Part 4 - The Pension Regulator’s AI plan - the rock upon which AI stands

Welcome to part 4 of our exploration of The Pensions Regulators (TPR) AI Plan. This time, we’re thinking about data. At the end of the my previous Part 3 blog, I noted that data is the foundation that everything else rests on. TPR’s plan says that “AI technologies function by learning from large, high-quality datasets”.  

The important piece to note here is “high-quality”. 

Why is this important? 

Almost all AI models will only ever be as good as the data and processes they are trained on. Many of you will have had experience of automated systems suffering from the occasional case of “Rubbish in = Rubbish out”, and it’s no different for AI. 

Anyone who’s worked in the LGPS is familiar with the nature of the broad pensions data challenge. Decades of records… Dozens or hundreds of employers… each with their individual scheme rules and payroll quirks. Historic aggregations, ageing systems, and more recently the significant data demands of the McCloud judgement. 

TPR’s plan builds on the work that the LGPS is already doing around data readiness, dashboards and agreeing common and open standards. Funds are clearly expected to have a data strategy, resource it appropriately and, as with GDPR, ensure their downstream suppliers meet similarly high standards. 

Pension Dashboards have also raised awareness of the importance of more standardised data. Nonetheless, AI raises the stakes even further, because the effect of poor-quality data, when processed by an AI model, is much harder to detect than with a more transparent deterministic process such as dashboards.  

AI behaves differently

AI is - as technical specialists would put it - “non-deterministic”, meaning that the same input will not always produce exactly the same output. This variability can make it harder to spot a gradual drift in results, including patterns that may disadvantage particular individuals or groups. 

This “model bias” is likely to become a much bigger issue as AI use becomes widespread. It’s not something we have traditionally had to think about in quite the same way as conventional software, because those systems are usually designed to be right or wrong in a fairly black and white sense. If a traditional software system receives invalid data, it will usually reject it, display an error message, or in the worst case, crash entirely. 

Bias in AI is rarely (if ever) the result of some sort of explicit malicious action. Far more often, it’s a result of subtleties in the material used to train the AI models. Assumptions evident in historic data, that might have seemed perfectly acceptable a couple of decades ago, might be (sometimes shockingly) outdated now. Prejudices that are learned faithfully from this imperfect historical data can end up being applied today without challenge, in part at least because that information has been invisibly encoded into a huge matrix of numbers, rather than human-readable rules. 

What could possibly go wrong!? 

I think that TPR’s AI plan is honest with its warning that AI could widen existing inequalities, and it’s promised to look at outcomes for people with protected characteristics. AI means that matters now more than ever. An example is the case against Workday in the USA. 

Lead plaintiff Derek Mobley alleged that Workday's AI recommendation and CV screening tools systematically filtered out qualified applicants who were black, disabled, or over the age of 40. Many candidates reported receiving rejections within minutes at odd hours, indicating zero human review. As a result the judge ruled that AI software vendors can be held legally liable as "agents" of employers under anti-discrimination laws if their software performs traditional screening tasks. 

For the LGPS, this might mean looking closely at things like part-time service records, which are inconsistent in a lot of older data and also disproportionately affect women. If you feed that data into an AI tool you don’t necessarily get an obvious “hard” error. The risk is that results are unlikely to be mathematically “wrong” but potentially reflect a more subtle and systemic unfairness.  

Data quality isn’t only about accuracy. It’s also about fairness. 

All that said, you’ve probably already done much more of the groundwork than you think. The data scores you’ve produced for the dashboard, your common data and your scheme-specific data, are as good a measure of “AI-readiness” as any, so you might want to start with benchmarking and improving those. 

A sensible place to begin is with the common data: names, addresses, National Insurance numbers and key dates, because those are the data almost every tool already relies on. Then consider the more common things you already know can be problematic such as GMPE, additional contributions and scheme-specific “quirks”.  

When you’re checking your data, pay closest attention to the records most likely to carry bias in their history. Part-time and term-time service. Breaks in service. Transfers in from years ago. Those are often where biases might hide, and it’s those areas where a carelessly developed tool left to operate without human oversight will likely do the most damage.  

And this is not just within fund or employer data. 

With so much data in circulation, there’s an industry-wide angle to this too. TPR is currently exploring whether it can open up its own datasets, and is pushing for open standards. This suggests that data quality is not just an issue for individual organisations but is, at least partially, an industry-wide problem that may be helped by industry-wide collaboration. 

Looking optimistically at this, improving confidence in our data gives us a platform on which the industry can confidently innovate. 

If we as an industry can collectively improve our data, understand where bias may be hiding and keep people firmly in the loop, the whole conversation changes. It stops being about what might go wrong and starts being about what we might create. 

And, as the management consultant Peter Drucker famously said, "the best way to predict the future is to create it.” 

To discuss any of the information in more detail, please get in touch, we'd love to hear from you.

 

This blog is based upon our understanding of events as at the date of publication. It is a general summary of topical matters and should not be regarded as financial advice. It should not be considered a substitute for professional advice on specific circumstances and objectives. Where this blog refers to legal matters please note that Hymans Robertson LLP is not qualified to provide legal opinion and therefore you may wish to obtain independent legal advice to consider any relevant law and/or regulation. Please read our Terms of Use - Hymans Robertson.