Introduction

Currently the text boxes in our web applications are black boxes. We can’t see what’s inside them and we only have primitive tools to work with them. We can use the number of characters for validation, check they start with a capital letter

Language models offer the chance to work with the meaning of text. With a better tools for extracting, comparing and manipulating the meaning of text we can make our form creator better and improve the forms our services make for our form fillers.

For example, we can use language models we might be able to:

  • label question text with the type of question it is asking
  • rephrase question text to make it easier to understand
  • summarise long text
  • extract the structured data out of the text
  • translate text to other languages

The problems with language models

LLMs are not perfect. There are strong reasons to be sceptical of them, they are:

  • prone to errors such as hallucinations, where they invent text without any input
  • present security issues like prompt injection
  • prone to bias, where they are trained on data that is not representative of the real world
  • computationally expensive to train and run
  • incredibly complicated and hard to reason about. They are hard to test and represent a large attack surface for abuse
  • a relatively new technology and the world is still learning how to use them
  • at a point in hype cycle where they are massively overhyped - it’s easy to get carried away with the possibilities and many people do

I think it’s still worth exploring the possibilities of LLMs:

  • they offer new ways to interact with users
  • given they hype around them, even if we don’t want to use them, we should be aware of them and their limitations so that we can clearly explain the problems they create
  • they are a new technology and we should be exploring them so that we can explain them to our colleagues and users
  • they are already in widespread use. They will be in Office, the Google suite and are already in search engines. We should look to find patterns and techniques that make them accessible and usable for all of our users
  • they are powerful tools and civil servants should have access to them