CSW asked a cross-section of officials to play around with ChatGPT and establish how good it is at some of the everyday tasks they perform in their jobs. Here, Henry from the Department for Business and Trade sees how well it can craft a ministerial submission.
As a former employee at the Office for AI, I have a keen interest in artificial intelligence, and was interested in the opportunity CSW offered to explore the potential of ChatGPT to undertake some civil service tasks.
I began by experimenting with the ministerial submission format. Having checked whether it knew what a ministerial submission looked like, and then tweaked its understanding, I asked it to write a submission about a fictional policy area (interestingly, it chose digital skills, something I’m really interested in).
It was perhaps unsurprising that three skills-related reflections emerged for me. First, effectively prompting LLMs is a new skill that needs rapid development. Understanding how to prompt them effectively to get the output that we want will be a new skill that we will all have to learn.
Second, the skill needed to work with these models’ changes from “write first, then read” to “read first, then improve”. LLMs quickly produce a partially complete output, but human oversight is still required to check for errors (such as the outdated government department, and bold confidence in naming an individual in the example to the left) and refine the output.
“How would a minister feel about ChatGPT writing a letter in their style? Or a constituent about receiving one? How to handle ethical issues is not a trivial challenge”
Third, we need sufficient understanding to know when, and when not, to use these tools: how would a minister feel, for example, about ChatGPT writing a letter in their style? Or a constituent about receiving one? How to handle ethical, privacy and data-security issues is not a trivial challenge. I was certainly not comfortable with the idea of putting any “Official” level information into public LLMs.
Doing some more research around the topic, you can see how AI “co-pilots” could be powerful tools for civil service tasks, provided they operate within secure IT systems. Potential applications that spring to mind include sorting consultation responses, creating PowerPoint presentations based on policy papers, and adapting writing styles to suit specific ministers. This could free up time, allowing civil servants to prioritise engaging with the communities their work impacts upon, deepening their knowledge of their policy areas.
While I am optimistic about the productivity benefits that these type of LLMs will bring, we must make sure we have the skills and the appreciation of the ethical, privacy and security risks before fully integrating them into our work.
The opportunities are vast, but in my view we must ensure that AI is used responsibly and securely within the civil service to truly harness its potential. If you want to read more on some of these topics, the Department for Science, Innovation and Technology recently published its white paper: A Pro-Innovation Approach to AI Regulation.
Much of the above was written by ChatGPT. I jotted down some of my initial reflections on the minsub which went to nearly 600 words, then asked ChatGPT to summarise it all in 300 words in the style of CSW. I then just made a few tweaks at the end… I’ll let you judge how well it picked up the style!
💬 Make up an imaginary policy idea, and write a ministerial submission for approval to implement your policy idea. The submission must be no more than two pages (annexes are not required)
💬 Title: Implementation of a Nationwide Digital Skills Training Programme
Submission reference number: DSTP-2023/03-28
Date: 28th March 2023
To: The Rt Hon Jane Smith MP, Minister for Digital, Culture, Media, and Sport
From: John Doe, Director of Digital Skills and Inclusion
Copy to: Permanent Secretary, Director General of Digital and Media
Timing: Urgent – Response required by 10th April 2023
Summary: This submission seeks approval for the implementation of a nationwide Digital Skills Training Programme (DST) to improve digital literacy and bridge the digital divide in the UK.
The programme aims to target vulnerable and underserved communities, such as low-income households, the elderly, and people with disabilities. The estimated budget for the programme is £150 million over three years, with an expected outcome of training 1 million people in basic digital skills. We recommend piloting the programme in selected areas, followed by a phased nationwide rollout.
Background: The UK’s digital economy is rapidly growing, with an increasing number of services and opportunities moving online. However, a significant portion of the population lacks basic digital skills, leading to a digital ...