Harnessing Innovative Language Models in an AI-Generated Internet Era  

Artificial intelligence (AI) is coming to pension engagement tools. But what will it look like, how will it shape outcomes and what challenges may it face? 

In London on 2 October, Professor Michael Wooldridge, Professor of AI at the University of Oxford, and Director for AI at the Alan Turing Institute, London, discussed the evolution of AI and machine learning from inception through to the power of large language models (LLMs) today.

This story begins in the aftermath of World War II, when the advent of the digital computer sparked the idea of an ‘electronic brain’, Wooldridge said.  Machine learning has been the major AI development this century, with predictions that, in the next decade, most of the content on the internet will be AI-generated, and increasingly indistinguishable from its human-made counterparts. 

Large tech companies have thrown resources into building enormous AI neural networks with vast amounts of training data. We’ve seen a rise in ‘big data’ – sets of data so complex and vast they cannot be processed by traditional data management systems, propelled further by social media and the internet. Consequently. there is evermore data to prompt, train and test machine learning technologies, fuelling stronger algorithms and patterns. We see the results in everyday actions such as predictive text messages.

It isn’t all smooth progress for AI. Wooldridge says that LLMs “have no concept of truth”, and so may fill in gaps by making assumptions based on bias or even lies within the data the model has tracked. While guard rails exist to prevent toxic data being used, these protections are not perfect. 

Nonetheless, LLMs present a wealth of opportunities. These include productivity, with unstructured data such as emails and meeting minutes being analysed and helping to reduce the burden of mundane administrative tasks, or “firing the creative process” by asking AI for ideas and discovering yet unknown categories of content. Wooldridge surmised: “We will find them weird, but our kids will take them for granted.”

Emerging capabilities may be able to teach AI a form of logic, while multi-modal AI can process and integrate multiple types of data, such as audio and video, simultaneously.

However, in pensions, technology must be in the service of what members do and need. Mercer’s Tom Higham and Stephen Coates explained that to create the LLM behind Ask My Pension, its AI-powered pension assistant, Mercer needed expert questions, insightful answers and many hours of testing and iteration.

Mercer’s main target when developing an AI model was helping the model mimic a human interaction seamlessly.

Higham and Coates demonstrated the Ask My Pension AI prototype, which they had integrated into WhatsApp, and provided audience members with the opportunity to use the AI with their own phones. Coates explained that Mercer’s prototype is the first of its type. Ultimately it will be integrated into other social media channels depending on customers’ preferences.

Coates explained: "The latest AI technology is giving us the chance to revert to what humans love to do: converse." We can engage in two-way, personal conversations with every saver, on their terms, in their own time, through the channels they like to use."

One current benefit of AI is helping to identify vulnerable customers via call handling, and thus “nudging members in action.” Aimée Denham, Trustee Director at Vidett said: “AI is good at noticing vulnerable customers and responding quickly to short-term vulnerabilities, such as following a bereavement. It harnesses dynamic information in a quick and efficient way.”

The more data we can feed to AI, the more developers can innovate and build more tailored communications, such as cost-benefit analyses. Andy Parker, Partner at Barnett Waddingham, added that in the long term, regulation and compliance may need to evolve further to deal with the likelihood that LLMs will need increasingly more data.

According to Felicia Meyerowitz Singh, co-founder of Engage Smarter, in time customers and members will use AI in much the same way as we have become familiar with using the internet.

However, challenges remain, not least “the issue of trust.” Wooldridge highlighted humans’ desire to hold others accountable, saying: “People feel more frustrated about machine error. For instance, they may not be happy with the idea of a machine deciding whether or not you get a mortgage.” Meyerowitz Singh added that the financial services sector is held to a higher standard than AI developers, and so transparency is critical to make the benefits of these technologies clear.

Effective, personalised AI language models can transform the way people interact with and plan for their future – a crucial factor in the pension landscape. By turning data into insight and information into education, AI switches the dialogue around from reactive to proactive, leading to better real-time engagement with members.

According to Wooldridge, AI will succeed by taking on relatively mundane or cumbersome tasks, allowing humans to focus on more complex or strategic needs. In the world of pensions, this means empowering people to learn, engage and make better decisions.