1
00:00:00,661 --> 00:00:12,981
And I get to introduce Travis Arndorfer, Creative Director with over 20 years experience leading strategy and creative work for organizations ranging from global brands to regional leaders across multiple industries.

2
00:00:13,701 --> 00:00:25,061
Travis has also founded and led an agency AI lab where he develops practical frameworks that integrate artificial intelligence into marketing workflows while maintaining strong strategic thinking and creative quality.

3
00:00:25,621 --> 00:00:31,861
His work focuses on using AI as a true partner, enhancing ideas rather than simply increasing output.

4
00:00:32,501 --> 00:00:42,901
In today's session, he will share the strategic stack framework, offering a clear approach to creating AI-assisted marketing that stands out, remains thoughtful, and delivers real impact.

5
00:00:43,141 --> 00:00:45,221
Please join me in welcoming Travis.

6
00:00:49,741 --> 00:00:50,181
Thank you.

7
00:00:50,341 --> 00:01:00,581
And thank you all for joining me here in this last session as we kind of tackle what I think of as a little bit of the elephant in the room when we talk AI.

8
00:01:00,661 --> 00:01:08,181
It's pretty exciting what AI can do, but also it can create a lot of troubles for ourselves.

9
00:01:08,341 --> 00:01:10,261
AI slot, as you've probably seen, is

10
00:01:11,381 --> 00:01:12,981
It's just everywhere.

11
00:01:13,541 --> 00:01:18,741
It's in really big, really expensive brand ads.

12
00:01:18,741 --> 00:01:24,981
I don't know if you saw Coca-Cola came out in the holiday season of 24.

13
00:01:25,461 --> 00:01:33,941
They did an AI version of their annual ad and it went over like a lead balloon because there were all kinds of broken things with it.

14
00:01:33,941 --> 00:01:35,621
Just it was very sloppy.

15
00:01:35,621 --> 00:01:40,101
So last year they doubled down, said, okay, we're going to give you 2 ads.

16
00:01:40,661 --> 00:01:43,621
And it was, in many cases, a repeat.

17
00:01:43,621 --> 00:01:51,301
We had a number of tires shifting, the truck was morphing, the physics were kind of wonky.

18
00:01:52,101 --> 00:01:53,381
It was a hot mess.

19
00:01:54,821 --> 00:02:00,581
To the point where there was some call for a boycott of Coke because of it.

20
00:02:01,821 --> 00:02:05,781
And it's interesting, and I think these first two, there's plenty of big brands I could have picked on.

21
00:02:05,781 --> 00:02:08,341
I could have picked on Nike, I could have picked on J.Crew.

22
00:02:09,141 --> 00:02:10,461
You name it.

23
00:02:11,221 --> 00:02:13,461
It's surprising, but many of them have stepped in it.

24
00:02:14,341 --> 00:02:35,701
But I picked these first two because I think what they're illustrating is something that we're seeing, that the more emotional something is when we're doing marketing that's trying to connect with people, and especially in an emotional space, and we're using AI to pull on it, and we do it in a sloppy way, consumers and buyers react vehemently.

25
00:02:37,221 --> 00:02:38,741
They do not like it.

26
00:02:40,981 --> 00:02:43,141
Coke is, I think, maybe still learning this.

27
00:02:43,141 --> 00:02:55,381
They actually, to the Hubbub this year, they came out with a rebuttal, which kind of a behind the scenes that was done by Notebook LM's voice and narrating it.

28
00:02:55,381 --> 00:02:59,301
So it did not go over real well for them.

29
00:02:59,461 --> 00:03:00,821
McDonald's stepped in it too.

30
00:03:00,821 --> 00:03:05,461
In the Netherlands, they actually had a really good idea in terms of a strategic

31
00:03:06,181 --> 00:03:12,341
starting place that the holidays are a really busy time of year and they can actually be sort of a struggle for us.

32
00:03:13,301 --> 00:03:16,581
But the execution was, again, a hot mess.

33
00:03:18,101 --> 00:03:25,301
And we had all kinds of just uncanny faces and broken physics.

34
00:03:25,781 --> 00:03:31,381
And it was so bad that within three days, they took it down.

35
00:03:32,741 --> 00:03:34,101
And I got Marvel up here.

36
00:03:34,101 --> 00:03:40,101
Marvel would tell you that the Fantastic Four poster did not have AI in it.

37
00:03:41,461 --> 00:03:45,861
Fans thought differently and they saw the same face in the crowd a couple of times.

38
00:03:46,581 --> 00:03:50,581
Okay, that can happen with some other things too, but it doesn't sound good.

39
00:03:51,221 --> 00:03:53,341
The fingers were wrong and some other things.

40
00:03:53,341 --> 00:03:55,861
That starts to smack a little bit more of AI.

41
00:03:56,261 --> 00:04:00,661
But the reason I put that one up here is because it doesn't really matter anymore.

42
00:04:01,501 --> 00:04:03,021
The bar is starting to change.

43
00:04:03,021 --> 00:04:06,501
Whether you use AI or not, people are sensitized to it.

44
00:04:06,501 --> 00:04:09,061
They're already on edge about it.

45
00:04:09,061 --> 00:04:13,861
And when they see slop, they're reacting to it really strong, strongly.

46
00:04:13,861 --> 00:04:18,061
I think we're getting a little feedback.

47
00:04:18,061 --> 00:04:18,981
Am I too close to this?

48
00:04:21,381 --> 00:04:21,781
Apologies.

49
00:04:21,781 --> 00:04:22,621
All right.

50
00:04:23,461 --> 00:04:26,981
We're seeing AI slop in all kinds of critical publications.

51
00:04:27,781 --> 00:04:35,941
was in the Maha report where we had, I think it was 7 different sources that we were trying to base the science on weren't there.

52
00:04:35,941 --> 00:04:37,061
They don't exist.

53
00:04:37,621 --> 00:04:47,181
There were dozens of other sources that had the appearance of coming right from ChatGPT, but they at least did exist.

54
00:04:47,541 --> 00:04:50,661
We're seeing it in tons of court cases.

55
00:04:50,661 --> 00:04:55,941
I think about, I'm not sure, about 90% of those 1300 plus cases

56
00:04:56,421 --> 00:04:57,541
were in the last year.

57
00:04:58,501 --> 00:05:02,661
So the momentum there is a little frightening.

58
00:05:03,381 --> 00:05:05,541
It's showing up in scientific journals.

59
00:05:05,861 --> 00:05:14,701
There's been over, I think it's like 335 scientific journal articles that have been entangled and eventually retracted because of it.

60
00:05:14,701 --> 00:05:21,941
And I think one of the insidious things there is that the average time to retract it is like 550 days.

61
00:05:21,941 --> 00:05:23,061
It's like a year and a half.

62
00:05:23,461 --> 00:05:29,141
So in that time, there's plenty of time for other articles to pick it up, cite it, and get into that cycle.

63
00:05:29,781 --> 00:05:33,181
And then there's all kinds of other articles.

64
00:05:33,181 --> 00:05:38,261
We see it from daily newspapers, the Sports Illustrated, it's all over.

65
00:05:39,141 --> 00:05:50,421
If your inbox is anything like mine, you're seeing it on a very personal level all the time too, from co-workers, from maybe even from clients or customers.

66
00:05:51,341 --> 00:05:57,141
And right now, it's believed that over half of LinkedIn articles that are longer than 100 words, that's...

67
00:05:57,261 --> 00:06:04,421
that are long form in LinkedIn world, few sentences, are believed to be written by AI.

68
00:06:05,261 --> 00:06:11,461
And then we're seeing it just in all kinds of, I would say, more everyday uses in small business like Logo.

69
00:06:12,021 --> 00:06:17,981
And again, I point out the Salty Otter because, again, I feel like it's a cautionary tale.

70
00:06:17,981 --> 00:06:20,341
This is a bar and grill in Santa Cruz.

71
00:06:21,141 --> 00:06:25,301
The owner had a dream to have this bar, but

72
00:06:25,941 --> 00:06:28,901
She's good at cooking, but maybe didn't have the design skills.

73
00:06:28,901 --> 00:06:33,901
So went to AI and AI the Otter later put in the text.

74
00:06:36,261 --> 00:06:39,381
Her Yelp page was just trash.

75
00:06:39,381 --> 00:06:44,581
It was one star review after one star review citing the logo.

76
00:06:45,901 --> 00:06:52,501
Now, four years ago, I swear, I've eaten in places that have this basic logo with the sun kind of set thing.

77
00:06:53,701 --> 00:06:54,661
It's not great.

78
00:06:55,381 --> 00:06:58,581
But again, I look at it as we are seeing a change.

79
00:06:58,581 --> 00:07:00,021
There's a shift in the bar.

80
00:07:00,021 --> 00:07:15,781
People are very sensitized, very on edge around sloppy use of AI, especially when it's seen as a cost cutting or human cutting, if you will, technique.

81
00:07:15,941 --> 00:07:19,061
I mean, that was a lot of the blowback with Coke is that

82
00:07:19,781 --> 00:07:29,141
in defense of it, the creative directors came out and were talking about how, oh, they were able to do it in 30 days instead of a year and all this money they've saved.

83
00:07:29,141 --> 00:07:31,861
And it's like, okay, he just stepped a bit deeper.

84
00:07:33,861 --> 00:07:41,541
So when we look at AI slop, a lot of us can see it, kind of smell it when we start to look at it.

85
00:07:41,861 --> 00:07:48,261
I like to think of it as the ultra-processed food of content because it looks like it should be good.

86
00:07:48,861 --> 00:07:53,461
Until we start to get into it and look at the nutritional analysis, there's not much there.

87
00:07:54,661 --> 00:08:04,981
AI is really good, and in fact, in many cases, better than humans at the contextual syntax and structure.

88
00:08:04,981 --> 00:08:07,541
So things look really, really good.

89
00:08:08,021 --> 00:08:09,781
They feel very official.

90
00:08:09,781 --> 00:08:11,461
It feels like it should be right.

91
00:08:12,141 --> 00:08:15,461
And we'll get to that in a little bit on the trouble that causes for us.

92
00:08:16,261 --> 00:08:23,541
But when we start to dig into it, the content length is there, the substance is often really light.

93
00:08:26,021 --> 00:08:36,261
And there's, a lot of slop happens because we think we're going to save time or money, but we maybe overlook some of the other costs that are hidden in there.

94
00:08:37,061 --> 00:08:43,781
Perhaps a little ancillary to our talk here today is, you know, I think the 1st place I look at is the work slop that maybe

95
00:08:44,981 --> 00:08:54,021
one person gets something done a little faster, but if they pass a riddle on to five or 10 other people, where is that being accounted for?

96
00:08:54,021 --> 00:08:59,621
And I expect we'll see a lot more about the hidden costs of WorkSlop.

97
00:09:00,341 --> 00:09:12,661
As Neil was just talking about, if you get crosswise with EEAT because you have whiz-banged yourself some content, that stuff can be down-ranked and you'll actually show up less.

98
00:09:13,061 --> 00:09:20,061
And I think for most of us here as marketers, we're trying to connect to our buyers and to our consumers.

99
00:09:20,061 --> 00:09:24,581
We're trying to make a meaningful connection so we can change their behavior.

100
00:09:25,381 --> 00:09:32,341
And if we invite or invoke disgust or anger, we're just shooting ourselves in the foot.

101
00:09:33,981 --> 00:09:37,061
And then the last point here, it's kind of a couple.

102
00:09:38,901 --> 00:09:42,861
It's easier than ever to flood the market with content.

103
00:09:42,861 --> 00:09:54,421
So we're seeing a higher bar of clutter, but we're also, as like the salty otter, it's sensitizing people and changing how they're reacting to what's out there.

104
00:09:55,141 --> 00:10:05,821
I had a creative director friend reach out to me a couple months ago and say, we had a client that pushed back on one of our articles and said, you know, this is AI's thought, like, what are you doing?

105
00:10:06,021 --> 00:10:07,381
Like, I thought we were friends.

106
00:10:07,741 --> 00:10:09,821
Like, why would you do this to us?

107
00:10:09,821 --> 00:10:11,941
Like, what's going on?

108
00:10:13,261 --> 00:10:17,061
And as they dig into it, it was more of a classic problem.

109
00:10:18,421 --> 00:10:24,501
Client hadn't maybe been forthcoming with content that was needed to produce a good article.

110
00:10:24,901 --> 00:10:27,781
Account service got nervous about the looming deadline.

111
00:10:28,141 --> 00:10:36,101
Next thing you know, what we have gets shooed along, and the writer does her level best, but the article ends up feeling fluffy, and the next thing you know,

112
00:10:38,101 --> 00:10:40,581
the client is kind of hot about it.

113
00:10:40,581 --> 00:10:42,901
And it tends to jump levels.

114
00:10:42,901 --> 00:10:46,661
It isn't that like, hey, this doesn't seem like it has enough stuff in it this time.

115
00:10:46,661 --> 00:10:49,621
It's like it's more of a breach of trust.

116
00:10:50,021 --> 00:10:58,261
So I think it's really important to keep that in mind, that whether we're using AI or not, it's affecting all of the work that we do and how it's being judged.

117
00:10:59,781 --> 00:11:06,021
And perhaps I'm a little naive, but I think most marketers turning out slop most aren't

118
00:11:06,501 --> 00:11:07,461
They aren't trying to.

119
00:11:09,461 --> 00:11:11,061
So why does it happen?

120
00:11:11,101 --> 00:11:16,261
Let's take a look a little bit at some of the things that create the tendencies for SLOP.

121
00:11:16,741 --> 00:11:22,421
And I think a good first place to start is with the tool, is with the LLMs.

122
00:11:23,781 --> 00:11:29,861
They're designed to predict that next most likely token.

123
00:11:29,861 --> 00:11:31,621
So the math behind them

124
00:11:32,981 --> 00:11:45,301
pushes them into, I mean, it constrains them to this, that to go to that comfortable middle, that common, and to avoid things that are more distinct and more surprising or rare.

125
00:11:45,701 --> 00:11:50,421
So there's some of that tendency is baked into the tool.

126
00:11:51,701 --> 00:11:59,101
But I think maybe what we don't realize is there's more of that baked in than what we had realized.

127
00:11:59,781 --> 00:12:03,141
So when a model comes out, it goes through alignment training.

128
00:12:03,781 --> 00:12:09,061
And there's a lot of research now showing, or some research showing now that there's an alignment tax of sorts.

129
00:12:09,301 --> 00:12:12,221
So alignment is to make the model safe.

130
00:12:12,221 --> 00:12:19,621
So it's producing things that are answers that we're not eating rocks, we're not doing unsafe things.

131
00:12:21,301 --> 00:12:22,421
It's those guardrails.

132
00:12:22,901 --> 00:12:29,301
And it has the overall effect of all of the queries that come from that model then,

133
00:12:29,781 --> 00:12:31,661
start to become more homogeneous.

134
00:12:32,261 --> 00:12:43,861
And it can go from rates, so a single cluster rate is like if you were to sample a model several times, how many of those times does it come back to the same basic answer?

135
00:12:44,501 --> 00:12:50,821
Before training, before the alignment training, it's typically about half a percent to a percent.

136
00:12:51,461 --> 00:12:54,181
After training, it can skyrocket.

137
00:12:54,421 --> 00:12:56,901
We can be, you know, as high as 79%.

138
00:12:57,621 --> 00:13:04,661
So there's real things going on with the tools that predispose some of this.

139
00:13:04,741 --> 00:13:19,621
It isn't all tools leave marks, all tools have ways at which they work, and good operators can learn to use those in ways that don't leave unwanted artifacts.

140
00:13:20,581 --> 00:13:27,261
Those things about the tools are real, and they do cause some of that generic sizing that we see in AI slot.

141
00:13:27,261 --> 00:13:29,661
But it's not the whole story.

142
00:13:29,661 --> 00:13:31,061
It doesn't make it inevitable.

143
00:13:32,061 --> 00:13:35,621
I think there's something more going on that has to do with us.

144
00:13:37,061 --> 00:13:39,941
This is a wonderful study that came out in March.

145
00:13:40,421 --> 00:13:50,181
Shaw and Navya out of the Wharton Business School looked at how we use AI and how we make decisions.

146
00:13:50,661 --> 00:13:59,861
So it's kind of building off of, if you're a fan of Daniel Kahneman's work with thinking fast and slow, it's building off of this, but kind of modernizing it.

147
00:14:00,261 --> 00:14:09,621
It's a wonderful study, but it found that four out of five people fouled AI down faulty logic, failed to override it.

148
00:14:10,541 --> 00:14:16,101
And there's lots of other kind of interesting tidbits within that study, but I think

149
00:14:16,621 --> 00:14:18,741
This should be a little bit of the wake-up calls.

150
00:14:18,741 --> 00:14:25,781
These are smart folks, and four out of five times, they went with AI when they shouldn't have.

151
00:14:27,701 --> 00:14:33,381
So to kind of unpack that, I think what starts to happen is what I call a cognitive cascade.

152
00:14:33,381 --> 00:14:35,541
There are things about the way our brain is built.

153
00:14:36,341 --> 00:14:41,781
Our brains are very metabolically expensive, and we

154
00:14:42,741 --> 00:14:51,861
We have evolutionarily designed to take some measures to stem that expenditure when it seems reasonable.

155
00:14:52,821 --> 00:14:59,101
And this fluency bias is that kind of the tip of the iceberg of that slide down the cognitive cascade.

156
00:14:59,101 --> 00:15:06,021
And that is when things look like a duck, quack like a duck, we're like, duck, and we move on.

157
00:15:07,541 --> 00:15:20,581
There's another bias here called the automation bias that we also have a tendency to prefer or to overvalue things that come from a machine that kind of like, poof, look at that.

158
00:15:21,221 --> 00:15:24,101
And those things together start to disarm us.

159
00:15:25,701 --> 00:15:35,381
And then we start to, and this is usually kind of a behind the scenes subconscious process, we start to see it as an authority.

160
00:15:35,781 --> 00:15:37,061
We start to defer.

161
00:15:37,941 --> 00:15:39,941
and we start to become passive.

162
00:15:40,741 --> 00:16:05,861
And interestingly, we get really confident about it, that Shaw and Navi's study showed that there's a double digit, I think it's like 11 percentage point increase in confidence when people used AI, even when in certain cases they had a 12, like a 12% chance of being right, like they under time pressure,

163
00:16:06,741 --> 00:16:10,261
You know, we tend to perform less well.

164
00:16:10,981 --> 00:16:14,101
And in those cases, it even went up higher.

165
00:16:14,501 --> 00:16:18,181
So this is where the cascade starts.

166
00:16:18,701 --> 00:16:20,181
But there's more to the picture.

167
00:16:20,341 --> 00:16:31,301
So our brains are wired in such a way that we tend to value things that we struggle for, where there's some amount of effort.

168
00:16:32,501 --> 00:16:33,701
It increases the value.

169
00:16:33,701 --> 00:16:36,661
It increases our investment in it and our ownership.

170
00:16:37,861 --> 00:16:44,421
And when AI can remove all the friction, subconsciously, we don't invest.

171
00:16:46,741 --> 00:16:52,181
And without those cues, we don't do the quality control.

172
00:16:52,181 --> 00:16:56,741
There's a direct relationship between a sense of psychological ownership

173
00:16:57,301 --> 00:17:25,301
and the quality control, if you will, that we put into something. If I value it and I feel like it's mine, I'm going to shine that up. I'm going to make sure it looks good. If it's not mine, I don't super care. I'll super care. And that's where we get into this metacognitive laziness and where we just fail to critically think and to catch those things. And that's

174
00:17:26,061 --> 00:17:54,181
where you see really smart people putting their names essentially on really sloppy work. So that's the kind of the cognitive cascade behind the scenes. But when it all comes together, the machine side and the human side, I think one of the worst parts about it is that it tends to cause this collapse of tasks, this task collapse, or like what I like to call task turduckens,

175
00:17:54,461 --> 00:18:22,021
where you start to do one thing, but it's really 10 things all wrapped up into one. And that's where it gets especially dangerous because we are outsourcing tons of decisions that the machine will make for us, and we'll start to regress toward that mean, toward that average. So let's look at an example. We're a small manufacturer. We produce fasteners that help assembly. It makes it easier, faster, and better in the long term.

176
00:18:22,661 --> 00:18:51,141
and it helps with automation. But our sales have been flat to a little bit down, and the sales team wants to get in front of more people faster. So we want to do a nurture series. Let's see if we can get more of our subscribers to request a demo. So create a nurture series to drive that request and use a professional tone, highlight our key benefits, and include a clear call to action in each email.

177
00:18:52,901 --> 00:19:18,261
Okay, as I look at this, let's start to unpack it. We've got one big request, that is, let's create a nurture series to drive to demo. And if you think about what it was going to take to do that, I think, well, I've got two main things. I've got to figure out what that overall strategy and architecture is. Like, what is it that I need to overcome? What's that story? How does my product help?

178
00:19:19,301 --> 00:19:43,221
How are we going to break that out? What's the cadence? Some of those more architectural kind of things. And then I need to write each email. And I can further unpack each of those things. And we can keep doing that. If you look at the architecture, again, I've got to understand what the audience needs to hear, what it is that motivates them, what they're overcoming, and how it is

179
00:19:43,541 --> 00:20:07,221
that our product help maps to that so that we can create that story. And then, again, that strategy, I need to figure out, well, what's going to be the right way to tell that story? In what order does that story get told? And how does it break out across our emails? We could keep doing this. I did this, and this is directional.

180
00:20:08,261 --> 00:20:35,701
You and I will probably disagree a little bit about some of the things that might need to go into this. But I think there's a whole host of things lurking behind this at this point where, you know, I don't even have any idea who's on that list, how they got there, what we've done with them, what they're expecting. Wow, there's just, there's really just tons. And if I think, okay, I'm going to throw a lot of great data at it,

181
00:20:36,461 --> 00:21:04,821
and I'm going to solve some of that stuff. So I've tried to be kind of generous here about what if I gave it a great brief, audience profile, CRM data, our brand tone and messaging guide, I give at the product messaging house. There's probably a lot of raw information there to make decisions, but there's still a lot of decisions looming, and there's probably still a lot of information I don't have. So even at this point,

182
00:21:05,221 --> 00:21:32,101
If I slap an LLM on the back and say, go get them, it will. And I have outsourced tons of decisions, all points at which things get rounded off and become more and more generic. Countering that is deceptively simple. It's taking small passes. Instead of doing, it's really, it's fighting that big problem of task collapse.

183
00:21:32,661 --> 00:22:01,141
It's breaking it down and taking each task one at a time. In each pass, you're going to have a human aim it, decide what it is we're trying to do, and then refine it, kind of call it. I like to think of it, you know, as like you might have been taught to do a good paint job. You know, you're going to take it in layers. The first thing you're going to do is prepare that surface.

184
00:22:01,581 --> 00:22:30,101
whether that's sanding it or degreasing it or getting it ready to accept the primer. You're going to put on that primer, and then you're going to sand it down and get it smooth so that the next surface, the next layer can be nice and smooth and repeat that. It's the same kind of thing here. We're just doing it at the speed of AI. There's just this big propensity to use AI like the easy button, and that just

185
00:22:30,341 --> 00:22:59,061
It really does take us to slop town, like it's the direct ticket. So this loop is fairly straightforward. It's just a very iterative thing. If you think to what JC was saying this morning about the human role becoming around judgment, accountability, and decisions, that's what this is. We're using AI to do the work, but we're the ones that are guiding it and deciding

186
00:22:59,821 --> 00:23:27,861
What is good? And then, and very importantly, at the end, we're refining it. We're culling, you know, many of the models today are machines. They just will overproduce for a variety of reasons. And it's really important, again, in each layer to take out what isn't part of your vision for good work.

187
00:23:29,061 --> 00:23:54,021
And it is a deceptively simple thing, but at every stage when we're doing this, that's a point at which we are imparting our own sensibilities and we are steering away from that average kind of generic approach to things. Each part of this, again, seems very simple. It is very simple.

188
00:23:54,581 --> 00:24:18,101
But each part of this plays an important role in that cognitive cascade and in fighting those root causes of failure and why we end up with sloppy stuff. So those small tasks, those small passes are really important to fighting task collapse. That's probably, I would say, the number one thing when you're using AI. If we just, even if we go back to that example,

189
00:24:19,141 --> 00:24:46,701
and we just took it in tiny passes, but we had AI do it all, I guarantee you would end up with better work. It's that we end up doing this in big chunks that round off that unique stuff faster. So taking it in small passes helps us stay with that task collapse. And then the other parts, the aiming, the refining, these are all parts of us being active and

190
00:24:48,021 --> 00:25:10,261
fighting us, becoming passive, and surrendering. It's what you would call productive friction. Again, to create value and ownership at a psychological level, I have to have some skin in this game. And by going through these small passes, I get that. Now, if you've ever wrangled with an SS or an LLM, I mean,

191
00:25:10,981 --> 00:25:38,741
There's certainly some friction there at times. And so these small passes really help keep us on our toes in terms of being vigilant and quality controlling things so that we don't end up off course. Now, we've talked a little bit about inputs. I think these are very important, and I think we have to think about them a little bit differently when we're using AI.

192
00:25:39,061 --> 00:26:07,461
It is great to have these messaging houses, personas and stuff, but it's less about outsourcing than it is about keeping things on track. And then to some degree, it's about managing, as a human operator, managing my decision fatigue so that I don't, I'm not always recreating wheels. I'm not, that I can attend to the decisions I need to at this time right now. That I've collected examples of our good blog posts, of our good emails.

193
00:26:07,861 --> 00:26:34,941
of our voice and so that it's easy to refer both myself and the machine back to those things so that we have good anchors. Like I said, every tool has a way of, it has qualities. You know, if you use it one way, it's great. You use it another, you'll encourage tear out in your product or whatever. Well, we don't have time to go into

194
00:26:36,181 --> 00:27:03,381
all of these into depth, I think it's very important to avail yourself to learning these tools. And I really like the word tool. And I know that we bandied about a lot, right? Oh, AI is a tool. It's a tool. I used to use and love the analogy of AI being your smart intern. It's like your intern that came from Oxford.

195
00:27:03,941 --> 00:27:33,381
or your smart assistant, you send it out, you do something, it comes back, you review the work, and then you recalibrate it and send it back out. I like that, but I don't, my issue with it is that we start to, it kind of primes us in terms of that cognitive cascade of thinking it as an authority, as something that an entity that can just do. Whereas when we think about it as a tool,

196
00:27:33,861 --> 00:27:57,181
And I think that that sets us up better to constrain the way that we use it so that we're getting the kind of results that we want. So we can tame some of that filler that we were talking about, you know, that ornamentation that LLMs have a tendency to do by using

197
00:27:57,261 --> 00:28:05,461
using these rigid formats like JSON so that it's just putting the specific things we want out there.

198
00:28:05,781 --> 00:28:07,701
And reading it's not real fun for us.

199
00:28:07,701 --> 00:28:10,741
It's sort of like the health food of content.

200
00:28:10,981 --> 00:28:12,301
But it's a means to an end here.

201
00:28:12,301 --> 00:28:19,701
It's a way to keep the LM from covering things up with some of that ornamented filter.

202
00:28:20,821 --> 00:28:23,941
Likewise, we can use prompting strategies

203
00:28:24,261 --> 00:28:32,261
like chain of thought with some verification steps within it to keep things more factual and keep it on track.

204
00:28:33,061 --> 00:28:44,181
We can use the LLM's own probability talents to prompt for more diverse or more unlikely answers.

205
00:28:44,501 --> 00:28:50,101
But it all comes down to how we as the human operators use it.

206
00:28:50,501 --> 00:28:51,141
I think it's

207
00:28:52,101 --> 00:28:53,381
I think that's just it.

208
00:28:53,541 --> 00:28:55,141
There isn't really more to it.

209
00:28:57,221 --> 00:29:04,181
The markers that I see consistently turning out better work aren't using better tools.

210
00:29:04,181 --> 00:29:06,061
They're not, they're just not.

211
00:29:07,381 --> 00:29:17,141
They've found a tool stack and maybe that shifts, but they're learning the tools and they're using it just as they might have before to craft good work.

212
00:29:17,861 --> 00:29:19,701
They're just doing it at the speed of AI.

213
00:29:20,061 --> 00:29:21,701
They're taking smaller passes.

214
00:29:22,061 --> 00:29:29,621
They're learning when to use the tool, because sometimes the best tool is a different tool.

215
00:29:30,061 --> 00:29:35,941
If I need to find and replace something, well, maybe that's maybe I should do that in another tool.

216
00:29:37,221 --> 00:29:46,341
But this is what it really comes down to, is learning those tools and using them in a way that can bring your vision to life.

217
00:29:48,421 --> 00:29:50,421
I don't really have more about that.

218
00:29:50,421 --> 00:29:50,981
It's just

219
00:29:52,021 --> 00:29:54,021
It's up to us to stop the slop.

220
00:29:54,581 --> 00:30:00,101
If we value quantity over quality, then that's what we'll get.

221
00:30:00,981 --> 00:30:17,781
If we want really distinct marketing communications, if we want to stand out in the market and yet do it at speed, we can do that as long as we use AI in a way that actively fights some of those failure points.

222
00:30:18,181 --> 00:30:24,261
that cognitive cascade to stop the task collapse and to put us back in control.

223
00:30:24,421 --> 00:30:33,301
Because that's really what happens here is that we become the passenger, we become passive, and we just get lazy.

224
00:30:34,341 --> 00:30:36,501
And it really doesn't need to be that way.

225
00:30:37,221 --> 00:30:47,221
If we regain the control with the craft and just take small passes through things, we can turn out really great work

226
00:30:48,021 --> 00:30:49,541
even using AI.

227
00:30:53,061 --> 00:30:54,501
What questions do you have?

228
00:31:00,551 --> 00:31:00,791
Yeah.

229
00:31:00,791 --> 00:31:03,671
You know, every marketing salesperson is different.

230
00:31:03,751 --> 00:31:07,751
Some people may love to hand draw everything, and somebody else loves AI or whatever.

231
00:31:08,311 --> 00:31:12,471
But how do you-- this message is, I thought, really great and so timely.

232
00:31:13,191 --> 00:31:19,111
How would you help your team understand this, these concepts?

233
00:31:22,261 --> 00:31:26,901
I'm the one that made the Kool-Aid, so I've had lots of it.

234
00:31:27,061 --> 00:31:28,821
So take that all with a grain of salt.

235
00:31:30,661 --> 00:31:40,821
For me, I think there's a lot of an aha in understanding some of that behind-the-scenes stuff that goes on with our brains.

236
00:31:42,981 --> 00:31:46,341
And I think knowing that can help arm people

237
00:31:46,741 --> 00:31:50,821
to see it and even to have a lexicon to talk about it with each other.

238
00:31:51,141 --> 00:31:58,741
I think part of what happens too, and I don't have the answer here, is it's hard to collaborate with AI and with a team.

239
00:31:58,821 --> 00:32:14,581
So if I'm working on a project with a team and I or someone else goes off and does part of it and they involve AI and there's a lot of back and forth and stuff, it's tricky to bring them up to speed in a way that they can be an intellectual partner

240
00:32:15,301 --> 00:32:20,741
and challenge me in a productive way and make sure that I didn't slide off the rails.

241
00:32:20,741 --> 00:32:27,701
Because it happens easier than it should, it seems like, at times.

242
00:32:28,501 --> 00:32:29,981
So I think that's part of it.

243
00:32:29,981 --> 00:32:36,181
And just acknowledging that this is what can happen to us, and it's just a natural thing.

244
00:32:36,181 --> 00:32:37,141
We just need to fight it.

245
00:32:37,141 --> 00:32:43,461
We just need to be aware, I think is a huge first step into stopping it.

246
00:32:51,141 --> 00:33:07,301
Anyone else or, you know, does anyone have any case studies about instances in your workplace where you've seen slop or seen slop stop that you've had success at stopping it?

247
00:33:08,421 --> 00:33:14,341
One of the challenges that I have is like because AI it's really good, but it's like not perfect.

248
00:33:14,781 --> 00:33:16,061
It's like getting stuck in the.

249
00:33:17,301 --> 00:33:22,141
can't really, like, trust it completely to fully let the thing out and do what it's supposed to do.

250
00:33:22,741 --> 00:33:31,461
So some days, like, I've noticed that I've gotten stuck in the, well, it's got to be perfect before we can even use it.

251
00:33:32,341 --> 00:33:40,261
And getting over that hump to, like, maybe 90% or 95% is actually good enough.

252
00:33:41,301 --> 00:33:44,901
How do you balance the excellence versus the good enough

253
00:33:47,461 --> 00:33:50,341
Yeah, I think that it depends.

254
00:33:50,341 --> 00:33:55,981
The first thing to do is decide when good enough is good enough and when it needs to be excellent.

255
00:33:55,981 --> 00:34:06,501
And I know that sounds kind of silly, but there are plenty of times where I need it to be well thought through and a good quality, but maybe I don't need it to be award-winning.

256
00:34:06,981 --> 00:34:10,101
And then there's other times where I've got more at stake and

257
00:34:10,501 --> 00:34:12,701
I really need this to be well polished.

258
00:34:12,741 --> 00:34:19,301
And I think that's probably the first step is kind of identifying how big is our target here.

259
00:34:20,341 --> 00:34:25,861
And when it is a smaller target, when we're trying to hit that more excellent thing,

260
00:34:27,301 --> 00:34:33,781
Again, some bias, but I think we have to inject ourselves at some point and be that quality control filter.

261
00:34:34,101 --> 00:34:38,901
And it might mean that, Neil was suggesting that it's great for outlines and getting started.

262
00:34:38,901 --> 00:34:43,941
That might be where I start with it and maybe then I take it and write the thing and do the parts.

263
00:34:44,421 --> 00:34:51,381
Or maybe I work in a different way and I, you know, I do some strategy with it up front and I have it write the drafts.

264
00:34:51,861 --> 00:35:00,421
And then I take it back and I'm the one that gets in there and throws out sense and structure that I think stinks and whatever else.

265
00:35:02,421 --> 00:35:13,141
I don't know if that's a great answer, but I do think that at some point we just have to, maybe 95% from the machine or 85% from the machine really fast is great.

266
00:35:14,661 --> 00:35:18,261
In fact, maybe that's huge.

267
00:35:19,941 --> 00:35:30,741
Maybe trying to get to 100% is, again, maybe not the right target for this specific tool, depending, of course, on what exactly you're trying to do with it.

268
00:35:31,941 --> 00:35:32,061
Yeah.

269
00:35:32,061 --> 00:35:40,181
From your perspective, do you see this slot problem getting worse, getting better, stopping, changing?

270
00:35:41,141 --> 00:35:44,181
Where are we in another week, another month, another year?

271
00:35:44,421 --> 00:35:45,461
Where do you see who's going?

272
00:35:46,741 --> 00:35:51,061
Predictions in the AI sphere are probably not the wise.

273
00:35:52,181 --> 00:35:54,341
I'm not, I don't know for sure.

274
00:35:54,581 --> 00:36:02,981
But I will tell you that I think some of, you know, we talked early on about the machine side.

275
00:36:05,061 --> 00:36:09,381
I think some of those kinds of things are going to continue to be ironed out, you know,

276
00:36:10,341 --> 00:36:22,501
The reason why there was such a big range, like 28.5% to 79%, is it varies depending on the exact model and the recipe and the exact way that they do the training.

277
00:36:23,061 --> 00:36:35,221
And the smart cats at OpenAI and Anthropic and other places are trying to figure out how to do some of these things, like the alignment training.

278
00:36:35,621 --> 00:36:39,221
and do it with less impact to the diversity on the model.

279
00:36:39,381 --> 00:36:42,981
I think some of those things, I just presume, will get ironed out.

280
00:36:44,501 --> 00:36:53,861
I wouldn't be surprised, though, if that, to some degree, intensifies some of the challenges with the cognitive cascade.

281
00:36:54,461 --> 00:37:04,341
I think right now, there's about, there's some research about AI imposter syndrome emerging, that people feeling inferior, it's about half.

282
00:37:05,301 --> 00:37:08,821
right now that would identify that their LLM is smarter than they are.

283
00:37:11,301 --> 00:37:26,821
So I think the danger is that part of it will get better and part of it might get worse because it will be even more tempting for us to trust it as an expert and to surrender to it.

284
00:37:27,221 --> 00:37:27,421
Yeah.

285
00:37:27,461 --> 00:37:32,541
I feel like the biggest problem with

286
00:37:38,741 --> 00:37:51,701
So like, I mean, maybe it's just a devil's advocate thing, but like just to layer things and do it slowly and like inject yourself through the process that you're describing, is that just a sophisticated way to not get caught?

287
00:37:51,781 --> 00:38:07,301
Or how do you draw the line where like, I'm using AI to help me, but like I'm also beginning to, and I don't know, maybe the way to illustrate it would be like that auto logo of the restaurant you gave earlier, like, okay, if that restaurant is going to use

288
00:38:07,861 --> 00:38:13,981
AI hopefully with that process, what might that look like it's the right?

289
00:38:16,261 --> 00:38:35,461
In the case of the salty otter, I don't know, because again, we've probably all eaten at places that have that basic logo because so many entrepreneurs in that situation come get there because they have great passion and talents around culinary interests.

290
00:38:35,461 --> 00:38:35,941
And

291
00:38:36,461 --> 00:38:46,821
They may not be native designers or writers or whatever, and they may not have a large budget and prioritize where they're spending their money.

292
00:38:46,821 --> 00:38:57,181
There's some of that I think we need to be aware of, and I don't have a great answer for how that, how we're going to get a, how an

293
00:38:57,461 --> 00:39:03,461
there would get around it other than to spend more attention to it.

294
00:39:04,181 --> 00:39:25,941
I don't, again, I don't know if that's maybe the, the answer you're looking for, but I guess like how do you, how do you use AI even if it's not SLOP, if like the SLOP is basically a red flag,

295
00:39:27,941 --> 00:39:28,141
Yeah.

296
00:39:29,141 --> 00:39:37,341
In fact, when you were saying that in the first part, my mind goes to, early on, we were talking about publications.

297
00:39:37,341 --> 00:39:54,981
I can't tell you how many printed publications, court citations and stuff, where people left things like, oh, this would make a nice introduction for your speech or whatever, or, obvious

298
00:39:55,581 --> 00:39:56,581
parts in there.

299
00:39:58,421 --> 00:40:00,821
And the answer to that is read your work.

300
00:40:02,101 --> 00:40:14,261
that's back to that judgment and accountability and decision making of one of the best things that we can do is invest ourselves and our time and look at that final outcome.

301
00:40:14,301 --> 00:40:24,981
When I'm using something, it's just like if I'm using a router or something and I've jammed the stock through, if I don't look at that and see if I've done a good job,

302
00:40:25,701 --> 00:40:27,061
I just throw it in a pile.

303
00:40:28,181 --> 00:40:33,381
I'm sort of opening myself up to that slop.

304
00:40:33,381 --> 00:40:34,741
I mean, I'm allowing that.

305
00:40:37,301 --> 00:40:37,541
Yeah.

306
00:40:37,541 --> 00:40:44,221
Can you touch a little bit more about the laziness, kind of what I call it, that human fatigue.

307
00:40:44,261 --> 00:40:51,861
How leaders can be hyperbiginous that it's like backing the-- taking a sample of pilot.

308
00:40:52,501 --> 00:40:55,221
that 90% of their work is automated.

309
00:40:55,621 --> 00:41:00,741
They just take off, and then the planes does everything, and we have planes that even land.

310
00:41:01,301 --> 00:41:10,501
But then in that area, they get fatigued, or they actually trust the system, and that's how they make errors.

311
00:41:10,501 --> 00:41:16,541
How leaders can be hyper-vigilant in that end while you touch the point of laziness?

312
00:41:17,861 --> 00:41:33,061
Yeah, I think the laziness you're talking about again is that metacognitive laziness that sets in once we have that confluence of factors where I've kind of become passive, I'm in the passenger seat, and things are almost frictionless.

313
00:41:34,181 --> 00:41:35,461
I think that's a key part.

314
00:41:35,461 --> 00:41:40,501
That's part of that disarming of the mental systems that would otherwise be in place.

315
00:41:40,901 --> 00:41:45,061
And so while I'm not an engineer and can't tell you exactly

316
00:41:45,661 --> 00:41:46,501
how you might do it.

317
00:41:46,501 --> 00:42:06,261
I know that there are plenty of techniques that we can build positive friction into our workflows and into the way tools are made such that we can avoid that sort of uneffortful me getting sleepy at the wheel, not paying attention kind of thing.

318
00:42:06,261 --> 00:42:11,861
So I think it's, again, it's about being aware that there's a trap in things being

319
00:42:12,901 --> 00:42:15,701
completely effortless and frictionless.

320
00:42:16,821 --> 00:42:21,221
But that's not all, like, again, it's like that deferred inefficiency.

321
00:42:21,461 --> 00:42:23,221
It's not looking at the whole picture.

322
00:42:23,221 --> 00:42:33,701
I'm just trying to make this one thing be the most efficient possible without realizing that then might cascade or defer problems elsewhere.

323
00:42:35,981 --> 00:42:36,101
Yeah.

324
00:42:37,381 --> 00:42:44,221
Maybe really content specific, but should we all just accept that the M-dash is dead now because of AI?

325
00:42:44,741 --> 00:42:48,501
Well, they've done an M-dash fix, so supposedly that's not coming out as more.

326
00:42:48,901 --> 00:42:51,621
And again, I personally, I love the M-dash.

327
00:42:51,621 --> 00:42:53,141
I was always a fan.

328
00:42:53,301 --> 00:42:58,101
I probably used it as a lazy writer crutch too many times myself.

329
00:42:58,101 --> 00:42:59,861
It's a beautiful thing when used well.

330
00:43:00,181 --> 00:43:02,021
I would say no, embrace it.

331
00:43:03,861 --> 00:43:16,421
The interesting thing is, and we can get into the different things, apparently there was also a time when the LLMs were using certain verbs, delve and things in excess, and it was changing.

332
00:43:17,861 --> 00:43:19,941
again, it's that overuse.

333
00:43:19,941 --> 00:43:23,701
And so I think we're going to continue to see patterns of those kinds of things.

334
00:43:24,781 --> 00:43:27,061
And like I said, I think you should own it.

335
00:43:27,501 --> 00:43:28,501
You should grab onto it.

336
00:43:28,901 --> 00:43:30,381
Use it well, and it will be good.

337
00:43:31,181 --> 00:43:45,581
I think that's the key thing with all of it is just, it's just trying to reclaim the active part of craft and in creating good work versus that kind of the passive stuff that we inadvertently get lulled into.

338
00:43:45,581 --> 00:43:49,301
Yeah, Nick.

339
00:43:49,301 --> 00:43:58,901
I think your kind of woodworking analogy with the router is interesting because I think of AI and I think, you know, thinking of it as a tool makes a lot of sense.

340
00:43:59,301 --> 00:44:05,901
Because if you think of like, say you're going to build a cabinet or something like that and you need to cut a piece of wood, you can use a hand saw to do that.

341
00:44:05,901 --> 00:44:11,061
And if you do that, you're probably going to be a lot more meticulous, a lot more careful, a lot more thoughtful.

342
00:44:11,381 --> 00:44:14,101
Or you can use an electric saw where you go faster.

343
00:44:14,421 --> 00:44:18,381
But at the end of the day, you still have to make sure that it's the right cut in the right place.

344
00:44:18,381 --> 00:44:24,261
And I think a lot of this slop can also be summed up by only a poor craftsman blames his tools.

345
00:44:24,301 --> 00:44:28,741
And so as you're working through this, you think about, okay, whether it is, you know,

346
00:44:29,381 --> 00:44:38,181
something that AI did or something that I did manually, at the end of the day, if you're the craftsman, you're still responsible for how everything comes together.

347
00:44:39,301 --> 00:44:39,621
Yeah.

348
00:44:40,901 --> 00:44:41,541
I like that.

349
00:44:44,261 --> 00:44:45,261
I know we're about out of time.

350
00:44:45,261 --> 00:44:47,381
Anything else before we wrap?

351
00:44:47,381 --> 00:44:49,421
All right.

352
00:44:49,661 --> 00:44:50,421
Well, thanks again.

353
00:44:50,421 --> 00:44:51,621
I really appreciate you being here.