Building trust in algorithms is essential. Not (just) because regulators want it, but because it is good for customers and business. The good news is that with the right approach and tooling, it is also achievable.
My first brush with the impact of opaque algorithms came early in my data career. Years ago, I found myself in front of a major banking regulator, explaining why a heavily-used algorithm had been misfiring spectacularly. I have two abiding memories from that experience: the sinking feeling when we realised we had missed something so basic (simple, rule-based logic!), and the fear of “what else might we be missing elsewhere”.
Since then, the importance of getting data and algorithms right has received much more recognition, at least in regulated industries such as banking, insurance, pharmaceuticals and healthcare. But the bar is also much higher now: more data, from more sources, in more formats, feeding more algorithms, with higher stakes. With increased use of Artificial Intelligence/ Machine Learning (“AI/ML”), today’s algorithms are also more powerful and difficult to understand.
A false dichotomy
At this point in the conversation, I get one of two reactions. One is of distrust in AI/ML and a belief that it should have little role to play in regulated industries. Another is of nonchalance; after all, most of us feel comfortable using ‘black-boxes’ (e.g., airplanes, smartphones) in our daily lives without being able to explain how they work. Why hold AI/ML to special standards?
Both make valid points. But the skeptics miss out on the very real opportunity cost of not using AI/ML – whether it is living with historical biases in human decision-making or simply not being able to do things that are too complex for a human to do, at scale. For example, the use of alternative data and AI/ML has helped bring financial services to many who have never had access before.
On the other hand, cheerleaders for unfettered use of AI/ML might be overlooking the fact that a human being (often with limited understanding of AI/ML) is always accountable for and/ or impacted by the algorithm. And fairly or otherwise, AI/ML models do elicit concerns around their opacity – among regulators, senior managers, customers and broader society. In many situations, ensuring that the human can understand the basis of algorithmic decisions is a necessity, not a luxury. My own experience with the regulator all those years ago would have been far more uncomfortable if we had not been able to quickly work out why the algorithm was not working as designed.
A way forward
Reconciling these seemingly conflicting requirements is possible. But it requires serious commitment from business and data/ analytics leaders – not (just) because regulators demand it, but because it is good for their customers and their business, and the only way to start capturing the full value from AI/ML.
1. ‘Heart’, not just ‘Head’: It is relatively easy to get people excited about experimenting with AI/ML. But when it comes to actually trusting the model to make decisions for us, we humans are likely to put up our defences. Convincing a loan approver, insurance under-writer, medical doctor or front-line sales-person to trust an AI/ML model – over their own knowledge or intuition – is as much about the ‘heart’ as the ‘head’. Helping them understand, on their own terms, how the alternative is at least as good as their current way of doing things, is crucial.
2. A Broad Church: Even in industries/ organisations that recognise the importance of governing AI/ML, there is a tendency to define it narrowly. For example, in Financial Services, one might argue that “an ML model is just another model” and expect existing Model Risk teams to deal with any incremental risks from AI/ML.
There are two issues with this approach:
- AI/ML models tend to require greater focus on model quality (e.g., with respect to stability, overfitting and unjust bias) than their traditional alternatives. The pace at which such models are expected to be introduced and re-calibrated is also much higher, stretching traditional model risk management approaches.
- There are second order implications of poorly designed AI/ML models. While not unique to AI/ML, these risks become accentuated due to model complexity, greater dependence on (high-volume, often non-traditional) data and ubiquitous adoption. One example is poor customer experience (e.g., badly communicated decisions) and unfair treatment (e.g., unfair denial of service, discrimination, misselling, inappropriate investment recommendations). Another is around the stability, integrity and competitiveness of financial markets (e.g., unintended collusion with other market players). Obligations under data privacy, sovereignty and security requirements could also become more challenging.
The only way to respond holistically is to bring together a broad coalition – of data managers and scientists, technologists, specialists from risk, compliance, operations and cyber-security, and business leaders.
3. Automate, Automate, Automate: A key driver for the adoption and effectiveness of AI/ ML is scalability. The techniques used to manage traditional models are often inadequate in the face of more data-hungry, widely used and rapidly refreshed AI/ML models. Whether it is during the development and testing phase, formal assessment/ validation or ongoing post-production monitoring, it is impossible to govern AI/ML at scale using manual processes alone.
So, somewhat counter-intuitively, we need more automation if we are to build and sustain trust in AI/ML. As the humans accountable for the outcomes of AI/ ML models, we can only be ‘in charge’ if we have the tools to provide us reliable intelligence on them – before and after they go into production.
I have heard people say “AI is too important to be left to the experts”. Perhaps. But I am yet to come across an AI/ML practitioner who is not keenly aware of the importance of making their models reliable and safe. What I have noticed is that they often lack suitable tools – to support them in analysing and monitoring models, and to enable conversations to build trust with stakeholders.
After spending years contributing to policy conversations world-wide on Responsible AI/ML and framing the institutional response within a global bank, I was keen to find an opportunity to create a sustainable capability in this space. TruEra’s Model Intelligence platform is designed with exactly that mission in mind. With reliable AI/ML explainability at its core, the platform can provide meaningful insights on models throughout their lifecycle. It is already beginning to help data scientists at multiple firms assess/improve model performance and stability, and detect potential issues around unjust bias. And with the right context, these insights can be highly relevant to other key stakeholders (business owners, validation teams, customers, operators, regulators).
While technology in this space is still evolving, TruEra has begun distinguishing itself in several ways. It operates independently of the ML platform used to build the models, and can therefore be used with both in-house and third-party models. The core explainability capability is unique, and underpinned by six years of research at Carnegie Mellon university. The team has extensive hands-on AI/ML experience, and is deeply committed to the mission of increasing trust in algorithms. I feel privileged to be able to join them on this important journey.