WEBVTT
1
00:00:03.480 --> 00:00:08.199
Welcome to Transform Now, the podcast
brought to you by Robotic process automation pioneer
2
00:00:08.560 --> 00:00:13.240
S S and C Blue Prison.
Digital transformation has the potential to reshape the
3
00:00:13.240 --> 00:00:18.320
way companies service their customers, engage
their employees, and manage their operations.
4
00:00:18.760 --> 00:00:23.640
Whether you're looking to develop strategies,
tactics, or best practices to positively impact
5
00:00:23.640 --> 00:00:27.839
the future of work, or you're
curious to see how other companies have successfully
6
00:00:27.920 --> 00:00:32.920
navigated their digital transformation programs, than
this podcast is for you. We're here
7
00:00:32.960 --> 00:00:38.079
to help you Transform Now. Hello
everyone, I'm Brad Hairston with S S
8
00:00:38.159 --> 00:00:43.840
and C Blue Prism. Welcome to
the Transform Now podcast. Today, I'm
9
00:00:43.880 --> 00:00:48.479
happy to have two guests on the
show. The first is Chris Johannesson,
10
00:00:48.600 --> 00:00:55.240
an automation and AI expert and the
editor of the Journal of AI, Robotics
11
00:00:55.320 --> 00:01:00.840
and Workplace Automation. My second guest
is Dr lu By Kenenheimer, CTO of
12
00:01:00.880 --> 00:01:06.599
the America's at S S SC Blue
Prison. Both Christ and Lee have been
13
00:01:06.599 --> 00:01:11.280
on the podcast before, and today
I will be talking with them about the
14
00:01:11.319 --> 00:01:19.120
ethics of automation and AI. Gentlemen, welcome. Let's start with some introductions.
15
00:01:19.200 --> 00:01:22.879
Chris, why don't you start us
off. Hello everybody, and hello
16
00:01:22.920 --> 00:01:27.599
again Brad Chris Johannison. I'm,
as Brad mentioned, editor of the Journal
17
00:01:27.640 --> 00:01:34.799
of AI, Robotics and Workplace Automation. That's journal launched about a year ago,
18
00:01:34.040 --> 00:01:40.120
published by Henry Stewart Publications out of
the UK, and we just had
19
00:01:40.200 --> 00:01:46.560
a special edition on ethics and AI. So this topic is top of mind.
20
00:01:46.640 --> 00:01:51.079
There's some other things going on in
the news as well that continue to
21
00:01:51.120 --> 00:01:55.680
bring the ifs to the forefront.
That's just a part time role that I
22
00:01:55.719 --> 00:02:00.159
serve in this journal. I've been
doing management consulting work and you automate shin
23
00:02:00.280 --> 00:02:05.480
and not AI space for the past
several years with firms such as Mackenzie and
24
00:02:05.599 --> 00:02:09.680
see It Partners, and before that
some of the early work in automation and
25
00:02:09.840 --> 00:02:15.240
AI going back to my time with
GE and including time that I spent with
26
00:02:15.599 --> 00:02:22.680
eBay Enterprise as well as leading a
marketing practice on behalf of Omncommon Philadelphia where
27
00:02:22.719 --> 00:02:25.840
we deviled in some a little bit
of AI as part of marketing automation.
28
00:02:27.520 --> 00:02:31.479
Awesome, thank you, lu Let's
Brent said. I'm Dr Lou Bockenhammer,
29
00:02:31.560 --> 00:02:36.439
the CTO of the Americas for Blue
Prism. I've been with Essence and see
30
00:02:36.479 --> 00:02:43.280
Blue Prism for almost four and a
half years now, and it's been an
31
00:02:43.280 --> 00:02:47.400
exciting time for both automation and AI. Before that, I was with IBM
32
00:02:47.479 --> 00:02:52.560
S Watson and Cloud Platform team working
with some of the AI and m L
33
00:02:52.599 --> 00:02:59.439
tools over there, and currently we're
at a I want to say an inflection
34
00:02:59.520 --> 00:03:02.840
point in how machine learning and AI
are being used in your day to day
35
00:03:02.840 --> 00:03:09.800
automations that make this conversation particularly relevant. So thanks for having me. It's
36
00:03:09.800 --> 00:03:13.840
great to have both of you on
the show today. Thank you for being
37
00:03:13.879 --> 00:03:16.960
here. Well. Now that we
are in full swing in the era of
38
00:03:17.000 --> 00:03:23.759
intelligent automation, the topic of ethics
at the intersection of r p A and
39
00:03:23.000 --> 00:03:30.479
AI has definitely taken on more importance. Just this month, the White House
40
00:03:30.560 --> 00:03:36.520
Office of Science and Technology took this
topic center stage by releasing its first ever
41
00:03:37.199 --> 00:03:44.080
AI Bill of Rights, which defines
five principles for making automated systems work for
42
00:03:44.120 --> 00:03:47.960
the American people. We will certainly
talk about that today. If you look
43
00:03:47.960 --> 00:03:53.680
at just where ethics and AI have
emerged in recent history, there have been
44
00:03:53.719 --> 00:04:01.319
several now infamous examples of truly unethical
AI. There's the two thousand fourteen example,
45
00:04:01.360 --> 00:04:10.520
where an AI powered recruiting software ended
up overwhelmingly preferring male candidates. There's
46
00:04:10.520 --> 00:04:16.279
the chatbot created back in which,
after being released on Twitter, started quickly
47
00:04:16.360 --> 00:04:26.519
spewing racist and misogynist messages. And
in six an algorithm called compass was used
48
00:04:26.519 --> 00:04:30.319
in the u S court systems to
predict the likelihood of offenders committing another crime.
49
00:04:30.839 --> 00:04:36.959
A journalist came forward and exposed that
this model predicted twice as many false
50
00:04:38.000 --> 00:04:43.759
positives for minority offenders. So,
of course all of these are AI gone
51
00:04:43.800 --> 00:04:49.000
wrong stories. They do not diminish
the many, many ways that AI has
52
00:04:49.040 --> 00:04:56.120
had a very positive impact on our
society and in enhanced automation for that matter.
53
00:04:56.959 --> 00:04:59.680
So I really want to get into
the dialogue with both of you to
54
00:04:59.759 --> 00:05:03.079
talk through this kind of stuff.
Lou let me start with you. Now
55
00:05:03.120 --> 00:05:11.160
that machine learning models are enabling decisions
in the automation of more complex business processes,
56
00:05:11.560 --> 00:05:16.560
do you see ethics becoming a bigger
concern for automation teams and their executive
57
00:05:16.560 --> 00:05:23.680
sponsors? Oh? Absolutely. If
you think about automation, you really can't
58
00:05:23.720 --> 00:05:30.519
get past the need to be including
machine learning and artificial intelligence because otherwise automation
59
00:05:30.519 --> 00:05:33.920
are just hands on a keyboard with
no brain behind them. But then you
60
00:05:34.040 --> 00:05:39.879
hit all the issues with AI or
machine learning that you brought up before.
61
00:05:40.240 --> 00:05:44.959
I think these fall into two different
categories. The first side you can think
62
00:05:44.959 --> 00:05:49.360
of as issues with the model themselves, how they were built, how they
63
00:05:49.360 --> 00:05:55.000
were trained, what sort of bias
they might have inherently, and the second
64
00:05:55.319 --> 00:06:00.319
category being issues with how the models
are being used, so that could be
65
00:06:00.360 --> 00:06:05.720
a for example, mass usage of
facial recognition technology. Typically when we are
66
00:06:05.800 --> 00:06:13.000
looking at automation use cases, we're
normally looking at the first category, when
67
00:06:13.000 --> 00:06:16.519
we're looking at the in any inherent
bias in the model, and that previously
68
00:06:16.560 --> 00:06:21.319
had been an issue. There was
a situation where some algorithms have been put
69
00:06:21.360 --> 00:06:28.319
into place to help approve mortgage loans
that again turned out to you have a
70
00:06:28.560 --> 00:06:33.120
racial bias most of the time when
you're looking at customers. Are two raised
71
00:06:33.240 --> 00:06:39.000
to handle this, and the first
is to be very careful with what sort
72
00:06:39.160 --> 00:06:43.519
of data you use, what your
corpusive data is that you're used to train
73
00:06:43.839 --> 00:06:47.759
that machine learning model. And the
second is to make sure that you have
74
00:06:48.160 --> 00:06:55.720
auditability at every step. We typically
see We typically used to see machine learning
75
00:06:55.720 --> 00:07:00.120
engines that were more of a black
box. But nowadays, just as you
76
00:07:00.120 --> 00:07:05.600
were asking, customers are much more
concerned with maintaining ethics and as a result,
77
00:07:06.000 --> 00:07:10.959
they find it very important to make
sure that any machine learning or AI
78
00:07:11.079 --> 00:07:15.560
solution they put into place has a
full audit trail to explain how it made
79
00:07:15.639 --> 00:07:19.920
every decision it made. Interesting,
Chris, let me ask you, from
80
00:07:19.959 --> 00:07:27.839
your perspective, which industries do you
see leading the charge in addressing ethics and
81
00:07:27.959 --> 00:07:32.680
automation and AI. There's a bit
of a classical model out there about technology
82
00:07:33.000 --> 00:07:36.199
adoption and something goes back to an
old book. I think it came out
83
00:07:36.199 --> 00:07:42.800
in early nineteen sixties, a book
I believed title was around the diffusions of
84
00:07:42.800 --> 00:07:46.040
innovations as every call top of my
head. So even though we cited a
85
00:07:46.040 --> 00:07:51.720
few technology for examples earlier, certainly
companies like that they're right there on the
86
00:07:51.800 --> 00:07:56.879
vanguard of addressing the ethics and automation
and AI not only for their own you
87
00:07:57.079 --> 00:08:01.920
use, but many of these corporations
are not only using this stuff for their
88
00:08:01.920 --> 00:08:07.800
own purposes, but they're trying to
bring software applications out to the market that
89
00:08:07.240 --> 00:08:11.959
leverage automation and the AI. So
they have to be able to educate everybody
90
00:08:11.959 --> 00:08:16.160
else. Okay, where give us
some ideas guidelines you think about with that
91
00:08:16.279 --> 00:08:22.759
too. Tesla an autopilot feature,
the driver assistance feature, and automobiles.
92
00:08:22.360 --> 00:08:26.879
Once in a while you still see
something in the press that a automated car,
93
00:08:26.519 --> 00:08:30.680
maybe not Tesla's, but somebody else
is testing out, like an automated
94
00:08:30.720 --> 00:08:35.639
pizza delivery car. I believe in
in Arizona got in an accident and heard
95
00:08:35.720 --> 00:08:41.720
somebody or even more recently the in
the press the automated chess player who had
96
00:08:41.720 --> 00:08:46.639
a robotic arm and pitched some kids
fingers. I believe that happened overseas somewhere.
97
00:08:46.639 --> 00:08:52.120
So the technology firms are certainly there
with that too. As Dr Lu
98
00:08:52.240 --> 00:08:56.200
mentioned, in financial services regulated industries, because they are always very risk averse,
99
00:08:56.879 --> 00:09:01.240
they're certainly kind of up there on
a bank as well, with being
100
00:09:01.279 --> 00:09:07.120
aware of risk conbides because they don't
want to get in any regulatory staff.
101
00:09:07.120 --> 00:09:11.519
Foods that's bad press could be operational
risk for that, let alone other larger
102
00:09:11.679 --> 00:09:16.440
financial risk. Healthcare, of course, once the while you read about something
103
00:09:16.440 --> 00:09:22.039
called robotic surgery and other health healthcare
innovation. Is anything directly touching a person
104
00:09:22.879 --> 00:09:28.279
they're certainly gonna be weighing in more
of these ethical concerns than say, somebody
105
00:09:28.360 --> 00:09:33.080
trying to test out a new way
to automate paper scanning to reduce errors and
106
00:09:33.279 --> 00:09:39.279
skin or paper faster with less error
tracking. A person maybe hurt by that
107
00:09:39.399 --> 00:09:45.799
in some way depending on an error
something being misread towards an application safer alone,
108
00:09:46.720 --> 00:09:50.120
but the piece of paper itself is
not so when it comes to touching
109
00:09:50.159 --> 00:09:54.120
people or or you can even see
with animals or the environment, that's going
110
00:09:54.200 --> 00:09:58.559
to be an issue. And also, like Dr lu mentioned, it's just
111
00:09:58.639 --> 00:10:03.559
becoming a larger issue with these E
s G concerns. The S and e
112
00:10:03.759 --> 00:10:07.480
s G as societal or social.
With that comes ethics. Are you doing
113
00:10:07.480 --> 00:10:11.799
the right thing by your people,
the people that you're provided services or products
114
00:10:11.799 --> 00:10:18.000
too, or the environment in the
world in general. So an the application
115
00:10:18.039 --> 00:10:22.600
of AI that you can be framed
against those contexts as well. Chris,
116
00:10:22.639 --> 00:10:28.000
you mentioned self driving cars. It
reminded me of a Sixty Minutes episode that
117
00:10:28.519 --> 00:10:35.559
focused on self driving cars and they
were interviewing a professor from Carnegie Mellon.
118
00:10:35.960 --> 00:10:41.600
I recall the interviewer asking him,
are these self driving cars? These automated
119
00:10:43.159 --> 00:10:48.399
machines. Are they are they benevolent? Can they make benevolent decisions? And
120
00:10:48.440 --> 00:10:54.480
I remember they got into this debate
and the professor said, they're only benevolent
121
00:10:54.559 --> 00:10:58.720
if we program them to be benevolent. But there's so many ethical things we
122
00:10:58.799 --> 00:11:05.720
have to consider in writing the code
or defining the business rules. It's a
123
00:11:05.799 --> 00:11:11.000
classic trolley problem, right, and
the Dolly problem was more of a thought
124
00:11:11.039 --> 00:11:13.879
experiment back in the day. But
now when you actually have to program it
125
00:11:13.919 --> 00:11:20.519
in for your literal trolley self driving
cars, that you have a hold host
126
00:11:20.600 --> 00:11:26.080
of ethics, it's strange to think
that the ethics questions might be what delays
127
00:11:26.120 --> 00:11:33.039
it further than the technology issues that
it has. Christ let me ask you
128
00:11:33.320 --> 00:11:37.720
about the automation life cycle. When
do you think in the automation life cycle,
129
00:11:39.159 --> 00:11:45.279
ethics needs to be taken into consideration? Honestly? Right at the outset,
130
00:11:45.559 --> 00:11:50.200
ideally something you see is kind of
a bit more commonplace again and regulated
131
00:11:50.320 --> 00:11:54.919
industries. Everybody has to take some
kind of compliance course that you understand different
132
00:11:54.960 --> 00:12:00.600
ethical risk and legal risk. I
know that's a little bit of a novel
133
00:12:00.840 --> 00:12:03.639
way of thinking about automation and a
on it's something that's been rattling in my
134
00:12:03.679 --> 00:12:09.039
head honestly a little bit, my
non robotic kid twi O. There.
135
00:12:09.600 --> 00:12:13.919
Ideally, when you're getting educated on
the potential for leveraging automation and AI as
136
00:12:13.960 --> 00:12:18.960
a way to improve a task word
process, there should be some kind of
137
00:12:18.000 --> 00:12:24.600
training that people read through at the
very least be aware of organizations, policies
138
00:12:24.639 --> 00:12:28.440
and procedures which typically align with some
regulatory regime depending on what part of the
139
00:12:28.440 --> 00:12:33.879
world you're in, but hopefully something
a little bit more rost robust than that,
140
00:12:33.039 --> 00:12:37.360
so that you can avoid kind of
those unknown so to speak, and
141
00:12:39.159 --> 00:12:43.399
not hit an ethical ship wire of
some kind, especially when you're working with
142
00:12:43.440 --> 00:12:50.519
anything mutchy people, whether it's empathies
and really with customer data. There certainly
143
00:12:50.519 --> 00:12:56.000
should be like a milestone in a
project review or approval process or even to
144
00:12:56.120 --> 00:12:58.799
go live just to kind of check
off once again to make sure that it's
145
00:12:58.840 --> 00:13:03.639
a kay. Do you even collect
the data to do something like that in
146
00:13:03.679 --> 00:13:07.759
the first place? Part of the
recipes We all know of automation that you
147
00:13:07.799 --> 00:13:11.519
need to have some data and formulas, and people tend to focus on the
148
00:13:11.600 --> 00:13:15.919
formulas right at the outside. Or
should you even be collecting that data in
149
00:13:15.960 --> 00:13:20.039
the first place could be an ethic
culture boire, let alone the the usage
150
00:13:20.039 --> 00:13:24.840
of that data in a biased formula. Some of these things could be pretty
151
00:13:24.879 --> 00:13:30.039
simple checkerles like oh yeah, like
in my paper skin an example, this
152
00:13:30.120 --> 00:13:33.240
looks pretty trivial. This should be
approved and I can sign off. But
153
00:13:33.360 --> 00:13:37.279
for something else, you might need
some further review as part of the process
154
00:13:37.360 --> 00:13:43.519
before you can even start running your
playbook and kicking off the automation project,
155
00:13:45.240 --> 00:13:50.000
whether you're doing a traditional way or
in some agile fashion. Lou, I
156
00:13:50.000 --> 00:13:54.519
want to return to you earlier.
You briefly mentioned a couple of steps do
157
00:13:54.600 --> 00:14:00.320
you think companies can take to make
sure automation is done ethically? Could you
158
00:14:00.360 --> 00:14:05.000
elaborate on those a bit more so. The first is to make sure that
159
00:14:05.519 --> 00:14:13.240
whenever you're using a machine learning engine, it's been built to avoid bias and
160
00:14:13.360 --> 00:14:18.200
to be ethical in its own right. For example, that means, if
161
00:14:18.240 --> 00:14:24.799
you're building out a machine learning engine, think about what sort of data you're
162
00:14:24.879 --> 00:14:31.720
using to train it. This is
where you see a lot of AI implementations
163
00:14:31.759 --> 00:14:37.559
fail because a lot of the available
data folks will go for it's typically more
164
00:14:37.639 --> 00:14:41.559
likely to be from white males,
even if you're doing something as simple as
165
00:14:41.559 --> 00:14:46.960
training handwriting, you're more likely to
get samples from white males and you are
166
00:14:48.600 --> 00:14:54.240
from females or from minorities. And
as you're setting up these pieces, the
167
00:14:54.360 --> 00:14:58.600
data you're using to train it is
going to determine the outcomes it can use.
168
00:15:00.360 --> 00:15:03.480
Similarly, you know when you mentioned
the system the court was using to
169
00:15:03.600 --> 00:15:11.320
predict offenders, that is based on
data that already had some of those inherent
170
00:15:11.360 --> 00:15:15.879
biases in it. When you're looking
at historical data, especially the farther back
171
00:15:15.919 --> 00:15:20.440
you go, the more likely you
are to see those sorts of concerns.
172
00:15:20.440 --> 00:15:24.320
So the first piece is to make
sure that the data you're using and that
173
00:15:24.399 --> 00:15:28.879
corpus of knowledge is what you want
it to be and there isn't bias in
174
00:15:28.919 --> 00:15:35.720
it. The next piece is to
make sure that you have the auditibility required.
175
00:15:35.679 --> 00:15:41.039
So depending on what algorithm you have, you should have a methodology for
176
00:15:43.320 --> 00:15:48.440
looking back and being able to give
an explanation for how every decision was made.
177
00:15:48.159 --> 00:15:52.039
That's not just to maintain the ethics
portion, but also just to protect
178
00:15:52.120 --> 00:15:58.440
yourself and your business for when you're
implementing these sorts of solutions. Thank you,
179
00:15:58.559 --> 00:16:03.440
lou ris Let. Me give you
a chance to weigh and on how
180
00:16:03.480 --> 00:16:10.279
companies can create the transparency that's needed
to reduce ethical risk and bias in AI
181
00:16:11.559 --> 00:16:15.879
well, especially when something is of
a sensitive nature. Of the transparency would
182
00:16:15.919 --> 00:16:19.720
certainly be driven from the top.
As we mentioned our top of the US
183
00:16:19.759 --> 00:16:25.159
example, it just came out in
the press. The AI bill of rights.
184
00:16:25.559 --> 00:16:30.080
Organizations typically stopped start at the very
top with some sort of a blueprint,
185
00:16:30.960 --> 00:16:34.879
some kind of framework, potentially a
company framework might be some sort of
186
00:16:34.919 --> 00:16:40.200
maturity model. That's something we typically
like to drive as part of the work
187
00:16:40.200 --> 00:16:44.320
that I've done in management consultancy and
elsewhere, out of a framework from some
188
00:16:44.399 --> 00:16:48.480
kind of maturity model where you might
look at okay, ethically immature, too
189
00:16:48.600 --> 00:16:52.120
immature, and then an alignment with
that. Of course, again I refer
190
00:16:52.200 --> 00:16:59.159
often to policies and protagers that aligned
not only with regulatory and compliance rules depending
191
00:16:59.200 --> 00:17:00.840
on the part of the world or
in, but also privacy rules, which
192
00:17:00.840 --> 00:17:06.759
can be a little problematic in some
parts of the world. Privacy as very
193
00:17:06.799 --> 00:17:10.920
straightforward as we see in Europe.
In the United States, there's still a
194
00:17:10.960 --> 00:17:15.440
bit of a patchwork of privacy regulations
with it seems like once in a while
195
00:17:15.519 --> 00:17:18.200
you'll read in the press another U
S state has come up where it's coming
196
00:17:18.240 --> 00:17:22.440
up with it don't take on privacy. So this can create a bit of
197
00:17:22.480 --> 00:17:26.960
a ladline for people to walk through, so speak especially with ethical concerns here.
198
00:17:27.799 --> 00:17:32.680
But ideally you have to write subject
matter experts in place, some people
199
00:17:32.759 --> 00:17:38.359
with legal expertise, obviously ethics or
compliance or regulatory expertise. It's very rare
200
00:17:38.400 --> 00:17:45.079
that you hear of organizations having professional
ethicist or philosophers even that do this sort
201
00:17:45.079 --> 00:17:48.039
of work. So typically you see
people out of a compliance space that can
202
00:17:48.039 --> 00:17:52.960
help with guidelines and then these help
you in turn create the operating models that
203
00:17:52.359 --> 00:17:56.759
from that you get playbooks to help
create something. You run books for after
204
00:17:56.920 --> 00:18:00.559
something is created, to keep it
up and running. Part and parcel with
205
00:18:00.599 --> 00:18:03.640
all this a little bit more bottom
up, I mentioned a little bit earlier
206
00:18:03.680 --> 00:18:07.559
the idea of education in the workplace. You see this a lot more in
207
00:18:07.559 --> 00:18:12.599
irregulated industries, but we're starting to
see the idea of AI literacy starting to
208
00:18:12.640 --> 00:18:18.240
be discussed. And depending on where
you see AI sitting in an organization,
209
00:18:18.440 --> 00:18:23.240
it could be its own literacy program
or that may flow into a data literacy
210
00:18:23.319 --> 00:18:27.039
program. The idea that AI should
be part of the remit of the data
211
00:18:27.079 --> 00:18:32.119
analytics team because there are the people
that work with data knows that, people
212
00:18:32.119 --> 00:18:37.079
that do models, and typically there
are people working most closely with your data
213
00:18:37.160 --> 00:18:41.960
scientists. They can provide some guidance
to the people who may not be data
214
00:18:41.000 --> 00:18:45.359
science but might be doing more simpler
forms of automation like you see in a
215
00:18:45.400 --> 00:18:49.680
process automation department. If you're not
in a regulated industry, there's can be
216
00:18:49.720 --> 00:18:52.119
a tach with good models to look
at, say, to look at what
217
00:18:52.160 --> 00:18:56.960
people are doing in financial services,
are healthcare, entertain utilities or ease in
218
00:18:57.000 --> 00:19:03.200
government? Well, definitely don't want
to ignore the AI Bill of Rights that
219
00:19:03.559 --> 00:19:07.359
the US government recently released. Loue, what do you think about it?
220
00:19:07.519 --> 00:19:15.039
Specifically the five principles they defined for
organizations to take into consideration. So first
221
00:19:15.440 --> 00:19:22.000
let's start by reiterating for listeners what
those five principles are. The first one
222
00:19:22.319 --> 00:19:30.000
is safe and effective systems, So
any system that being used should be effective
223
00:19:30.160 --> 00:19:37.319
and safe, war folks, pretty
self explanatory. Next, algorithmic discrimination protections.
224
00:19:37.359 --> 00:19:40.279
This is some of what we were
talking about earlier, making sure that
225
00:19:40.599 --> 00:19:48.200
any algorithm being used isn't discriminatory and
follows standard laws and regulations. Data privacy
226
00:19:48.480 --> 00:19:52.640
is the third one, which is
get a crucial component of any form of
227
00:19:52.960 --> 00:19:59.400
data enterprise. Fourth, we have
notice an explanation, so if your data
228
00:19:59.519 --> 00:20:06.839
is being used in machine learning model, you should have transparency to understand how
229
00:20:06.880 --> 00:20:11.960
and why it's being used. And
finally we have human alternatives consideration and fallback.
230
00:20:12.400 --> 00:20:17.240
I should point out a couple of
things about these. First, A
231
00:20:17.279 --> 00:20:22.240
lot of these are just basic common
sense. It makes sense not just for
232
00:20:22.319 --> 00:20:26.200
the consumer but also for the business
themselves to protect themselves. Right. No,
233
00:20:26.319 --> 00:20:32.000
but no business wants to go show
up in the news for having a
234
00:20:32.119 --> 00:20:37.559
racist Twitter bot. That's terrible press. At the same point, looking at
235
00:20:37.599 --> 00:20:41.880
human alternatives necessary you call into a
contact center, folks get frustrated. They
236
00:20:41.920 --> 00:20:45.680
don't want to be talking to a
robot anymore. You have to be able
237
00:20:45.720 --> 00:20:49.799
to press the road to talk to
a human that. These aren't anything revolutionary.
238
00:20:49.880 --> 00:20:55.559
This is the general things at businesses
for the most part, are already
239
00:20:55.599 --> 00:21:00.640
doing. But aside from that,
a lot of this is echoing a very
240
00:21:00.680 --> 00:21:04.960
similar piece of actual legislation out of
the EU called the AI Act, which
241
00:21:06.079 --> 00:21:11.559
puts some teeth behind this and until
we see some of that in the United
242
00:21:11.559 --> 00:21:18.599
States. Remember that this is just
die blinds for folks to use. So
243
00:21:18.559 --> 00:21:22.759
with all that in mind, I'd
say that a lot of this is general
244
00:21:22.839 --> 00:21:27.960
best practice that most folks should be
following. Anyways. One piece i'd say
245
00:21:29.000 --> 00:21:34.359
in here that I see as potentially
changing the way things are done today is
246
00:21:34.440 --> 00:21:41.240
the notice an explanation component because a
lot of what data has used today you
247
00:21:41.279 --> 00:21:45.359
don't always get notice, and that
again is going to depend on where you
248
00:21:45.440 --> 00:21:52.519
live and what regional privacy laws there
are. But even then it's most likely
249
00:21:52.599 --> 00:21:56.559
not going to make a huge change
in a consumer's life. There'll just be
250
00:21:56.640 --> 00:22:00.240
one more agreement you check at the
beginning, if you know what mean.
251
00:22:00.240 --> 00:22:03.440
Yeah, I would agree with Dr
Lew there as I was reading through this
252
00:22:03.519 --> 00:22:08.599
Bill of Rights myself and I just
thought about gul everybody. I'm sure it
253
00:22:08.720 --> 00:22:12.079
might not be along on this podcast, but I have some young folks in
254
00:22:12.160 --> 00:22:17.720
my family and I see them on
TikTok. Nothing in the TikTok experience,
255
00:22:17.759 --> 00:22:22.720
at least from my recollection of just
looking at it just briefly to something pop
256
00:22:22.799 --> 00:22:27.279
up and saying we're powered by automated
system and some of the content you are
257
00:22:27.400 --> 00:22:33.279
seeing might be automatically suggested to you
based off of data and imparted form by
258
00:22:33.359 --> 00:22:41.039
what you're looking at. We're looking
for TikTok. TikTok's becoming an interesting topic
259
00:22:41.079 --> 00:22:47.799
of conversation in the marketing space because
there's a recent studies showing that the younger
260
00:22:47.839 --> 00:22:51.240
generations, even the youngest generations,
the one that they are just old enough
261
00:22:51.279 --> 00:22:55.799
to have some kind of bodied power, are starting to use TikTok as much
262
00:22:55.920 --> 00:23:00.839
as as Google to People typically think
of Google by default because they're the primary
263
00:23:00.880 --> 00:23:04.720
search engine, but now you see
people searching for things in other forms as
264
00:23:04.759 --> 00:23:10.119
well, and it might not be
aware of that if you're using Amazon as
265
00:23:10.119 --> 00:23:15.279
your primary way to search for products, unless you read the disclaimers down there
266
00:23:15.319 --> 00:23:18.359
at the bottom of the footter.
If something doesn't automatically pop up in front
267
00:23:18.400 --> 00:23:26.119
of you and tell you that you're
using an automated system and these other caveats,
268
00:23:26.240 --> 00:23:27.960
it could be it could be interesting, I'll put it that way.
269
00:23:29.079 --> 00:23:33.039
And on the other side is the
usability factor part of why so many people
270
00:23:33.079 --> 00:23:38.400
have drawn to digital experiences versus the
human alternatives. Will say, paper and
271
00:23:38.440 --> 00:23:42.359
pen and things like that are just
calling somebody up on the phone is the
272
00:23:42.440 --> 00:23:47.200
friction of the digital experience. And
you see some of these stories in the
273
00:23:47.240 --> 00:23:52.720
last couple of years from overseas with
the stricter especially in Western Europe, the
274
00:23:52.799 --> 00:23:59.680
stricter privacy rules, some people saying
that they're practically making certain websites unusable.
275
00:24:00.519 --> 00:24:04.000
So there's going to be I think
a little bit of trial and error tests
276
00:24:04.039 --> 00:24:08.640
and learn to kind of navigate what
this blueprint actually implies, at least for
277
00:24:08.640 --> 00:24:15.279
the United States. Yeah, let
me ask you. Some have called for
278
00:24:15.359 --> 00:24:23.200
an ethics framework as the foundation upon
which any AI technology is created and implemented.
279
00:24:23.279 --> 00:24:27.759
What are your thoughts on this.
I think it makes sense depending on
280
00:24:27.960 --> 00:24:32.640
what sort of framework you put in
place, and I think any regulation like
281
00:24:32.680 --> 00:24:37.680
this is going to have to come
from a government level. Now, if
282
00:24:37.720 --> 00:24:41.880
you think about the basics that we
were discussing earlier, like the AI Bill
283
00:24:41.920 --> 00:24:48.440
of Rights. Again, I don't
see anyone complaining about this, but putting
284
00:24:48.480 --> 00:24:52.960
it into a codified form means that
your compliance folks, who, again,
285
00:24:52.000 --> 00:24:57.599
as Chris said, are generally going
to be the folks looking at ethics within
286
00:24:57.720 --> 00:25:06.160
an enterprise, and most of them
are not hiring philosophers have something to benchmark
287
00:25:06.279 --> 00:25:11.119
against. So in general, it's
gonna make a lot of sense because it
288
00:25:11.160 --> 00:25:17.079
gives companies that the backbone to look
at when they're trying to make these sorts
289
00:25:17.079 --> 00:25:23.759
of decisions. One fairly recent example
of an ethics framework that I read about
290
00:25:25.039 --> 00:25:27.039
I wanted to bring it up and
Chris, I'll direct this one to you.
291
00:25:29.039 --> 00:25:34.839
In Google announced that it had created
a council to develop an ethics framework,
292
00:25:34.920 --> 00:25:42.359
primarily to address AI powered facial recognition
and fairness in machine learning. However,
293
00:25:42.480 --> 00:25:48.720
within ten days of its establishment,
the council was dissolved due to suspected
294
00:25:48.759 --> 00:25:55.160
bias amongst one of the council members. It raises the question, how can
295
00:25:55.240 --> 00:26:00.359
people most or you could even say
all of whom some sort of biases.
296
00:26:00.880 --> 00:26:07.359
How can they effectively play a role
in developing a framework to avoid biases in
297
00:26:07.440 --> 00:26:11.440
AI? And maybe the toughest question
we've had today, So sorry for directing
298
00:26:11.440 --> 00:26:17.200
it your way. That's okay,
because I don't think it's too hard.
299
00:26:17.319 --> 00:26:21.559
I mean, again, we mentioned
this idea of of a blueprint at a
300
00:26:21.640 --> 00:26:27.400
high government level regulations and compliance.
They serve as guard rails. From my
301
00:26:27.480 --> 00:26:32.240
vantage point, you would kind of
look at the creation of such a framework
302
00:26:33.000 --> 00:26:40.799
using the classical model of people processing
technology and people might even have like these
303
00:26:40.880 --> 00:26:45.839
higher level considerations as within you know, the notion of B s G and
304
00:26:45.880 --> 00:26:52.480
the particular process. Certain processes,
depending on whether or not there across industry,
305
00:26:52.519 --> 00:26:57.960
like a marketing process might have a
little bit of additional ethical review to
306
00:26:59.079 --> 00:27:03.680
it because they're that notion of targeting
people automatically, they might not be aware
307
00:27:03.720 --> 00:27:07.400
of it. Do you need to
present some kind of notice that's happening,
308
00:27:07.960 --> 00:27:11.079
whether it's with an AD or is
you mentioned a offer force and kind of
309
00:27:11.079 --> 00:27:15.440
say a lending product or maybe some
other product. When you look at something
310
00:27:15.519 --> 00:27:19.160
on the shopping website and you go
into your Facebook and you get an AD
311
00:27:19.200 --> 00:27:25.400
for within minutes and sometimes it's mind
blowing the speed and how fast that,
312
00:27:25.440 --> 00:27:29.400
I mean, you can kind of
creep you out. Creating a framework again,
313
00:27:29.720 --> 00:27:33.680
certainly you'd have some initial the consultative
group to get together legal experts,
314
00:27:33.680 --> 00:27:41.240
compliance experts, subject matter experts in
different functions to start waiting through and seeing
315
00:27:41.440 --> 00:27:47.720
does this that within an existing regulatory
and compliance our seam and a risk calculator
316
00:27:47.759 --> 00:27:49.960
that we have at the enterprise level, or do we need to create or
317
00:27:51.000 --> 00:27:56.319
modify something new, right, And
just to add on there, I'd say
318
00:27:56.359 --> 00:28:00.880
this is actually probably one of the
simplest questions you've asked, Brad, because
319
00:28:00.559 --> 00:28:06.200
we're not reinventing the wheel here.
This isn't the first time we've needed to
320
00:28:06.960 --> 00:28:14.039
put a group of people together to
figure out how to enforce basic ethics.
321
00:28:14.559 --> 00:28:18.119
If you think about pretty much any
HR policy, a lot of the same
322
00:28:18.680 --> 00:28:23.839
factors were brought into play. So
you need a diverse group of individuals forming
323
00:28:23.920 --> 00:28:27.119
a committee to figure this out.
This is nothing new, We're just you're
324
00:28:27.160 --> 00:28:33.680
doing it for a new purpose.
Very good point, Mr Leeu Blue.
325
00:28:33.759 --> 00:28:37.160
Let me ask you about GPT three. As you know, r P A
326
00:28:37.279 --> 00:28:44.559
combined with natural language processing skills has
been playing a very big role in areas
327
00:28:44.720 --> 00:28:51.839
like contact centers as automation has become
integral in in that function. GPT three
328
00:28:52.119 --> 00:28:57.160
is one of the largest NLP language
models in use and widely considered a breakthrough
329
00:28:57.279 --> 00:29:03.079
in AI. It can generate sentences
and even write article summaries or generate full
330
00:29:03.119 --> 00:29:07.079
stories that are creative in nature just
based on a prompt of a few lines.
331
00:29:07.400 --> 00:29:14.799
Pretty amazing stuff. However, even
GPT three has proven in in recent
332
00:29:14.920 --> 00:29:21.920
times to generate problematic content if it's
fed with harmful or malicious prompts. So
333
00:29:22.000 --> 00:29:26.880
the question for you, Lou,
it's just a situation where AI algorithms will
334
00:29:26.920 --> 00:29:32.839
have to be written to regulate AI
algorithms. That's a that's a tricky question
335
00:29:32.960 --> 00:29:37.680
because if you think about the GPT
three example you just gave, it was
336
00:29:38.079 --> 00:29:47.799
generating offensive text when given of offensive
prompt. So the the phrasing here is
337
00:29:47.960 --> 00:29:52.119
we don't give it an offensive prompt, and some of that is just going
338
00:29:52.160 --> 00:29:56.559
to be basic what I like to
call a potty mouth filter. And I
339
00:29:56.559 --> 00:30:00.759
actually didn't come up with that term. That was from IBM way back when
340
00:30:02.640 --> 00:30:07.640
Watson was on Jeopardy. They are
doing a training session and they had the
341
00:30:07.680 --> 00:30:14.440
executive producers of Jeopardy in the room
and the category was four letter words,
342
00:30:14.640 --> 00:30:19.480
and Watson dropped the F bomb and
they're like, well, it's a good
343
00:30:19.480 --> 00:30:25.519
thing. This isn't on national television. This is a practice zoo need to
344
00:30:25.559 --> 00:30:29.880
fix that, so they put in
a potty mouth filter. I think some
345
00:30:29.920 --> 00:30:33.039
of the same things just need to
go into place here and not necessarily on
346
00:30:33.079 --> 00:30:37.200
the output, but sometimes looking at
the input itself. The same way where
347
00:30:37.880 --> 00:30:44.839
you have safe search options for any
other category if you're doing things, the
348
00:30:44.920 --> 00:30:51.319
same principles apply here where you want
to make sure that any use of machine
349
00:30:51.359 --> 00:30:56.039
learning or AI is going to be
safe and is going to be efficient.
350
00:30:57.240 --> 00:31:04.079
Gotcha, gotcha, Chris. As
we automate more and more business processes,
351
00:31:04.079 --> 00:31:11.519
will ethical concerns continue to increase?
In your opinion, I think it depends
352
00:31:11.759 --> 00:31:18.920
on the application of automation within particular
functional areas and businesses. You think about.
353
00:31:19.119 --> 00:31:22.680
Once upon a time, human resources, for example, this wasn't too
354
00:31:22.759 --> 00:31:29.960
much of a concern because human resources
and through you had humans reading resumes,
355
00:31:30.000 --> 00:31:33.680
and you had humans doing a lot
of work with personal data to set up
356
00:31:33.720 --> 00:31:37.920
like your benefits and your training and
all the learning and development different things.
357
00:31:37.039 --> 00:31:42.279
But now you see the challenge of
ethics coming and play the the example with
358
00:31:42.759 --> 00:31:48.240
automated silvering of resumes based off of
gender being potentially sticky. I've heard some
359
00:31:48.359 --> 00:31:53.240
things like that have happened in the
past on a more manual basis, where
360
00:31:53.279 --> 00:31:59.400
people's first names could be filtered out
by people. And you see some of
361
00:31:59.440 --> 00:32:01.079
these studies lolat around on the internet
once in a while in the last year
362
00:32:01.119 --> 00:32:07.119
or two, where there's different bias
just in this selection of candidate. So
363
00:32:07.279 --> 00:32:10.319
that's an area we see ethics being
a concern in the automation space too.
364
00:32:10.680 --> 00:32:15.680
There's the kind of the rising interest
in this topic of process mining and process
365
00:32:15.720 --> 00:32:20.359
analytics. You already seen a little
bit of that tripping some ethical trip wire
366
00:32:20.519 --> 00:32:23.640
so speak. I believe there are
certain parts of Europe that are more regulated
367
00:32:23.680 --> 00:32:28.279
than others, and I already braced
that as a flag to be addressed that
368
00:32:28.400 --> 00:32:32.000
you need people to opt in to
have what they're doing at work tracked.
369
00:32:32.359 --> 00:32:37.160
Where in the US, it's just
assume that you signed off a part of
370
00:32:37.160 --> 00:32:40.759
your employment agreement, and if you're
walking around the hallways or in anabilits,
371
00:32:40.839 --> 00:32:45.680
or you're using certain computer systems or
anything you're on your screen or your voicemail,
372
00:32:45.359 --> 00:32:50.680
those things could potentially be fodder for
some kind of process mining and process
373
00:32:50.759 --> 00:32:55.160
analysis exercise. Certainly couldn't be much
more sensitivity when you add the additional layers
374
00:32:55.359 --> 00:33:01.440
of financial services and healthcare and energy
thing or things that are sensitive at a
375
00:33:01.440 --> 00:33:07.839
more macro level, such a government
whatsoever, think about it like this brand
376
00:33:07.119 --> 00:33:14.640
all these processes whenever you're automating an
existing process. The ethics concerns were already
377
00:33:14.880 --> 00:33:19.359
thought about. It's just a matter
of who's completing it. Is it a
378
00:33:19.400 --> 00:33:23.160
person or is it an automation.
And I would argue that the most part,
379
00:33:23.480 --> 00:33:28.720
you have a lot more transparency with
an automation than you do with a
380
00:33:28.799 --> 00:33:34.839
human being because of the auditability.
That's exactly, you don't know if the
381
00:33:35.480 --> 00:33:39.839
banker you happen to be talking to
has any form of bias, and there's
382
00:33:39.880 --> 00:33:45.200
no way to track how those decisions
were made, whereas with an automation you
383
00:33:45.359 --> 00:33:52.359
have a deeper level of insight and
auditability. Right, good point. So,
384
00:33:52.480 --> 00:33:55.680
Louf, you've been on the podcast
in the past to discuss the next
385
00:33:55.720 --> 00:34:01.799
phase of automation, one that is
likely to be much more autonomous. How
386
00:34:01.839 --> 00:34:08.400
does this present even more ethical challenges? There's a decent amount of them,
387
00:34:08.440 --> 00:34:13.239
I'll be honest, because if you
remember the last time I was on,
388
00:34:13.320 --> 00:34:19.400
we talked about how the ultimate goal
of automation was too automate as much of
389
00:34:19.440 --> 00:34:22.599
it as possible, with the eventual
goal of the world. Where you were
390
00:34:22.639 --> 00:34:28.199
doing something that was repetitive, the
CEE could get a pop up that says,
391
00:34:28.199 --> 00:34:30.719
hey, it looks like Rad's doing
a repetitive task. Would you like
392
00:34:30.760 --> 00:34:35.079
to automate, it clicks okay,
and it's done. Now. When we
393
00:34:35.119 --> 00:34:37.760
first talked about that, it was
it seemed like kind of a pipe dream.
394
00:34:37.800 --> 00:34:43.440
What with the technology we've put in
place since then, we have automated
395
00:34:43.559 --> 00:34:50.760
enough to shorten development cycles by anywhere
from so we're a decent ways down that
396
00:34:50.800 --> 00:34:57.119
path. But a key portion of
that is task mining, which uses desktop
397
00:34:57.199 --> 00:35:02.760
based recorders to monitor what human employ
these are doing and separate out multitasking and
398
00:35:02.800 --> 00:35:09.199
digitize that into a model right.
Doing that involves a lot of the privacy
399
00:35:09.239 --> 00:35:14.880
concerns that Chris was just mentioning.
You need to make sure that when you're
400
00:35:15.000 --> 00:35:22.159
monitoring folks, you're monitoring them in
a controlled and governed fashion. So at
401
00:35:22.320 --> 00:35:24.159
S, S and C Blueprism,
the way we do this is by applying
402
00:35:24.199 --> 00:35:30.679
a whole bunch of privacy controls.
Sometimes the individual users have the ability to
403
00:35:30.719 --> 00:35:36.760
start or stop the recording functionality.
You could set it up to obscure sensitive
404
00:35:36.840 --> 00:35:43.559
data, redact the I I,
or also only look at specific applications or
405
00:35:43.719 --> 00:35:46.880
exclude applications, so you can say
I only want to record what they're doing
406
00:35:46.920 --> 00:35:53.760
in these five specific applications or work, or you could say I want to
407
00:35:53.800 --> 00:36:00.400
exclude anything from Facebook dot Com because
that clearly won't be relevant. But the
408
00:36:00.440 --> 00:36:06.159
fact is you have to have consideration
for pieces like that and for the privacy
409
00:36:06.159 --> 00:36:10.119
of not just workers, but the
customers will be having their data run through
410
00:36:10.159 --> 00:36:16.719
these systems. Any final thoughts today
before we bring our podcast to conclusion.
411
00:36:19.119 --> 00:36:22.079
So one of the things folks often
don't think about, and this is an
412
00:36:22.119 --> 00:36:28.000
area that was addressed in the Journal
of the Aerobotics and Workplace Automation recently,
413
00:36:28.840 --> 00:36:37.599
is the ability for automation to improve
upon the existing ethics that are in our
414
00:36:37.639 --> 00:36:44.679
workplace. For example, automations can
do things like monitor performance refused to any
415
00:36:44.719 --> 00:36:52.039
instances or bias, or help with
recruiting and give that layer of insight that
416
00:36:52.119 --> 00:36:58.840
otherwise we wouldn't have. So in
general, I think the application of AI,
417
00:36:59.000 --> 00:37:04.920
machine learning, and anomation to the
workplace will improve the ethics that are
418
00:37:04.960 --> 00:37:09.559
being maintained rather than complicate them.
I would second that it's certainly one of
419
00:37:09.599 --> 00:37:16.519
the One of the positives of automation
is that it improves quality, reduces errors,
420
00:37:17.519 --> 00:37:22.239
and if you have these ethical oversight
boards just to make sure that things
421
00:37:22.239 --> 00:37:25.480
do go off the rails in the
wrong direction in the long run, when
422
00:37:25.679 --> 00:37:30.559
you're ensuring that people and or divibul
workers were both are doing what they're supposed
423
00:37:30.599 --> 00:37:35.360
to be doing, and you have
some automated checks in there, you will
424
00:37:35.639 --> 00:37:39.639
reduce risk. So and you're seeing
some of that, especially in financial services
425
00:37:39.679 --> 00:37:45.639
where there's large scale applications of pilots
going on with automated scanning of call centered
426
00:37:45.639 --> 00:37:52.039
transcripts and other forms of communications,
and make sure that the traders aren't communicating
427
00:37:52.039 --> 00:37:54.400
with each other in forms that they're
not supposed to, which I know is
428
00:37:54.440 --> 00:37:59.639
something that's happened in some of the
European financial markets not too long ago.
429
00:38:00.960 --> 00:38:05.159
Thank you both for ending us on
a positive note on the ethics topic.
430
00:38:05.400 --> 00:38:10.039
As I mentioned at the beginning,
this conversation about ethics and automation and AI
431
00:38:10.159 --> 00:38:15.760
is an important one to have,
but it should not reflect poorly on either
432
00:38:15.840 --> 00:38:20.679
of these areas. They're doing a
ton of good in the world as we
433
00:38:20.760 --> 00:38:23.039
know well. Gentlemen, I want
to thank you both for being on the
434
00:38:23.079 --> 00:38:29.400
podcast today. It's been a fascinating
conversation. I've certainly learned a lot just
435
00:38:29.519 --> 00:38:32.639
listening to both of you, so
thank you for sharing your perspective. You
436
00:38:32.679 --> 00:38:36.880
guys have a lot to offer on
this topic, and I know it's going
437
00:38:36.920 --> 00:38:40.760
to be informative for our listeners.
In the history of the podcast, I
438
00:38:40.800 --> 00:38:45.519
believe this is the first time we've
ever talked about ethics in intelligent automation.
439
00:38:46.079 --> 00:38:52.519
So thank you for helping us cover
it sufficiently. Brad, appreciate the time
440
00:38:52.559 --> 00:38:58.039
today. Always happy to you jump
onto these. I think the future is
441
00:38:58.480 --> 00:39:04.719
exciting and application of machine learning and
automation will just continue to make our lives
442
00:39:04.880 --> 00:39:08.960
better and easier. I echo with
Dr Luke said, thank you right,
443
00:39:09.039 --> 00:39:13.480
not only on the behalf of myself, but I'll speak up. I'll behalf
444
00:39:13.519 --> 00:39:17.519
of Henry's publications. It was a
lot of fun pioneering the issue on this
445
00:39:17.599 --> 00:39:21.960
topic. There are certainly some books
coming out here and there in the press
446
00:39:22.639 --> 00:39:27.239
lately that are starting to delve into
the topics of ethical machines and ethics and
447
00:39:27.320 --> 00:39:32.000
AI and gaining away chrom maybe the
fearmongerine that goes way back to Isaac Eisamouth
448
00:39:32.079 --> 00:39:37.679
and I robot and science fiction and
the scary robots and more towards the positive
449
00:39:37.719 --> 00:39:42.800
aspects of your rewards off automation and
the I all the best tea boat,
450
00:39:43.039 --> 00:39:46.559
gentlemen. Thanks for tuning in to
Transform Now. For more insightful discussions on
451
00:39:46.639 --> 00:39:51.639
digital transformation and more, check out
our podcast channel, will you'll find all
452
00:39:51.679 --> 00:39:54.760
of our previous episodes and to make
sure you never missed an episode, subscribe
453
00:39:54.800 --> 00:39:58.920
to the show in your favorite podcast
player, and if you like what you've
454
00:39:58.920 --> 00:40:02.480
heard, please leave was a review. For more information about digital transformation and
455
00:40:02.559 --> 00:40:07.519
the future of work, check out
blueprism dot com to learn how s s
456
00:40:07.559 --> 00:40:12.599
and see Blueprism's digital workforce is enabling
enterprise transformation now
1
00:00:03.480 --> 00:00:08.199
Welcome to Transform Now, the podcast
brought to you by Robotic process automation pioneer
2
00:00:08.560 --> 00:00:13.240
S S and C Blue Prison.
Digital transformation has the potential to reshape the
3
00:00:13.240 --> 00:00:18.320
way companies service their customers, engage
their employees, and manage their operations.
4
00:00:18.760 --> 00:00:23.640
Whether you're looking to develop strategies,
tactics, or best practices to positively impact
5
00:00:23.640 --> 00:00:27.839
the future of work, or you're
curious to see how other companies have successfully
6
00:00:27.920 --> 00:00:32.920
navigated their digital transformation programs, than
this podcast is for you. We're here
7
00:00:32.960 --> 00:00:38.079
to help you Transform Now. Hello
everyone, I'm Brad Hairston with S S
8
00:00:38.159 --> 00:00:43.840
and C Blue Prism. Welcome to
the Transform Now podcast. Today, I'm
9
00:00:43.880 --> 00:00:48.479
happy to have two guests on the
show. The first is Chris Johannesson,
10
00:00:48.600 --> 00:00:55.240
an automation and AI expert and the
editor of the Journal of AI, Robotics
11
00:00:55.320 --> 00:01:00.840
and Workplace Automation. My second guest
is Dr lu By Kenenheimer, CTO of
12
00:01:00.880 --> 00:01:06.599
the America's at S S SC Blue
Prison. Both Christ and Lee have been
13
00:01:06.599 --> 00:01:11.280
on the podcast before, and today
I will be talking with them about the
14
00:01:11.319 --> 00:01:19.120
ethics of automation and AI. Gentlemen, welcome. Let's start with some introductions.
15
00:01:19.200 --> 00:01:22.879
Chris, why don't you start us
off. Hello everybody, and hello
16
00:01:22.920 --> 00:01:27.599
again Brad Chris Johannison. I'm,
as Brad mentioned, editor of the Journal
17
00:01:27.640 --> 00:01:34.799
of AI, Robotics and Workplace Automation. That's journal launched about a year ago,
18
00:01:34.040 --> 00:01:40.120
published by Henry Stewart Publications out of
the UK, and we just had
19
00:01:40.200 --> 00:01:46.560
a special edition on ethics and AI. So this topic is top of mind.
20
00:01:46.640 --> 00:01:51.079
There's some other things going on in
the news as well that continue to
21
00:01:51.120 --> 00:01:55.680
bring the ifs to the forefront.
That's just a part time role that I
22
00:01:55.719 --> 00:02:00.159
serve in this journal. I've been
doing management consulting work and you automate shin
23
00:02:00.280 --> 00:02:05.480
and not AI space for the past
several years with firms such as Mackenzie and
24
00:02:05.599 --> 00:02:09.680
see It Partners, and before that
some of the early work in automation and
25
00:02:09.840 --> 00:02:15.240
AI going back to my time with
GE and including time that I spent with
26
00:02:15.599 --> 00:02:22.680
eBay Enterprise as well as leading a
marketing practice on behalf of Omncommon Philadelphia where
27
00:02:22.719 --> 00:02:25.840
we deviled in some a little bit
of AI as part of marketing automation.
28
00:02:27.520 --> 00:02:31.479
Awesome, thank you, lu Let's
Brent said. I'm Dr Lou Bockenhammer,
29
00:02:31.560 --> 00:02:36.439
the CTO of the Americas for Blue
Prism. I've been with Essence and see
30
00:02:36.479 --> 00:02:43.280
Blue Prism for almost four and a
half years now, and it's been an
31
00:02:43.280 --> 00:02:47.400
exciting time for both automation and AI. Before that, I was with IBM
32
00:02:47.479 --> 00:02:52.560
S Watson and Cloud Platform team working
with some of the AI and m L
33
00:02:52.599 --> 00:02:59.439
tools over there, and currently we're
at a I want to say an inflection
34
00:02:59.520 --> 00:03:02.840
point in how machine learning and AI
are being used in your day to day
35
00:03:02.840 --> 00:03:09.800
automations that make this conversation particularly relevant. So thanks for having me. It's
36
00:03:09.800 --> 00:03:13.840
great to have both of you on
the show today. Thank you for being
37
00:03:13.879 --> 00:03:16.960
here. Well. Now that we
are in full swing in the era of
38
00:03:17.000 --> 00:03:23.759
intelligent automation, the topic of ethics
at the intersection of r p A and
39
00:03:23.000 --> 00:03:30.479
AI has definitely taken on more importance. Just this month, the White House
40
00:03:30.560 --> 00:03:36.520
Office of Science and Technology took this
topic center stage by releasing its first ever
41
00:03:37.199 --> 00:03:44.080
AI Bill of Rights, which defines
five principles for making automated systems work for
42
00:03:44.120 --> 00:03:47.960
the American people. We will certainly
talk about that today. If you look
43
00:03:47.960 --> 00:03:53.680
at just where ethics and AI have
emerged in recent history, there have been
44
00:03:53.719 --> 00:04:01.319
several now infamous examples of truly unethical
AI. There's the two thousand fourteen example,
45
00:04:01.360 --> 00:04:10.520
where an AI powered recruiting software ended
up overwhelmingly preferring male candidates. There's
46
00:04:10.520 --> 00:04:16.279
the chatbot created back in which,
after being released on Twitter, started quickly
47
00:04:16.360 --> 00:04:26.519
spewing racist and misogynist messages. And
in six an algorithm called compass was used
48
00:04:26.519 --> 00:04:30.319
in the u S court systems to
predict the likelihood of offenders committing another crime.
49
00:04:30.839 --> 00:04:36.959
A journalist came forward and exposed that
this model predicted twice as many false
50
00:04:38.000 --> 00:04:43.759
positives for minority offenders. So,
of course all of these are AI gone
51
00:04:43.800 --> 00:04:49.000
wrong stories. They do not diminish
the many, many ways that AI has
52
00:04:49.040 --> 00:04:56.120
had a very positive impact on our
society and in enhanced automation for that matter.
53
00:04:56.959 --> 00:04:59.680
So I really want to get into
the dialogue with both of you to
54
00:04:59.759 --> 00:05:03.079
talk through this kind of stuff.
Lou let me start with you. Now
55
00:05:03.120 --> 00:05:11.160
that machine learning models are enabling decisions
in the automation of more complex business processes,
56
00:05:11.560 --> 00:05:16.560
do you see ethics becoming a bigger
concern for automation teams and their executive
57
00:05:16.560 --> 00:05:23.680
sponsors? Oh? Absolutely. If
you think about automation, you really can't
58
00:05:23.720 --> 00:05:30.519
get past the need to be including
machine learning and artificial intelligence because otherwise automation
59
00:05:30.519 --> 00:05:33.920
are just hands on a keyboard with
no brain behind them. But then you
60
00:05:34.040 --> 00:05:39.879
hit all the issues with AI or
machine learning that you brought up before.
61
00:05:40.240 --> 00:05:44.959
I think these fall into two different
categories. The first side you can think
62
00:05:44.959 --> 00:05:49.360
of as issues with the model themselves, how they were built, how they
63
00:05:49.360 --> 00:05:55.000
were trained, what sort of bias
they might have inherently, and the second
64
00:05:55.319 --> 00:06:00.319
category being issues with how the models
are being used, so that could be
65
00:06:00.360 --> 00:06:05.720
a for example, mass usage of
facial recognition technology. Typically when we are
66
00:06:05.800 --> 00:06:13.000
looking at automation use cases, we're
normally looking at the first category, when
67
00:06:13.000 --> 00:06:16.519
we're looking at the in any inherent
bias in the model, and that previously
68
00:06:16.560 --> 00:06:21.319
had been an issue. There was
a situation where some algorithms have been put
69
00:06:21.360 --> 00:06:28.319
into place to help approve mortgage loans
that again turned out to you have a
70
00:06:28.560 --> 00:06:33.120
racial bias most of the time when
you're looking at customers. Are two raised
71
00:06:33.240 --> 00:06:39.000
to handle this, and the first
is to be very careful with what sort
72
00:06:39.160 --> 00:06:43.519
of data you use, what your
corpusive data is that you're used to train
73
00:06:43.839 --> 00:06:47.759
that machine learning model. And the
second is to make sure that you have
74
00:06:48.160 --> 00:06:55.720
auditability at every step. We typically
see We typically used to see machine learning
75
00:06:55.720 --> 00:07:00.120
engines that were more of a black
box. But nowadays, just as you
76
00:07:00.120 --> 00:07:05.600
were asking, customers are much more
concerned with maintaining ethics and as a result,
77
00:07:06.000 --> 00:07:10.959
they find it very important to make
sure that any machine learning or AI
78
00:07:11.079 --> 00:07:15.560
solution they put into place has a
full audit trail to explain how it made
79
00:07:15.639 --> 00:07:19.920
every decision it made. Interesting,
Chris, let me ask you, from
80
00:07:19.959 --> 00:07:27.839
your perspective, which industries do you
see leading the charge in addressing ethics and
81
00:07:27.959 --> 00:07:32.680
automation and AI. There's a bit
of a classical model out there about technology
82
00:07:33.000 --> 00:07:36.199
adoption and something goes back to an
old book. I think it came out
83
00:07:36.199 --> 00:07:42.800
in early nineteen sixties, a book
I believed title was around the diffusions of
84
00:07:42.800 --> 00:07:46.040
innovations as every call top of my
head. So even though we cited a
85
00:07:46.040 --> 00:07:51.720
few technology for examples earlier, certainly
companies like that they're right there on the
86
00:07:51.800 --> 00:07:56.879
vanguard of addressing the ethics and automation
and AI not only for their own you
87
00:07:57.079 --> 00:08:01.920
use, but many of these corporations
are not only using this stuff for their
88
00:08:01.920 --> 00:08:07.800
own purposes, but they're trying to
bring software applications out to the market that
89
00:08:07.240 --> 00:08:11.959
leverage automation and the AI. So
they have to be able to educate everybody
90
00:08:11.959 --> 00:08:16.160
else. Okay, where give us
some ideas guidelines you think about with that
91
00:08:16.279 --> 00:08:22.759
too. Tesla an autopilot feature,
the driver assistance feature, and automobiles.
92
00:08:22.360 --> 00:08:26.879
Once in a while you still see
something in the press that a automated car,
93
00:08:26.519 --> 00:08:30.680
maybe not Tesla's, but somebody else
is testing out, like an automated
94
00:08:30.720 --> 00:08:35.639
pizza delivery car. I believe in
in Arizona got in an accident and heard
95
00:08:35.720 --> 00:08:41.720
somebody or even more recently the in
the press the automated chess player who had
96
00:08:41.720 --> 00:08:46.639
a robotic arm and pitched some kids
fingers. I believe that happened overseas somewhere.
97
00:08:46.639 --> 00:08:52.120
So the technology firms are certainly there
with that too. As Dr Lu
98
00:08:52.240 --> 00:08:56.200
mentioned, in financial services regulated industries, because they are always very risk averse,
99
00:08:56.879 --> 00:09:01.240
they're certainly kind of up there on
a bank as well, with being
100
00:09:01.279 --> 00:09:07.120
aware of risk conbides because they don't
want to get in any regulatory staff.
101
00:09:07.120 --> 00:09:11.519
Foods that's bad press could be operational
risk for that, let alone other larger
102
00:09:11.679 --> 00:09:16.440
financial risk. Healthcare, of course, once the while you read about something
103
00:09:16.440 --> 00:09:22.039
called robotic surgery and other health healthcare
innovation. Is anything directly touching a person
104
00:09:22.879 --> 00:09:28.279
they're certainly gonna be weighing in more
of these ethical concerns than say, somebody
105
00:09:28.360 --> 00:09:33.080
trying to test out a new way
to automate paper scanning to reduce errors and
106
00:09:33.279 --> 00:09:39.279
skin or paper faster with less error
tracking. A person maybe hurt by that
107
00:09:39.399 --> 00:09:45.799
in some way depending on an error
something being misread towards an application safer alone,
108
00:09:46.720 --> 00:09:50.120
but the piece of paper itself is
not so when it comes to touching
109
00:09:50.159 --> 00:09:54.120
people or or you can even see
with animals or the environment, that's going
110
00:09:54.200 --> 00:09:58.559
to be an issue. And also, like Dr lu mentioned, it's just
111
00:09:58.639 --> 00:10:03.559
becoming a larger issue with these E
s G concerns. The S and e
112
00:10:03.759 --> 00:10:07.480
s G as societal or social.
With that comes ethics. Are you doing
113
00:10:07.480 --> 00:10:11.799
the right thing by your people,
the people that you're provided services or products
114
00:10:11.799 --> 00:10:18.000
too, or the environment in the
world in general. So an the application
115
00:10:18.039 --> 00:10:22.600
of AI that you can be framed
against those contexts as well. Chris,
116
00:10:22.639 --> 00:10:28.000
you mentioned self driving cars. It
reminded me of a Sixty Minutes episode that
117
00:10:28.519 --> 00:10:35.559
focused on self driving cars and they
were interviewing a professor from Carnegie Mellon.
118
00:10:35.960 --> 00:10:41.600
I recall the interviewer asking him,
are these self driving cars? These automated
119
00:10:43.159 --> 00:10:48.399
machines. Are they are they benevolent? Can they make benevolent decisions? And
120
00:10:48.440 --> 00:10:54.480
I remember they got into this debate
and the professor said, they're only benevolent
121
00:10:54.559 --> 00:10:58.720
if we program them to be benevolent. But there's so many ethical things we
122
00:10:58.799 --> 00:11:05.720
have to consider in writing the code
or defining the business rules. It's a
123
00:11:05.799 --> 00:11:11.000
classic trolley problem, right, and
the Dolly problem was more of a thought
124
00:11:11.039 --> 00:11:13.879
experiment back in the day. But
now when you actually have to program it
125
00:11:13.919 --> 00:11:20.519
in for your literal trolley self driving
cars, that you have a hold host
126
00:11:20.600 --> 00:11:26.080
of ethics, it's strange to think
that the ethics questions might be what delays
127
00:11:26.120 --> 00:11:33.039
it further than the technology issues that
it has. Christ let me ask you
128
00:11:33.320 --> 00:11:37.720
about the automation life cycle. When
do you think in the automation life cycle,
129
00:11:39.159 --> 00:11:45.279
ethics needs to be taken into consideration? Honestly? Right at the outset,
130
00:11:45.559 --> 00:11:50.200
ideally something you see is kind of
a bit more commonplace again and regulated
131
00:11:50.320 --> 00:11:54.919
industries. Everybody has to take some
kind of compliance course that you understand different
132
00:11:54.960 --> 00:12:00.600
ethical risk and legal risk. I
know that's a little bit of a novel
133
00:12:00.840 --> 00:12:03.639
way of thinking about automation and a
on it's something that's been rattling in my
134
00:12:03.679 --> 00:12:09.039
head honestly a little bit, my
non robotic kid twi O. There.
135
00:12:09.600 --> 00:12:13.919
Ideally, when you're getting educated on
the potential for leveraging automation and AI as
136
00:12:13.960 --> 00:12:18.960
a way to improve a task word
process, there should be some kind of
137
00:12:18.000 --> 00:12:24.600
training that people read through at the
very least be aware of organizations, policies
138
00:12:24.639 --> 00:12:28.440
and procedures which typically align with some
regulatory regime depending on what part of the
139
00:12:28.440 --> 00:12:33.879
world you're in, but hopefully something
a little bit more rost robust than that,
140
00:12:33.039 --> 00:12:37.360
so that you can avoid kind of
those unknown so to speak, and
141
00:12:39.159 --> 00:12:43.399
not hit an ethical ship wire of
some kind, especially when you're working with
142
00:12:43.440 --> 00:12:50.519
anything mutchy people, whether it's empathies
and really with customer data. There certainly
143
00:12:50.519 --> 00:12:56.000
should be like a milestone in a
project review or approval process or even to
144
00:12:56.120 --> 00:12:58.799
go live just to kind of check
off once again to make sure that it's
145
00:12:58.840 --> 00:13:03.639
a kay. Do you even collect
the data to do something like that in
146
00:13:03.679 --> 00:13:07.759
the first place? Part of the
recipes We all know of automation that you
147
00:13:07.799 --> 00:13:11.519
need to have some data and formulas, and people tend to focus on the
148
00:13:11.600 --> 00:13:15.919
formulas right at the outside. Or
should you even be collecting that data in
149
00:13:15.960 --> 00:13:20.039
the first place could be an ethic
culture boire, let alone the the usage
150
00:13:20.039 --> 00:13:24.840
of that data in a biased formula. Some of these things could be pretty
151
00:13:24.879 --> 00:13:30.039
simple checkerles like oh yeah, like
in my paper skin an example, this
152
00:13:30.120 --> 00:13:33.240
looks pretty trivial. This should be
approved and I can sign off. But
153
00:13:33.360 --> 00:13:37.279
for something else, you might need
some further review as part of the process
154
00:13:37.360 --> 00:13:43.519
before you can even start running your
playbook and kicking off the automation project,
155
00:13:45.240 --> 00:13:50.000
whether you're doing a traditional way or
in some agile fashion. Lou, I
156
00:13:50.000 --> 00:13:54.519
want to return to you earlier.
You briefly mentioned a couple of steps do
157
00:13:54.600 --> 00:14:00.320
you think companies can take to make
sure automation is done ethically? Could you
158
00:14:00.360 --> 00:14:05.000
elaborate on those a bit more so. The first is to make sure that
159
00:14:05.519 --> 00:14:13.240
whenever you're using a machine learning engine, it's been built to avoid bias and
160
00:14:13.360 --> 00:14:18.200
to be ethical in its own right. For example, that means, if
161
00:14:18.240 --> 00:14:24.799
you're building out a machine learning engine, think about what sort of data you're
162
00:14:24.879 --> 00:14:31.720
using to train it. This is
where you see a lot of AI implementations
163
00:14:31.759 --> 00:14:37.559
fail because a lot of the available
data folks will go for it's typically more
164
00:14:37.639 --> 00:14:41.559
likely to be from white males,
even if you're doing something as simple as
165
00:14:41.559 --> 00:14:46.960
training handwriting, you're more likely to
get samples from white males and you are
166
00:14:48.600 --> 00:14:54.240
from females or from minorities. And
as you're setting up these pieces, the
167
00:14:54.360 --> 00:14:58.600
data you're using to train it is
going to determine the outcomes it can use.
168
00:15:00.360 --> 00:15:03.480
Similarly, you know when you mentioned
the system the court was using to
169
00:15:03.600 --> 00:15:11.320
predict offenders, that is based on
data that already had some of those inherent
170
00:15:11.360 --> 00:15:15.879
biases in it. When you're looking
at historical data, especially the farther back
171
00:15:15.919 --> 00:15:20.440
you go, the more likely you
are to see those sorts of concerns.
172
00:15:20.440 --> 00:15:24.320
So the first piece is to make
sure that the data you're using and that
173
00:15:24.399 --> 00:15:28.879
corpus of knowledge is what you want
it to be and there isn't bias in
174
00:15:28.919 --> 00:15:35.720
it. The next piece is to
make sure that you have the auditibility required.
175
00:15:35.679 --> 00:15:41.039
So depending on what algorithm you have, you should have a methodology for
176
00:15:43.320 --> 00:15:48.440
looking back and being able to give
an explanation for how every decision was made.
177
00:15:48.159 --> 00:15:52.039
That's not just to maintain the ethics
portion, but also just to protect
178
00:15:52.120 --> 00:15:58.440
yourself and your business for when you're
implementing these sorts of solutions. Thank you,
179
00:15:58.559 --> 00:16:03.440
lou ris Let. Me give you
a chance to weigh and on how
180
00:16:03.480 --> 00:16:10.279
companies can create the transparency that's needed
to reduce ethical risk and bias in AI
181
00:16:11.559 --> 00:16:15.879
well, especially when something is of
a sensitive nature. Of the transparency would
182
00:16:15.919 --> 00:16:19.720
certainly be driven from the top.
As we mentioned our top of the US
183
00:16:19.759 --> 00:16:25.159
example, it just came out in
the press. The AI bill of rights.
184
00:16:25.559 --> 00:16:30.080
Organizations typically stopped start at the very
top with some sort of a blueprint,
185
00:16:30.960 --> 00:16:34.879
some kind of framework, potentially a
company framework might be some sort of
186
00:16:34.919 --> 00:16:40.200
maturity model. That's something we typically
like to drive as part of the work
187
00:16:40.200 --> 00:16:44.320
that I've done in management consultancy and
elsewhere, out of a framework from some
188
00:16:44.399 --> 00:16:48.480
kind of maturity model where you might
look at okay, ethically immature, too
189
00:16:48.600 --> 00:16:52.120
immature, and then an alignment with
that. Of course, again I refer
190
00:16:52.200 --> 00:16:59.159
often to policies and protagers that aligned
not only with regulatory and compliance rules depending
191
00:16:59.200 --> 00:17:00.840
on the part of the world or
in, but also privacy rules, which
192
00:17:00.840 --> 00:17:06.759
can be a little problematic in some
parts of the world. Privacy as very
193
00:17:06.799 --> 00:17:10.920
straightforward as we see in Europe.
In the United States, there's still a
194
00:17:10.960 --> 00:17:15.440
bit of a patchwork of privacy regulations
with it seems like once in a while
195
00:17:15.519 --> 00:17:18.200
you'll read in the press another U
S state has come up where it's coming
196
00:17:18.240 --> 00:17:22.440
up with it don't take on privacy. So this can create a bit of
197
00:17:22.480 --> 00:17:26.960
a ladline for people to walk through, so speak especially with ethical concerns here.
198
00:17:27.799 --> 00:17:32.680
But ideally you have to write subject
matter experts in place, some people
199
00:17:32.759 --> 00:17:38.359
with legal expertise, obviously ethics or
compliance or regulatory expertise. It's very rare
200
00:17:38.400 --> 00:17:45.079
that you hear of organizations having professional
ethicist or philosophers even that do this sort
201
00:17:45.079 --> 00:17:48.039
of work. So typically you see
people out of a compliance space that can
202
00:17:48.039 --> 00:17:52.960
help with guidelines and then these help
you in turn create the operating models that
203
00:17:52.359 --> 00:17:56.759
from that you get playbooks to help
create something. You run books for after
204
00:17:56.920 --> 00:18:00.559
something is created, to keep it
up and running. Part and parcel with
205
00:18:00.599 --> 00:18:03.640
all this a little bit more bottom
up, I mentioned a little bit earlier
206
00:18:03.680 --> 00:18:07.559
the idea of education in the workplace. You see this a lot more in
207
00:18:07.559 --> 00:18:12.599
irregulated industries, but we're starting to
see the idea of AI literacy starting to
208
00:18:12.640 --> 00:18:18.240
be discussed. And depending on where
you see AI sitting in an organization,
209
00:18:18.440 --> 00:18:23.240
it could be its own literacy program
or that may flow into a data literacy
210
00:18:23.319 --> 00:18:27.039
program. The idea that AI should
be part of the remit of the data
211
00:18:27.079 --> 00:18:32.119
analytics team because there are the people
that work with data knows that, people
212
00:18:32.119 --> 00:18:37.079
that do models, and typically there
are people working most closely with your data
213
00:18:37.160 --> 00:18:41.960
scientists. They can provide some guidance
to the people who may not be data
214
00:18:41.000 --> 00:18:45.359
science but might be doing more simpler
forms of automation like you see in a
215
00:18:45.400 --> 00:18:49.680
process automation department. If you're not
in a regulated industry, there's can be
216
00:18:49.720 --> 00:18:52.119
a tach with good models to look
at, say, to look at what
217
00:18:52.160 --> 00:18:56.960
people are doing in financial services,
are healthcare, entertain utilities or ease in
218
00:18:57.000 --> 00:19:03.200
government? Well, definitely don't want
to ignore the AI Bill of Rights that
219
00:19:03.559 --> 00:19:07.359
the US government recently released. Loue, what do you think about it?
220
00:19:07.519 --> 00:19:15.039
Specifically the five principles they defined for
organizations to take into consideration. So first
221
00:19:15.440 --> 00:19:22.000
let's start by reiterating for listeners what
those five principles are. The first one
222
00:19:22.319 --> 00:19:30.000
is safe and effective systems, So
any system that being used should be effective
223
00:19:30.160 --> 00:19:37.319
and safe, war folks, pretty
self explanatory. Next, algorithmic discrimination protections.
224
00:19:37.359 --> 00:19:40.279
This is some of what we were
talking about earlier, making sure that
225
00:19:40.599 --> 00:19:48.200
any algorithm being used isn't discriminatory and
follows standard laws and regulations. Data privacy
226
00:19:48.480 --> 00:19:52.640
is the third one, which is
get a crucial component of any form of
227
00:19:52.960 --> 00:19:59.400
data enterprise. Fourth, we have
notice an explanation, so if your data
228
00:19:59.519 --> 00:20:06.839
is being used in machine learning model, you should have transparency to understand how
229
00:20:06.880 --> 00:20:11.960
and why it's being used. And
finally we have human alternatives consideration and fallback.
230
00:20:12.400 --> 00:20:17.240
I should point out a couple of
things about these. First, A
231
00:20:17.279 --> 00:20:22.240
lot of these are just basic common
sense. It makes sense not just for
232
00:20:22.319 --> 00:20:26.200
the consumer but also for the business
themselves to protect themselves. Right. No,
233
00:20:26.319 --> 00:20:32.000
but no business wants to go show
up in the news for having a
234
00:20:32.119 --> 00:20:37.559
racist Twitter bot. That's terrible press. At the same point, looking at
235
00:20:37.599 --> 00:20:41.880
human alternatives necessary you call into a
contact center, folks get frustrated. They
236
00:20:41.920 --> 00:20:45.680
don't want to be talking to a
robot anymore. You have to be able
237
00:20:45.720 --> 00:20:49.799
to press the road to talk to
a human that. These aren't anything revolutionary.
238
00:20:49.880 --> 00:20:55.559
This is the general things at businesses
for the most part, are already
239
00:20:55.599 --> 00:21:00.640
doing. But aside from that,
a lot of this is echoing a very
240
00:21:00.680 --> 00:21:04.960
similar piece of actual legislation out of
the EU called the AI Act, which
241
00:21:06.079 --> 00:21:11.559
puts some teeth behind this and until
we see some of that in the United
242
00:21:11.559 --> 00:21:18.599
States. Remember that this is just
die blinds for folks to use. So
243
00:21:18.559 --> 00:21:22.759
with all that in mind, I'd
say that a lot of this is general
244
00:21:22.839 --> 00:21:27.960
best practice that most folks should be
following. Anyways. One piece i'd say
245
00:21:29.000 --> 00:21:34.359
in here that I see as potentially
changing the way things are done today is
246
00:21:34.440 --> 00:21:41.240
the notice an explanation component because a
lot of what data has used today you
247
00:21:41.279 --> 00:21:45.359
don't always get notice, and that
again is going to depend on where you
248
00:21:45.440 --> 00:21:52.519
live and what regional privacy laws there
are. But even then it's most likely
249
00:21:52.599 --> 00:21:56.559
not going to make a huge change
in a consumer's life. There'll just be
250
00:21:56.640 --> 00:22:00.240
one more agreement you check at the
beginning, if you know what mean.
251
00:22:00.240 --> 00:22:03.440
Yeah, I would agree with Dr
Lew there as I was reading through this
252
00:22:03.519 --> 00:22:08.599
Bill of Rights myself and I just
thought about gul everybody. I'm sure it
253
00:22:08.720 --> 00:22:12.079
might not be along on this podcast, but I have some young folks in
254
00:22:12.160 --> 00:22:17.720
my family and I see them on
TikTok. Nothing in the TikTok experience,
255
00:22:17.759 --> 00:22:22.720
at least from my recollection of just
looking at it just briefly to something pop
256
00:22:22.799 --> 00:22:27.279
up and saying we're powered by automated
system and some of the content you are
257
00:22:27.400 --> 00:22:33.279
seeing might be automatically suggested to you
based off of data and imparted form by
258
00:22:33.359 --> 00:22:41.039
what you're looking at. We're looking
for TikTok. TikTok's becoming an interesting topic
259
00:22:41.079 --> 00:22:47.799
of conversation in the marketing space because
there's a recent studies showing that the younger
260
00:22:47.839 --> 00:22:51.240
generations, even the youngest generations,
the one that they are just old enough
261
00:22:51.279 --> 00:22:55.799
to have some kind of bodied power, are starting to use TikTok as much
262
00:22:55.920 --> 00:23:00.839
as as Google to People typically think
of Google by default because they're the primary
263
00:23:00.880 --> 00:23:04.720
search engine, but now you see
people searching for things in other forms as
264
00:23:04.759 --> 00:23:10.119
well, and it might not be
aware of that if you're using Amazon as
265
00:23:10.119 --> 00:23:15.279
your primary way to search for products, unless you read the disclaimers down there
266
00:23:15.319 --> 00:23:18.359
at the bottom of the footter.
If something doesn't automatically pop up in front
267
00:23:18.400 --> 00:23:26.119
of you and tell you that you're
using an automated system and these other caveats,
268
00:23:26.240 --> 00:23:27.960
it could be it could be interesting, I'll put it that way.
269
00:23:29.079 --> 00:23:33.039
And on the other side is the
usability factor part of why so many people
270
00:23:33.079 --> 00:23:38.400
have drawn to digital experiences versus the
human alternatives. Will say, paper and
271
00:23:38.440 --> 00:23:42.359
pen and things like that are just
calling somebody up on the phone is the
272
00:23:42.440 --> 00:23:47.200
friction of the digital experience. And
you see some of these stories in the
273
00:23:47.240 --> 00:23:52.720
last couple of years from overseas with
the stricter especially in Western Europe, the
274
00:23:52.799 --> 00:23:59.680
stricter privacy rules, some people saying
that they're practically making certain websites unusable.
275
00:24:00.519 --> 00:24:04.000
So there's going to be I think
a little bit of trial and error tests
276
00:24:04.039 --> 00:24:08.640
and learn to kind of navigate what
this blueprint actually implies, at least for
277
00:24:08.640 --> 00:24:15.279
the United States. Yeah, let
me ask you. Some have called for
278
00:24:15.359 --> 00:24:23.200
an ethics framework as the foundation upon
which any AI technology is created and implemented.
279
00:24:23.279 --> 00:24:27.759
What are your thoughts on this.
I think it makes sense depending on
280
00:24:27.960 --> 00:24:32.640
what sort of framework you put in
place, and I think any regulation like
281
00:24:32.680 --> 00:24:37.680
this is going to have to come
from a government level. Now, if
282
00:24:37.720 --> 00:24:41.880
you think about the basics that we
were discussing earlier, like the AI Bill
283
00:24:41.920 --> 00:24:48.440
of Rights. Again, I don't
see anyone complaining about this, but putting
284
00:24:48.480 --> 00:24:52.960
it into a codified form means that
your compliance folks, who, again,
285
00:24:52.000 --> 00:24:57.599
as Chris said, are generally going
to be the folks looking at ethics within
286
00:24:57.720 --> 00:25:06.160
an enterprise, and most of them
are not hiring philosophers have something to benchmark
287
00:25:06.279 --> 00:25:11.119
against. So in general, it's
gonna make a lot of sense because it
288
00:25:11.160 --> 00:25:17.079
gives companies that the backbone to look
at when they're trying to make these sorts
289
00:25:17.079 --> 00:25:23.759
of decisions. One fairly recent example
of an ethics framework that I read about
290
00:25:25.039 --> 00:25:27.039
I wanted to bring it up and
Chris, I'll direct this one to you.
291
00:25:29.039 --> 00:25:34.839
In Google announced that it had created
a council to develop an ethics framework,
292
00:25:34.920 --> 00:25:42.359
primarily to address AI powered facial recognition
and fairness in machine learning. However,
293
00:25:42.480 --> 00:25:48.720
within ten days of its establishment,
the council was dissolved due to suspected
294
00:25:48.759 --> 00:25:55.160
bias amongst one of the council members. It raises the question, how can
295
00:25:55.240 --> 00:26:00.359
people most or you could even say
all of whom some sort of biases.
296
00:26:00.880 --> 00:26:07.359
How can they effectively play a role
in developing a framework to avoid biases in
297
00:26:07.440 --> 00:26:11.440
AI? And maybe the toughest question
we've had today, So sorry for directing
298
00:26:11.440 --> 00:26:17.200
it your way. That's okay,
because I don't think it's too hard.
299
00:26:17.319 --> 00:26:21.559
I mean, again, we mentioned
this idea of of a blueprint at a
300
00:26:21.640 --> 00:26:27.400
high government level regulations and compliance.
They serve as guard rails. From my
301
00:26:27.480 --> 00:26:32.240
vantage point, you would kind of
look at the creation of such a framework
302
00:26:33.000 --> 00:26:40.799
using the classical model of people processing
technology and people might even have like these
303
00:26:40.880 --> 00:26:45.839
higher level considerations as within you know, the notion of B s G and
304
00:26:45.880 --> 00:26:52.480
the particular process. Certain processes,
depending on whether or not there across industry,
305
00:26:52.519 --> 00:26:57.960
like a marketing process might have a
little bit of additional ethical review to
306
00:26:59.079 --> 00:27:03.680
it because they're that notion of targeting
people automatically, they might not be aware
307
00:27:03.720 --> 00:27:07.400
of it. Do you need to
present some kind of notice that's happening,
308
00:27:07.960 --> 00:27:11.079
whether it's with an AD or is
you mentioned a offer force and kind of
309
00:27:11.079 --> 00:27:15.440
say a lending product or maybe some
other product. When you look at something
310
00:27:15.519 --> 00:27:19.160
on the shopping website and you go
into your Facebook and you get an AD
311
00:27:19.200 --> 00:27:25.400
for within minutes and sometimes it's mind
blowing the speed and how fast that,
312
00:27:25.440 --> 00:27:29.400
I mean, you can kind of
creep you out. Creating a framework again,
313
00:27:29.720 --> 00:27:33.680
certainly you'd have some initial the consultative
group to get together legal experts,
314
00:27:33.680 --> 00:27:41.240
compliance experts, subject matter experts in
different functions to start waiting through and seeing
315
00:27:41.440 --> 00:27:47.720
does this that within an existing regulatory
and compliance our seam and a risk calculator
316
00:27:47.759 --> 00:27:49.960
that we have at the enterprise level, or do we need to create or
317
00:27:51.000 --> 00:27:56.319
modify something new, right, And
just to add on there, I'd say
318
00:27:56.359 --> 00:28:00.880
this is actually probably one of the
simplest questions you've asked, Brad, because
319
00:28:00.559 --> 00:28:06.200
we're not reinventing the wheel here.
This isn't the first time we've needed to
320
00:28:06.960 --> 00:28:14.039
put a group of people together to
figure out how to enforce basic ethics.
321
00:28:14.559 --> 00:28:18.119
If you think about pretty much any
HR policy, a lot of the same
322
00:28:18.680 --> 00:28:23.839
factors were brought into play. So
you need a diverse group of individuals forming
323
00:28:23.920 --> 00:28:27.119
a committee to figure this out.
This is nothing new, We're just you're
324
00:28:27.160 --> 00:28:33.680
doing it for a new purpose.
Very good point, Mr Leeu Blue.
325
00:28:33.759 --> 00:28:37.160
Let me ask you about GPT three. As you know, r P A
326
00:28:37.279 --> 00:28:44.559
combined with natural language processing skills has
been playing a very big role in areas
327
00:28:44.720 --> 00:28:51.839
like contact centers as automation has become
integral in in that function. GPT three
328
00:28:52.119 --> 00:28:57.160
is one of the largest NLP language
models in use and widely considered a breakthrough
329
00:28:57.279 --> 00:29:03.079
in AI. It can generate sentences
and even write article summaries or generate full
330
00:29:03.119 --> 00:29:07.079
stories that are creative in nature just
based on a prompt of a few lines.
331
00:29:07.400 --> 00:29:14.799
Pretty amazing stuff. However, even
GPT three has proven in in recent
332
00:29:14.920 --> 00:29:21.920
times to generate problematic content if it's
fed with harmful or malicious prompts. So
333
00:29:22.000 --> 00:29:26.880
the question for you, Lou,
it's just a situation where AI algorithms will
334
00:29:26.920 --> 00:29:32.839
have to be written to regulate AI
algorithms. That's a that's a tricky question
335
00:29:32.960 --> 00:29:37.680
because if you think about the GPT
three example you just gave, it was
336
00:29:38.079 --> 00:29:47.799
generating offensive text when given of offensive
prompt. So the the phrasing here is
337
00:29:47.960 --> 00:29:52.119
we don't give it an offensive prompt, and some of that is just going
338
00:29:52.160 --> 00:29:56.559
to be basic what I like to
call a potty mouth filter. And I
339
00:29:56.559 --> 00:30:00.759
actually didn't come up with that term. That was from IBM way back when
340
00:30:02.640 --> 00:30:07.640
Watson was on Jeopardy. They are
doing a training session and they had the
341
00:30:07.680 --> 00:30:14.440
executive producers of Jeopardy in the room
and the category was four letter words,
342
00:30:14.640 --> 00:30:19.480
and Watson dropped the F bomb and
they're like, well, it's a good
343
00:30:19.480 --> 00:30:25.519
thing. This isn't on national television. This is a practice zoo need to
344
00:30:25.559 --> 00:30:29.880
fix that, so they put in
a potty mouth filter. I think some
345
00:30:29.920 --> 00:30:33.039
of the same things just need to
go into place here and not necessarily on
346
00:30:33.079 --> 00:30:37.200
the output, but sometimes looking at
the input itself. The same way where
347
00:30:37.880 --> 00:30:44.839
you have safe search options for any
other category if you're doing things, the
348
00:30:44.920 --> 00:30:51.319
same principles apply here where you want
to make sure that any use of machine
349
00:30:51.359 --> 00:30:56.039
learning or AI is going to be
safe and is going to be efficient.
350
00:30:57.240 --> 00:31:04.079
Gotcha, gotcha, Chris. As
we automate more and more business processes,
351
00:31:04.079 --> 00:31:11.519
will ethical concerns continue to increase?
In your opinion, I think it depends
352
00:31:11.759 --> 00:31:18.920
on the application of automation within particular
functional areas and businesses. You think about.
353
00:31:19.119 --> 00:31:22.680
Once upon a time, human resources, for example, this wasn't too
354
00:31:22.759 --> 00:31:29.960
much of a concern because human resources
and through you had humans reading resumes,
355
00:31:30.000 --> 00:31:33.680
and you had humans doing a lot
of work with personal data to set up
356
00:31:33.720 --> 00:31:37.920
like your benefits and your training and
all the learning and development different things.
357
00:31:37.039 --> 00:31:42.279
But now you see the challenge of
ethics coming and play the the example with
358
00:31:42.759 --> 00:31:48.240
automated silvering of resumes based off of
gender being potentially sticky. I've heard some
359
00:31:48.359 --> 00:31:53.240
things like that have happened in the
past on a more manual basis, where
360
00:31:53.279 --> 00:31:59.400
people's first names could be filtered out
by people. And you see some of
361
00:31:59.440 --> 00:32:01.079
these studies lolat around on the internet
once in a while in the last year
362
00:32:01.119 --> 00:32:07.119
or two, where there's different bias
just in this selection of candidate. So
363
00:32:07.279 --> 00:32:10.319
that's an area we see ethics being
a concern in the automation space too.
364
00:32:10.680 --> 00:32:15.680
There's the kind of the rising interest
in this topic of process mining and process
365
00:32:15.720 --> 00:32:20.359
analytics. You already seen a little
bit of that tripping some ethical trip wire
366
00:32:20.519 --> 00:32:23.640
so speak. I believe there are
certain parts of Europe that are more regulated
367
00:32:23.680 --> 00:32:28.279
than others, and I already braced
that as a flag to be addressed that
368
00:32:28.400 --> 00:32:32.000
you need people to opt in to
have what they're doing at work tracked.
369
00:32:32.359 --> 00:32:37.160
Where in the US, it's just
assume that you signed off a part of
370
00:32:37.160 --> 00:32:40.759
your employment agreement, and if you're
walking around the hallways or in anabilits,
371
00:32:40.839 --> 00:32:45.680
or you're using certain computer systems or
anything you're on your screen or your voicemail,
372
00:32:45.359 --> 00:32:50.680
those things could potentially be fodder for
some kind of process mining and process
373
00:32:50.759 --> 00:32:55.160
analysis exercise. Certainly couldn't be much
more sensitivity when you add the additional layers
374
00:32:55.359 --> 00:33:01.440
of financial services and healthcare and energy
thing or things that are sensitive at a
375
00:33:01.440 --> 00:33:07.839
more macro level, such a government
whatsoever, think about it like this brand
376
00:33:07.119 --> 00:33:14.640
all these processes whenever you're automating an
existing process. The ethics concerns were already
377
00:33:14.880 --> 00:33:19.359
thought about. It's just a matter
of who's completing it. Is it a
378
00:33:19.400 --> 00:33:23.160
person or is it an automation.
And I would argue that the most part,
379
00:33:23.480 --> 00:33:28.720
you have a lot more transparency with
an automation than you do with a
380
00:33:28.799 --> 00:33:34.839
human being because of the auditability.
That's exactly, you don't know if the
381
00:33:35.480 --> 00:33:39.839
banker you happen to be talking to
has any form of bias, and there's
382
00:33:39.880 --> 00:33:45.200
no way to track how those decisions
were made, whereas with an automation you
383
00:33:45.359 --> 00:33:52.359
have a deeper level of insight and
auditability. Right, good point. So,
384
00:33:52.480 --> 00:33:55.680
Louf, you've been on the podcast
in the past to discuss the next
385
00:33:55.720 --> 00:34:01.799
phase of automation, one that is
likely to be much more autonomous. How
386
00:34:01.839 --> 00:34:08.400
does this present even more ethical challenges? There's a decent amount of them,
387
00:34:08.440 --> 00:34:13.239
I'll be honest, because if you
remember the last time I was on,
388
00:34:13.320 --> 00:34:19.400
we talked about how the ultimate goal
of automation was too automate as much of
389
00:34:19.440 --> 00:34:22.599
it as possible, with the eventual
goal of the world. Where you were
390
00:34:22.639 --> 00:34:28.199
doing something that was repetitive, the
CEE could get a pop up that says,
391
00:34:28.199 --> 00:34:30.719
hey, it looks like Rad's doing
a repetitive task. Would you like
392
00:34:30.760 --> 00:34:35.079
to automate, it clicks okay,
and it's done. Now. When we
393
00:34:35.119 --> 00:34:37.760
first talked about that, it was
it seemed like kind of a pipe dream.
394
00:34:37.800 --> 00:34:43.440
What with the technology we've put in
place since then, we have automated
395
00:34:43.559 --> 00:34:50.760
enough to shorten development cycles by anywhere
from so we're a decent ways down that
396
00:34:50.800 --> 00:34:57.119
path. But a key portion of
that is task mining, which uses desktop
397
00:34:57.199 --> 00:35:02.760
based recorders to monitor what human employ
these are doing and separate out multitasking and
398
00:35:02.800 --> 00:35:09.199
digitize that into a model right.
Doing that involves a lot of the privacy
399
00:35:09.239 --> 00:35:14.880
concerns that Chris was just mentioning.
You need to make sure that when you're
400
00:35:15.000 --> 00:35:22.159
monitoring folks, you're monitoring them in
a controlled and governed fashion. So at
401
00:35:22.320 --> 00:35:24.159
S, S and C Blueprism,
the way we do this is by applying
402
00:35:24.199 --> 00:35:30.679
a whole bunch of privacy controls.
Sometimes the individual users have the ability to
403
00:35:30.719 --> 00:35:36.760
start or stop the recording functionality.
You could set it up to obscure sensitive
404
00:35:36.840 --> 00:35:43.559
data, redact the I I,
or also only look at specific applications or
405
00:35:43.719 --> 00:35:46.880
exclude applications, so you can say
I only want to record what they're doing
406
00:35:46.920 --> 00:35:53.760
in these five specific applications or work, or you could say I want to
407
00:35:53.800 --> 00:36:00.400
exclude anything from Facebook dot Com because
that clearly won't be relevant. But the
408
00:36:00.440 --> 00:36:06.159
fact is you have to have consideration
for pieces like that and for the privacy
409
00:36:06.159 --> 00:36:10.119
of not just workers, but the
customers will be having their data run through
410
00:36:10.159 --> 00:36:16.719
these systems. Any final thoughts today
before we bring our podcast to conclusion.
411
00:36:19.119 --> 00:36:22.079
So one of the things folks often
don't think about, and this is an
412
00:36:22.119 --> 00:36:28.000
area that was addressed in the Journal
of the Aerobotics and Workplace Automation recently,
413
00:36:28.840 --> 00:36:37.599
is the ability for automation to improve
upon the existing ethics that are in our
414
00:36:37.639 --> 00:36:44.679
workplace. For example, automations can
do things like monitor performance refused to any
415
00:36:44.719 --> 00:36:52.039
instances or bias, or help with
recruiting and give that layer of insight that
416
00:36:52.119 --> 00:36:58.840
otherwise we wouldn't have. So in
general, I think the application of AI,
417
00:36:59.000 --> 00:37:04.920
machine learning, and anomation to the
workplace will improve the ethics that are
418
00:37:04.960 --> 00:37:09.559
being maintained rather than complicate them.
I would second that it's certainly one of
419
00:37:09.599 --> 00:37:16.519
the One of the positives of automation
is that it improves quality, reduces errors,
420
00:37:17.519 --> 00:37:22.239
and if you have these ethical oversight
boards just to make sure that things
421
00:37:22.239 --> 00:37:25.480
do go off the rails in the
wrong direction in the long run, when
422
00:37:25.679 --> 00:37:30.559
you're ensuring that people and or divibul
workers were both are doing what they're supposed
423
00:37:30.599 --> 00:37:35.360
to be doing, and you have
some automated checks in there, you will
424
00:37:35.639 --> 00:37:39.639
reduce risk. So and you're seeing
some of that, especially in financial services
425
00:37:39.679 --> 00:37:45.639
where there's large scale applications of pilots
going on with automated scanning of call centered
426
00:37:45.639 --> 00:37:52.039
transcripts and other forms of communications,
and make sure that the traders aren't communicating
427
00:37:52.039 --> 00:37:54.400
with each other in forms that they're
not supposed to, which I know is
428
00:37:54.440 --> 00:37:59.639
something that's happened in some of the
European financial markets not too long ago.
429
00:38:00.960 --> 00:38:05.159
Thank you both for ending us on
a positive note on the ethics topic.
430
00:38:05.400 --> 00:38:10.039
As I mentioned at the beginning,
this conversation about ethics and automation and AI
431
00:38:10.159 --> 00:38:15.760
is an important one to have,
but it should not reflect poorly on either
432
00:38:15.840 --> 00:38:20.679
of these areas. They're doing a
ton of good in the world as we
433
00:38:20.760 --> 00:38:23.039
know well. Gentlemen, I want
to thank you both for being on the
434
00:38:23.079 --> 00:38:29.400
podcast today. It's been a fascinating
conversation. I've certainly learned a lot just
435
00:38:29.519 --> 00:38:32.639
listening to both of you, so
thank you for sharing your perspective. You
436
00:38:32.679 --> 00:38:36.880
guys have a lot to offer on
this topic, and I know it's going
437
00:38:36.920 --> 00:38:40.760
to be informative for our listeners.
In the history of the podcast, I
438
00:38:40.800 --> 00:38:45.519
believe this is the first time we've
ever talked about ethics in intelligent automation.
439
00:38:46.079 --> 00:38:52.519
So thank you for helping us cover
it sufficiently. Brad, appreciate the time
440
00:38:52.559 --> 00:38:58.039
today. Always happy to you jump
onto these. I think the future is
441
00:38:58.480 --> 00:39:04.719
exciting and application of machine learning and
automation will just continue to make our lives
442
00:39:04.880 --> 00:39:08.960
better and easier. I echo with
Dr Luke said, thank you right,
443
00:39:09.039 --> 00:39:13.480
not only on the behalf of myself, but I'll speak up. I'll behalf
444
00:39:13.519 --> 00:39:17.519
of Henry's publications. It was a
lot of fun pioneering the issue on this
445
00:39:17.599 --> 00:39:21.960
topic. There are certainly some books
coming out here and there in the press
446
00:39:22.639 --> 00:39:27.239
lately that are starting to delve into
the topics of ethical machines and ethics and
447
00:39:27.320 --> 00:39:32.000
AI and gaining away chrom maybe the
fearmongerine that goes way back to Isaac Eisamouth
448
00:39:32.079 --> 00:39:37.679
and I robot and science fiction and
the scary robots and more towards the positive
449
00:39:37.719 --> 00:39:42.800
aspects of your rewards off automation and
the I all the best tea boat,
450
00:39:43.039 --> 00:39:46.559
gentlemen. Thanks for tuning in to
Transform Now. For more insightful discussions on
451
00:39:46.639 --> 00:39:51.639
digital transformation and more, check out
our podcast channel, will you'll find all
452
00:39:51.679 --> 00:39:54.760
of our previous episodes and to make
sure you never missed an episode, subscribe
453
00:39:54.800 --> 00:39:58.920
to the show in your favorite podcast
player, and if you like what you've
454
00:39:58.920 --> 00:40:02.480
heard, please leave was a review. For more information about digital transformation and
455
00:40:02.559 --> 00:40:07.519
the future of work, check out
blueprism dot com to learn how s s
456
00:40:07.559 --> 00:40:12.599
and see Blueprism's digital workforce is enabling
enterprise transformation now
