WEBVTT

00:00:00.000 --> 00:00:03.680
GenAI is a probability engine.
If you're not the probable answer,

00:00:03.680 --> 00:00:06.320
then you're probably not the possible
answer either.

00:00:07.680 --> 00:00:09.840
We're joined by Rebecca from The
Brandtech Group,

00:00:09.840 --> 00:00:10.600
welcome.

00:00:10.600 --> 00:00:12.000
Thank you so much. Rebecca.

00:00:12.320 --> 00:00:17.160
Bias in AI has been a problem for
marketing leaders especially ever since

00:00:17.160 --> 00:00:18.080
its inception.

00:00:18.600 --> 00:00:21.720
Why is it proving to be such a tough
problem to solve even now,

00:00:21.720 --> 00:00:22.560
so many years on?

00:00:23.520 --> 00:00:25.760
I think the challenge is the opacity.

00:00:25.800 --> 00:00:28.800
So, for most marketing leaders,
they're not building AI,

00:00:28.800 --> 00:00:30.000
they're leveraging AI.

00:00:30.720 --> 00:00:34.840
And so the people building AI,
the responsibility can sit with them in

00:00:34.840 --> 00:00:36.960
some way to kind of find a solution.

00:00:37.560 --> 00:00:40.160
And it's really baked into the very
fabric of the industry.

00:00:40.160 --> 00:00:45.120
So, you may not know this,
but almost 92% of AI software developers

00:00:45.120 --> 00:00:46.400
identify as male.

00:00:46.600 --> 00:00:49.640
Only 1% of investment went to female
founders in this space.

00:00:49.920 --> 00:00:53.360
So I just think if probability is not on
your side, you know,

00:00:53.360 --> 00:00:57.040
GenAI is a probability engine.
If you're not the probable answer,

00:00:57.040 --> 00:01:01.120
then you're probably not the possible
answer either because nobody who's

00:01:01.120 --> 00:01:05.640
building AI is focused on making kind of
fringe representation part of the core.

00:01:06.000 --> 00:01:07.080
And that's really the challenge.

00:01:07.320 --> 00:01:11.600
What sort of strategies are you seeing
brands or platforms deploy in order to

00:01:11.600 --> 00:01:13.120
try and mitigate that bias?

00:01:13.120 --> 00:01:16.920
What's working? At the moment,
I would say it's largely avoidance.

00:01:16.960 --> 00:01:20.520
So lots of brands we're seeing have a
strategy that says we won't generate

00:01:20.520 --> 00:01:24.480
synthetic humans or we will be very
careful about a representation of real people.

00:01:24.480 --> 00:01:26.360
They'll stick to low-risk use cases.

00:01:26.600 --> 00:01:27.800
That is really loosening now.

00:01:27.800 --> 00:01:31.320
I think the evolution of the frontier
video models means that you really have

00:01:31.320 --> 00:01:34.760
to find a new position that you're
comfortable with ethically for your kind

00:01:34.760 --> 00:01:36.000
of responsible AI practice.

00:01:36.600 --> 00:01:40.200
And then when it comes to then really
considering the inclusion of humans,

00:01:40.200 --> 00:01:42.520
then you need a technical level of
mitigation.

00:01:42.520 --> 00:01:44.440
You need something baked into the
technology.

00:01:44.440 --> 00:01:48.720
You can't just put that responsibility at
the door of prompters to always remember

00:01:48.720 --> 00:01:52.320
to be conscious prompting,
to have a full grasp of inclusivity every

00:01:52.320 --> 00:01:53.600
time they they generate.

00:01:54.080 --> 00:01:58.040
And what would your advice be to
marketing leaders who are trying to scale

00:01:58.040 --> 00:02:01.320
GenAI, but, you know,
trying to do it responsibly and yet not

00:02:01.320 --> 00:02:03.000
trying to slow down innovation?

00:02:03.520 --> 00:02:06.040
Yeah,
that's where I think you want to find a

00:02:06.040 --> 00:02:10.440
great technical mitigation and then
deploy it at scale inside of your bussiness.

00:02:10.640 --> 00:02:14.200
So the way that we've tried to kind of
overcome this challenge is we've

00:02:14.200 --> 00:02:16.080
prototyped a tool called Bias Breaker.

00:02:16.560 --> 00:02:20.320
So it's a level of technical mitigation
that sits on top of any image or video

00:02:20.320 --> 00:02:21.160
generation model.

00:02:21.360 --> 00:02:24.560
We configured . . .
starting with six dice for the six most

00:02:24.560 --> 00:02:28.920
common forms of inclusivity: gender, age,
ethnicity, religion, disability, et cetera.

00:02:29.680 --> 00:02:32.200
And when you enter a simple prompt,
we roll those dice,

00:02:32.200 --> 00:02:34.560
we give you up to two forms of
inclusivity back in,

00:02:34.560 --> 00:02:37.640
woven into your prompt--more
sophisticated prompt that you can then

00:02:37.640 --> 00:02:38.840
use in any of those models.

00:02:38.880 --> 00:02:43.240
And I think just being able to bring that
level of randomized representation

00:02:43.240 --> 00:02:47.720
without having to bring your own sort of
tokenistic or unconscious bias to the

00:02:47.720 --> 00:02:52.400
process is a really powerful way to bake
the solution into the technology itself.

00:02:52.720 --> 00:02:53.640
Hugely promising.

00:02:53.640 --> 00:02:56.000
Rebecca,
thanks so much for joining us here in Cannes.

00:02:56.000 --> 00:02:56.800
My absolute pleasure.