1
00:00:00,000 --> 00:00:09,960
Welcome to Artificially Intelligent Marketing, a weekly podcast where we stay on top of the

2
00:00:09,960 --> 00:00:15,700
latest trends, tips and tools in the world of marketing AI, helping you get the best

3
00:00:15,700 --> 00:00:18,520
results from your marketing efforts.

4
00:00:18,520 --> 00:00:22,440
Now let's join our hosts, Paul Avery and Martin Broadhurst.

5
00:00:22,440 --> 00:00:23,440
Hello everybody.

6
00:00:23,440 --> 00:00:27,800
Welcome to episode 29 of Artificially Intelligent Marketing.

7
00:00:27,800 --> 00:00:31,960
It's Paul Avery here, joined as always by my beautiful and wonderful co-host, Martin

8
00:00:31,960 --> 00:00:32,960
Broadhurst.

9
00:00:32,960 --> 00:00:33,960
Martin, how are you?

10
00:00:33,960 --> 00:00:36,800
Beautiful and wonderful, as ever.

11
00:00:36,800 --> 00:00:37,800
Thank you for noticing.

12
00:00:37,800 --> 00:00:39,720
Do you know what?

13
00:00:39,720 --> 00:00:42,320
A lot of people are going to listen to this on the audio and so I thought it was only

14
00:00:42,320 --> 00:00:47,720
right that they should know the wondrousness that they could experience if they decided

15
00:00:47,720 --> 00:00:49,080
to look at the video version.

16
00:00:49,080 --> 00:00:50,080
Yeah, yeah.

17
00:00:50,080 --> 00:00:55,000
Well, they can always get a photo of my face off the internet and print it out and put

18
00:00:55,000 --> 00:01:00,480
it in their wallet for just a quick glance when they feel like a bit of beauty in their

19
00:01:00,480 --> 00:01:01,480
life.

20
00:01:01,480 --> 00:01:05,640
I'll feel better because then it's not just me that has that in my wallet.

21
00:01:05,640 --> 00:01:12,840
So moving swiftly on, lots of wonderful stories from the world of AI and marketing this week

22
00:01:12,840 --> 00:01:19,360
and we have a bonus little interview at the end of today's episode as well.

23
00:01:19,360 --> 00:01:23,240
So without further ado, let's get into the first story, Martin.

24
00:01:23,240 --> 00:01:31,480
So Google has joined some of the other big tech players by offering legal cover for generative

25
00:01:31,480 --> 00:01:33,240
AI users.

26
00:01:33,240 --> 00:01:40,120
So as expectations of copyright disputes rises, Google Cloud is now taking steps to protect

27
00:01:40,120 --> 00:01:48,160
users against claims made if they use generative AI tools.

28
00:01:48,160 --> 00:01:53,600
So they are offering an indemnification policy against potential infringement claims.

29
00:01:53,600 --> 00:01:58,280
The promise covers two specific areas.

30
00:01:58,280 --> 00:02:06,360
So Google's own training data and any generated outputs such as text, images or audio created

31
00:02:06,360 --> 00:02:12,040
through services like MusicML or Google Workspace.

32
00:02:12,040 --> 00:02:18,240
Basically if you face any intellectual property lawsuits at all over content created with

33
00:02:18,240 --> 00:02:24,800
Google AI, Google AI is saying that they will take on the legal risk.

34
00:02:24,800 --> 00:02:30,160
So this addresses something of a significant pain point for many enterprises, something

35
00:02:30,160 --> 00:02:34,800
that has slowed enterprise adoption of generative AI models.

36
00:02:34,800 --> 00:02:41,080
No one wants to take on any unintended liability and Google is doing what it can to mitigate

37
00:02:41,080 --> 00:02:48,000
that and insists that this shows their shared fate with the customers.

38
00:02:48,000 --> 00:02:51,000
So the indemnity does depend on responsible use.

39
00:02:51,000 --> 00:02:54,640
It doesn't cover intentional infringement, which makes sense.

40
00:02:54,640 --> 00:03:00,760
You wouldn't expect Google to be covering people recklessly using the models.

41
00:03:00,760 --> 00:03:06,040
They themselves say that it provides a balanced practical coverage, which I'm sure they would

42
00:03:06,040 --> 00:03:11,000
say as opposed to it being unbalanced and highly impractical.

43
00:03:11,000 --> 00:03:15,760
That was the name of my second album as well, balanced practical coverage.

44
00:03:15,760 --> 00:03:17,800
It didn't do very well.

45
00:03:17,800 --> 00:03:25,960
Well yeah, so I mentioned at the start that they're joining big tech players such as Microsoft

46
00:03:25,960 --> 00:03:38,160
and Adobe who have also made similar pledges trying to spur innovation by easing IP concerns.

47
00:03:38,160 --> 00:03:40,200
And these risks certainly do remain.

48
00:03:40,200 --> 00:03:46,560
You see the likes of Sarah Silverman's lawsuit still going through the courts, many authors

49
00:03:46,560 --> 00:03:48,240
taking action against open AI.

50
00:03:48,240 --> 00:03:51,520
So we can expect more of this in the future and you can understand why companies would

51
00:03:51,520 --> 00:03:56,800
be a little bit wary about dipping their toe in the generative AI pond.

52
00:03:56,800 --> 00:03:59,240
Yeah, it's interesting.

53
00:03:59,240 --> 00:04:04,960
So we onboarded or in the process of onboarding a new client, signing all the contracts and

54
00:04:04,960 --> 00:04:07,720
et cetera that comes with that.

55
00:04:07,720 --> 00:04:14,680
And there is a clause been inserted into the contract about the use of LLMs effectively

56
00:04:14,680 --> 00:04:19,440
forbidding their use for any practical purpose for us to deliver what we're going to deliver

57
00:04:19,440 --> 00:04:20,600
as a marketing agency.

58
00:04:20,600 --> 00:04:27,840
That's the first time that I've seen that, but I feel it's very much in line with we

59
00:04:27,840 --> 00:04:30,320
can't afford to get in trouble down the line.

60
00:04:30,320 --> 00:04:35,640
I think it's a protectionism as it relates to things like copyright infringement versus

61
00:04:35,640 --> 00:04:41,600
a we want to make sure we're getting for value for money out of the agency aspect.

62
00:04:41,600 --> 00:04:43,920
Although I sort of understand that as well.

63
00:04:43,920 --> 00:04:48,680
I actually think it's driven by all of these concerns that the bigger companies have about

64
00:04:48,680 --> 00:04:52,240
what legal risks they could be opening themselves up to.

65
00:04:52,240 --> 00:04:57,280
Yeah, lots of companies approaching this whole space with an abundance of caution.

66
00:04:57,280 --> 00:04:58,720
Agreed.

67
00:04:58,720 --> 00:05:05,960
And I think until we have some clarification on what a misdemeanor looks like and how it's

68
00:05:05,960 --> 00:05:09,480
punished, we're not really going to know, are we?

69
00:05:09,480 --> 00:05:13,960
Even Adobe, Microsoft and all these others that are offering and Google now offering

70
00:05:13,960 --> 00:05:20,000
to cover you if you are pulled up for your use of LLMs, we still won't really know until

71
00:05:20,000 --> 00:05:21,000
it happens.

72
00:05:21,000 --> 00:05:22,000
Righty-o.

73
00:05:22,000 --> 00:05:24,280
So that's a good story.

74
00:05:24,280 --> 00:05:28,240
And of course, if you are a large brand, you're probably living and breathing this.

75
00:05:28,240 --> 00:05:34,280
If you're a small brand, maybe that gives you a bit of an opportunity to really leverage

76
00:05:34,280 --> 00:05:39,640
these tools without some of the more risk averse controls that are put in place quite

77
00:05:39,640 --> 00:05:41,560
understandably by large companies.

78
00:05:41,560 --> 00:05:47,120
But yeah, as a small company, maybe that creates a little window of opportunity.

79
00:05:47,120 --> 00:05:48,120
Good stuff.

80
00:05:48,120 --> 00:05:49,120
Thank you, Martin.

81
00:05:49,120 --> 00:05:54,720
Let's look at our second story this week, which is for Skyscanner, which is setting

82
00:05:54,720 --> 00:06:00,200
a new standard in customer centricity with its generative AI-powered discovery tool,

83
00:06:00,200 --> 00:06:05,920
which has been launched as a beta across Australia, India and Singapore.

84
00:06:05,920 --> 00:06:11,960
The tool, which is fueled by OpenAI's chat GPT, is called Dream and Discover with AI.

85
00:06:11,960 --> 00:06:15,480
That feels like a podcast episode waiting to happen.

86
00:06:15,480 --> 00:06:20,880
And it aims to meet a specific consumer need, in this case, travel inspiration.

87
00:06:20,880 --> 00:06:22,840
So what's interesting about this?

88
00:06:22,840 --> 00:06:27,800
Well, it's Skyscanner's sort of conscious approach to addressing the needs of customers

89
00:06:27,800 --> 00:06:31,560
through a sort of a more interactive chat style AI.

90
00:06:31,560 --> 00:06:36,240
So in essence, if it is a traveler, you can engage with the tool by asking open any questions

91
00:06:36,240 --> 00:06:42,080
like best cities for cultural tours or where should I take a city break if I'm a bit of

92
00:06:42,080 --> 00:06:43,320
a foodie.

93
00:06:43,320 --> 00:06:48,560
And then what the tool do is it will produce tailored travel ideas, of course, complete

94
00:06:48,560 --> 00:06:53,400
with optimal flight options for those getting the little commercial aspect in there.

95
00:06:53,400 --> 00:06:57,640
So it's not a pure customer support play here.

96
00:06:57,640 --> 00:07:01,920
There is that sort of sales aspect in there as well.

97
00:07:01,920 --> 00:07:06,280
Although of course, the goal here is to try and where appropriate, create a bit more of

98
00:07:06,280 --> 00:07:12,520
a friction free and fun experience for a customer to find somewhere interesting to go and make

99
00:07:12,520 --> 00:07:15,320
it easy for them to book.

100
00:07:15,320 --> 00:07:22,120
Piero Sierra, Skyscanner's chief product officer mentions that 56% of users come to Skyscanner

101
00:07:22,120 --> 00:07:23,920
for inspiration.

102
00:07:23,920 --> 00:07:28,080
And that is why they're using this beta launch to try and gather valuable insights about

103
00:07:28,080 --> 00:07:33,680
how their customers would engage with such tools when it comes to booking holidays, etc.,

104
00:07:33,680 --> 00:07:35,080
compared to existing methods.

105
00:07:35,080 --> 00:07:41,160
So there's logic there in terms of having an understanding of customers that not every

106
00:07:41,160 --> 00:07:43,680
search on Skyscanner is transactional.

107
00:07:43,680 --> 00:07:48,600
Some of it's discovery, and this is enabling that discovery.

108
00:07:48,600 --> 00:07:51,960
And so from that perspective, it's a bit of a masterclass really in consumer first strategy

109
00:07:51,960 --> 00:07:56,280
because they're using generative AI in a very targeted manner.

110
00:07:56,280 --> 00:08:01,040
There's not any enhancing their product offering, but also it's going to allow them to gather

111
00:08:01,040 --> 00:08:06,760
precious behavioral data about how their customers use the tool that can inform future campaigns,

112
00:08:06,760 --> 00:08:11,000
both using the AI, but also for all their other marketing and sales.

113
00:08:11,000 --> 00:08:16,000
So it's an interesting approach, Mal, what do you think about this approach and its applicability

114
00:08:16,000 --> 00:08:17,000
in other areas?

115
00:08:17,000 --> 00:08:25,720
Yeah, I think it's a really smart move introducing conversational AI when you are effectively

116
00:08:25,720 --> 00:08:26,720
a marketplace, right?

117
00:08:26,720 --> 00:08:28,120
And that's what Skyscanner is.

118
00:08:28,120 --> 00:08:35,760
And when you have a plethora of options and you can sell people whatever they want really

119
00:08:35,760 --> 00:08:42,280
within a particular domain, they can offer you travel anywhere in the world.

120
00:08:42,280 --> 00:08:47,480
Helping inspire people, helping people make those decisions is a shrewd move.

121
00:08:47,480 --> 00:08:54,520
I'm really interested to know just how much fine-tuning they've done.

122
00:08:54,520 --> 00:09:00,120
They've clearly integrated their database, hence why they can put these itineraries together

123
00:09:00,120 --> 00:09:05,520
with the travel plans connected to the best flight options, the cheapest flight options,

124
00:09:05,520 --> 00:09:07,360
all of that kind of stuff.

125
00:09:07,360 --> 00:09:12,400
But I'd be interested to know how much fine-tuning has gone into the model to come up with the

126
00:09:12,400 --> 00:09:20,280
itineraries themselves, or is it pretty much a base model giving you an itinerary not too

127
00:09:20,280 --> 00:09:26,440
dissimilar to what you might get if you were to go into ChatGPT and say, give me a weekend

128
00:09:26,440 --> 00:09:36,200
break to a city in Europe for a foodie that is interested in seafood, right?

129
00:09:36,200 --> 00:09:41,640
ChatGPT would give you that itinerary as it is, that would be very good.

130
00:09:41,640 --> 00:09:46,840
Have they done anything over and above that with the outputs and with the training of

131
00:09:46,840 --> 00:09:47,840
the model?

132
00:09:47,840 --> 00:09:52,040
That's a great question because if you wanted to emulate this in your own business for your

133
00:09:52,040 --> 00:09:55,720
own customers, you're going to want to know how easy or hard is it.

134
00:09:55,720 --> 00:10:02,360
In essence, have Skyscanner skinned ChatGPT and gave it access to their flight database,

135
00:10:02,360 --> 00:10:06,780
or have they really gone the hard yards of trying to fine-tune and train the model on

136
00:10:06,780 --> 00:10:12,560
additional data that they know about holidays and great destinations?

137
00:10:12,560 --> 00:10:16,040
I can imagine that this is a domain where, like you say, ChatGPT could probably do a

138
00:10:16,040 --> 00:10:20,800
good job already, but if you operate in a domain where the training materials for ChatGPT

139
00:10:20,800 --> 00:10:27,200
are going to be less detailed, then if you wanted to achieve a similar output, you probably

140
00:10:27,200 --> 00:10:34,080
would need to fine-tune the model or prepare it in a way to be able to actually answer

141
00:10:34,080 --> 00:10:36,080
consumer questions.

142
00:10:36,080 --> 00:10:37,960
Interesting.

143
00:10:37,960 --> 00:10:43,280
Thank you, Skyscanner, for giving us marketers a use case to think, oh, how could I better

144
00:10:43,280 --> 00:10:48,580
serve my customers and interact with my customers and gather valuable data about them if I created

145
00:10:48,580 --> 00:10:54,680
a chatbot just for them, helping them find the best thing that we do to solve their problem.

146
00:10:54,680 --> 00:10:56,400
Next story.

147
00:10:56,400 --> 00:11:01,120
Next story looks at the potential downsides of fine-tuning.

148
00:11:01,120 --> 00:11:07,920
This is an interesting piece of research that came out this week where everyone's very excited

149
00:11:07,920 --> 00:11:14,800
about using the likes of ChatGPT and GPT 3.5 Turbo, and all of these other large language

150
00:11:14,800 --> 00:11:21,960
models, and fine-tuning them so that companies can customize these models to make them better

151
00:11:21,960 --> 00:11:25,640
able to answer questions based on their own company data.

152
00:11:25,640 --> 00:11:31,380
However, scientists found that both malicious hackers and well-meaning users can actually

153
00:11:31,380 --> 00:11:38,160
undermine model safety and the guardrails that are put around models during the fine-tuning

154
00:11:38,160 --> 00:11:39,160
process.

155
00:11:39,160 --> 00:11:49,040
With just a handful of carefully crafted examples, models like GPT 3.5 Turbo can easily learn

156
00:11:49,040 --> 00:11:51,240
harmful responses.

157
00:11:51,240 --> 00:11:59,340
Even tuning a model on normal data, not giving it malicious prompts, but training it on fairly

158
00:11:59,340 --> 00:12:06,120
normal data can degrade protections against bias.

159
00:12:06,120 --> 00:12:14,960
This is a very new discovery in a very nascent field, fine-tuning large language models.

160
00:12:14,960 --> 00:12:16,760
This really matters, right?

161
00:12:16,760 --> 00:12:23,680
Lots of companies are jumping onto this bandwagon and wanting to make the use of large language

162
00:12:23,680 --> 00:12:25,680
models.

163
00:12:25,680 --> 00:12:36,000
We have to be thinking about making sure that we're testing for safety, pre-fine-tuning,

164
00:12:36,000 --> 00:12:41,840
and once we've done the fine-tuning of our model, we have to do an audit of the safety

165
00:12:41,840 --> 00:12:43,160
after the fact as well.

166
00:12:43,160 --> 00:12:50,120
We cannot assume that any protections that were there beforehand remain in place afterwards.

167
00:12:50,120 --> 00:12:57,800
I think this is a really important takeaway for anyone looking to implement a fine-tuned

168
00:12:57,800 --> 00:12:59,520
LLM commercially, right?

169
00:12:59,520 --> 00:13:04,120
Because you're opening yourself up to risks here.

170
00:13:04,120 --> 00:13:12,720
You can have a conversation about what harm looks like with large language models.

171
00:13:12,720 --> 00:13:19,520
Should we have such high thresholds for reduction of harm with LLMs that people generally have?

172
00:13:19,520 --> 00:13:26,440
If you ask large language models how to create a bomb, for instance, they'll typically say,

173
00:13:26,440 --> 00:13:28,680
no, I can't tell you that.

174
00:13:28,680 --> 00:13:35,080
Particularly if they've been guardrailed and had the right safety mechanisms put in place.

175
00:13:35,080 --> 00:13:36,880
But you can Google that and find the same answer, right?

176
00:13:36,880 --> 00:13:42,120
So is there a disparity there in our expectations for the tech in the first place?

177
00:13:42,120 --> 00:13:48,880
But assuming that you're wanting to minimize risks and harm when deploying these models,

178
00:13:48,880 --> 00:13:55,440
you definitely need to think about the impact of what fine-tuning can have on those guardrails.

179
00:13:55,440 --> 00:14:02,840
I think it's really interesting because I think there's a philosophical question about

180
00:14:02,840 --> 00:14:07,800
the harm management aspects when you can find certain harmful information in lots of other

181
00:14:07,800 --> 00:14:09,400
sources really easily.

182
00:14:09,400 --> 00:14:12,360
But as marketers, there's a brand exposure impact, right?

183
00:14:12,360 --> 00:14:16,240
In Skyscanner's case, we don't think they fine-tuned the models perhaps, but if they

184
00:14:16,240 --> 00:14:21,360
had, maybe you could ask it, I want to have a great holiday, but I also want to smuggle

185
00:14:21,360 --> 00:14:22,960
some cocaine back.

186
00:14:22,960 --> 00:14:23,960
What's the best place to go, right?

187
00:14:23,960 --> 00:14:27,480
And then it would probably give you all the best places you could go.

188
00:14:27,480 --> 00:14:32,600
Maybe you can even ask it for advice on how to smuggle the cocaine.

189
00:14:32,600 --> 00:14:40,280
So yeah, I think it's more about the brand impact if that story went viral and the carnage

190
00:14:40,280 --> 00:14:47,120
that would cause that will keep certainly the commercial use of fine-tuning of models

191
00:14:47,120 --> 00:14:52,640
very rigorous in following the process that you outlined, Martin, to make sure that your

192
00:14:52,640 --> 00:14:57,000
brand doesn't get trawled through the mud because of something that the bot spouts out

193
00:14:57,000 --> 00:14:58,000
that it shouldn't have.

194
00:14:58,000 --> 00:15:01,920
Yeah, it's definitely one to watch as we see more and more of this being rolled out, particularly

195
00:15:01,920 --> 00:15:09,120
with the open source models like Llama, which are, you know, heavily, they're going to be

196
00:15:09,120 --> 00:15:11,520
heavily deployed and fine-tuned models.

197
00:15:11,520 --> 00:15:12,520
Yeah.

198
00:15:12,520 --> 00:15:19,120
I'm sure it's only like the 0.01%, but there are definitely folks out there who are taking

199
00:15:19,120 --> 00:15:23,840
extreme joy in figuring out how to game these models.

200
00:15:23,840 --> 00:15:30,400
And my prediction would be the first few companies that start releasing these bots at any sort

201
00:15:30,400 --> 00:15:34,800
of reasonable scale, they're going to get tested.

202
00:15:34,800 --> 00:15:36,660
And then they're going to say things they shouldn't.

203
00:15:36,660 --> 00:15:42,840
And then that stuff's going to go viral on X, formerly known as Twitter.

204
00:15:42,840 --> 00:15:45,960
And it's going to be a cautionary lesson for everyone else who comes afterwards.

205
00:15:45,960 --> 00:15:47,400
Yeah, absolutely.

206
00:15:47,400 --> 00:15:51,800
And teams deploying these models need to have that in their risk management process.

207
00:15:51,800 --> 00:15:52,800
Absolutely.

208
00:15:52,800 --> 00:15:53,800
Right.

209
00:15:53,800 --> 00:15:55,160
Next story then.

210
00:15:55,160 --> 00:16:00,400
So this one's about some recent survey data.

211
00:16:00,400 --> 00:16:06,760
So less than a year ago, CEOs had lukewarm feelings towards AI with only 23% considering

212
00:16:06,760 --> 00:16:10,960
its use to bolster productivity or mitigate labor costs.

213
00:16:10,960 --> 00:16:16,240
If you fast forward to today, and over half of them, 53%, are already harnessing AI for

214
00:16:16,240 --> 00:16:23,520
business use, making a staggering 130% uptick in interest within just eight months.

215
00:16:23,520 --> 00:16:25,760
And yes, there's quite a few percentages in there.

216
00:16:25,760 --> 00:16:28,840
So 23% became 53%.

217
00:16:28,840 --> 00:16:31,480
That's a big jump.

218
00:16:31,480 --> 00:16:37,440
So it does look like AI is now becoming a mainstream tool, at least in the mind of executives.

219
00:16:37,440 --> 00:16:43,760
However, it's not all rosy, because despite growing interest, CEOs continue to tread carefully,

220
00:16:43,760 --> 00:16:48,720
avoiding concerns about bugs, privacy issues, hallucinations, etc.

221
00:16:48,720 --> 00:16:56,160
So even within this expanding landscape, the uptake across various industries is varying

222
00:16:56,160 --> 00:16:57,160
a lot.

223
00:16:57,160 --> 00:17:02,560
Meanwhile, while 70% of CEOs in fields like advertising, marketing, and PR see clear business

224
00:17:02,560 --> 00:17:07,400
value, those in sectors like construction and consumer manufacturing are less convinced.

225
00:17:07,400 --> 00:17:13,880
Not surprising, Martin, given where a lot of the early impact of generative AI has been.

226
00:17:13,880 --> 00:17:22,120
No, not surprising at all, but really encouraging that executives are overall embracing this.

227
00:17:22,120 --> 00:17:26,520
We see this in the work that we do with leadership teams.

228
00:17:26,520 --> 00:17:32,580
They get a sniff of it typically on their own or someone signposts them to this tech,

229
00:17:32,580 --> 00:17:34,780
and then they go, oh, this is amazing.

230
00:17:34,780 --> 00:17:36,100
This is brilliant.

231
00:17:36,100 --> 00:17:37,100
We need to roll this out.

232
00:17:37,100 --> 00:17:44,680
So I think there is clearly an appetite at the individual level, the ability for individuals

233
00:17:44,680 --> 00:17:46,040
to become more productive.

234
00:17:46,040 --> 00:17:53,320
And we're still all figuring out the kind of company level use cases for this technology.

235
00:17:53,320 --> 00:18:00,920
But yeah, it's encouraging to see CEOs, 53% of them, embracing generative AI.

236
00:18:00,920 --> 00:18:08,680
Yeah, I agree, and when you look at the issues with the tools, like their propensity to hallucinate,

237
00:18:08,680 --> 00:18:14,280
like how challenging it can be to get them to do certain things that you want them to

238
00:18:14,280 --> 00:18:22,560
be able to do reliably, they still can save people time in so many areas that even in

239
00:18:22,560 --> 00:18:28,280
this sort of gray area that we're in right now, where there's some benefits, but oh,

240
00:18:28,280 --> 00:18:30,600
you've got to be mindful of all of these drawbacks.

241
00:18:30,600 --> 00:18:35,080
I still think there's enough advantages to be able to roll them out in a number of areas.

242
00:18:35,080 --> 00:18:40,200
And as we've said again and again, and as many else have said on the social sphere,

243
00:18:40,200 --> 00:18:42,940
this is the worst these tools will ever be.

244
00:18:42,940 --> 00:18:44,460
So they're only going to get better from here.

245
00:18:44,460 --> 00:18:47,040
And there are those clear use cases where there are benefits.

246
00:18:47,040 --> 00:18:49,400
And it's almost like sliding scales.

247
00:18:49,400 --> 00:18:54,800
You've got certain tasks where they just obviously save so much time, even with some of the issues

248
00:18:54,800 --> 00:18:57,360
with them, you should start using them today.

249
00:18:57,360 --> 00:19:01,820
There's a few where it might be about net zero once you look at the amount of prompt

250
00:19:01,820 --> 00:19:07,620
engineering to get what you want, or QC and QA of the output to make sure it's half decent.

251
00:19:07,620 --> 00:19:11,840
And there are definitely those areas where either the tools are not up to snuff yet,

252
00:19:11,840 --> 00:19:16,160
or you have to spend so much time trying to get the output actually takes you longer than

253
00:19:16,160 --> 00:19:18,800
if you'd have just done it human manually.

254
00:19:18,800 --> 00:19:25,120
But I think the bars there on that graph, or on that bar chart are basically shifting

255
00:19:25,120 --> 00:19:30,200
all the time towards ever more applications in a, this can do it really well and it saves

256
00:19:30,200 --> 00:19:32,200
us lots of time section.

257
00:19:32,200 --> 00:19:33,520
Yeah.

258
00:19:33,520 --> 00:19:40,320
And I think the more people use the tools day to day, going back to that Ethan Molyk

259
00:19:40,320 --> 00:19:47,640
paper that we saw recently where he was talking about the Boston Consulting Group consultants,

260
00:19:47,640 --> 00:19:54,800
the way that they use these tools is either like a Centaur or a Cyborg and the Centaurs

261
00:19:54,800 --> 00:19:59,120
are the people that kind of, they do the human bit of the task and then they have a bit of

262
00:19:59,120 --> 00:20:04,400
the task that they give over to the AI and kind of bounce between the two.

263
00:20:04,400 --> 00:20:09,920
Whereas the Cyborgs work almost as if they're fully integrated with the AI, it is as if

264
00:20:09,920 --> 00:20:12,280
it is part of the same person.

265
00:20:12,280 --> 00:20:19,520
I think you end up being a Centaur or a Cyborg, but that just comes through familiarization

266
00:20:19,520 --> 00:20:22,600
with what the tools can do.

267
00:20:22,600 --> 00:20:28,520
Understanding the limits and understanding how your workflows, how you work, right?

268
00:20:28,520 --> 00:20:32,200
Your working practices and you start to integrate them more and more.

269
00:20:32,200 --> 00:20:39,160
And as the models become more capable, you test boundaries and introduce new techniques

270
00:20:39,160 --> 00:20:40,160
into what you do.

271
00:20:40,160 --> 00:20:46,440
Certainly I do every day, I find something new and I'm like, stick that on the list.

272
00:20:46,440 --> 00:20:47,440
Absolutely.

273
00:20:47,440 --> 00:20:52,320
You've got to be willing to jump in and get your hands dirty in your particular work type

274
00:20:52,320 --> 00:20:57,200
and the workflows that you operate in and think of creative ways to test these tools

275
00:20:57,200 --> 00:21:05,400
out to see if they help you be more creative, get better outputs, deliver things faster.

276
00:21:05,400 --> 00:21:12,240
Whilst combining that with tech scouting, listening to podcasts like this in terms of

277
00:21:12,240 --> 00:21:15,520
what different tech can do over time so you can wrap that into your workflows.

278
00:21:15,520 --> 00:21:21,400
And I think there's a fair bit of sort of the more operational type process mapping

279
00:21:21,400 --> 00:21:27,040
to really help identify which bits could be smart automated with AI, often identifying

280
00:21:27,040 --> 00:21:32,440
a number of areas which could just be automated, AI or not, and getting benefits regardless.

281
00:21:32,440 --> 00:21:40,320
So all of those things combined are the best way to combine AI with your current workflows

282
00:21:40,320 --> 00:21:41,680
to find those improvements.

283
00:21:41,680 --> 00:21:45,320
And outside of listening to this podcast, which of course you are, we thank you, thanks

284
00:21:45,320 --> 00:21:50,000
for being here, following someone like Ethan Moller on LinkedIn is a great way to see how

285
00:21:50,000 --> 00:21:54,240
people are testing these tools to see what can be done and perhaps where they're not

286
00:21:54,240 --> 00:21:55,880
quite right yet.

287
00:21:55,880 --> 00:22:02,800
Yeah, that example of the smart autonomous process mapping as well as interesting point

288
00:22:02,800 --> 00:22:09,280
that you make takes me back to Zapier Canvas, the tool that they released last week.

289
00:22:09,280 --> 00:22:11,080
I got early access to it.

290
00:22:11,080 --> 00:22:13,960
I encourage people to have a play with it.

291
00:22:13,960 --> 00:22:19,200
If you can sign up and get early access to it, simply map out a process.

292
00:22:19,200 --> 00:22:24,920
It doesn't have to be something that is integrated with tech at that point, but it helps if you've

293
00:22:24,920 --> 00:22:30,680
got some steps that are and it will use AI to tell you which bits can be fully smart

294
00:22:30,680 --> 00:22:31,680
integrated.

295
00:22:31,680 --> 00:22:33,360
So you don't even have to do the thinking there yourself.

296
00:22:33,360 --> 00:22:34,360
It's great.

297
00:22:34,360 --> 00:22:35,360
Highly recommend it.

298
00:22:35,360 --> 00:22:38,360
Oh, I've got to get me on that and have a little play.

299
00:22:38,360 --> 00:22:40,620
Maybe we'll, we haven't done a tool of the week for a while.

300
00:22:40,620 --> 00:22:45,320
Maybe you can do a deep dive for us on an upcoming episode of mine.

301
00:22:45,320 --> 00:22:48,780
But for now, we're going to go on to the next story.

302
00:22:48,780 --> 00:22:55,280
So this one is less marketing focused and more of a reality check for the UK government

303
00:22:55,280 --> 00:22:59,640
courtesy of some researchers at the University of Cambridge.

304
00:22:59,640 --> 00:23:04,160
So the UK government has a vision to become a global leader in generative AI, but this

305
00:23:04,160 --> 00:23:10,800
has come under scrutiny from these researchers who published a recent report outlining some

306
00:23:10,800 --> 00:23:15,480
of the key challenges for the UK in this endeavour.

307
00:23:15,480 --> 00:23:21,560
Now primarily they point out that there is a lack of substantial capital investment and

308
00:23:21,560 --> 00:23:28,440
importantly the computing power in the UK to develop generative AI technologies.

309
00:23:28,440 --> 00:23:36,440
In the report, they say that the computing cost for chat GPT alone is estimated at $33

310
00:23:36,440 --> 00:23:39,240
million per month.

311
00:23:39,240 --> 00:23:46,200
Whereas the UK's Frontier AI Task Force has been earmarked an initial budget of a hundred

312
00:23:46,200 --> 00:23:50,640
million pounds, which wouldn't go too far.

313
00:23:50,640 --> 00:23:56,280
Now obviously this is a fund that's going to be investing in seed funding, different

314
00:23:56,280 --> 00:24:02,160
ventures, but it goes to show the difference in scale in terms of becoming a real power

315
00:24:02,160 --> 00:24:03,160
player.

316
00:24:03,160 --> 00:24:12,240
Last week we discussed OpenAI's, the budget required if they were to hit the scale of

317
00:24:12,240 --> 00:24:18,480
just one tenth of the queries of Google, which I believe was $46 billion in compute power

318
00:24:18,480 --> 00:24:19,480
required.

319
00:24:19,480 --> 00:24:22,720
So the numbers are absolutely massive.

320
00:24:22,720 --> 00:24:27,720
And speaking of that compute power, the UK just doesn't have the infrastructure.

321
00:24:27,720 --> 00:24:37,320
It's lacking any major clusters of GPUs, the graphic processing units that power AI compute.

322
00:24:37,320 --> 00:24:43,320
They effectively do all of the important heavy number crunching in the big machine learning

323
00:24:43,320 --> 00:24:44,320
models.

324
00:24:44,320 --> 00:24:51,200
And the UK has a 900 million pound supercomputer, which is devoted to AI research.

325
00:24:51,200 --> 00:24:54,800
This is in the pipeline, it's going to be heading to Bristol, but it isn't expected

326
00:24:54,800 --> 00:25:00,600
to be operational until 2026, which feels like light years away.

327
00:25:00,600 --> 00:25:04,920
We'll all be living in a computer simulation by then, Martin.

328
00:25:04,920 --> 00:25:07,880
We'll be running on GPU clusters by the time.

329
00:25:07,880 --> 00:25:08,880
Yeah, absolutely.

330
00:25:08,880 --> 00:25:11,040
It just feels so far away.

331
00:25:11,040 --> 00:25:18,800
And I think for the UK to be competitive in this field, it just seems we don't have the

332
00:25:18,800 --> 00:25:23,840
resources to be truly competitive on a global stage.

333
00:25:23,840 --> 00:25:25,320
However, it's not all bad news.

334
00:25:25,320 --> 00:25:31,000
The report does advocate a different path for global leadership for the UK when it comes

335
00:25:31,000 --> 00:25:32,000
to AI.

336
00:25:32,000 --> 00:25:36,320
And that's one that's grounded in real world applications.

337
00:25:36,320 --> 00:25:41,240
So in the report they call for tax incentives for companies that are developing or incorporating

338
00:25:41,240 --> 00:25:46,920
AI, and also the development of a solid legal and ethical framework to foster trust.

339
00:25:46,920 --> 00:25:52,040
So becoming more of the, it's almost like a soft power play.

340
00:25:52,040 --> 00:25:54,920
It's the soft skills of AI.

341
00:25:54,920 --> 00:25:58,260
We can lead the world in this domain.

342
00:25:58,260 --> 00:26:04,880
So as the UK focuses on its strengths in cyber security, fintech, health tech, it might be

343
00:26:04,880 --> 00:26:12,500
able to carve out a bit of a niche in the world of AI, just not in the way originally

344
00:26:12,500 --> 00:26:18,080
envisioned in the UK government's AI white paper.

345
00:26:18,080 --> 00:26:23,520
So what do we think that means then, a range of different incentives to try and encourage

346
00:26:23,520 --> 00:26:27,520
British businesses to explore how AI can increase efficiency and productivity?

347
00:26:27,520 --> 00:26:35,800
Yeah, I think they've got to create the landscape that enables companies to experiment and play

348
00:26:35,800 --> 00:26:38,480
with tech developed outside of the UK, right?

349
00:26:38,480 --> 00:26:43,640
So it's not about becoming the place where the foundational models are created.

350
00:26:43,640 --> 00:26:51,680
It's about becoming the place where the foundational models are implemented and made useful in

351
00:26:51,680 --> 00:26:54,240
real world commercial applications.

352
00:26:54,240 --> 00:26:55,600
Interesting.

353
00:26:55,600 --> 00:27:02,600
Well, we're trying to do our bit in the UK here for helping companies to realize the

354
00:27:02,600 --> 00:27:07,120
power of AI with just those types of use cases in mind.

355
00:27:07,120 --> 00:27:11,640
So if you're interested in figuring out how to do that, little plug for our training and

356
00:27:11,640 --> 00:27:15,720
consultancy services there, do you get in touch on the artificially intelligent marketing

357
00:27:15,720 --> 00:27:16,720
website.

358
00:27:16,720 --> 00:27:21,120
But outside of that, I think we'll slip into our next story, Martin.

359
00:27:21,120 --> 00:27:22,680
And it's 11 Labs.

360
00:27:22,680 --> 00:27:27,640
So regular listeners to the podcast would know we're a bit of a fan of 11 Labs and its

361
00:27:27,640 --> 00:27:33,640
voice synthesis engine driven by AI creating really natural language sounding voices.

362
00:27:33,640 --> 00:27:40,160
Well, Martin has had the chance to test out their much anticipated AI dubbing system.

363
00:27:40,160 --> 00:27:45,520
So they already offer 11 Labs voice cloning technology and the latest feature enables

364
00:27:45,520 --> 00:27:51,220
any person's voice to be authentically dubbed into over 20 languages.

365
00:27:51,220 --> 00:27:56,480
So this is much like Hey Jen, which we've spoken about previously on the podcast.

366
00:27:56,480 --> 00:28:04,280
So it's kind of a bit me too in terms of offering that, but 11 Labs is extremely cost effective

367
00:28:04,280 --> 00:28:05,280
to work with.

368
00:28:05,280 --> 00:28:10,720
It has a free option and I think the paid option is like $20 a month or something crazily

369
00:28:10,720 --> 00:28:11,720
low.

370
00:28:11,720 --> 00:28:15,320
You can actually get it for $5 a month at their base level.

371
00:28:15,320 --> 00:28:16,840
There you go.

372
00:28:16,840 --> 00:28:19,640
So you've been playing with this, Martin.

373
00:28:19,640 --> 00:28:20,640
What was your thoughts?

374
00:28:20,640 --> 00:28:24,560
Yeah, it's super easy to use.

375
00:28:24,560 --> 00:28:31,160
Literally upload a video, choose a language, hit dub and it runs the process.

376
00:28:31,160 --> 00:28:35,440
Doesn't take long for it to do so it's very, very simple.

377
00:28:35,440 --> 00:28:43,360
However, the simplicity also is reflected in the quality of the output.

378
00:28:43,360 --> 00:28:50,600
It isn't at the level of Hey Jen, the lip sync element that Hey Jen gets right.

379
00:28:50,600 --> 00:28:51,680
It just isn't there.

380
00:28:51,680 --> 00:29:00,160
It's much more like an old fashioned dubbed movie, like watching an old martial arts kung

381
00:29:00,160 --> 00:29:03,840
fu movie from the 1970s or something like that.

382
00:29:03,840 --> 00:29:06,240
It doesn't even attempt to do that.

383
00:29:06,240 --> 00:29:11,200
It just puts on new audio over the top of the piece.

384
00:29:11,200 --> 00:29:14,760
So if you've got something where the person is speaking in the video, it doesn't do a

385
00:29:14,760 --> 00:29:15,760
great job.

386
00:29:15,760 --> 00:29:18,000
If it's a voiceover, it does a pretty good job, right?

387
00:29:18,000 --> 00:29:22,120
So if you've just recorded yourself doing a voiceover on a video and you want that translating

388
00:29:22,120 --> 00:29:25,600
into another language, it appears to do a pretty good job.

389
00:29:25,600 --> 00:29:31,680
All caveats apply such as when I listened back to it, the example that I did for myself

390
00:29:31,680 --> 00:29:36,160
was translated into German and Italian.

391
00:29:36,160 --> 00:29:37,160
It sounds okay.

392
00:29:37,160 --> 00:29:41,400
In fact, I think we have even got a bit of a clip that we could play.

393
00:29:41,400 --> 00:29:43,400
Is that available Paul?

394
00:29:43,400 --> 00:29:46,240
Yeah, let's play that now.

395
00:29:46,240 --> 00:29:54,840
In the report it says initial estimates suggest that if chat GPT were to scale to just a tenth

396
00:29:54,840 --> 00:30:04,120
of Google's search queries, it would require a staggering 48.1 billion dollars worth of

397
00:30:04,120 --> 00:30:12,080
GPU, initially and approximately 16 billion annually for maintenance.

398
00:30:12,080 --> 00:30:17,840
In the report it says that initial estimates suggest that if chat GPT were to scale to

399
00:30:17,840 --> 00:30:30,200
just a tenth of Google's search queries, it would require a staggering 48.1 billion dollars

400
00:30:30,200 --> 00:30:39,960
worth of GPU, initially and approximately 16 billion annually for maintenance.

401
00:30:39,960 --> 00:30:40,960
So there you go.

402
00:30:40,960 --> 00:30:41,960
Me in Italian.

403
00:30:41,960 --> 00:30:46,480
That was a clip from Artificially Intelligent Marketing last week.

404
00:30:46,480 --> 00:30:52,760
We took a small snippet of the episode and played that.

405
00:30:52,760 --> 00:30:53,760
What are your thoughts?

406
00:30:53,760 --> 00:30:59,480
I think it does a decent enough job, but I don't speak Italian, so I would like one of

407
00:30:59,480 --> 00:31:04,200
our Italian listeners to judge, but how about yourself?

408
00:31:04,200 --> 00:31:05,200
Do you know what?

409
00:31:05,200 --> 00:31:09,480
Listening to it, obviously the lip sync feature that blows you away when you see, hey Jen's

410
00:31:09,480 --> 00:31:10,480
not there.

411
00:31:10,480 --> 00:31:12,000
So we'll just put that to one side.

412
00:31:12,000 --> 00:31:18,920
In terms of the audio itself, it does maintain the sort of cadence of speech in English.

413
00:31:18,920 --> 00:31:23,960
So like even though it doesn't lip sync, the audio is synced with when your lips move,

414
00:31:23,960 --> 00:31:25,600
which I think is interesting.

415
00:31:25,600 --> 00:31:32,440
And there's some sort of emotional quality and tonality that's still there.

416
00:31:32,440 --> 00:31:38,520
As I said to you when you first sent it to me, it sounds a bit like you, but not really.

417
00:31:38,520 --> 00:31:44,200
I think the Italian accent that's applied is probably necessary in order to have it

418
00:31:44,200 --> 00:31:45,200
speaking effectively.

419
00:31:45,200 --> 00:31:48,600
Although of course, Italian listeners, we want to know your thoughts because you're

420
00:31:48,600 --> 00:31:51,520
better qualified to speak on that than we are.

421
00:31:51,520 --> 00:31:57,080
But in the end, for me, it sounds like a voice actor that sounds a bit like you, but who

422
00:31:57,080 --> 00:32:05,640
is speaking in Italian and potentially is Italian versus me thinking, crumbs, does Martin

423
00:32:05,640 --> 00:32:06,640
learn Italian?

424
00:32:06,640 --> 00:32:07,960
Do you know what I mean?

425
00:32:07,960 --> 00:32:10,040
So that's not there.

426
00:32:10,040 --> 00:32:14,400
Having said that, I obviously know you and have known you for a while.

427
00:32:14,400 --> 00:32:20,040
Does anybody else care how close the voice is for different marketing and sales applications?

428
00:32:20,040 --> 00:32:21,040
Almost certainly not.

429
00:32:21,040 --> 00:32:22,040
Right?

430
00:32:22,040 --> 00:32:23,040
No.

431
00:32:23,040 --> 00:32:26,600
So from that perspective, I can really see the value.

432
00:32:26,600 --> 00:32:31,960
And we've talked previously about how long is it before this technology makes its way

433
00:32:31,960 --> 00:32:38,600
into Zoom and Microsoft Teams and Google Meet so that you can speak in one language, but

434
00:32:38,600 --> 00:32:42,600
the other person hears you speaking in their language.

435
00:32:42,600 --> 00:32:45,480
And I think we're getting ever closer.

436
00:32:45,480 --> 00:32:49,840
And I think people would absolutely be fine with the lack of lip syncing for that type

437
00:32:49,840 --> 00:32:52,400
of application anyway.

438
00:32:52,400 --> 00:32:57,800
And as you said, if it's just voiceovers and stuff, then all of your corporate videos where

439
00:32:57,800 --> 00:33:02,120
you're showing your facilities, if you operate in a global market, like for example, some

440
00:33:02,120 --> 00:33:08,240
of our contract research and contract manufacturing clients often do, you can now localize that

441
00:33:08,240 --> 00:33:13,760
in all these languages, which for many companies would have been probably not worth the cost

442
00:33:13,760 --> 00:33:18,240
previously, if I'm honest, because of all the voice acting work that would have been

443
00:33:18,240 --> 00:33:19,240
required.

444
00:33:19,240 --> 00:33:24,600
But now if that's basically click of a button and the outputs of quality, big asterisk there,

445
00:33:24,600 --> 00:33:31,360
as we said, not strongest German and Italian speakers to really judge, then that does open

446
00:33:31,360 --> 00:33:36,680
up a new avenue of opportunities to really localize content at scale in a way that just

447
00:33:36,680 --> 00:33:40,040
has never been possible before.

448
00:33:40,040 --> 00:33:44,980
And adding to that, like you said a moment ago, this is the worst that this model will

449
00:33:44,980 --> 00:33:47,280
ever be, right?

450
00:33:47,280 --> 00:33:50,600
This is, it's only going to get better from here on out.

451
00:33:50,600 --> 00:33:51,600
Absolutely.

452
00:33:51,600 --> 00:33:57,840
I mean, I guess it opens up that little Pandora's box that we talk about often, how misused

453
00:33:57,840 --> 00:33:58,840
this could be.

454
00:33:58,840 --> 00:34:07,080
I mean, one assumes I could have pushed that video into 11 Labs and had you translated.

455
00:34:07,080 --> 00:34:11,280
I don't know how easy it is to change the content of what's translated.

456
00:34:11,280 --> 00:34:16,560
So maybe I could have you speaking something you didn't originally say in Italian, perhaps

457
00:34:16,560 --> 00:34:19,240
to make you look especially good or especially bad.

458
00:34:19,240 --> 00:34:21,800
Yeah, there was no way to edit what was said.

459
00:34:21,800 --> 00:34:25,920
So it's not got any like descript style text editing.

460
00:34:25,920 --> 00:34:26,920
Okay.

461
00:34:26,920 --> 00:34:29,360
However, they do have a detector.

462
00:34:29,360 --> 00:34:35,520
So if you put any audio into the website and it will tell you whether or not it was made

463
00:34:35,520 --> 00:34:36,880
with 11 Labs.

464
00:34:36,880 --> 00:34:42,720
So I wonder whether they've got some sort of watermark in the audio.

465
00:34:42,720 --> 00:34:43,720
Interesting.

466
00:34:43,720 --> 00:34:48,440
But then your average user is not necessarily going to know that and necessarily getting

467
00:34:48,440 --> 00:34:53,760
produced with 11 Labs, which is going to kill a lot of commercial applications, I would

468
00:34:53,760 --> 00:34:55,040
have thought.

469
00:34:55,040 --> 00:34:57,080
So I can imagine them not introducing that.

470
00:34:57,080 --> 00:35:00,640
But I think it's another one of those areas that as marketers, we just need to keep an

471
00:35:00,640 --> 00:35:04,000
eye on how this develops.

472
00:35:04,000 --> 00:35:07,760
Is it seen as something that's going to be trusted by customers or would they rather

473
00:35:07,760 --> 00:35:09,000
you just didn't bother?

474
00:35:09,000 --> 00:35:10,000
Right.

475
00:35:10,000 --> 00:35:16,000
I remember when automated meeting scheduling tools came out and you really had to handle

476
00:35:16,000 --> 00:35:21,560
them very carefully when you were using them, because you could send the message, I'm really

477
00:35:21,560 --> 00:35:25,800
too busy to schedule a meeting with you, because you're not important enough for me to do it

478
00:35:25,800 --> 00:35:26,800
manually.

479
00:35:26,800 --> 00:35:28,400
So I've got to do it automatically with this link.

480
00:35:28,400 --> 00:35:34,040
And you really had to try and play up the benefits to the other person to try and sort

481
00:35:34,040 --> 00:35:38,840
of get over that technical barrier until they became basically accepted, which is where

482
00:35:38,840 --> 00:35:41,320
I feel they probably are now.

483
00:35:41,320 --> 00:35:45,200
Will we see something when it comes to the adoption of this type of stuff, which is at

484
00:35:45,200 --> 00:35:48,440
first people are like, no, I didn't.

485
00:35:48,440 --> 00:35:51,560
I'm quite happy just reading the subtitles, other than knowing that you didn't put any

486
00:35:51,560 --> 00:35:54,480
real effort into translating this for me.

487
00:35:54,480 --> 00:35:55,480
I don't know, right?

488
00:35:55,480 --> 00:35:58,000
It'd be interesting to see how some of that plays out.

489
00:35:58,000 --> 00:35:59,000
Right.

490
00:35:59,000 --> 00:36:00,000
Final story.

491
00:36:00,000 --> 00:36:07,600
So this week, some eagle-eyed reporters spotted that OpenAI has recently changed its core

492
00:36:07,600 --> 00:36:09,600
values on its website.

493
00:36:09,600 --> 00:36:15,400
Previously, their website had listed that their focus was on attributes like being audacious,

494
00:36:15,400 --> 00:36:17,600
thoughtful and impact driven.

495
00:36:17,600 --> 00:36:24,000
But suddenly we've seen a shift to a new set of values, such as AGI focus, intense and

496
00:36:24,000 --> 00:36:28,600
scrappy and scale.

497
00:36:28,600 --> 00:36:35,480
So clearly a shift in priorities, should I say.

498
00:36:35,480 --> 00:36:40,880
If the company can change its values like that on a whim though, people were asking,

499
00:36:40,880 --> 00:36:44,360
were they ever core values to begin with?

500
00:36:44,360 --> 00:36:49,480
I think that's probably a little unfair given the growth and development of the company

501
00:36:49,480 --> 00:36:52,280
over the past five years.

502
00:36:52,280 --> 00:36:59,480
But it is interesting for us as marketers to consider the role that values play and

503
00:36:59,480 --> 00:37:07,360
how it will shape the direction of a business and it should act as a lens through which

504
00:37:07,360 --> 00:37:10,040
we can prioritize decisions.

505
00:37:10,040 --> 00:37:18,000
Clearly in this instance, OpenAI are focused very firmly on artificial general intelligence

506
00:37:18,000 --> 00:37:20,840
and achieving scale.

507
00:37:20,840 --> 00:37:25,240
So this will ultimately guide decision-making, it's going to influence culture and it's

508
00:37:25,240 --> 00:37:34,240
going to be a yardstick against which external entities measure the value, the integrity

509
00:37:34,240 --> 00:37:37,200
as well of the business.

510
00:37:37,200 --> 00:37:44,520
Sorry, I think it's not just assuming that they've implemented these values as you would

511
00:37:44,520 --> 00:37:51,120
hope, they've communicated these internally and rolled them out as part of a wider piece.

512
00:37:51,120 --> 00:37:53,320
It's not just a change in wording.

513
00:37:53,320 --> 00:38:01,840
I think we as a market should also understand that this change in focus will have long-term

514
00:38:01,840 --> 00:38:07,400
impacts on the goals and the mission of the organization, much of which I think has been

515
00:38:07,400 --> 00:38:14,400
communicated by Sam Altman over the past few months anyway.

516
00:38:14,400 --> 00:38:19,880
All of which is an interesting segue and the timing of the story was particularly good

517
00:38:19,880 --> 00:38:30,240
as this week I had a very enlightening discussion with Olivia Gamblin, who is the CEO and founder

518
00:38:30,240 --> 00:38:34,440
of Ethical Intelligence.

519
00:38:34,440 --> 00:38:42,960
She is an ethicist by training and we sat down this week to discuss things like how

520
00:38:42,960 --> 00:38:52,080
corporate values can help companies when developing AI tools, AI products or incorporating AI into

521
00:38:52,080 --> 00:38:54,240
the way their business works.

522
00:38:54,240 --> 00:38:57,120
Let's pick it up with Olivia.

523
00:38:57,120 --> 00:38:59,840
Welcome to Artificially Intelligent Marketing, Olivia Gamblin.

524
00:38:59,840 --> 00:39:04,000
Thank you so much for having me here today, Martin.

525
00:39:04,000 --> 00:39:09,320
So yeah, you're an entrepreneur and an ethicist and many people listening to this podcast

526
00:39:09,320 --> 00:39:13,680
will be familiar with the day-to-day life of an entrepreneur, but maybe less so of a

527
00:39:13,680 --> 00:39:14,680
professional ethicist.

528
00:39:14,680 --> 00:39:17,680
Tell us, what does that look like?

529
00:39:17,680 --> 00:39:19,680
What do your days entail?

530
00:39:19,680 --> 00:39:25,440
Well, they're quite busy lately, especially since the onset of generative AI.

531
00:39:25,440 --> 00:39:29,200
But I do say I do really wear two different hats.

532
00:39:29,200 --> 00:39:34,040
One is the entrepreneur, one is the ethicist and that ethicist hat, oftentimes I'm the

533
00:39:34,040 --> 00:39:37,880
first ethicist that people will meet, which is very fun for me.

534
00:39:37,880 --> 00:39:43,080
They're usually surprised to hear that I am not going to sit there and split hairs with

535
00:39:43,080 --> 00:39:47,000
them about good and bad and tell them they're a bad person.

536
00:39:47,000 --> 00:39:51,340
Instead, actually as an ethicist, my role is a very supportive role.

537
00:39:51,340 --> 00:39:54,740
So I specifically work with decision makers.

538
00:39:54,740 --> 00:39:59,260
So I'm in with management and leadership teams working through what I call a decision

539
00:39:59,260 --> 00:40:05,420
analysis process of applying ethics where we are analyzing the decisions they're making

540
00:40:05,420 --> 00:40:11,200
around their strategy and crucial decision making for their technology, specifically

541
00:40:11,200 --> 00:40:12,640
AI.

542
00:40:12,640 --> 00:40:17,680
And I'm assessing to see its alignment with the ethical values that they've set out for

543
00:40:17,680 --> 00:40:23,820
the company that they've set out either by regulation or their own company values.

544
00:40:23,820 --> 00:40:35,400
So my work really as a day-to-day as an ethicist is, let's say, sitting as the consciousness

545
00:40:35,400 --> 00:40:37,200
underneath everything.

546
00:40:37,200 --> 00:40:42,240
You know the binding nemo, you have that one scene where Dory is saying, nemo, I am your

547
00:40:42,240 --> 00:40:44,400
conscience.

548
00:40:44,400 --> 00:40:48,160
Sometimes that's kind of how I feel where I'm sitting in these meetings and people are

549
00:40:48,160 --> 00:40:52,160
talking and they're getting excited about ideas and I'm going, yes, that's great.

550
00:40:52,160 --> 00:40:54,760
But maybe there's a better decision that we can be making here.

551
00:40:54,760 --> 00:40:58,000
What if we tweaked it in this direction from the background?

552
00:40:58,000 --> 00:41:02,240
So I'm not there to force any frameworks or ideas on anyone.

553
00:41:02,240 --> 00:41:08,520
I'm more like a living, breathing moral compass for a company and their AI.

554
00:41:08,520 --> 00:41:13,600
That's a nice way of framing it.

555
00:41:13,600 --> 00:41:19,240
So what took you into the realms of AI specifically then?

556
00:41:19,240 --> 00:41:23,940
The world of ethics can take you into many areas, I'm sure.

557
00:41:23,940 --> 00:41:28,680
So what was it that appealed and how did you land in the tech space?

558
00:41:28,680 --> 00:41:30,020
I think probably a combination.

559
00:41:30,020 --> 00:41:33,280
So I grew up in the Silicon Valley, so I grew up around AI.

560
00:41:33,280 --> 00:41:39,480
I like to joke that that is my first second language as I speak techie.

561
00:41:39,480 --> 00:41:43,160
So I think there was always that focus of, well, of course, the only industry that exists

562
00:41:43,160 --> 00:41:44,160
is technology.

563
00:41:44,160 --> 00:41:49,400
At least when you grow up in that area, that's kind of what you come to believe.

564
00:41:49,400 --> 00:41:55,200
And I think though later on in life as I started actually practicing as an ethicist and looking

565
00:41:55,200 --> 00:42:01,760
at what field I wanted to work in, the field of AI and data attracted me because it's quite

566
00:42:01,760 --> 00:42:06,420
an interesting mirror to our own selves, our own humanity.

567
00:42:06,420 --> 00:42:11,120
When you're working, say for example, as an ethicist in medicine or business or politics,

568
00:42:11,120 --> 00:42:16,160
you're looking at specific use cases, which is still the same case with AI.

569
00:42:16,160 --> 00:42:19,560
But with AI, it's almost like that use case is pointed back at the humans.

570
00:42:19,560 --> 00:42:25,320
So we're having to address a lot of these very difficult ethical questions that we've

571
00:42:25,320 --> 00:42:28,880
been struggling with since the ancient philosophers.

572
00:42:28,880 --> 00:42:35,200
And now it's coming back in just in a different format and in a different clothing, let's

573
00:42:35,200 --> 00:42:36,200
say.

574
00:42:36,200 --> 00:42:38,000
So I find it fascinating.

575
00:42:38,000 --> 00:42:44,120
It brings the work of an ethicist to a whole different scale and realm that really attracted

576
00:42:44,120 --> 00:42:46,320
me, I have to say.

577
00:42:46,320 --> 00:42:52,200
And I imagine given the breadth and the way that technology touches so many aspects of

578
00:42:52,200 --> 00:42:59,880
our lives, you're still getting involved in the political, the policy, the governance,

579
00:42:59,880 --> 00:43:01,680
all of that still very much applies, right?

580
00:43:01,680 --> 00:43:02,680
Oh yes.

581
00:43:02,680 --> 00:43:04,480
I like to joke with my fellow ethicists.

582
00:43:04,480 --> 00:43:05,760
We have to be ethicists.

583
00:43:05,760 --> 00:43:06,960
We have to be politicians.

584
00:43:06,960 --> 00:43:08,660
We have to be business analysts.

585
00:43:08,660 --> 00:43:09,660
We have to be engineers.

586
00:43:09,660 --> 00:43:15,240
We have to be, gosh, anything and everything because we have to have such a strong working

587
00:43:15,240 --> 00:43:19,800
knowledge of so many different fields that impact our day-to-day work.

588
00:43:19,800 --> 00:43:20,800
Yeah.

589
00:43:20,800 --> 00:43:27,040
And I can imagine that multifacetedness is in high demand right now.

590
00:43:27,040 --> 00:43:31,120
So you spoke about the rise of generative AI.

591
00:43:31,120 --> 00:43:38,680
Can you just talk to me about your current take on generative AI and the ethical considerations

592
00:43:38,680 --> 00:43:41,920
or implications of this new tech?

593
00:43:41,920 --> 00:43:43,000
Yeah.

594
00:43:43,000 --> 00:43:48,640
So I'll talk about it from kind of a business perspective and then from an ethicist perspective

595
00:43:48,640 --> 00:43:49,640
specifically.

596
00:43:49,640 --> 00:43:56,640
Again, as an ethicist, I work actually quite often on business ethics as well as AI ethics.

597
00:43:56,640 --> 00:43:59,280
Those are closely intertwined.

598
00:43:59,280 --> 00:44:10,560
But the business perspective around generative AI is it's slower for let's say in-depth adoption

599
00:44:10,560 --> 00:44:16,120
than I think it's made out to be, meaning I've seen and worked with a lot of companies

600
00:44:16,120 --> 00:44:21,720
that are still wrapping their minds around what is the use case for this technology.

601
00:44:21,720 --> 00:44:22,720
Very fascinating.

602
00:44:22,720 --> 00:44:28,260
And on an individual level, we can come up with uses for chat GPT, but that doesn't necessarily

603
00:44:28,260 --> 00:44:33,600
directly translate into this is useful for the internal operations of a company.

604
00:44:33,600 --> 00:44:37,800
There's also a lot of security considerations and data privacy considerations.

605
00:44:37,800 --> 00:44:47,040
So the adoption of generative AI is a lot slower internally than it's made out to be,

606
00:44:47,040 --> 00:44:51,720
which I know sometimes takes people by surprise because it's the hype right now.

607
00:44:51,720 --> 00:44:57,440
And then as an ethicist, I would say it's a very, very fascinating tool.

608
00:44:57,440 --> 00:45:03,640
To me, it's very cool to see and be able to prompt say chat GPT with a question and know

609
00:45:03,640 --> 00:45:10,560
that the answer that I'm getting back is like this giant, I think I heard it once described

610
00:45:10,560 --> 00:45:17,040
as a word calculator, meaning it's going through all of these different scenarios and predictions

611
00:45:17,040 --> 00:45:20,720
and possibilities statistics of what is the next possible word.

612
00:45:20,720 --> 00:45:23,840
And that's fascinating to me because the amount of data that it had to process to be able

613
00:45:23,840 --> 00:45:31,720
to come back with that answer to show that this is generally on average what a person's

614
00:45:31,720 --> 00:45:34,680
response will be to this question.

615
00:45:34,680 --> 00:45:36,480
Fascinates me, absolutely.

616
00:45:36,480 --> 00:45:44,600
But also as an ethicist, I think there's a little bit of, my concern comes into the point

617
00:45:44,600 --> 00:45:51,120
that we forget the limitations of this technology in the sense that it is really good at giving

618
00:45:51,120 --> 00:45:52,640
us a good average.

619
00:45:52,640 --> 00:45:56,800
I heard described once as a mediocre consultant, which is true.

620
00:45:56,800 --> 00:45:57,800
It's a mediocre consultant.

621
00:45:57,800 --> 00:45:59,960
You're like, yeah, that's good enough.

622
00:45:59,960 --> 00:46:03,520
And so you can use it for things that are, that you only really need something that's

623
00:46:03,520 --> 00:46:05,400
good enough.

624
00:46:05,400 --> 00:46:09,160
But the ideation and the creation still belongs to the people.

625
00:46:09,160 --> 00:46:15,680
And my only concern is that we forget that in this process where the default becomes

626
00:46:15,680 --> 00:46:16,880
we'll chat GPT set it.

627
00:46:16,880 --> 00:46:19,040
So that's got to be the best answer.

628
00:46:19,040 --> 00:46:20,040
That's not true.

629
00:46:20,040 --> 00:46:24,480
It's the best mediocre answer, but now it's up to you as the person to then push it to

630
00:46:24,480 --> 00:46:25,480
the next level.

631
00:46:25,480 --> 00:46:30,880
Yeah, I 100% agree with that and it's something that we've spoken about on the podcast recently

632
00:46:30,880 --> 00:46:36,800
is that it's actually great for if you get a problem solving, maybe you've got a particular

633
00:46:36,800 --> 00:46:38,240
situation with a client.

634
00:46:38,240 --> 00:46:41,880
This is a kind of real world consultancy or agency scenario.

635
00:46:41,880 --> 00:46:44,760
I've had recently where I've had a situation.

636
00:46:44,760 --> 00:46:51,000
I've thrown the situation into chat GPT and said, give me all the things I need to consider

637
00:46:51,000 --> 00:46:57,440
with this particular thing and it's kind of spat out 10, 12 considerations.

638
00:46:57,440 --> 00:47:03,480
Eight of them I'd already considered, two or three of them are kind of nonsense and

639
00:47:03,480 --> 00:47:04,680
almost irrelevant.

640
00:47:04,680 --> 00:47:08,080
But one of them is, is the kind of nugget of gold that I'd overlooked.

641
00:47:08,080 --> 00:47:12,720
Like it was, oh yeah, that's giving me a bit of a sense check there.

642
00:47:12,720 --> 00:47:20,280
But if you were, if you were treating it as infallible as the single source of truth,

643
00:47:20,280 --> 00:47:25,320
then you're going to incorporate those two or three options that were irrelevant or a

644
00:47:25,320 --> 00:47:26,320
bit of a nonsense.

645
00:47:26,320 --> 00:47:31,360
So yeah, really important to bring that discretionary element to it.

646
00:47:31,360 --> 00:47:40,880
I am, at the opening you mentioned about values and the ethical values of an organisation.

647
00:47:40,880 --> 00:47:45,920
And I think that's a really interesting perspective.

648
00:47:45,920 --> 00:47:52,000
Can you talk about the work that you do in terms of helping organisations distill or

649
00:47:52,000 --> 00:47:55,480
kind of articulate their values?

650
00:47:55,480 --> 00:48:03,960
And then how do we align the practices, the use of technology to those values?

651
00:48:03,960 --> 00:48:04,960
Yeah.

652
00:48:04,960 --> 00:48:11,920
So I use a very specific technique actually when working with clients to distill their,

653
00:48:11,920 --> 00:48:14,480
call it like a foundational value set.

654
00:48:14,480 --> 00:48:17,920
These are the foundational values that they're going to be working off of.

655
00:48:17,920 --> 00:48:23,440
So we're pulling in information from three different sources.

656
00:48:23,440 --> 00:48:26,640
First one being regulation or policy that's existing.

657
00:48:26,640 --> 00:48:31,340
So looking at that country that the company is operating within, what are the legal systems

658
00:48:31,340 --> 00:48:33,440
that it needs to adhere to?

659
00:48:33,440 --> 00:48:37,760
So for example, if you, I'm in Belgium right now.

660
00:48:37,760 --> 00:48:42,560
If you are operating in Belgium, then you are going to be looking to the EU AI Act.

661
00:48:42,560 --> 00:48:49,000
It does have policies and specific values that are outlined within that regulation and

662
00:48:49,000 --> 00:48:50,000
that legislation.

663
00:48:50,000 --> 00:48:54,480
So the first input is regulation.

664
00:48:54,480 --> 00:48:56,320
It's on the country level.

665
00:48:56,320 --> 00:48:59,000
Second input, you're looking at industry standards.

666
00:48:59,000 --> 00:49:05,680
So the clearest example of this one would be a company operating in the healthcare sector.

667
00:49:05,680 --> 00:49:09,240
They're going to be looking at things like the Hippocratic oath and medical ethics.

668
00:49:09,240 --> 00:49:13,400
These are longstanding values and standards within the industry.

669
00:49:13,400 --> 00:49:19,600
Some industries have more robust and easily identifiable value sets than others, but there

670
00:49:19,600 --> 00:49:25,840
is always this kind of, think like standard practice within an industry.

671
00:49:25,840 --> 00:49:27,320
So that's the second input.

672
00:49:27,320 --> 00:49:31,680
Third input is actually the company itself, the company values, which the company should

673
00:49:31,680 --> 00:49:33,860
have values at this point in time.

674
00:49:33,860 --> 00:49:38,440
And these are ones that you can either find written on the wall or say, here are company

675
00:49:38,440 --> 00:49:42,360
values and this is what our brand embodies and our mission, distilling out from there

676
00:49:42,360 --> 00:49:45,040
one of the values there.

677
00:49:45,040 --> 00:49:49,640
Those three inputs, there is a process to actually triangulate between the three.

678
00:49:49,640 --> 00:49:52,160
We are looking for commonalities between them.

679
00:49:52,160 --> 00:49:58,960
So when you find a value such as, I'll use a Buzz One trust that hits on all three of

680
00:49:58,960 --> 00:50:01,680
those value sets, that's a core value.

681
00:50:01,680 --> 00:50:05,600
That means no matter where you're looking, you need to have that.

682
00:50:05,600 --> 00:50:11,080
Let's say you have a value like transparency that's only showing up on two of those inputs.

683
00:50:11,080 --> 00:50:16,840
That's still an important value, but you're less likely going to be using that as a core

684
00:50:16,840 --> 00:50:20,360
value and more of one that helps you with prioritization.

685
00:50:20,360 --> 00:50:25,440
And let's say detailed decision making, that one will be weighing in.

686
00:50:25,440 --> 00:50:32,620
So it's actually a very clear process of bringing together this core foundational value set.

687
00:50:32,620 --> 00:50:35,960
And then to answer the second part of your question, when it comes to decision making

688
00:50:35,960 --> 00:50:39,200
off of that, there are two approaches.

689
00:50:39,200 --> 00:50:44,760
So you're looking at a value, let's say trust again, and you can either take a risk based

690
00:50:44,760 --> 00:50:47,440
approach or an innovation based approach.

691
00:50:47,440 --> 00:50:51,720
The risk based approach is you are trying to protect for that value.

692
00:50:51,720 --> 00:50:56,480
You are trying to prevent that value from going wrong.

693
00:50:56,480 --> 00:51:00,400
So actually, let's use privacy.

694
00:51:00,400 --> 00:51:02,720
This one's an easier example case.

695
00:51:02,720 --> 00:51:04,600
So privacy, you're protecting for privacy.

696
00:51:04,600 --> 00:51:06,800
That means that you're following the GDPR.

697
00:51:06,800 --> 00:51:12,420
You're ensuring that you're compliant in terms of data collection policies and practices.

698
00:51:12,420 --> 00:51:18,200
It's very focused on, I am protecting and preventing from things going wrong.

699
00:51:18,200 --> 00:51:22,280
The innovation side though, you're looking more at alignment and design.

700
00:51:22,280 --> 00:51:27,940
So that's something where you're going to see privacy by design, meaning you are incorporating

701
00:51:27,940 --> 00:51:33,040
elements that are respectful of the user's privacy from the very start, from the very

702
00:51:33,040 --> 00:51:34,600
design itself.

703
00:51:34,600 --> 00:51:40,040
So both of these approaches are very important and it's good to have a balance between the

704
00:51:40,040 --> 00:51:44,680
two, but companies will naturally tend towards one versus the other.

705
00:51:44,680 --> 00:51:51,480
So you protect or you align and that depends on the company itself and the objectives around

706
00:51:51,480 --> 00:51:52,480
this.

707
00:51:52,480 --> 00:51:57,000
But to summarize, you have an easy process for finding your values and then you're either

708
00:51:57,000 --> 00:52:00,400
looking to protect or align with those values.

709
00:52:00,400 --> 00:52:07,040
Having sat in multiple sessions with leadership teams, looking at values, I feel like you're

710
00:52:07,040 --> 00:52:10,480
downplaying how easy it is to arrive at those set of values.

711
00:52:10,480 --> 00:52:13,000
I know what those workshops can be like.

712
00:52:13,000 --> 00:52:16,280
I always have this problem because I say, oh, it's easy.

713
00:52:16,280 --> 00:52:19,320
And then I had a friend saying, you really have to stop saying that because it's not

714
00:52:19,320 --> 00:52:20,320
easy.

715
00:52:20,320 --> 00:52:21,320
It's easy for you.

716
00:52:21,320 --> 00:52:22,320
It's like, oh yeah, that's fair.

717
00:52:22,320 --> 00:52:23,740
I've done this enough.

718
00:52:23,740 --> 00:52:24,740
It's easy for me.

719
00:52:24,740 --> 00:52:25,740
It's very fun for me.

720
00:52:25,740 --> 00:52:30,440
I think it's really fascinating conversations and you really have to pay attention as people

721
00:52:30,440 --> 00:52:34,360
are talking to pull out the common threads.

722
00:52:34,360 --> 00:52:35,360
But that is not easy.

723
00:52:35,360 --> 00:52:36,360
Let me rephrase that.

724
00:52:36,360 --> 00:52:39,760
It is not easy, but there is a clear process.

725
00:52:39,760 --> 00:52:42,480
Yeah, I grant you that.

726
00:52:42,480 --> 00:52:49,400
I love that little framework of the regulatory, the industry and the corporate.

727
00:52:49,400 --> 00:52:52,800
I think that's a really nice methodology.

728
00:52:52,800 --> 00:52:58,000
So yeah, thanks for sharing that.

729
00:52:58,000 --> 00:53:08,440
So when it comes to taking the next steps with this, you have developed something called

730
00:53:08,440 --> 00:53:10,200
the ethics maturity continuum.

731
00:53:10,200 --> 00:53:11,200
Is that right?

732
00:53:11,200 --> 00:53:15,080
Is that the next stage of the process in terms of the engagement that you would take someone

733
00:53:15,080 --> 00:53:16,560
with or have I misunderstood?

734
00:53:16,560 --> 00:53:19,640
Can you talk us through what that continuum is?

735
00:53:19,640 --> 00:53:25,160
Yeah, so the ethics maturity continuum is designed specifically to be used by either

736
00:53:25,160 --> 00:53:30,200
venture capitalists and investors or by startups.

737
00:53:30,200 --> 00:53:33,700
There are two different layers.

738
00:53:33,700 --> 00:53:35,200
Let's call that to the continuum.

739
00:53:35,200 --> 00:53:37,360
One is for younger startups.

740
00:53:37,360 --> 00:53:43,600
So you're looking C to series B. And then the second layer is for series C and above

741
00:53:43,600 --> 00:53:45,680
in terms of investment rounds.

742
00:53:45,680 --> 00:53:48,680
And this is specifically for startups.

743
00:53:48,680 --> 00:53:54,760
You can actually take this continuum and use it within teams within larger companies.

744
00:53:54,760 --> 00:53:59,000
But the continuum is basically designed to help catch blind spots.

745
00:53:59,000 --> 00:54:05,160
So it's got five different value pillars and with each of those prompting questions.

746
00:54:05,160 --> 00:54:12,040
And you are self-assessing, but you're looking at, yes, I have this in place or I don't have

747
00:54:12,040 --> 00:54:14,320
that in place or this sounds this.

748
00:54:14,320 --> 00:54:19,120
I identify most with this description versus this one.

749
00:54:19,120 --> 00:54:27,160
And what it does is it, I know I'm talking very, very vaguely, but it allows you to understand

750
00:54:27,160 --> 00:54:30,880
your risk score when you're just starting out.

751
00:54:30,880 --> 00:54:38,720
And these five pillars have been focused on what are the main risk spots basically that

752
00:54:38,720 --> 00:54:44,640
a startup will encounter when first developing AI and AI product.

753
00:54:44,640 --> 00:54:50,720
So the continuum allows venture capitalists to either assess for ethical risk in a potential

754
00:54:50,720 --> 00:54:56,540
investment or it allows startups to actually be able to catch a blind spot and be able

755
00:54:56,540 --> 00:54:59,920
to showcase to investors, stakeholders and so on.

756
00:54:59,920 --> 00:55:02,320
Look, this is where we land within the continuum.

757
00:55:02,320 --> 00:55:03,320
Interesting.

758
00:55:03,320 --> 00:55:06,840
I guess I think I misunderstood the application of it there.

759
00:55:06,840 --> 00:55:12,280
So that's, that is an interesting one to me because immediately I think to myself, well,

760
00:55:12,280 --> 00:55:18,200
I can see how, you know, even mature teams that haven't been through that process themselves,

761
00:55:18,200 --> 00:55:22,160
that sounds like it would be, it would be useful there as well.

762
00:55:22,160 --> 00:55:25,640
I mean, ethics is, you know, you talk about blind spots, right?

763
00:55:25,640 --> 00:55:31,600
I would say that ethics is often a blind spot.

764
00:55:31,600 --> 00:55:37,520
It doesn't come up all the time in conversations.

765
00:55:37,520 --> 00:55:45,640
I mean, certainly with product teams, marketing teams that I engage with, it's a regular blind

766
00:55:45,640 --> 00:55:46,640
spot, right?

767
00:55:46,640 --> 00:55:52,120
There's, they understand the basics of things like respecting people's privacy, but it tends

768
00:55:52,120 --> 00:55:57,840
to be very much on that kind of what do we have to do on a legal compliance perspective,

769
00:55:57,840 --> 00:56:03,760
what the kind of wider ethical, what are our values, how do we want to operate basis?

770
00:56:03,760 --> 00:56:10,920
So yeah, I maybe understood the ethics maturity continuum to be applicable to wider teams,

771
00:56:10,920 --> 00:56:13,680
but no, that's interesting that it's for startups.

772
00:56:13,680 --> 00:56:18,840
Yeah, it's, it was specifically designed for startups, but there are different frameworks

773
00:56:18,840 --> 00:56:22,900
out there that do exist for teams beyond startups.

774
00:56:22,900 --> 00:56:28,200
You can find frameworks and processes that are openly accessible.

775
00:56:28,200 --> 00:56:34,400
You do have to, let's say, customize them to your own, to your own needs.

776
00:56:34,400 --> 00:56:39,080
But the purpose of those is to actually be able to walk teams through specific decision

777
00:56:39,080 --> 00:56:43,900
making or specific value implementation.

778
00:56:43,900 --> 00:56:46,640
For example, there's a lot of fairness frameworks out there.

779
00:56:46,640 --> 00:56:51,600
Fairness is a very difficult value to implement, but there's a lot of different fairness frameworks

780
00:56:51,600 --> 00:56:57,040
that allow a company to see this, these are the questions and when and where we need to

781
00:56:57,040 --> 00:57:03,440
ask these questions to ensure that we are doing to the best of our ability, elimination

782
00:57:03,440 --> 00:57:04,800
of unwanted bias.

783
00:57:04,800 --> 00:57:08,720
Yeah, that's a, an important consideration.

784
00:57:08,720 --> 00:57:15,500
And just, I can see how that's very applicable to companies developing the tech.

785
00:57:15,500 --> 00:57:21,960
If we were to look at companies that are, take a typical medium sized enterprise at

786
00:57:21,960 --> 00:57:28,680
the moment that is, they're not developing any tech themselves, but they're looking to

787
00:57:28,680 --> 00:57:34,020
implement these existing technologies into their workflows.

788
00:57:34,020 --> 00:57:40,080
Maybe they're looking at the likes of, you know, open AI, seeing all of the hype around

789
00:57:40,080 --> 00:57:42,880
the products that they're putting out and you know, they've done their, they've done

790
00:57:42,880 --> 00:57:48,080
a bit of due diligence and they've, they've seen that open AI has spent a lot of money

791
00:57:48,080 --> 00:57:51,400
on red teaming and things like that.

792
00:57:51,400 --> 00:57:53,920
And they've gone, yeah, people are, people are on this.

793
00:57:53,920 --> 00:57:55,640
Someone's thought about this.

794
00:57:55,640 --> 00:57:56,640
It's a safe product.

795
00:57:56,640 --> 00:57:59,320
What would you say to those teams?

796
00:57:59,320 --> 00:58:07,640
How should they approach the, the implementation of AI tech into their workflows?

797
00:58:07,640 --> 00:58:13,300
So I would say you want to focus on two points here.

798
00:58:13,300 --> 00:58:21,040
One actually being during the procurement process of truly analyzing a supplier.

799
00:58:21,040 --> 00:58:27,520
I've seen companies have a lot of success in actually developing what is an ethics procurement

800
00:58:27,520 --> 00:58:35,040
questionnaire that allows them to assess not necessarily, and I want to stress this, not

801
00:58:35,040 --> 00:58:41,960
necessarily if the company has done all of the, let's say right things, meaning having

802
00:58:41,960 --> 00:58:48,400
a red team, having a responsible AI policy and so on, but more looking at assessing if

803
00:58:48,400 --> 00:58:55,440
that supplier company, the values of that supplier company align with the company that

804
00:58:55,440 --> 00:58:57,720
is procuring the technology.

805
00:58:57,720 --> 00:59:06,000
Because you can have a company that is developing, let's say a safe product and they are going

806
00:59:06,000 --> 00:59:12,440
through the right processes, but they prioritize transparency over privacy.

807
00:59:12,440 --> 00:59:19,800
And you are a health tech company and you're procuring this, this, this information, you're

808
00:59:19,800 --> 00:59:28,040
procuring this, this, let's say AI and you need to prioritize privacy over transparency

809
00:59:28,040 --> 00:59:32,480
because of the sensitivity of the data that goes through, that's going to go through that

810
00:59:32,480 --> 00:59:35,160
system for you.

811
00:59:35,160 --> 00:59:43,280
That although small difference in prioritization can result in very different outcomes for

812
00:59:43,280 --> 00:59:44,280
a company.

813
00:59:44,280 --> 00:59:49,760
So with those procurement processes, you're looking for alignment in, let's say prioritization

814
00:59:49,760 --> 00:59:50,760
of values.

815
00:59:50,760 --> 00:59:56,720
Of course, also checking to make sure that they've done the right things, but even more

816
00:59:56,720 --> 00:59:59,720
so looking at if those values are in alignment.

817
00:59:59,720 --> 01:00:00,720
That's one.

818
01:00:00,720 --> 01:00:06,440
One point, the second point here for teams that are looking to implement procured AI

819
01:00:06,440 --> 01:00:11,840
is ensuring that, and it sounds very simple, but ensuring there is a human in the loop

820
01:00:11,840 --> 01:00:14,200
monitoring the progress.

821
01:00:14,200 --> 01:00:19,160
So there's this misconception of I am buying a full package.

822
01:00:19,160 --> 01:00:20,160
It's good.

823
01:00:20,160 --> 01:00:21,160
It's done.

824
01:00:21,160 --> 01:00:22,720
I can just put it running.

825
01:00:22,720 --> 01:00:24,600
No problem.

826
01:00:24,600 --> 01:00:25,600
Not how AI works.

827
01:00:25,600 --> 01:00:31,240
When a model is live, when you've been, when you have deployed your system, it's now taking

828
01:00:31,240 --> 01:00:37,120
in live data and that live data is going to, let's say influence the model.

829
01:00:37,120 --> 01:00:41,840
So you do need someone that's actually monitoring the output closely to ensure, Hey, this is,

830
01:00:41,840 --> 01:00:43,160
this is drifting.

831
01:00:43,160 --> 01:00:44,160
It happens all the time.

832
01:00:44,160 --> 01:00:45,160
You have model drift.

833
01:00:45,160 --> 01:00:49,040
This is drifting from what we originally intended it for.

834
01:00:49,040 --> 01:00:50,040
It's drifting.

835
01:00:50,040 --> 01:00:53,240
We're not getting the results that we wanted out of it versus yes, it's good.

836
01:00:53,240 --> 01:00:54,640
It's an alignment.

837
01:00:54,640 --> 01:00:55,800
It's doing what we wanted it to.

838
01:00:55,800 --> 01:00:58,340
We haven't caught any, any challenges.

839
01:00:58,340 --> 01:01:02,200
You need someone in that, that monitoring position.

840
01:01:02,200 --> 01:01:07,040
That's actually one of the most important positions is monitoring that, that human in

841
01:01:07,040 --> 01:01:13,160
the loop, making sure that the deployed system is continuing to perform as you expect it

842
01:01:13,160 --> 01:01:19,520
to, because it will change and you need to be able to, to identify that as fast as you

843
01:01:19,520 --> 01:01:23,400
can and catch it before it's gone too far, let's say.

844
01:01:23,400 --> 01:01:28,480
And I think that goes to that earlier point about putting too much trust and faith into

845
01:01:28,480 --> 01:01:29,480
chat GPT.

846
01:01:29,480 --> 01:01:31,080
It's the same thing with any model, isn't it?

847
01:01:31,080 --> 01:01:34,360
You just kind of set it up and go, Oh no, it knows what it's doing.

848
01:01:34,360 --> 01:01:36,320
We've, we've, we've built it.

849
01:01:36,320 --> 01:01:37,320
Off it goes.

850
01:01:37,320 --> 01:01:38,320
I'm nice.

851
01:01:38,320 --> 01:01:42,480
I think that's, that's a really important, important point.

852
01:01:42,480 --> 01:01:47,440
You just going back to something you mentioned earlier about the, the regulatory landscape.

853
01:01:47,440 --> 01:01:52,120
You mentioned the EU AI act.

854
01:01:52,120 --> 01:01:58,660
That's going through European parliament as we speak.

855
01:01:58,660 --> 01:01:59,660
What are your thoughts on it?

856
01:01:59,660 --> 01:02:03,120
It takes a person centered risk-based approach.

857
01:02:03,120 --> 01:02:05,680
Do you, do you think it's, it's good?

858
01:02:05,680 --> 01:02:10,440
Do you think it goes far enough to view any, any critiques of it?

859
01:02:10,440 --> 01:02:12,800
Well it is risk-based.

860
01:02:12,800 --> 01:02:19,000
I as a personal preference tend to prefer innovation based, but that's my personal preference

861
01:02:19,000 --> 01:02:20,840
as, as just an ethicist.

862
01:02:20,840 --> 01:02:29,520
I'm very much a design thinker, but I would say with the EU AI act, I am concerned with

863
01:02:29,520 --> 01:02:36,560
some of the risk categorizations because you can develop a system in a low risk category

864
01:02:36,560 --> 01:02:44,960
and then let's say apply it into a high risk category and potentially.

865
01:02:44,960 --> 01:02:50,200
So examples like that, there's, there's a few holes that I, that I think will be covered

866
01:02:50,200 --> 01:02:58,520
once the, once the regulation is in effect, but kind of like the GDPR with legitimate

867
01:02:58,520 --> 01:03:02,440
interest hole where everything's suddenly legitimate interest.

868
01:03:02,440 --> 01:03:05,880
I think we're going to see very similar aspects of that with the AI act.

869
01:03:05,880 --> 01:03:15,720
I'm also curious to see how it works in terms of application.

870
01:03:15,720 --> 01:03:21,600
I know for example, a lot of member states still struggle with actually having their

871
01:03:21,600 --> 01:03:26,120
data protection officers, DPOs in place for the GDPR.

872
01:03:26,120 --> 01:03:30,520
We're still struggling with GDPR implementation.

873
01:03:30,520 --> 01:03:37,960
The one place that I find it falling short and because all of the things that I mentioned

874
01:03:37,960 --> 01:03:42,840
that can be corrected over time and we do just need to basically, well, I'll speak techie

875
01:03:42,840 --> 01:03:48,800
here deploy it so that we can actually see how it goes to be able to, to, to correct

876
01:03:48,800 --> 01:03:50,320
for those, those blind spots.

877
01:03:50,320 --> 01:03:57,680
Yeah, the, the hole that I see that I think has me most concerned is actually the startup

878
01:03:57,680 --> 01:04:03,680
ecosystem here in, here in Europe and any startup that's looking to, to operate within

879
01:04:03,680 --> 01:04:05,200
the European market.

880
01:04:05,200 --> 01:04:09,800
Not because regulation is, is a constraint on innovation.

881
01:04:09,800 --> 01:04:17,120
I think quite the opposite, but instead because this regulation, this, this, this AI act,

882
01:04:17,120 --> 01:04:25,760
it puts a heavy emphasis on documentation and auditing that costs time, money, resources

883
01:04:25,760 --> 01:04:27,600
that costs a lot.

884
01:04:27,600 --> 01:04:32,840
And so I have heard investors specifically say post AI act, we're not investing in any

885
01:04:32,840 --> 01:04:35,560
high risk categories.

886
01:04:35,560 --> 01:04:49,400
So my fear, unless the EU sets up specific, let's say institutions or resources for startups,

887
01:04:49,400 --> 01:04:53,600
they're going to cut off a lot of innovation that's going, that could happen in the higher

888
01:04:53,600 --> 01:04:58,400
risk that could happen totally fine, responsibly in the high risk, but only because startups

889
01:04:58,400 --> 01:05:04,080
do not have the resources to comply with the documentation required for those risk categories.

890
01:05:04,080 --> 01:05:06,120
So that's, that's my big concern.

891
01:05:06,120 --> 01:05:12,320
I haven't seen enough movement there in terms of covering the potential consequences for

892
01:05:12,320 --> 01:05:13,320
that.

893
01:05:13,320 --> 01:05:20,400
Yeah, that's, I think the auditability, the, the being able to interrogate the models and

894
01:05:20,400 --> 01:05:24,880
understand how it's arrived at decisions and what have you is such a big part of it.

895
01:05:24,880 --> 01:05:25,880
Yeah.

896
01:05:25,880 --> 01:05:31,200
I think that you can see why, why people would be concerned if they were building from the

897
01:05:31,200 --> 01:05:35,800
ground up, like you say, getting the systems and processes in place to do that could be

898
01:05:35,800 --> 01:05:38,000
quite a serious ask.

899
01:05:38,000 --> 01:05:39,080
Right.

900
01:05:39,080 --> 01:05:47,240
We are heading towards the end of, of this podcast, but I would like to jump into the

901
01:05:47,240 --> 01:05:53,000
role of marketing teams specifically as it relates to responsible AI.

902
01:05:53,000 --> 01:05:56,080
Is that something that you've looked at in great detail?

903
01:05:56,080 --> 01:06:02,760
Have you got any advice for the marketing professionals working in and around this,

904
01:06:02,760 --> 01:06:03,760
this tech?

905
01:06:03,760 --> 01:06:05,840
Yes, absolutely.

906
01:06:05,840 --> 01:06:10,800
And marketing is one of the interesting ones because it's often overlooked as having any

907
01:06:10,800 --> 01:06:14,400
type of impact or role in responsible AI.

908
01:06:14,400 --> 01:06:21,200
What I say responsibly, I mean, the industry of, of let's say AI governance, ethics, safety,

909
01:06:21,200 --> 01:06:23,000
risk, et cetera, et cetera.

910
01:06:23,000 --> 01:06:27,880
This all fits under the term responsibly AI and typically people think, oh, responsibly

911
01:06:27,880 --> 01:06:33,840
AI is for tech teams, but marketing teams and marketing professionals have a very important

912
01:06:33,840 --> 01:06:38,320
role to play really in two different directions.

913
01:06:38,320 --> 01:06:41,000
The first one is in communication.

914
01:06:41,000 --> 01:06:48,760
So when it comes to responsible AI, there is this risk of something called ethics washing

915
01:06:48,760 --> 01:06:53,680
where a company says, look at us, we're so ethical and then you lift up the, the hood,

916
01:06:53,680 --> 01:06:56,120
you look behind the curtain, you're like, woof, what's going on here?

917
01:06:56,120 --> 01:06:57,120
That is not right.

918
01:06:57,120 --> 01:07:02,600
It's kind of like, it's also known as blue washing, which is, it's similar to, if you've

919
01:07:02,600 --> 01:07:07,080
ever heard the term green washing where companies say, look, we're so environmentally friendly.

920
01:07:07,080 --> 01:07:11,360
And then you pull back the curtain, you're like, you're dumping the gallons of oil into

921
01:07:11,360 --> 01:07:12,360
the ocean.

922
01:07:12,360 --> 01:07:14,760
I don't call that environmentally friendly.

923
01:07:14,760 --> 01:07:19,760
Similar with ethics where, where a company can say, look, we have our values.

924
01:07:19,760 --> 01:07:22,160
And then you look and you're like, well, but you don't use them.

925
01:07:22,160 --> 01:07:25,360
So that's something called ethics washing.

926
01:07:25,360 --> 01:07:30,760
And what that does is it creates a very intense mistrust with, with a customer base.

927
01:07:30,760 --> 01:07:35,840
So customers can see through, you can say, we respect your privacy.

928
01:07:35,840 --> 01:07:39,280
And then a customer is actually going through the user experience and they're having to

929
01:07:39,280 --> 01:07:42,200
submit like their mother's maiden name.

930
01:07:42,200 --> 01:07:46,080
And they're thinking, that's not, no, why do you need that?

931
01:07:46,080 --> 01:07:49,460
You're telling me one thing and you're acting in a, in another way.

932
01:07:49,460 --> 01:07:51,240
That doesn't sit right with me.

933
01:07:51,240 --> 01:07:58,960
So for marketing professionals, being able to accurately represent and communicate with

934
01:07:58,960 --> 01:08:05,000
a user base, how the company approaches responsibly, or how the company approaches ethics is incredibly

935
01:08:05,000 --> 01:08:09,360
important because that's what's going to be a huge trust builder.

936
01:08:09,360 --> 01:08:13,560
You can have all the work being done behind the scenes with the tech teams, ensuring that

937
01:08:13,560 --> 01:08:16,760
there are strong ethical solutions in place.

938
01:08:16,760 --> 01:08:21,600
But if you're not being able to communicate with that user base through your marketing,

939
01:08:21,600 --> 01:08:27,880
that that is done in a way that is holistic, then you are both going to miss out on the

940
01:08:27,880 --> 01:08:32,520
benefits of responsibility, but also you risk creating this kind of mistrust with your user

941
01:08:32,520 --> 01:08:33,520
base.

942
01:08:33,520 --> 01:08:39,200
So on one hand, marketers have this very important role of that trust building communication

943
01:08:39,200 --> 01:08:45,520
with user bases, but then market marketers can also have the ability to be a fantastic

944
01:08:45,520 --> 01:08:47,320
resource for feedback.

945
01:08:47,320 --> 01:08:54,360
So for example, if they're getting feedback from their, from the audience saying, you,

946
01:08:54,360 --> 01:09:00,800
we don't trust you, or we think that you don't have very fair practices, that's a great point

947
01:09:00,800 --> 01:09:04,880
in time to feed that information back to your tech and your product and your leadership

948
01:09:04,880 --> 01:09:05,940
teams.

949
01:09:05,940 --> 01:09:10,080
So marketing really actually has a very important role, they're the trust builders and they're

950
01:09:10,080 --> 01:09:18,520
also that insight point of insight for teams to be able to better connect with their user

951
01:09:18,520 --> 01:09:19,520
base.

952
01:09:19,520 --> 01:09:24,680
And this is all under the umbrella of responsible AI and successful implementation of ethics

953
01:09:24,680 --> 01:09:26,520
and values and technology.

954
01:09:26,520 --> 01:09:33,120
So a big role for marketing to play in the communication and interface between the customers

955
01:09:33,120 --> 01:09:36,680
and the tech teams, an important role to play.

956
01:09:36,680 --> 01:09:37,680
Fantastic.

957
01:09:37,680 --> 01:09:39,880
So, well, thank you for joining us.

958
01:09:39,880 --> 01:09:44,520
I believe, am I right in saying you've got a book in development?

959
01:09:44,520 --> 01:09:45,520
Yes.

960
01:09:45,520 --> 01:09:46,520
Yes.

961
01:09:46,520 --> 01:09:51,280
I am currently at the phase of dear God, what am I doing?

962
01:09:51,280 --> 01:09:55,800
So I'm about two thirds of the way through, but my book is coming.

963
01:09:55,800 --> 01:10:04,120
I'm with Kogan Publishing and publication is set for summer, June 2024.

964
01:10:04,120 --> 01:10:08,480
My title is Responsible AI, Implement an Ethical Approach in Your Organization.

965
01:10:08,480 --> 01:10:09,480
Big surprise.

966
01:10:09,480 --> 01:10:16,080
Well, I would love to have you back on closer to release date to talk about responsible

967
01:10:16,080 --> 01:10:19,900
AI and the book launch in general.

968
01:10:19,900 --> 01:10:21,800
Where can people find you online?

969
01:10:21,800 --> 01:10:28,240
Absolutely, you can check me out at oliviagamblin.com and I am actually going to be starting a newsletter

970
01:10:28,240 --> 01:10:35,240
soon all around this idea of pursuit of good tech and the crosshairs of ethics, design,

971
01:10:35,240 --> 01:10:37,480
thinking and leadership in AI.

972
01:10:37,480 --> 01:10:39,880
So we'll see how that goes.

973
01:10:39,880 --> 01:10:41,880
But you can sign up on my website.

974
01:10:41,880 --> 01:10:45,480
You can also find me on LinkedIn under the name Olivia Gamblin.

975
01:10:45,480 --> 01:10:47,980
I do respond to messages on LinkedIn.

976
01:10:47,980 --> 01:10:51,480
It sometimes takes me a while, but I do respond.

977
01:10:51,480 --> 01:10:52,480
Fantastic.

978
01:10:52,480 --> 01:10:56,480
Well, thank you for joining us on artificially intelligent marketing.

979
01:10:56,480 --> 01:10:58,880
Thank you for having me here today, Martin.

980
01:10:58,880 --> 01:11:01,680
Well, thanks for hosting great interview there, Martin.

981
01:11:01,680 --> 01:11:04,720
I hope the listeners enjoyed that and thank you for your time this week.

982
01:11:04,720 --> 01:11:06,940
As always, great to hang out with you.

983
01:11:06,940 --> 01:11:08,380
Good to see you too.

984
01:11:08,380 --> 01:11:13,600
Looking forward to next week's episode where there'll be even more great AI content.

985
01:11:13,600 --> 01:11:15,600
See you later, Paul.

986
01:11:15,600 --> 01:11:16,600
Cheers.

987
01:11:16,600 --> 01:11:17,600
Bye.

988
01:11:17,600 --> 01:11:23,240
Thank you for listening to artificially intelligent marketing to stay on top of the latest trends,

989
01:11:23,240 --> 01:11:27,140
tips and tools in the world of marketing AI.

990
01:11:27,140 --> 01:11:28,920
Be sure to subscribe.

991
01:11:28,920 --> 01:11:48,920
We look forward to seeing you again next week.