Article 27: Algorithmic System Integrity: Explainability (Part 4)

Spoken by a human version of this article.

TL;DR (TL;DL?)

• Explainability is necessary to build trust in AI systems.
• There is no universally accepted definition of explainability.
• So we focus on key considerations that don't require us to select any particular definition.

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

Yusuf (00:14):
Okay, we are onto algorithmic system integrity,
explainability, part four.
The too long read version ofthis is explainability is
necessary to build trust in AIsystems.
There is no universally accepteddefinition of explainability, so
we focus on key considerationsthat don't require us to select

(00:34):
any particular definition.
In the previous article previousepisode, we explored the concept
of explainability itsimportance, and four challenges.
We then addressed the first andsecond challenges being
complexity and complicatedprocesses.
Before we dive into the thirdchallenge, that's privacy and
confidentiality.
Let's pause to considerexplainability in practical

(00:57):
terms.
Expandability helps build trustin AI systems, but there is no
single definition.
This brief article is astraightforward discussion
focused on the underlying intentand practical implications, so a
prevailing definition.
There are several definitionsthat have been proposed, some
with much deliberation.

(01:18):
Despite these efforts, there isno universal consensus on a
single definition.
So rather than trying to createyet another definition, we'll
focus instead on key practicalconsiderations.
There's five of them.
And by focusing on these, we canestablish a foundation.
So here are the five keyconsiderations.
First one is be able to explain.

(01:39):
None of this will matter if wecan't explain our systematic
decisions.
That's global or individualdecisions.
Those are local.
This includes using some of thesolutions we've already
described to address complexityand complicated processes.
The next is considering thecontext, so other systems
producing fully automateddecisions, or are there humans

(02:00):
in the loop are systematicoutputs used as inputs to manual
processes, but not directly thefinal decisions.
A note here, if the output isfed into a human process, but
used as the default decision, itcould be considered a systematic
decision in practice.
The third is.
Considering the purpose, so whatis the algorithmic system being

(02:23):
used for?
For example, fraud detectionsystems require internal
explanations, but may limitexternal transparency to avoid
exposing logic to fraudsters.
The purpose influences both thetype and level of explanations
needed.
The fourth is considering theaudience's needs.

(02:43):
So stakeholders have varyingneeds.
Developers need technicaldetails to debug and refine end
user.
Employees need to understand howdecisions are influenced, but
not all the technical specifics.
Auditors and regulators needevidence of compliance and
transparency.
Customers need plain languageexplanations so they can, that

(03:04):
they can act on.
Tailoring the explanationsensures they are relevant and
understandable to each audience.
Then we decide whether what andhow.
So with the core mattersconsidered and an ability to
explain, we ask, under whatcircumstances will we provide
explanations?

(03:26):
Where we decide to explain whatinformation will be, will we
provide, and then how do wecommunicate the explanations
effectively.
So those are the fiveconsiderations.
And next, now that we've laidthis groundwork, we are better
prepared to tackle thecomplexities of privacy and
confidentiality, ensuring thatAI systems are both transparent

(03:46):
and secure.
So the next article, nextepisode, will delve into
privacy.
Thanks for listening.