Article 28. Algorithmic System Integrity: Explainability (Part 5) - Privacy and Confidentiality

Spoken by a human version of this article.

TL;DR (TL;DL?)

• Algorithmic systems create challenges in balancing explainability with privacy and confidentiality.
• Key challenges include protecting sensitive information, preserving proprietary algorithms, and securing fraud detection systems.
• Focusing on what audiences need, with a few specific considerations, can help address these.

To subscribe to the weekly articles: https://riskinsights.com.au/blog#subscribe

About this podcast

A podcast for Financial Services leaders, where we discuss fairness and accuracy in the use of data, algorithms, and AI.

Hosted by Yusuf Moolla.
Produced by Risk Insights (riskinsights.com.au).

Episode Transcript

(00:13):
This is part five of explainability,so algorithmic system integrity, colon
explainability, part five, and thisone's about privacy and confidentiality.
The TL;DR algorithmic systems createchallenges in balancing explainability
with privacy and confidentiality.
The key challenges include protectingsensitive information, preserving

(00:35):
proprietary algorithms, andsecuring fraud detection systems.
Focusing on what audiences needwith a few specific considerations
can help address these.
So in a previous article,previous episode, we explored
the concept of explainability itsimportance, and four challenges.
We then addressed the first andsecond challenges being complexity and

(00:57):
complicated processes, and we consideredexplainability in practical terms.
In this episode, we discussthe third challenge.
Which is privacy and confidentiality.
So to recap that challenge, algorithmicsystems create unique challenges
when balancing explainabilitywith privacy and confidentiality.
Among these are three items.

(01:19):
The first is protectingsensitive customer information.
So providing detailed explanationsof AI decisions can risk exposing
sensitive customer information,for instance, in banking, revealing
how a credit score is calculated.
Could inadvertently disclosepersonal financial data.
The second is preservingproprietary algorithms.

(01:42):
detailed explanations canreveal proprietary algorithms
enabling competitors to shortcutthe development process.
Perhaps more importantly, it canenable customers or prospective
customers to try to game the system.
The third is securing fraud detectionsystems, and there revealing how potential

(02:03):
fraud is identified can be risky.
Bad actors can use this knowledgeto manipulate the system.
Fraud detection systems typicallyrely on identifying specific patterns,
if frauds does know what we'relooking for and how they could alter
their behavior or the informationthey provide to evade detection.

(02:24):
One way to identify solutions to thesechallenges is to tailor our explanations
based on who we are explaining to.
Customers or prospective customers.
Here we want to enhance trustand satisfaction and meet
compliance expectations.
We use clear, concise languagewith explanations that
can be easily understood.
We avoid technical jargon.

(02:46):
Our default expectation is that customerswill do the right thing, but we don't want
to enable them to game the system either.
So there's a few things we can do here,and that includes limiting explanations,
especially for fraud systems, explainingthe decision, but not necessarily
the detailed sequence of steps.

(03:06):
Keeping records of data providedpreviously so that we can identify
changes that don't make sense.
For frontline staff, we wantto enable frontline staff
to address customer queries.
They may get the same infothat we give to customers.
If we provide more information tofrontline staff, we need to be clear

(03:28):
about how to use and communicate it.
We also need to prevent leakage and misusewith developers.
So developers are interested in systemimprovement and troubleshooting,
so they need the details.
But we restrict this to authorizedpersonnel only with the usual mechanisms
to prevent leakage and misuse.

(03:50):
Ideally, we keep models assimple as possible, and we
don't overcomplicate processes.
For senior management, they need toensure that systems meet customer
expectations, support businessobjectives, and enable compliance.
Often this means high level explanationsthat focus transparently, but simply

(04:13):
on enabling leaders to understand.
This can sometimes meangetting into the details.
Then regulators and auditors, weneed to demonstrate compliance with
laws and regulations with decisionsthat are transparent and justifiable.
We provide explanations thatdemonstrate compliance with
regulations using unambiguous language.

(04:36):
The details will vary depending onspecific needs, but this group may
also need to dive into the details.
The next article will delveinto the fourth challenge.
Making sure that theexplanations can be understood.