Beyond the hype, how can AI really help councils day-to-day? For example, how can users write more effective prompts, or incorporate new tools in existing processes? For our latest blog, and the first in a series, our Graduate Consultant and AI expert, Nima Karshenas gives a practical guide to some of the ways local government can use these new systems to improve efficiency, streamline tasks and turn them into helpful tools.
AI-generated image showing a woman at a desktop computer wearing headphones
With the unprecedented growth of AI (artificial intelligence) and LLM (large language model) industries, unlocked by advances in hardware, we are increasingly seeing these tools incorporated into different aspects of our lives. The buzz around the potential of AI and the amount of information out there can be overwhelming – particularly when new services suggest it can be essentially ‘thrown’ at tasks to instantly transform and improve existing processes. However, as with all previous transformative technologies, it takes time, due diligence and a fundamental understanding of the technology to see beyond the haze of extreme narratives.
Here at DG Cities, we will be taking a closer look in the coming months at how AI can be used in specific contexts to transform processes in local authorities, ensuring local government can deliver its services more robustly and efficiently in the face of immense resource pressures. But we want it to be as useful as possible. So before we dive into future plans, this blog will explore how LLMs such as ChatGPT and Claude 3.5 can be immediately incorporated into individuals’ workflows, with a particular focus on tasks undertaken at the council.
How do LLMs work and where should we trust them?
At their core, and especially in earlier iterations like GPT 3.5, which many people still use after exceeding their GPT-4o daily limit, LLMs are master imitators. They regurgitate text in the same manner as the data on which they have been trained. These datasets comprise a vast portion of the internet, which LLMs first learn to imitate and are then fine-tuned to answer both general and task-specific questions.
Despite their immense power, LLMs are not reasoners. They do not follow a strict model or process to ensure the accuracy of information. This means that they can suffer from what is referred to in the industry as ‘hallucinations’, where quotes and figures are made up, and have no basis in any kind of official source. Furthermore, being frozen to the data on which they are trained means they often do not use the most up-to-date information.
Although LLMs are beginning to incorporate more sophisticated ‘helper’ models that utilise external tools, such as documents, web browsing, and calculators to support LLM responses, reduce hallucinatory behaviour, and provide the latest information, accuracy is still not guaranteed. Therefore, human oversight is essential, and LLMs should be seen as the starting point, not the final step, of any task they are used for.
Where are LLMs most useful for councils?
Considering all this, which specific tasks do we see current LLMs having the most positive impact on everyday work?
Research: LLM web search tools such as Perplexity.ai serve as an excellent starting point for research. They scan the internet and provide sources for statements, but still sometimes suffer from hallucinations. To truly understand a topic, use the information as a starting point and delve deeper by conducting your own searches or asking Perplexity more specific questions – and always remember to verify information that you have drawn.
Translation: DeepL is a state-of-the-art translation tool that can be particularly useful for understanding research in other languages, corresponding with foreign stakeholders, or reducing the effort required by council employees for whom English is a second language.
Text summarisation: This feature helps you understand long documents by extracting key information. However, hallucinations still apply. If you are not specific about the information you want to extract, key segments might be missed, and the likelihood of hallucinations increases. Use it to get a high-level understanding of a document, not a final comprehension. If the understanding is critical, further checks will still be necessary.
Create slide decks from reports: LLMs can help set up a skeleton of a slide deck by taking information from a document. This saves significant time in the initial stages of collecting key information and structuring it.
Writing Support: LLMs can help rewrite text to match a certain tone or target audience, useful for emails, blogs, and reports. Let LLMs be the cure for writer’s block, but do not rely on them for the final version. If everyone used LLMs to write, text would lose personality and become monotonous, and would not provide or at least imply the absence of a bespoke response that service users expect. Keep things interesting by incorporating your personal touch!
Data Analysis: With the sheer volume, complexity and variety of data that flows through a council, it has been previously arduous to find the answers to questions we have of our own data. With LLMs, data analysis becomes more intuitive and accessible. By simply asking questions in plain English, users can quickly retrieve key insights from their datasets in an instant. These insights can then be visualised through graphs and charts, helping provide our decision makers with a complete, evidence-informed picture of their area of interest.
There are also benefits more specific to a local authority’s IT and Data Analysis departments:
Unlocking the power of programming: LLMs write and run code based on natural language prompts, making it easier to develop and test applications that can automate and streamline council processes. Through description of a task, LLMs such as ChatGPT can instantly provide code to tackle the problem. This allows users, even with limited programming knowledge, to develop and test applications rapidly, and to eliminate repetitive and routine coding tasks, freeing up time for true innovation.
Data Cleaning & Preparation: LLMs, including ChatGPT, can automate data cleaning tasks, such as joining datasets, fixing incomplete data, removing duplicates, and adding new information. This ensures the data used in analysis is accurate and reliable. By improving data quality, we open up the possibility of more sophisticated tools to come in and make use of that data in the future, easing the transition to smarter council communication and decision-making.
Guide: how to write better prompts
Hold your LLM’s hands! The key to maintaining robustness and accuracy in LLMs is maximising the guidance provided in prompts. Chain of Thought is an effective technique for maximising guidance, and ensuring you are getting the quality of results you are looking for.
Step 1: Define the Objective Clearly state what you want the LLM to achieve. This sets the context for the prompt and helps the model understand your end goal.
Step 2: Break Down the Task Break the task into smaller, more manageable steps. This approach makes sure the model can follow the logical progression of the task.
Step 3: Provide Context and Examples: Give the LLM context and examples relevant to the task. This helps the model understand how to handle each step.
Remember to use clear and specific language. Craft prompts using precise language to minimise ambiguity.
Tips for teams using LLMs
Create a repository for prompts for commonly-used tasks: Having a well-organised repository of effective prompts for different tasks can save time and ensure consistency in the output. Prompts are hard to get right off the bat, so having a collaborative repository enables teams to iterate, and improve their prompts to suit their desired team outputs over time.
And finally…
While LLMs offer incredible capabilities that can transform how local authorities operate, it is crucial to use them wisely. They should be seen as powerful tools that augment our efforts, not replace them. With careful application, ethical use and human oversight, LLMs can significantly enhance efficiency and effectiveness in our daily operations.
Local authorities need guidelines, which they may have introduced – people need to proceed with caution when they are using LLM's to draft letters or other forms of communication with, as it isn't infallible and communications need to go through a filter of sorts to ensure managers are comfortable with the language and responses being made. Follow us for more practical and forward-looking guidance on the specific applications, risks and opportunities for AI in local government over the next few months.
Nima Karshenas is a Graduate Consultant at DG Cities working across a range of technology-led projects. He has a Masters in Electrical and Electronic Engineering from Imperial College London, with a particular research interest and in-depth technical knowledge of AI, Statistics, and Signal Processing in projects that promote social impact and sustainability.