character in Hannah and Her Sisters, as he goes through fearing imminent death, to eventual death, trying on religions and eventually watching the Marx Brothers. Put simply: we get to have fun.
en.wikipedia.org/wiki/Hannah_...
Sunday question: Have you stopped having existential crises and if so, how did you resolve them?
Asking because my youngest is really suffering, finding it hard to believe they and their life matters at all when the universe is going to end one day.
hang on
For ‘AI’ to reliably function as advertised, it would need such a huge amount of human effort in data management, quality assurance and change control that it could never be cost-effective.
The true cost of fit-for-purpose data is starting to emerge, and it is way higher than anyone anticipated.
I spoke to a local women’s Institute yesterday about the history of witches. I was in two minds about whether to go ahead given the behaviour of the organisation, but in fact, it turned out that many individuals wanted to tell me that they disapproved strongly of the stance taken by leadership.
Photos of Jeni, Adam, Tim and Emily above the words Connected by Data. The photos are licensed to Paul Clarke by CC-BY.
In this blog @jenitennison.com shares why we're winding up in 2027.
"When I set up Connected by Data I also stated my intention that this organisation would have a maximum lifespan of 3-5 years. We are now at the end of our fourth year..."
Read the full blog at: connectedbydata.org/blog/2026/03...
🧑🎓 Join us to explore the methods and findings from the Generative AI in Education: Have your Say distributed dialogue.
🔗 Register here: connectedbydata.org/...
🗓️ Wednesday 22 April
🕰️ 15:00 - 16:00
💻 Zoom
(1/2)
Ouch (on many fronts)! Hope you have a good birthday though :)
What’s changing on data.gov.uk and why
dataingovernment.blog.gov.uk/2026/03/25/w...
#opendata
In an AI age it will be increasingly important to audit the process and outputs from the tools. So this is a timely toolkit from @nao.org.uk and others with a super comprehensive spreadsheet which will be of interest to anyone overseeing / auditing AI usage
www.nao.org.uk/insights/ai-...
Trans, non-binary and gender-diverse adults feel there are big gaps in what we know about transition-related healthcare. Many healthcare professionals feel this too.
Complete the national consultation to help us breach the gap 👇
https://wp.lancs.ac.uk/transadultspsp/
Ooh, where's that puzzle from? Looks like a great challenge!
I love this! Public sector AI innovation to address a clear problem, prioritising learning, listening to users and respecting professional standards. Great to see this written up so clearly too.
Long list of the session recordings available at the URL in the post.
Over half way through uploading recordings from @pairs.site 2026: critical presentations on participatory approaches to AI development, governance and resistance. www.pairs.site/Session-reco...
Most talks are < 15 minutes, and packed with ideas & insights. Papers & slides linked where available.
Maybe they were envisioning a sovereign ha-ha, but really couldn't use that term!
en.wikipedia.org/wiki/Ha-ha
French streaming service Deezer said streams of AI-generated music on its platform were dominated by fraudsters, who upload and then repeatedly listen to thousands of songs to generate royalty payments to the detriment of legitimate artists.
AI drafts your job application and reads your job application. AI writes a report and then the reader has AI summarise it.
The AI on both sides of the transaction just got real with this neat fraud. You get AI to make garbage music, get AI to listen to it, and get paid!
www.ft.com/content/8a47...
Oh I thought I recognised your face, sorry for not saying hi!
Our next event!
🚨 6PM, WED 25 March 🚨
At the MoJ or online.
Lara Groves and Oliver Bruff, Ada Lovelace Institute, on AI transcription tools in social care
Owen Prendeville, on the ICO’s campaign to drive improvements for access to care records
luma.com/8i84tnwy
Please share! Thanks 🙏
It's still early access, so you can easily wait until they've actually balanced the gameplay and finished the art!
Yes, or rely *only* on synthetic data without awareness of shortcomings and supplementing with other testing approaches. I think it's probably better than nothing, but there are also other options than nothing!
New characters, new cards, new relics, new events, new enemies, same completely addictive game play, now with friends!!
Guys! Guys! You have to add Slay the Spire II to the list! It's not couch coop but we're successfully playing together in the same room on separate PCs/Macs and really enjoying it.
... and that I don't think are prevalent on the web. So there I think we really lack a ground truth on which to train synthetic data generators, and need to look at other ways to validate the synthetic data, and other ways to test/benchmark genAI (like community red-teaming).
(Which makes me think that raw data of actual queries to eg Citizens Advice and other support organisations is really valuable if only for training the generator of the synthetic data.)
The other example I saw involved interactions that weren't currently happening, or aren't recorded...
That's why the ODI example used Reddit as their source, but I'd still question how well Reddit Q&A reflects either likely queries to public services or the kinds of responses you'd want from public services!
I can totally see the reasons for this approach, but there's obviously a risk that the synthetic data itself contains a whole load of biases that the system is then trained to replicate. These are more acute if the synthetic data is generated from a source that isn't close to real interactions...
This is the second example I've seen in the last couple of weeks where LLM performance on a public-service-related task has been tested using benchmarks measuring performance against a synthetic dataset (ie one where the prompt/interaction has itself been generated by an LLM).
