Insights: The Milton Keynes AI Festival

How to build trustworthy and reliable AI

This year’s Milton Keynes Artificial Intelligence Festival took place in November. This week long event brought together tech experts and businesses to showcase cutting-edge technologies that will benefit the city in coming years. The festival included a series of events running across the city, hosted by a growing number of major organisations, including Bletchley Park, The Open University, Connected Places Catapult and HMGCC.

Siobhan King, Consultant here at Metataxis, attended the How to build trust in AI event hosted by PwC, as part of the AI Festival, and shares her recent learnings:

It was great to be part of the recent Milton Keynes AI Festival. I attended a ‘show and tell’ session delivered by PwC as part of the Festival. There were four speakers, each covering a different aspect of ways to build trustworthy AI. More than once, things took a turn for the recursive in a “if you type Google into Google” way.  PwC’s technical implementation and thinking about how to responsibly use AI is quite far along in terms of development, meaning that they are now at a stage where they are using AI to manage AI. As I walk through each presentation, you’ll soon see how:

Felicity Copeland, Director at PwC, talked through many of the challenges and opportunities associated with auditing AI. For AI to be trustworthy, there needs to be some accountability for the business decisions made, to manage the associated risks and to ensure the tools operate in a fair, unbiased and ethical manner. Being able to audit AI is essential to building that trust. The part of this discussion that really captured my imagination was when Copeland talked about how at PwC, they had successfully used an LLM to check the accuracy of LLM outputs. (Of course, as soon as I got home, I had to experiment with this myself on ChatGPT with disappointing results. It just goes to show how much better things work when your LLM is grounded.)

Mo Meskarian, another PwC Director, talked about the importance of constantly keeping on top of risk when working with AI. In this fast-moving area, things change quickly, and risks may evolve mid-project. (Case in point: the EU AI Act which Mo gave us much information about and briefly discussed an AI tool PwC has created to help navigate the Act.) Meskarian emphasised the importance of agile risk management as a response to this ever-changing environment. This gave a heavy sense of déjà vu for those of us in the audience familiar with Privacy by Design, and the term GDPR was murmured frequently in discussions over tea and sandwiches afterwards.

Having set the scene by demonstrating how you need to have the right strategies, governance and management in place, Alexandra Lochead, Manager in Risk – AI, Cloud and Data along with colleague Matthew Haughton, Senior Manager – Risk Digital, followed up with live demos of two of PwC’s AI tools. First up, we were shown ChatGPT Enterprise, which is a RAG enabled version of the popular public platform, leveraging PwC sources to answer questions.

It’s hard to comment from a short demo as an outsider on how accurate the tool was, but what was impressive was the consideration that has gone into the human side of the AI equation. This was most evident in the design of the prompt screen which used forms to help users formulate questions, suggested a set of prompts and linked to support as needed. What’s more, what ChatGPT Enterprise can achieve is quite sophisticated. When Lochead lamented that after years of working in Python, she has recently discovered she could use Enterprise to generate a GPT, you could see where things are going in terms of tools teach us how to formulate better prompts.

The second tool demonstrated was a great example of identifying a use case, assessing viability and risk, and developing a solution. Matthew Haughton talked through a GPT developed for the Risk and Compliance team to help users connect with the right information from their very densely populated intranet. The problem for users and practitioners alike was that there was just too much information, and it took time to find the correct answers about, for example, AML compliance. So, the PwC AI team decided to develop an AI solution to make the information more accessible. This was the perfect opportunity to step through the pathways that PwC advise their clients to follow for any AI project.

To prepare, the team took unstructured data from the Risk and Compliance intranet pages and “added some structure” to make this content available to the AI tool. The process of adding structure may seem mystical or even onerous – but it doesn’t have to be. Metataxis has been doing this sort of work for years: creating ontologies; data models; knowledge graphs and information architectures, all designed to help clients wrangle large amounts of data.

Haughton also reminded us that just because there are some risks associated with working with this type of information in this context, it doesn’t mean it is impossible to use AI as a solution. Haughton emphasised having strong guardrails in place, this is essential when the risk of getting things wrong is elevated. But it does not exclude certain use cases from being taken forward.

The top five take aways

1. Having robust processes in place for assessing AI from use case to post implementation is essential
2. To keep up with the pace of change, risk management needs to be a continuous activity – not a ‘one and done’ or annual event
3. Deadlines for complying with EU AI Act requirements are coming thick and fast with the first drop date being 1^st February 2025. (This is the ban of AI systems that pose unacceptable risks)
4. For an AI tool to be more accurate and effective, some form of structuring is needed to arrange unstructured data
5. Getting your AI culture right is critical. PwC were able to demonstrate a healthy combination of savvy, circumspection and curiosity to make theirs a safe workplace to try new things.

The question is: Where the leaders go, will the pack follow? My concern is the size of the gap between the leaders and everybody else. It’s all about perspective. The bar for entry is not too difficult once you’re over it, but if you’re not, it may feel insurmountable.

Need help “getting over the AI bar?” Talk to us.