WEBVTT
1
00:00:00.841 --> 00:00:06.506
Hey everyone, welcome to Future Ready Today, the number one podcast focused on the future of work.
2
00:00:07.187 --> 00:00:12.972
Every day I break down the top three stories that you need to be paying attention to and I break them down.
3
00:00:13.112 --> 00:00:19.598
I give you the insights, the trends, the key insights that you should be paying attention to so that you can be future ready.
4
00:00:20.139 --> 00:00:24.963
Today is Wednesday, May 13th, 2026 and the three stories that I want to get to today, number one.
5
00:00:25.724 --> 00:00:33.846
Goldman Sachs COO publicly calls his workforce a human assembly line and says that AI agents are the robots.
6
00:00:34.266 --> 00:00:40.227
Story number two, Amazon employees are gaming AI leaderboards by burning fake tokens.
7
00:00:40.868 --> 00:00:45.549
And story number three, Princeton ends a 133 years of trust-based exams because AI made cheating invisible.
8
00:00:55.971 --> 00:01:00.215
where they're not going to have proctors watching you take exams, again, because of the honor code.
9
00:01:00.856 --> 00:01:04.720
AI has abolished this 133-year-old tradition.
10
00:01:04.740 --> 00:01:05.240
And guess what?
11
00:01:05.881 --> 00:01:11.026
Now when you take an exam at Princeton, you're going to have somebody watching you actually take that test.
12
00:01:11.666 --> 00:01:13.468
So those are the top three stories of the day.
13
00:01:13.488 --> 00:01:14.329
We're going to break them down.
14
00:01:16.051 --> 00:01:28.562
talk about what these things actually mean, look at some other historical trends and stories and data to really help kind of connect what these things mean to the broader theme on the future of work.
15
00:01:29.042 --> 00:01:30.463
So let's get into the first story of the day.
16
00:01:30.503 --> 00:01:37.570
Goldman Sachs COO publicly calls his workforce a human assembly line, and he says that AI agents are the robots.
17
00:01:37.710 --> 00:01:37.850
Okay.
18
00:01:38.520 --> 00:01:40.422
This was in CNBC.
19
00:01:40.482 --> 00:01:42.343
This was an interview that he gave yesterday.
20
00:01:42.363 --> 00:01:44.185
And this is John Waldron.
21
00:01:44.365 --> 00:01:47.608
He's the chief operating officer over at Goldman Sachs.
22
00:01:48.669 --> 00:01:49.730
And he described...
23
00:01:51.576 --> 00:01:56.521
His exact words, quote, our human assembly lines will become more digitized.
24
00:01:57.121 --> 00:01:59.604
Digital agents will be our robots.
25
00:02:00.325 --> 00:02:11.156
He drew an explicit parallel to what happened in manufacturing decades ago when factories replaced workers with machines and said that banks simply haven't gone on that journey yet.
26
00:02:11.236 --> 00:02:11.517
He said,
27
00:02:12.077 --> 00:02:22.360
Until now, Goldman plans to deploy AI agents across the firm to automate workflows to cut costs and to boost productivity, what he called, quote, a big unlock.
28
00:02:23.080 --> 00:02:25.220
He insisted this won't mean mass layoffs.
29
00:02:25.761 --> 00:02:27.681
Headcount will remain roughly stable.
30
00:02:27.861 --> 00:02:38.484
New engineering jobs will be created and pushed back onto the media narrative, blaming AI for current industry layoffs, arguing those cuts still largely reflect post-pandemic growth.
31
00:02:39.364 --> 00:02:41.765
workforce corrections rather than AI displacement.
32
00:02:42.485 --> 00:02:47.907
But he also acknowledged candidly that he doesn't know dynamically how the overall headcount will change.
33
00:02:48.687 --> 00:02:53.369
And that honest uncertainty, I think, deserves some, you know, paying attention to.
34
00:02:54.269 --> 00:02:55.550
Now, the futurist lens here...
35
00:02:56.929 --> 00:03:01.273
is it's not so much that Goldman Sachs is deploying AI.
36
00:03:01.834 --> 00:03:05.217
I think every company around the world is in the process of deploying AI.
37
00:03:05.357 --> 00:03:11.823
I think for me, it's mainly about the language that Waldron used, because I think at his level as the COO,
38
00:03:12.664 --> 00:03:15.566
the language that's used is not accidental.
39
00:03:15.866 --> 00:03:20.169
And maybe he wasn't supposed to say that he views humans as, you know, the assembly line.
40
00:03:20.669 --> 00:03:26.232
Maybe he was just thinking it and, you know, the internal thought bubble made its way to the outside.
41
00:03:26.813 --> 00:03:37.860
But when the president and COO of one of the most powerful financial institutions on planet Earth describes his own workforce as an assembly line, I think he is...
42
00:03:39.770 --> 00:03:42.032
Maybe he's trying to normalize this framework.
43
00:03:42.792 --> 00:03:48.016
And the assembly line metaphor obviously has a history that's worth understanding.
44
00:03:48.036 --> 00:03:50.258
It goes back to Henry Ford.
45
00:03:50.858 --> 00:03:56.722
He didn't invent the assembly line, but Henry Ford, you could argue, he actually did perfect the assembly line.
46
00:03:57.463 --> 00:04:00.965
And when he did, it didn't just change how cars were actually made.
47
00:04:01.465 --> 00:04:05.448
It changed the fundamental relationship between a company and its workers.
48
00:04:06.169 --> 00:04:07.750
It transformed craftsmen.
49
00:04:08.465 --> 00:04:16.067
who made a living really focusing and owning a unique and specialized craft into component operators.
50
00:04:16.267 --> 00:04:18.388
You were just one piece on the assembly line.
51
00:04:19.348 --> 00:04:36.113
Now, of course, output went up, made the individual for the first time in history genuinely interchangeable, where if you were injured or you couldn't do a certain task on the assembly line, you were just swapped out, just like a part.
52
00:04:36.865 --> 00:04:38.326
It's like, okay, John can't come in today.
53
00:04:38.386 --> 00:04:43.148
All right, Jacob, go in there, take his spot, do the exact same thing that John was doing.
54
00:04:44.189 --> 00:04:45.189
And that's how it was viewed.
55
00:04:46.310 --> 00:04:49.051
Now, output in that situation went up, costs went down.
56
00:04:49.892 --> 00:05:01.158
The individual's bargaining power collapsed completely almost overnight because the skill that once made an individual irreplaceable kind of got absorbed into the system itself.
57
00:05:02.204 --> 00:05:14.274
And so what Waldron is telling us from Goldman Sachs, in you could say polite corporate language, is that the transition that happened in the assembly line is also coming for white-collar work.
58
00:05:15.375 --> 00:05:30.207
And at its height in the year 2000, and this was an article, I think I first came across this article, it was published almost a decade ago, Goldman's U.S. Cash Equities Trading Desk employed 600 traders.
59
00:05:31.559 --> 00:05:36.100
There were 600 traders who were on that cash equities trading desk.
60
00:05:37.220 --> 00:05:44.382
Today, and I don't even know if this number is even accurate because the last stat that I heard was from a few years ago, there are two.
61
00:05:45.803 --> 00:05:50.684
So we went from 600 to two, and I don't even know if it's still two, maybe it's one.
62
00:05:51.664 --> 00:05:55.185
But to go from 600 to two, wiping out 99.5% of everybody there,
63
00:06:00.159 --> 00:06:05.741
automated trading programs took the rest, supported by 200 engineers, computer engineers.
64
00:06:06.581 --> 00:06:18.664
Goldman's then CFO, Mary Chavez, revealed in 2017 that the firm had found a consistent ratio during the transition, four traders replaced by one computer engineer.
65
00:06:19.484 --> 00:06:25.966
At the time of the disclosure, 9,000 people, roughly a third of Goldman's entire staff, were already computer engineers.
66
00:06:27.062 --> 00:06:37.666
Goldman also mapped 146 distinct steps in a typical IPO process and said that many were, in Chavez's own words, begging to be automated.
67
00:06:38.666 --> 00:06:40.007
Okay, so you understand what they did there.
68
00:06:40.927 --> 00:06:53.192
They broke down, in this case, the steps that required to IPO, for a company to go public, 146 distinct steps.
69
00:06:53.232 --> 00:06:55.553
So they broke down that entire process of,
70
00:06:56.222 --> 00:07:03.468
which is a pretty complicated process, into these 146 steps and said, why do we need people doing this now?
71
00:07:03.708 --> 00:07:13.916
Now that we can break it down and standardize it and create a template for these 146 steps, we would love AI to be able to do it.
72
00:07:15.628 --> 00:07:33.972
Now, the difference now is that AI agents can attack the cognitive work, not just the kind of a template process work, but they can attack things like the pitch books, the client analysis, relationship intelligence, judgment calls, right?
73
00:07:34.052 --> 00:07:38.813
These are things that pure algorithmic trading in the past was never able to touch.
74
00:07:39.881 --> 00:07:50.449
Goldman's own research economist estimated earlier this year that AI threatens to automate tasks currently accounting for 25% of all American work hours.
75
00:07:51.250 --> 00:08:00.397
Bloomberg Intelligence projects that global investment banks may shed as many as 200,000 jobs over the next three to five years as they implement AI.
76
00:08:01.838 --> 00:08:07.743
But the interesting thing is that Goldman's own research team, they published a study in April.
77
00:08:08.698 --> 00:08:14.448
so just a month and a half or two ago, looking at 40 years of technological displacement.
78
00:08:15.470 --> 00:08:21.721
And what they found in their own analysis is that displaced workers don't just lose jobs.
79
00:08:22.777 --> 00:08:26.420
they experience what the researchers call occupational downgrading.
80
00:08:27.120 --> 00:08:34.165
They slide into roles requiring fewer analytical and interpersonal skills, not just fewer routine ones.
81
00:08:34.926 --> 00:08:45.694
And it's because the same technological forces that eliminated their old jobs also eroded the market value of their existing expertise.
82
00:08:46.852 --> 00:08:54.295
And what they found is that this displacement is somewhat slow, it's structural, and it's very hard to reverse once it actually takes hold.
83
00:08:55.195 --> 00:09:05.758
So Waldron's no mass layoffs assurance may be a genuine statement, but the assembly line analogy, if you kind of follow it all the way through, doesn't really promise job preservation.
84
00:09:06.099 --> 00:09:09.800
It promises role transformation, which is a very different thing.
85
00:09:11.140 --> 00:09:15.682
Now, confidentially, you know, in a recent CHRO meeting,
86
00:09:16.915 --> 00:09:19.739
that we had with a very, very large employer.
87
00:09:19.759 --> 00:09:33.377
And I won't mention who the company is, but at this executive, the executive at this company confidentially said that, because I asked them about the impact that AI is having on their workforce.
88
00:09:33.397 --> 00:09:33.517
Right.
89
00:09:35.402 --> 00:09:47.467
And she confidentially told us and the other CHROs in the room that it's not causing layoffs, but it is causing them to potentially keep their hiring more flat.
90
00:09:48.570 --> 00:10:01.777
So in other words, revising projections as far as how many people they want to bring into the company, kind of keeping it flat, not letting go of people, but also not hiring as aggressively as they thought they might potentially be doing.
91
00:10:02.418 --> 00:10:06.840
So this kind of aligns with what Goldman is saying here as well.
92
00:10:08.121 --> 00:10:14.064
Now, if you think about Ford, the workers at Ford didn't lose their jobs on day one.
93
00:10:15.293 --> 00:10:23.799
Okay, they slowly, they lost their craft identity, their autonomy, their wage premium, and then eventually they lost their leverage.
94
00:10:25.100 --> 00:10:30.023
So the reassurance and the risk can both be true at the same time.
95
00:10:30.043 --> 00:10:32.565
You can have the pros and cons happening at the same time.
96
00:10:33.345 --> 00:10:41.631
And that's what makes it so hard right now for so many leaders and for so many organizations to kind of navigate what the impact of AI is actually going to be.
97
00:10:42.711 --> 00:10:43.492
Now, story number two.
98
00:10:44.742 --> 00:10:55.325
And before I jump into story two, if you want more of these types of deep analysis, the deep dives, the breakdowns, go over to my newsletter.
99
00:10:55.425 --> 00:10:56.306
It's on Substack.
100
00:10:56.526 --> 00:10:59.387
It is called futureofworknewsletter.com.
101
00:11:00.967 --> 00:11:04.649
There is a free and a paid version that you can subscribe to.
102
00:11:05.870 --> 00:11:08.071
The free version, I do a weekly briefing.
103
00:11:08.331 --> 00:11:12.693
And the paid version, I do a deep dive, which I put out every one to two weeks.
104
00:11:13.254 --> 00:11:18.256
Really tackling a topic and, like it sounds, doing a very deep dive in it.
105
00:11:18.737 --> 00:11:20.518
Thousands of words, charts, graphs, facts.
106
00:11:21.118 --> 00:11:22.918
frameworks around a particular topic.
107
00:11:22.958 --> 00:11:28.980
The last one I did was how to actually engineer a jobs apocalypse if we wanted to do so.
108
00:11:29.460 --> 00:11:33.821
And it looked at the eight criteria that are required to make a jobs apocalypse happen.
109
00:11:34.321 --> 00:11:38.022
Again, you can learn more by going to futureofworknewsletter.com.
110
00:11:39.302 --> 00:11:44.744
So story two, Amazon employees, gaming AI leaderboards by burning fake tokens.
111
00:11:45.684 --> 00:11:47.784
So the Financial Times, they broke this story.
112
00:11:48.044 --> 00:11:49.665
I think it came out today.
113
00:11:51.719 --> 00:11:52.859
which I think is pretty interesting.
114
00:11:52.879 --> 00:11:56.482
And it's something that I've talked about on this podcast for a few months now.
115
00:11:57.102 --> 00:12:01.225
So Amazon deployed an internal AI tool called MeshClaw.
116
00:12:02.091 --> 00:12:21.706
And MeshClaw is, of course, inspired by the viral open source tool that so many people have talked about called OpenClaw that basically allows employees to create their AI agents that connect workflows to, I'm sorry, to connect workplace software and carry out tasks in kind of the workflow.
117
00:12:21.726 --> 00:12:23.387
So it automates some of the work that you're able to do.
118
00:12:24.328 --> 00:12:29.650
You kind of think of it as your own personal assistant or your own coworker, your own work assistant, right?
119
00:12:29.670 --> 00:12:33.391
So you can kind of plug them into your workflows, have them do things that you would usually do.
120
00:12:34.111 --> 00:12:42.354
So it can triage emails, it can initiate code deployments, it can interact with Slack, you know, genuinely useful stuff.
121
00:12:43.214 --> 00:12:50.277
But the problem was that Amazon simultaneously, while they introduced this, they also introduced targets.
122
00:12:51.400 --> 00:12:57.747
requiring more than 80% of developers to use AI every single week.
123
00:12:58.747 --> 00:13:05.433
And they started tracking AI token consumption, which many other companies have started doing, including JPMorgan Chase and Meta and several others.
124
00:13:06.694 --> 00:13:15.501
And so the token consumption is basically what a token, what that means is they're tracking the units of data that are being processed by AI models.
125
00:13:15.681 --> 00:13:17.723
And these units of data are known as tokens.
126
00:13:18.203 --> 00:13:22.887
And tokens, you can kind of think of as like three quarters or almost a full word, right?
127
00:13:23.708 --> 00:13:30.312
So every three quarters of a word that these AI systems are processing is considered one token.
128
00:13:31.332 --> 00:13:37.136
And what's happening is that a lot of these organizations are creating leaderboards to see how many tokens are you using.
129
00:13:37.916 --> 00:13:41.598
In other words, how much are you feeding into AI and how much is AI giving you back?
130
00:13:42.098 --> 00:13:43.439
So we're tracking your usage.
131
00:13:44.380 --> 00:13:48.462
And so companies created these internal leaderboards in the case of Amazon.
132
00:13:49.383 --> 00:13:49.963
So what happened?
133
00:13:51.229 --> 00:13:59.876
Well, in the case of Amazon, employees started using MeshClaw not to do actual work, but to inflate their token scores.
134
00:14:00.596 --> 00:14:07.662
They started automating unnecessary tasks, burning tokens on meaningless activity, gaming the leaderboard.
135
00:14:08.422 --> 00:14:15.908
One employee told the Financial Times, quote, there is so much pressure to use these tools that some people are just using MeshClaw to maximize their token usage.
136
00:14:16.709 --> 00:14:17.550
Another said, quote,
137
00:14:18.508 --> 00:14:19.689
Managers are looking at it.
138
00:14:20.270 --> 00:14:26.054
And when they track usage, it creates perverse incentives, and some people are very competitive about it.
139
00:14:26.995 --> 00:14:32.060
Amazon told employees that the token statistics would not factor into performance reviews.
140
00:14:32.880 --> 00:14:35.182
The employees, not surprisingly, didn't believe them.
141
00:14:35.763 --> 00:14:45.151
Meta employees have engaged in the same behavior on their own internal leaderboards, and the practice already has a name called, you got it, token maxing.
142
00:14:46.163 --> 00:14:47.264
Now, the futurist lens here.
143
00:14:47.845 --> 00:14:52.609
This story has a name, and the name is called Goodhart's Law.
144
00:14:53.570 --> 00:15:00.656
And the British economist Charles Goodhart, he first articulated this in 1975 in the context of monetary policy.
145
00:15:01.677 --> 00:15:08.604
And the law simply states that when a measure becomes the target, it ceases to be a good measure.
146
00:15:09.620 --> 00:15:12.043
And the mechanism is almost always the same thing.
147
00:15:12.824 --> 00:15:18.951
A well-intentioned metric is introduced, whether it's engagement, whether it's token maxing, whether it's, you know, whatever you want the metric to be.
148
00:15:19.852 --> 00:15:26.559
And people begin optimizing specifically for that metric rather than the underlying thing that the metric was actually meant to capture.
149
00:15:27.513 --> 00:15:30.396
And the number improves, meaning the metrics improving.
150
00:15:30.996 --> 00:15:37.942
But the reality it was supposed to represent actually starts to degrade or disappear entirely.
151
00:15:37.962 --> 00:15:39.703
And I'm going to give you a couple of historical examples.
152
00:15:40.844 --> 00:15:45.969
One of them, I think the most interesting and colorful example of this, is the cobra effect.
153
00:15:47.030 --> 00:15:50.112
And this is something that most people probably have never heard of or have talked about.
154
00:15:51.193 --> 00:15:52.634
But it's from British colonial India.
155
00:15:53.895 --> 00:15:54.096
And...
156
00:15:54.997 --> 00:16:01.120
This is where the government, they were very concerned about the large number of venomous cobras in Delhi.
157
00:16:02.060 --> 00:16:06.642
And so what they did is they offered a bounty to every dead cobra brought in.
158
00:16:07.483 --> 00:16:11.064
Now, obviously, enterprising locals, they saw this as a way to make money, and they thought, great.
159
00:16:11.204 --> 00:16:18.447
Well, if I can bring in dead cobras and I get kind of a bounty for it, well, it's in my interest to kill as many of these cobras as I can.
160
00:16:19.268 --> 00:16:20.328
So what do you think they started doing?
161
00:16:21.269 --> 00:16:23.570
They actually started breeding cobras.
162
00:16:24.272 --> 00:16:25.052
to collect the bounty.
163
00:16:25.852 --> 00:16:33.495
Because if you breed cobras, and then you kill the cobras and bring them in, you kind of have created your own personal ATM machine.
164
00:16:34.415 --> 00:16:41.137
And when the government discovered this, and then they canceled the program, what do you think the breeders then did?
165
00:16:42.057 --> 00:16:46.398
Well, they took all these cobras that they were breeding, and they just released them.
166
00:16:47.079 --> 00:16:48.519
Because now the cobras are worthless.
167
00:16:49.039 --> 00:16:50.119
So what's the point of killing them?
168
00:16:50.159 --> 00:16:51.300
What's the point of doing anything with them?
169
00:16:52.045 --> 00:16:52.945
You started breeding them.
170
00:16:53.225 --> 00:16:54.146
There's no more incentive.
171
00:16:54.246 --> 00:16:57.487
So now you let them out into the wild, let them out onto the streets.
172
00:16:58.707 --> 00:17:04.989
And so what ended up happening is they made the original problem worse than before the intervention.
173
00:17:05.630 --> 00:17:10.852
In other words, the metric became the goal, and then the goal destroyed the metric.
174
00:17:12.352 --> 00:17:14.853
This also happened in the Soviet Union.
175
00:17:16.058 --> 00:17:21.361
Now, my family is actually from the former USSR, from the Republic of Georgia.
176
00:17:21.741 --> 00:17:23.062
My grandmother, born in Ukraine.
177
00:17:23.262 --> 00:17:26.444
My mom and dad were both born in the Republic of Georgia in Belisi.
178
00:17:27.464 --> 00:17:34.368
And so, you know, I used to hear lots of different types of stories and examples of what was going on at the time.
179
00:17:35.609 --> 00:17:43.093
But one of the movements that actually came about in the 1930s was known as the Stakhanovite movement.
180
00:17:44.376 --> 00:17:46.198
And this was, again, in the 1930s.
181
00:17:46.518 --> 00:17:55.185
There was a famous coal miner named Alexei Stokhanov, who allegedly mined 14 times his daily quota in a single shift.
182
00:17:56.086 --> 00:18:03.012
And because he was able to mine so much coal, he was obviously portrayed as a national hero, basically saying, look how much this person can do.
183
00:18:03.393 --> 00:18:06.555
If he can mine that much coal, you should be able to mine that much coal too.
184
00:18:07.816 --> 00:18:11.499
And so the government used his achievement to raise quotas
185
00:18:12.680 --> 00:18:14.041
across all the different workers.
186
00:18:15.361 --> 00:18:27.546
Now, what the state propaganda did not publicize at the time, in Stalin's USSR at the time, is that Stokhanov actually had a team of assistants who were helping him out.
187
00:18:28.586 --> 00:18:32.968
And the feat was almost certainly staged for optics.
188
00:18:33.895 --> 00:18:41.046
And so the result was a workforce that quickly learned to optimize for quota numbers rather than for output.
189
00:18:41.066 --> 00:18:45.913
So managers kept quotas deliberately low so they could appear to overfill them.
190
00:18:45.933 --> 00:18:46.354
Okay.
191
00:18:46.784 --> 00:18:50.045
Workers gamed piece rate systems to maximize play.
192
00:18:50.725 --> 00:19:05.090
The numbers on paper and the output in reality diverged so far that the Soviet statistical agencies became internationally famous for producing figures that bore little relationship to what was actually happening in factories and on farms.
193
00:19:05.570 --> 00:19:06.730
You want a more recent example?
194
00:19:06.770 --> 00:19:08.431
Take a look at what happened with Wells Fargo.
195
00:19:09.576 --> 00:19:16.318
When the bank tied very aggressive cross-selling targets to performance evaluations, what did employees actually do?
196
00:19:16.838 --> 00:19:18.258
And of course, you probably remember this.
197
00:19:19.098 --> 00:19:23.199
Employees opened millions of unauthorized customer accounts to hit the numbers.
198
00:19:24.020 --> 00:19:32.602
And so the metric, which was accounts opened, completely decoupled from the thing it was supposed to measure, which was genuine customer relationships.
199
00:19:33.511 --> 00:19:41.675
And so the result was one of the largest banking scandals in American history, a $3 billion settlement, and years and years of reputational damage.
200
00:19:43.236 --> 00:19:47.878
And so now Amazon and a lot of these other organizations are doing the exact same thing with AI.
201
00:19:48.998 --> 00:19:52.260
And I think what makes the Amazon version particularly dangerous...
202
00:19:53.376 --> 00:19:57.859
is a finding from one of the most rigorous studies ever conducted on AI in 2025.
203
00:19:58.500 --> 00:20:01.182
This was a nonprofit called METR.
204
00:20:01.562 --> 00:20:02.523
M-E-T-R.
205
00:20:03.023 --> 00:20:08.727
They ran a randomized control trial using the same methodology used in clinical drug trials.
206
00:20:09.588 --> 00:20:18.971
And they did this on experienced software developers using frontier AI tools, including Cursor Pro, Cloud 3.5, and 3.7.
207
00:20:19.611 --> 00:20:22.593
And the result was unexpected and also pretty important.
208
00:20:23.313 --> 00:20:29.035
AI tools caused developers to take 19% longer to complete their tasks.
209
00:20:30.175 --> 00:20:31.596
Slower, not faster.
210
00:20:31.656 --> 00:20:34.597
Now, of course, you can make the argument the tools have gotten better since then.
211
00:20:35.637 --> 00:20:38.218
But here's the part that I think is really worth paying attention to.
212
00:20:39.394 --> 00:20:44.637
is that the developers actually thought that AI made them 20% faster.
213
00:20:45.938 --> 00:20:50.541
So they were wrong in the exact opposite direction and they had no idea.
214
00:20:51.742 --> 00:20:58.265
So their subjective experience of productivity and their actual productivity were moving in completely opposite directions.
215
00:20:59.226 --> 00:21:06.630
So it was taking them almost 20% longer to complete a task, but they thought it was taking them 20% less time to complete a task.
216
00:21:07.431 --> 00:21:08.872
And so I think this finding...
217
00:21:11.147 --> 00:21:15.808
shows that AI usage can produce wildly inflated self-assessments.
218
00:21:15.908 --> 00:21:18.409
And that's the problem with a lot of these things, self-assessments.
219
00:21:18.909 --> 00:21:20.889
This is the problem with engagement surveys.
220
00:21:20.929 --> 00:21:22.910
This is the problem with well-being studies.
221
00:21:22.950 --> 00:21:25.110
This is the problem with the four-day workweek studies.
222
00:21:26.010 --> 00:21:33.732
When you do these types of self-assessed studies, the data is not that valuable.
223
00:21:34.690 --> 00:21:46.378
I mean, obviously, if you tell me, for example, hey, Jacob, you can make the exact same amount of money, get the same speaking gigs, the same podcast views, the same everything, and only have to work two, three days a week, I'd say fantastic.
224
00:21:47.579 --> 00:21:49.500
And then, of course, I'm going to say I'm so much more happy.
225
00:21:50.260 --> 00:21:50.961
I'm much happier.
226
00:21:50.981 --> 00:21:51.681
I'm more productive.
227
00:21:52.282 --> 00:21:54.203
But does that mean I'm actually more productive?
228
00:21:54.603 --> 00:21:58.366
Am I actually, you know, are the numbers actually better from a business perspective?
229
00:21:59.287 --> 00:21:59.467
No.
230
00:21:59.487 --> 00:21:59.547
No.
231
00:22:00.458 --> 00:22:02.739
Right, subjectively, I might tell you I feel better.
232
00:22:03.559 --> 00:22:05.600
I think I'm more productive, I'm happier.
233
00:22:07.141 --> 00:22:07.581
But so what?
234
00:22:08.702 --> 00:22:12.243
Those types of self assessments do not mean much.
235
00:22:14.004 --> 00:22:18.967
So this is what I think a lot of organizations are missing.
236
00:22:18.987 --> 00:22:21.548
They're not paying attention to this, right?
237
00:22:21.848 --> 00:22:27.731
So a leaderboard, for example, an 80% adoption target,
238
00:22:28.620 --> 00:22:37.686
where a manager's watching token counts, you know, that creates, I think, the exact condition for Goodhart's law to operate at full force.
239
00:22:38.826 --> 00:22:44.430
Employees are optimizing for a number that was already, I think, a pretty poor proxy for value.
240
00:22:44.890 --> 00:22:47.932
Why does token usage mean that I'm adding more value?
241
00:22:48.592 --> 00:22:53.655
Why does token usage mean that I'm a more productive or a better employee or making more of an impact on the business?
242
00:22:53.695 --> 00:22:53.936
Why?
243
00:22:53.956 --> 00:22:54.296
It doesn't.
244
00:22:56.857 --> 00:22:57.098
So...
245
00:22:58.522 --> 00:23:00.443
I think what's happening is companies are actually making things worse.
246
00:23:01.624 --> 00:23:04.166
And the broader data out there confirms that it's not just Amazon.
247
00:23:05.147 --> 00:23:07.708
Thomson Reuters study from February earlier this year
248
00:23:08.931 --> 00:23:17.994
found that 82% of leaders say their organizations are either not collecting ROI metrics around AI usage or are unsure whether they are.
249
00:23:18.614 --> 00:23:28.757
A KPMG survey from September, 2025, found that 78% of leaders believe traditional business metrics are becoming insufficient to measure AI's impact.
250
00:23:30.218 --> 00:23:37.260
Companies are mandating the adoption of a technology whose value they largely cannot measure or don't know how to measure
251
00:23:38.904 --> 00:23:45.689
which is creating literally the exact environment where Goodhart's law takes over.
252
00:23:46.249 --> 00:23:48.811
The leaderboard is not the answer to the problem.
253
00:23:48.851 --> 00:23:52.374
The leaderboard is the acceleration of the problem.
254
00:23:53.895 --> 00:24:03.803
Token consumption, I think, is a treacherous metric for organizations to be optimizing for because it is invisible to common sense.
255
00:24:05.824 --> 00:24:08.126
A thousand tokens burned on a genuine customer...
256
00:24:10.029 --> 00:24:25.400
analysis or customer project and a thousand tokens burned on a mesh cloth agent that's running overnight, they might look identical on a leaderboard, on a dashboard, right?
257
00:24:25.440 --> 00:24:26.481
I'm burning a thousand tokens.
258
00:24:26.501 --> 00:24:27.662
You're burning a thousand tokens.
259
00:24:27.682 --> 00:24:29.563
You're doing who knows what with it.
260
00:24:29.623 --> 00:24:31.345
I'm actually trying to solve a customer problem.
261
00:24:33.066 --> 00:24:34.107
Is what we're doing the same?
262
00:24:34.127 --> 00:24:34.967
It's not.
263
00:24:35.008 --> 00:24:35.348
It's not.
264
00:24:36.397 --> 00:24:46.585
This is what a lot of researchers call a counterfeit reality, a data set that tells leadership the transformation is proceeding while the underlying behavior is actually degrading.
265
00:24:47.626 --> 00:24:53.070
And so the lesson for leaders is not that AI adoption metrics are necessarily bad or wrong.
266
00:24:53.610 --> 00:24:55.392
Some measurement of adoption is necessary.
267
00:24:55.412 --> 00:24:56.553
It's worth paying attention to.
268
00:24:57.033 --> 00:25:04.559
But the lesson is that activity metrics, the tokens consumed, the tools launched, the sessions logged, the percentage of employees who opened the app,
269
00:25:05.473 --> 00:25:06.854
These things are always going to get gamed.
270
00:25:07.094 --> 00:25:07.395
Always.
271
00:25:08.135 --> 00:25:08.636
They always have.
272
00:25:09.016 --> 00:25:09.577
They always will.
273
00:25:10.778 --> 00:25:16.242
The only metrics that resist Goodhart's Law are outcome metrics.
274
00:25:17.383 --> 00:25:18.985
Did the quality of the decisions improve?
275
00:25:19.946 --> 00:25:22.228
Did the customer problems get solved faster?
276
00:25:23.168 --> 00:25:24.970
Are we able to bring on more customers?
277
00:25:25.851 --> 00:25:27.352
Are we entering new markets?
278
00:25:27.812 --> 00:25:30.355
Did the work product actually get better?
279
00:25:31.359 --> 00:25:44.283
These are obviously much harder things to measure and they are much slower to see, but these are the only numbers that cannot be faked by any of these leaderboards, any of these dashboards like a mesh cloth agent that's burning tokens while an employee sleeps.
280
00:25:45.943 --> 00:25:49.304
So if you want to get serious about it, then get serious about it.
281
00:25:50.785 --> 00:25:52.425
What you measure grows.
282
00:25:53.445 --> 00:25:55.466
If you want to measure tokens, then guess what?
283
00:25:55.906 --> 00:25:59.227
You're going to see your employees burning through tokens like you wouldn't believe.
284
00:26:00.513 --> 00:26:02.756
But at the end of the day, who really cares?
285
00:26:04.198 --> 00:26:12.248
Story three, Princeton ends 133 years of trust-based exams because AI made cheating invisible.
286
00:26:13.392 --> 00:26:19.758
Now, before I get into this one, if you're watching this and you're listening to this, you're a chief human resource or chief people officer.
287
00:26:20.338 --> 00:26:26.323
We now actually have over 40 chief human resource officers at companies around the world who joined us.
288
00:26:26.363 --> 00:26:30.167
It's a big, exciting milestone for us.
289
00:26:32.028 --> 00:26:33.970
So we're very, very excited for this.
290
00:26:35.291 --> 00:26:37.573
Over 40 CHROs from companies around the world.
291
00:26:38.694 --> 00:26:38.914
And...
292
00:26:41.140 --> 00:26:46.943
If you want to learn more, you can check us out, futureofworkleaders.com.
293
00:26:46.963 --> 00:26:50.265
We get together every single month virtually.
294
00:26:50.285 --> 00:26:52.486
We get together one or two times a year in person.
295
00:26:53.207 --> 00:26:59.691
There's no reason for why you should be spending $20,000, $30,000, $40,000, $50,000 for some of these CHRO groups out there.
296
00:27:00.231 --> 00:27:00.791
You join us.
297
00:27:00.831 --> 00:27:03.373
You spend a tiny fraction of that amount.
298
00:27:04.193 --> 00:27:06.534
And I promise you it is well worth your while.
299
00:27:06.554 --> 00:27:09.355
We're specifically focused on the future of work.
300
00:27:09.775 --> 00:27:13.557
We're not talking about compensation and benefits and any of that kind of traditional stuff.
301
00:27:14.017 --> 00:27:16.078
We're talking more about the future of work.
302
00:27:16.098 --> 00:27:20.059
So if you want to learn more, check us out, futureofworkleaders.com.
303
00:27:20.139 --> 00:27:21.200
Or you can also email me.
304
00:27:21.900 --> 00:27:26.002
That is jacobatthefutureorganization.com.
305
00:27:26.642 --> 00:27:28.823
And you can learn a bit more there.
306
00:27:29.740 --> 00:27:31.001
Okay, so last story of the day.
307
00:27:31.521 --> 00:27:40.007
Princeton University faculty voted this week with a single dissenting vote to require proctors at all in-person examinations beginning July 1st.
308
00:27:40.748 --> 00:27:44.570
That ends 133 years of unbroken tradition.
309
00:27:45.211 --> 00:27:50.635
Princeton's honor system was established in 1893, not by administrators, but by students themselves.
310
00:27:51.435 --> 00:27:57.920
And these are people who petitioned the faculty to eliminate proctoring and take personal responsibility for academic integrity,
311
00:27:58.768 --> 00:28:04.271
The pledge, quote, I pledge my honor that I have not violated the honor code during this examination.
312
00:28:05.052 --> 00:28:08.614
This has been signed at the bottom of every Princeton exam since then.
313
00:28:09.254 --> 00:28:17.359
Faculty was explicitly banned from exam rooms under the university's rule for the entirety of that period.
314
00:28:18.140 --> 00:28:18.840
And what broke it?
315
00:28:19.381 --> 00:28:21.122
Well, you guessed it, AI.
316
00:28:22.806 --> 00:28:34.612
The dean of the college, Michael Gordon, wrote that generative AI programs and discrete personal devices have made it far more difficult for students to identify and report classmates suspected of cheating.
317
00:28:35.333 --> 00:28:45.258
The policy proposal stated plainly that AI tools on a small personal device have changed, quote, the external appearance of misconduct during an examination.
318
00:28:46.078 --> 00:28:50.020
end quote, making cheating much harder for other students to observe and hence report.
319
00:28:50.761 --> 00:28:59.426
Princeton's own senior survey found that nearly 30% of seniors reported cheating on an assignment or exam during their time there.
320
00:29:00.126 --> 00:29:05.549
44.6% reported knowing about the honor code violations they chose not to report.
321
00:29:05.649 --> 00:29:07.250
Only 0.4%, 0.4% reported a peer.
322
00:29:10.875 --> 00:29:15.924
One faculty member who supported the change said simply, quote, I think it's a shame, but it's necessary.
323
00:29:16.946 --> 00:29:22.896
Now the futurist lens here, I love that ending quote, I think it's a shame, but it's necessary.
324
00:29:23.957 --> 00:29:27.658
The shame is genuine, I think, and so is the necessity.
325
00:29:28.778 --> 00:29:31.379
So the Princeton honor system was not just a policy.
326
00:29:31.399 --> 00:29:32.619
This was a philosophy.
327
00:29:33.059 --> 00:29:41.300
This was an original idea proposed by the very students themselves at Princeton in 1893.
328
00:29:42.761 --> 00:29:43.161
Crazy.
329
00:29:43.341 --> 00:29:44.321
133 years ago.
330
00:29:45.941 --> 00:29:53.563
This was that a community's shared commitment to honesty could be stronger than any external enforcement mechanism.
331
00:29:54.643 --> 00:30:00.508
So if we all agree to be honorable, then no external force or mechanism is going to be able to break that honor code.
332
00:30:01.930 --> 00:30:06.814
And that trust is now gone.
333
00:30:08.396 --> 00:30:08.636
Right?
334
00:30:08.716 --> 00:30:16.043
So if everyone shared the same values and could see each other's behavior, the system ideally should be able to self-regulate without requiring any form of surveillance.
335
00:30:16.143 --> 00:30:16.463
Right?
336
00:30:16.948 --> 00:30:18.849
Kind of a radical idea, a crazy idea.
337
00:30:19.550 --> 00:30:21.991
Most institutions didn't trust students that way.
338
00:30:22.091 --> 00:30:25.013
Princeton did, and for 133 years, that's the way that things worked.
339
00:30:25.713 --> 00:30:30.997
But what AI has done is not just make cheating easier.
340
00:30:31.737 --> 00:30:33.138
I mean, cheating has always been possible.
341
00:30:35.508 --> 00:30:38.169
you know, in high school, in college.
342
00:30:38.189 --> 00:30:40.711
I mean, we've all seen cheating.
343
00:30:40.851 --> 00:30:43.332
We've all probably have cheated in some form of capacity.
344
00:30:43.352 --> 00:30:46.754
Try to like, you know, sneak at your neighbor's paper and see what's going on there.
345
00:30:48.075 --> 00:30:53.999
But what AI has actually done now is they've, I mean, they've made, it made cheating invisible.
346
00:30:55.559 --> 00:30:56.160
And it matters.
347
00:30:56.180 --> 00:30:56.500
It matters.
348
00:30:57.370 --> 00:31:00.152
A 2026 analysis made a striking observation.
349
00:31:01.052 --> 00:31:06.995
In 2012, 17% of students used phones to text answers during assessments.
350
00:31:07.696 --> 00:31:11.858
In 2026, 18% used AI to submit unedited work.
351
00:31:12.638 --> 00:31:16.680
The overall rate of outright academic dishonesty has barely moved.
352
00:31:17.981 --> 00:31:19.462
So what changed is not the...
353
00:31:21.319 --> 00:31:45.728
ethical disposition of students what changed is the behavioral signature of cheating so previous technologies calculator the internet smartphones these things all left traces all right someone typing on a phone under the desk you could see that someone glancing at a smart watch you could see that somebody taking out a calculator you could see that it was detectable
354
00:31:46.906 --> 00:31:54.713
The University of Reading found that 94% of AI-generated academic work goes undetected under current methods.
355
00:31:55.594 --> 00:32:04.202
Three-quarters of university CTOs now say AI is a moderate or significant risk to academic integrity at their institutions.
356
00:32:04.963 --> 00:32:10.648
And a growing number of universities across multiple continents are returning to paper-based exams.
357
00:32:11.772 --> 00:32:22.180
not out of nostalgia, but because paper is currently the only way, the only medium where you can be reasonably confident the person sitting in the exam is actually doing the work.
358
00:32:23.401 --> 00:32:25.443
I mean, when I was in college, all of my exams were on paper.
359
00:32:26.464 --> 00:32:28.405
Blue books, remember those Scantrons?
360
00:32:28.425 --> 00:32:35.771
All right, I mean, all of my exams when I was doing, I double majored in economics and psychology.
361
00:32:35.791 --> 00:32:37.632
It was business management, economics, and psychology.
362
00:32:37.653 --> 00:32:37.993
All of my
363
00:32:39.599 --> 00:32:43.981
Economics, the calculus, the business calculus, everything was done on paper.
364
00:32:45.222 --> 00:32:48.803
My economic history courses where you had to write essays, they were all done in blue books.
365
00:32:50.224 --> 00:32:53.686
So I never even had the experience of using these digital tools for the exams.
366
00:32:54.766 --> 00:32:58.108
Now, AI in a small device, it looks like thinking.
367
00:32:59.208 --> 00:33:01.369
There's no behavioral signature to observe.
368
00:33:01.770 --> 00:33:08.713
The pure accountability loop, which is the entire mechanism that Princeton was relying on, completely breaks it.
369
00:33:10.106 --> 00:33:11.127
The community still exists.
370
00:33:11.187 --> 00:33:13.429
The shared values kind of mostly still exist.
371
00:33:14.010 --> 00:33:20.297
But the ability to observe cheating and therefore to self-regulate, that's been removed.
372
00:33:22.279 --> 00:33:30.827
And I think Princeton's honor system is an unusually pure and well-documented version of a governance model that most organizations operate in some form.
373
00:33:31.740 --> 00:33:40.223
We set rules, we explain the rules, we explain the values behind them, we hire people who share those values and we trust the community, we trust our employees and our peers to self-govern.
374
00:33:42.144 --> 00:33:52.028
And then we have things like performance management, expense reporting, conflict of interest, disclosures, time tracking, quality, attestation and assessments.
375
00:33:53.568 --> 00:33:59.951
All of these rest on some of the assumptions and assertions of what Princeton built in 1893.
376
00:34:01.858 --> 00:34:11.142
The assumption underlying all these things is that human peers, managers, colleagues, they can observe behavior well enough that self-governance will function.
377
00:34:12.363 --> 00:34:26.610
And AI comes along and it's systematically disrupting that assumption, not because people are suddenly less ethical, although I suppose you could potentially make the argument for that, but because the observability
378
00:34:28.277 --> 00:34:39.121
the ability to see somebody using these tools and cheating and skirting the system is collapsing across every domain that AI touches.
379
00:34:40.841 --> 00:34:51.865
And so when you cannot actually see whether somebody's doing the work or an agent is doing the work, the trust architecture, that self-governance that...
380
00:34:53.445 --> 00:34:56.207
I just talked about at Princeton, it starts to erode.
381
00:34:56.907 --> 00:34:58.948
It erodes slowly at first and then faster.
382
00:34:59.889 --> 00:35:04.092
And so Princeton's response was to add external enforcement proctors.
383
00:35:04.732 --> 00:35:10.275
And that's because the community-based alternative no longer had the informational foundation it needed to function.
384
00:35:11.616 --> 00:35:17.420
So every organization, I think, is going to face a potential version of this decision internally.
385
00:35:18.721 --> 00:35:20.942
And the most important thing to understand is...
386
00:35:24.110 --> 00:35:36.115
It's not a moral failure that's gonna be forcing this, although it will become harder for an employee to act in a more moral way, but it's more of an architectural challenge.
387
00:35:37.436 --> 00:35:43.018
The community doesn't necessarily stop believing in honesty, you just lose the ability to observe it.
388
00:35:43.899 --> 00:35:49.121
And these are different responses that require different problems that require different responses.
389
00:35:49.141 --> 00:35:51.062
I mean, look at what organizations like Meta are doing.
390
00:35:52.302 --> 00:35:53.823
Look at what so many organizations are doing.
391
00:35:53.843 --> 00:35:55.663
They're trying to track employee keystrokes.
392
00:35:56.124 --> 00:36:06.848
They're trying to use AI to see if an employee is using certain words during calls or during drive-throughs.
393
00:36:06.888 --> 00:36:14.292
Like if a customer orders food through a drive-through window, we're using AI to like listen in to see if you're using the right keywords.
394
00:36:15.052 --> 00:36:21.015
So we're creating this kind of surveillance world to track, to monitor, to keep an eye on you.
395
00:36:21.855 --> 00:36:28.820
And so you can see that this really presses in on honesty, on the morality inside the organization.
396
00:36:28.840 --> 00:36:37.705
And I think this is going to lead to some very deep and structural philosophical debates and discussions around how moral can and should you be?
397
00:36:38.286 --> 00:36:40.507
How much surveillance should a company have?
398
00:36:40.627 --> 00:36:42.248
Can you bypass that surveillance?
399
00:36:42.268 --> 00:36:44.950
And you start to get into this situation.
400
00:36:45.816 --> 00:36:51.338
this battle that seems to transcend beyond just tools, beyond just governance.
401
00:36:51.478 --> 00:36:57.521
And it starts to, again, feel a little bit more philosophical in a lot of ways, right?
402
00:36:57.561 --> 00:36:58.581
Questioning morality.
403
00:36:59.781 --> 00:37:02.322
What's true inside your organization?
404
00:37:02.362 --> 00:37:02.783
What's not?
405
00:37:05.264 --> 00:37:08.645
So the next few years with these AI tools is going to be very interesting.
406
00:37:08.665 --> 00:37:13.567
And we already see so much pushback against AI, right?
407
00:37:14.630 --> 00:37:18.455
There was actually a, somebody might be able to fact check this.
408
00:37:19.596 --> 00:37:24.401
I can't remember who it was, but there was, somebody was giving a graduation speech at a university.
409
00:37:24.742 --> 00:37:26.043
And I just randomly saw a clip of it.
410
00:37:26.063 --> 00:37:27.585
I think it was on Instagram or something like that.
411
00:37:28.406 --> 00:37:35.234
And this lady gets up on stage and she is giving the commencement speech, I think a graduation speech, something.
412
00:37:36.566 --> 00:37:38.207
And she starts off her speech.
413
00:37:38.247 --> 00:37:42.468
And then one of the first things she says is that AI is the new industrial revolution.
414
00:37:42.508 --> 00:37:52.472
And you know, she's trying to talk about AI a little bit and you could just hear all the recent college grads start booing her and they're like, boo, boo.
415
00:37:52.532 --> 00:37:53.873
And she turns around, she's like freaking out.
416
00:37:53.893 --> 00:37:54.613
She doesn't know what to do.
417
00:37:54.653 --> 00:37:55.433
She's like, Oh my goodness.
418
00:37:55.493 --> 00:37:58.234
I, I guess I struck a curve, a chord.
419
00:37:58.854 --> 00:38:00.415
And she's like, can I, can I finish my speech?
420
00:38:00.435 --> 00:38:02.676
You know, she's looking behind her like what's going on and what do I do?
421
00:38:05.162 --> 00:38:20.348
So we see a lot of backlash, a lot of pushback, whether it's from unions, whether it's from employees entering the workforce, whether it's protests against data centers, whether it's, you know, throwing a Molotov cocktail in Sam Altman's house, whether it's removing people through elections on city councils.
422
00:38:21.609 --> 00:38:22.409
Massive pushback.
423
00:38:24.750 --> 00:38:30.473
So it's going to be very interesting to see how the next one to five years of this is going to play out.
424
00:38:31.412 --> 00:38:33.533
So those are the top stories of the day.
425
00:38:33.593 --> 00:38:34.874
I hope you enjoyed today's episode.
426
00:38:34.994 --> 00:38:42.698
If so, please consider rating and reviewing the podcast on Apple or on Spotify, whatever your preferred platform or channel is.
427
00:38:43.518 --> 00:38:46.720
It takes just a couple seconds and my team and I greatly appreciate it.
428
00:38:47.380 --> 00:38:50.481
I will be back tomorrow, of course, but I want to leave you with a quote.
429
00:38:50.541 --> 00:38:55.144
This one is from William Bruce Cameron, who's a sociologist or was a sociologist.
430
00:38:56.010 --> 00:39:00.692
And it's a quote that oftentimes gets misattributed to Albert Einstein, but it's actually not his.
431
00:39:01.032 --> 00:39:02.812
It's William Bruce Cameron's.
432
00:39:03.633 --> 00:39:06.494
Quote, not everything that can be counted counts.
433
00:39:07.174 --> 00:39:09.975
And not everything that counts can be counted.
434
00:39:10.915 --> 00:39:11.695
I'll say that one more time.
435
00:39:11.735 --> 00:39:13.436
Not everything that can be counted counts.
436
00:39:14.056 --> 00:39:16.617
And not everything that counts can be counted.
437
00:39:17.277 --> 00:39:19.898
Something for you to ponder and think about on this wonderful Wednesday.
438
00:39:20.578 --> 00:39:21.619
I will see you all tomorrow.
439
00:39:21.919 --> 00:39:22.799
Have a great rest of your day.
1
00:00:00.841 --> 00:00:06.506
Hey everyone, welcome to Future Ready Today, the number one podcast focused on the future of work.
2
00:00:07.187 --> 00:00:12.972
Every day I break down the top three stories that you need to be paying attention to and I break them down.
3
00:00:13.112 --> 00:00:19.598
I give you the insights, the trends, the key insights that you should be paying attention to so that you can be future ready.
4
00:00:20.139 --> 00:00:24.963
Today is Wednesday, May 13th, 2026 and the three stories that I want to get to today, number one.
5
00:00:25.724 --> 00:00:33.846
Goldman Sachs COO publicly calls his workforce a human assembly line and says that AI agents are the robots.
6
00:00:34.266 --> 00:00:40.227
Story number two, Amazon employees are gaming AI leaderboards by burning fake tokens.
7
00:00:40.868 --> 00:00:45.549
And story number three, Princeton ends a 133 years of trust-based exams because AI made cheating invisible.
8
00:00:55.971 --> 00:01:00.215
where they're not going to have proctors watching you take exams, again, because of the honor code.
9
00:01:00.856 --> 00:01:04.720
AI has abolished this 133-year-old tradition.
10
00:01:04.740 --> 00:01:05.240
And guess what?
11
00:01:05.881 --> 00:01:11.026
Now when you take an exam at Princeton, you're going to have somebody watching you actually take that test.
12
00:01:11.666 --> 00:01:13.468
So those are the top three stories of the day.
13
00:01:13.488 --> 00:01:14.329
We're going to break them down.
14
00:01:16.051 --> 00:01:28.562
talk about what these things actually mean, look at some other historical trends and stories and data to really help kind of connect what these things mean to the broader theme on the future of work.
15
00:01:29.042 --> 00:01:30.463
So let's get into the first story of the day.
16
00:01:30.503 --> 00:01:37.570
Goldman Sachs COO publicly calls his workforce a human assembly line, and he says that AI agents are the robots.
17
00:01:37.710 --> 00:01:37.850
Okay.
18
00:01:38.520 --> 00:01:40.422
This was in CNBC.
19
00:01:40.482 --> 00:01:42.343
This was an interview that he gave yesterday.
20
00:01:42.363 --> 00:01:44.185
And this is John Waldron.
21
00:01:44.365 --> 00:01:47.608
He's the chief operating officer over at Goldman Sachs.
22
00:01:48.669 --> 00:01:49.730
And he described...
23
00:01:51.576 --> 00:01:56.521
His exact words, quote, our human assembly lines will become more digitized.
24
00:01:57.121 --> 00:01:59.604
Digital agents will be our robots.
25
00:02:00.325 --> 00:02:11.156
He drew an explicit parallel to what happened in manufacturing decades ago when factories replaced workers with machines and said that banks simply haven't gone on that journey yet.
26
00:02:11.236 --> 00:02:11.517
He said,
27
00:02:12.077 --> 00:02:22.360
Until now, Goldman plans to deploy AI agents across the firm to automate workflows to cut costs and to boost productivity, what he called, quote, a big unlock.
28
00:02:23.080 --> 00:02:25.220
He insisted this won't mean mass layoffs.
29
00:02:25.761 --> 00:02:27.681
Headcount will remain roughly stable.
30
00:02:27.861 --> 00:02:38.484
New engineering jobs will be created and pushed back onto the media narrative, blaming AI for current industry layoffs, arguing those cuts still largely reflect post-pandemic growth.
31
00:02:39.364 --> 00:02:41.765
workforce corrections rather than AI displacement.
32
00:02:42.485 --> 00:02:47.907
But he also acknowledged candidly that he doesn't know dynamically how the overall headcount will change.
33
00:02:48.687 --> 00:02:53.369
And that honest uncertainty, I think, deserves some, you know, paying attention to.
34
00:02:54.269 --> 00:02:55.550
Now, the futurist lens here...
35
00:02:56.929 --> 00:03:01.273
is it's not so much that Goldman Sachs is deploying AI.
36
00:03:01.834 --> 00:03:05.217
I think every company around the world is in the process of deploying AI.
37
00:03:05.357 --> 00:03:11.823
I think for me, it's mainly about the language that Waldron used, because I think at his level as the COO,
38
00:03:12.664 --> 00:03:15.566
the language that's used is not accidental.
39
00:03:15.866 --> 00:03:20.169
And maybe he wasn't supposed to say that he views humans as, you know, the assembly line.
40
00:03:20.669 --> 00:03:26.232
Maybe he was just thinking it and, you know, the internal thought bubble made its way to the outside.
41
00:03:26.813 --> 00:03:37.860
But when the president and COO of one of the most powerful financial institutions on planet Earth describes his own workforce as an assembly line, I think he is...
42
00:03:39.770 --> 00:03:42.032
Maybe he's trying to normalize this framework.
43
00:03:42.792 --> 00:03:48.016
And the assembly line metaphor obviously has a history that's worth understanding.
44
00:03:48.036 --> 00:03:50.258
It goes back to Henry Ford.
45
00:03:50.858 --> 00:03:56.722
He didn't invent the assembly line, but Henry Ford, you could argue, he actually did perfect the assembly line.
46
00:03:57.463 --> 00:04:00.965
And when he did, it didn't just change how cars were actually made.
47
00:04:01.465 --> 00:04:05.448
It changed the fundamental relationship between a company and its workers.
48
00:04:06.169 --> 00:04:07.750
It transformed craftsmen.
49
00:04:08.465 --> 00:04:16.067
who made a living really focusing and owning a unique and specialized craft into component operators.
50
00:04:16.267 --> 00:04:18.388
You were just one piece on the assembly line.
51
00:04:19.348 --> 00:04:36.113
Now, of course, output went up, made the individual for the first time in history genuinely interchangeable, where if you were injured or you couldn't do a certain task on the assembly line, you were just swapped out, just like a part.
52
00:04:36.865 --> 00:04:38.326
It's like, okay, John can't come in today.
53
00:04:38.386 --> 00:04:43.148
All right, Jacob, go in there, take his spot, do the exact same thing that John was doing.
54
00:04:44.189 --> 00:04:45.189
And that's how it was viewed.
55
00:04:46.310 --> 00:04:49.051
Now, output in that situation went up, costs went down.
56
00:04:49.892 --> 00:05:01.158
The individual's bargaining power collapsed completely almost overnight because the skill that once made an individual irreplaceable kind of got absorbed into the system itself.
57
00:05:02.204 --> 00:05:14.274
And so what Waldron is telling us from Goldman Sachs, in you could say polite corporate language, is that the transition that happened in the assembly line is also coming for white-collar work.
58
00:05:15.375 --> 00:05:30.207
And at its height in the year 2000, and this was an article, I think I first came across this article, it was published almost a decade ago, Goldman's U.S. Cash Equities Trading Desk employed 600 traders.
59
00:05:31.559 --> 00:05:36.100
There were 600 traders who were on that cash equities trading desk.
60
00:05:37.220 --> 00:05:44.382
Today, and I don't even know if this number is even accurate because the last stat that I heard was from a few years ago, there are two.
61
00:05:45.803 --> 00:05:50.684
So we went from 600 to two, and I don't even know if it's still two, maybe it's one.
62
00:05:51.664 --> 00:05:55.185
But to go from 600 to two, wiping out 99.5% of everybody there,
63
00:06:00.159 --> 00:06:05.741
automated trading programs took the rest, supported by 200 engineers, computer engineers.
64
00:06:06.581 --> 00:06:18.664
Goldman's then CFO, Mary Chavez, revealed in 2017 that the firm had found a consistent ratio during the transition, four traders replaced by one computer engineer.
65
00:06:19.484 --> 00:06:25.966
At the time of the disclosure, 9,000 people, roughly a third of Goldman's entire staff, were already computer engineers.
66
00:06:27.062 --> 00:06:37.666
Goldman also mapped 146 distinct steps in a typical IPO process and said that many were, in Chavez's own words, begging to be automated.
67
00:06:38.666 --> 00:06:40.007
Okay, so you understand what they did there.
68
00:06:40.927 --> 00:06:53.192
They broke down, in this case, the steps that required to IPO, for a company to go public, 146 distinct steps.
69
00:06:53.232 --> 00:06:55.553
So they broke down that entire process of,
70
00:06:56.222 --> 00:07:03.468
which is a pretty complicated process, into these 146 steps and said, why do we need people doing this now?
71
00:07:03.708 --> 00:07:13.916
Now that we can break it down and standardize it and create a template for these 146 steps, we would love AI to be able to do it.
72
00:07:15.628 --> 00:07:33.972
Now, the difference now is that AI agents can attack the cognitive work, not just the kind of a template process work, but they can attack things like the pitch books, the client analysis, relationship intelligence, judgment calls, right?
73
00:07:34.052 --> 00:07:38.813
These are things that pure algorithmic trading in the past was never able to touch.
74
00:07:39.881 --> 00:07:50.449
Goldman's own research economist estimated earlier this year that AI threatens to automate tasks currently accounting for 25% of all American work hours.
75
00:07:51.250 --> 00:08:00.397
Bloomberg Intelligence projects that global investment banks may shed as many as 200,000 jobs over the next three to five years as they implement AI.
76
00:08:01.838 --> 00:08:07.743
But the interesting thing is that Goldman's own research team, they published a study in April.
77
00:08:08.698 --> 00:08:14.448
so just a month and a half or two ago, looking at 40 years of technological displacement.
78
00:08:15.470 --> 00:08:21.721
And what they found in their own analysis is that displaced workers don't just lose jobs.
79
00:08:22.777 --> 00:08:26.420
they experience what the researchers call occupational downgrading.
80
00:08:27.120 --> 00:08:34.165
They slide into roles requiring fewer analytical and interpersonal skills, not just fewer routine ones.
81
00:08:34.926 --> 00:08:45.694
And it's because the same technological forces that eliminated their old jobs also eroded the market value of their existing expertise.
82
00:08:46.852 --> 00:08:54.295
And what they found is that this displacement is somewhat slow, it's structural, and it's very hard to reverse once it actually takes hold.
83
00:08:55.195 --> 00:09:05.758
So Waldron's no mass layoffs assurance may be a genuine statement, but the assembly line analogy, if you kind of follow it all the way through, doesn't really promise job preservation.
84
00:09:06.099 --> 00:09:09.800
It promises role transformation, which is a very different thing.
85
00:09:11.140 --> 00:09:15.682
Now, confidentially, you know, in a recent CHRO meeting,
86
00:09:16.915 --> 00:09:19.739
that we had with a very, very large employer.
87
00:09:19.759 --> 00:09:33.377
And I won't mention who the company is, but at this executive, the executive at this company confidentially said that, because I asked them about the impact that AI is having on their workforce.
88
00:09:33.397 --> 00:09:33.517
Right.
89
00:09:35.402 --> 00:09:47.467
And she confidentially told us and the other CHROs in the room that it's not causing layoffs, but it is causing them to potentially keep their hiring more flat.
90
00:09:48.570 --> 00:10:01.777
So in other words, revising projections as far as how many people they want to bring into the company, kind of keeping it flat, not letting go of people, but also not hiring as aggressively as they thought they might potentially be doing.
91
00:10:02.418 --> 00:10:06.840
So this kind of aligns with what Goldman is saying here as well.
92
00:10:08.121 --> 00:10:14.064
Now, if you think about Ford, the workers at Ford didn't lose their jobs on day one.
93
00:10:15.293 --> 00:10:23.799
Okay, they slowly, they lost their craft identity, their autonomy, their wage premium, and then eventually they lost their leverage.
94
00:10:25.100 --> 00:10:30.023
So the reassurance and the risk can both be true at the same time.
95
00:10:30.043 --> 00:10:32.565
You can have the pros and cons happening at the same time.
96
00:10:33.345 --> 00:10:41.631
And that's what makes it so hard right now for so many leaders and for so many organizations to kind of navigate what the impact of AI is actually going to be.
97
00:10:42.711 --> 00:10:43.492
Now, story number two.
98
00:10:44.742 --> 00:10:55.325
And before I jump into story two, if you want more of these types of deep analysis, the deep dives, the breakdowns, go over to my newsletter.
99
00:10:55.425 --> 00:10:56.306
It's on Substack.
100
00:10:56.526 --> 00:10:59.387
It is called futureofworknewsletter.com.
101
00:11:00.967 --> 00:11:04.649
There is a free and a paid version that you can subscribe to.
102
00:11:05.870 --> 00:11:08.071
The free version, I do a weekly briefing.
103
00:11:08.331 --> 00:11:12.693
And the paid version, I do a deep dive, which I put out every one to two weeks.
104
00:11:13.254 --> 00:11:18.256
Really tackling a topic and, like it sounds, doing a very deep dive in it.
105
00:11:18.737 --> 00:11:20.518
Thousands of words, charts, graphs, facts.
106
00:11:21.118 --> 00:11:22.918
frameworks around a particular topic.
107
00:11:22.958 --> 00:11:28.980
The last one I did was how to actually engineer a jobs apocalypse if we wanted to do so.
108
00:11:29.460 --> 00:11:33.821
And it looked at the eight criteria that are required to make a jobs apocalypse happen.
109
00:11:34.321 --> 00:11:38.022
Again, you can learn more by going to futureofworknewsletter.com.
110
00:11:39.302 --> 00:11:44.744
So story two, Amazon employees, gaming AI leaderboards by burning fake tokens.
111
00:11:45.684 --> 00:11:47.784
So the Financial Times, they broke this story.
112
00:11:48.044 --> 00:11:49.665
I think it came out today.
113
00:11:51.719 --> 00:11:52.859
which I think is pretty interesting.
114
00:11:52.879 --> 00:11:56.482
And it's something that I've talked about on this podcast for a few months now.
115
00:11:57.102 --> 00:12:01.225
So Amazon deployed an internal AI tool called MeshClaw.
116
00:12:02.091 --> 00:12:21.706
And MeshClaw is, of course, inspired by the viral open source tool that so many people have talked about called OpenClaw that basically allows employees to create their AI agents that connect workflows to, I'm sorry, to connect workplace software and carry out tasks in kind of the workflow.
117
00:12:21.726 --> 00:12:23.387
So it automates some of the work that you're able to do.
118
00:12:24.328 --> 00:12:29.650
You kind of think of it as your own personal assistant or your own coworker, your own work assistant, right?
119
00:12:29.670 --> 00:12:33.391
So you can kind of plug them into your workflows, have them do things that you would usually do.
120
00:12:34.111 --> 00:12:42.354
So it can triage emails, it can initiate code deployments, it can interact with Slack, you know, genuinely useful stuff.
121
00:12:43.214 --> 00:12:50.277
But the problem was that Amazon simultaneously, while they introduced this, they also introduced targets.
122
00:12:51.400 --> 00:12:57.747
requiring more than 80% of developers to use AI every single week.
123
00:12:58.747 --> 00:13:05.433
And they started tracking AI token consumption, which many other companies have started doing, including JPMorgan Chase and Meta and several others.
124
00:13:06.694 --> 00:13:15.501
And so the token consumption is basically what a token, what that means is they're tracking the units of data that are being processed by AI models.
125
00:13:15.681 --> 00:13:17.723
And these units of data are known as tokens.
126
00:13:18.203 --> 00:13:22.887
And tokens, you can kind of think of as like three quarters or almost a full word, right?
127
00:13:23.708 --> 00:13:30.312
So every three quarters of a word that these AI systems are processing is considered one token.
128
00:13:31.332 --> 00:13:37.136
And what's happening is that a lot of these organizations are creating leaderboards to see how many tokens are you using.
129
00:13:37.916 --> 00:13:41.598
In other words, how much are you feeding into AI and how much is AI giving you back?
130
00:13:42.098 --> 00:13:43.439
So we're tracking your usage.
131
00:13:44.380 --> 00:13:48.462
And so companies created these internal leaderboards in the case of Amazon.
132
00:13:49.383 --> 00:13:49.963
So what happened?
133
00:13:51.229 --> 00:13:59.876
Well, in the case of Amazon, employees started using MeshClaw not to do actual work, but to inflate their token scores.
134
00:14:00.596 --> 00:14:07.662
They started automating unnecessary tasks, burning tokens on meaningless activity, gaming the leaderboard.
135
00:14:08.422 --> 00:14:15.908
One employee told the Financial Times, quote, there is so much pressure to use these tools that some people are just using MeshClaw to maximize their token usage.
136
00:14:16.709 --> 00:14:17.550
Another said, quote,
137
00:14:18.508 --> 00:14:19.689
Managers are looking at it.
138
00:14:20.270 --> 00:14:26.054
And when they track usage, it creates perverse incentives, and some people are very competitive about it.
139
00:14:26.995 --> 00:14:32.060
Amazon told employees that the token statistics would not factor into performance reviews.
140
00:14:32.880 --> 00:14:35.182
The employees, not surprisingly, didn't believe them.
141
00:14:35.763 --> 00:14:45.151
Meta employees have engaged in the same behavior on their own internal leaderboards, and the practice already has a name called, you got it, token maxing.
142
00:14:46.163 --> 00:14:47.264
Now, the futurist lens here.
143
00:14:47.845 --> 00:14:52.609
This story has a name, and the name is called Goodhart's Law.
144
00:14:53.570 --> 00:15:00.656
And the British economist Charles Goodhart, he first articulated this in 1975 in the context of monetary policy.
145
00:15:01.677 --> 00:15:08.604
And the law simply states that when a measure becomes the target, it ceases to be a good measure.
146
00:15:09.620 --> 00:15:12.043
And the mechanism is almost always the same thing.
147
00:15:12.824 --> 00:15:18.951
A well-intentioned metric is introduced, whether it's engagement, whether it's token maxing, whether it's, you know, whatever you want the metric to be.
148
00:15:19.852 --> 00:15:26.559
And people begin optimizing specifically for that metric rather than the underlying thing that the metric was actually meant to capture.
149
00:15:27.513 --> 00:15:30.396
And the number improves, meaning the metrics improving.
150
00:15:30.996 --> 00:15:37.942
But the reality it was supposed to represent actually starts to degrade or disappear entirely.
151
00:15:37.962 --> 00:15:39.703
And I'm going to give you a couple of historical examples.
152
00:15:40.844 --> 00:15:45.969
One of them, I think the most interesting and colorful example of this, is the cobra effect.
153
00:15:47.030 --> 00:15:50.112
And this is something that most people probably have never heard of or have talked about.
154
00:15:51.193 --> 00:15:52.634
But it's from British colonial India.
155
00:15:53.895 --> 00:15:54.096
And...
156
00:15:54.997 --> 00:16:01.120
This is where the government, they were very concerned about the large number of venomous cobras in Delhi.
157
00:16:02.060 --> 00:16:06.642
And so what they did is they offered a bounty to every dead cobra brought in.
158
00:16:07.483 --> 00:16:11.064
Now, obviously, enterprising locals, they saw this as a way to make money, and they thought, great.
159
00:16:11.204 --> 00:16:18.447
Well, if I can bring in dead cobras and I get kind of a bounty for it, well, it's in my interest to kill as many of these cobras as I can.
160
00:16:19.268 --> 00:16:20.328
So what do you think they started doing?
161
00:16:21.269 --> 00:16:23.570
They actually started breeding cobras.
162
00:16:24.272 --> 00:16:25.052
to collect the bounty.
163
00:16:25.852 --> 00:16:33.495
Because if you breed cobras, and then you kill the cobras and bring them in, you kind of have created your own personal ATM machine.
164
00:16:34.415 --> 00:16:41.137
And when the government discovered this, and then they canceled the program, what do you think the breeders then did?
165
00:16:42.057 --> 00:16:46.398
Well, they took all these cobras that they were breeding, and they just released them.
166
00:16:47.079 --> 00:16:48.519
Because now the cobras are worthless.
167
00:16:49.039 --> 00:16:50.119
So what's the point of killing them?
168
00:16:50.159 --> 00:16:51.300
What's the point of doing anything with them?
169
00:16:52.045 --> 00:16:52.945
You started breeding them.
170
00:16:53.225 --> 00:16:54.146
There's no more incentive.
171
00:16:54.246 --> 00:16:57.487
So now you let them out into the wild, let them out onto the streets.
172
00:16:58.707 --> 00:17:04.989
And so what ended up happening is they made the original problem worse than before the intervention.
173
00:17:05.630 --> 00:17:10.852
In other words, the metric became the goal, and then the goal destroyed the metric.
174
00:17:12.352 --> 00:17:14.853
This also happened in the Soviet Union.
175
00:17:16.058 --> 00:17:21.361
Now, my family is actually from the former USSR, from the Republic of Georgia.
176
00:17:21.741 --> 00:17:23.062
My grandmother, born in Ukraine.
177
00:17:23.262 --> 00:17:26.444
My mom and dad were both born in the Republic of Georgia in Belisi.
178
00:17:27.464 --> 00:17:34.368
And so, you know, I used to hear lots of different types of stories and examples of what was going on at the time.
179
00:17:35.609 --> 00:17:43.093
But one of the movements that actually came about in the 1930s was known as the Stakhanovite movement.
180
00:17:44.376 --> 00:17:46.198
And this was, again, in the 1930s.
181
00:17:46.518 --> 00:17:55.185
There was a famous coal miner named Alexei Stokhanov, who allegedly mined 14 times his daily quota in a single shift.
182
00:17:56.086 --> 00:18:03.012
And because he was able to mine so much coal, he was obviously portrayed as a national hero, basically saying, look how much this person can do.
183
00:18:03.393 --> 00:18:06.555
If he can mine that much coal, you should be able to mine that much coal too.
184
00:18:07.816 --> 00:18:11.499
And so the government used his achievement to raise quotas
185
00:18:12.680 --> 00:18:14.041
across all the different workers.
186
00:18:15.361 --> 00:18:27.546
Now, what the state propaganda did not publicize at the time, in Stalin's USSR at the time, is that Stokhanov actually had a team of assistants who were helping him out.
187
00:18:28.586 --> 00:18:32.968
And the feat was almost certainly staged for optics.
188
00:18:33.895 --> 00:18:41.046
And so the result was a workforce that quickly learned to optimize for quota numbers rather than for output.
189
00:18:41.066 --> 00:18:45.913
So managers kept quotas deliberately low so they could appear to overfill them.
190
00:18:45.933 --> 00:18:46.354
Okay.
191
00:18:46.784 --> 00:18:50.045
Workers gamed piece rate systems to maximize play.
192
00:18:50.725 --> 00:19:05.090
The numbers on paper and the output in reality diverged so far that the Soviet statistical agencies became internationally famous for producing figures that bore little relationship to what was actually happening in factories and on farms.
193
00:19:05.570 --> 00:19:06.730
You want a more recent example?
194
00:19:06.770 --> 00:19:08.431
Take a look at what happened with Wells Fargo.
195
00:19:09.576 --> 00:19:16.318
When the bank tied very aggressive cross-selling targets to performance evaluations, what did employees actually do?
196
00:19:16.838 --> 00:19:18.258
And of course, you probably remember this.
197
00:19:19.098 --> 00:19:23.199
Employees opened millions of unauthorized customer accounts to hit the numbers.
198
00:19:24.020 --> 00:19:32.602
And so the metric, which was accounts opened, completely decoupled from the thing it was supposed to measure, which was genuine customer relationships.
199
00:19:33.511 --> 00:19:41.675
And so the result was one of the largest banking scandals in American history, a $3 billion settlement, and years and years of reputational damage.
200
00:19:43.236 --> 00:19:47.878
And so now Amazon and a lot of these other organizations are doing the exact same thing with AI.
201
00:19:48.998 --> 00:19:52.260
And I think what makes the Amazon version particularly dangerous...
202
00:19:53.376 --> 00:19:57.859
is a finding from one of the most rigorous studies ever conducted on AI in 2025.
203
00:19:58.500 --> 00:20:01.182
This was a nonprofit called METR.
204
00:20:01.562 --> 00:20:02.523
M-E-T-R.
205
00:20:03.023 --> 00:20:08.727
They ran a randomized control trial using the same methodology used in clinical drug trials.
206
00:20:09.588 --> 00:20:18.971
And they did this on experienced software developers using frontier AI tools, including Cursor Pro, Cloud 3.5, and 3.7.
207
00:20:19.611 --> 00:20:22.593
And the result was unexpected and also pretty important.
208
00:20:23.313 --> 00:20:29.035
AI tools caused developers to take 19% longer to complete their tasks.
209
00:20:30.175 --> 00:20:31.596
Slower, not faster.
210
00:20:31.656 --> 00:20:34.597
Now, of course, you can make the argument the tools have gotten better since then.
211
00:20:35.637 --> 00:20:38.218
But here's the part that I think is really worth paying attention to.
212
00:20:39.394 --> 00:20:44.637
is that the developers actually thought that AI made them 20% faster.
213
00:20:45.938 --> 00:20:50.541
So they were wrong in the exact opposite direction and they had no idea.
214
00:20:51.742 --> 00:20:58.265
So their subjective experience of productivity and their actual productivity were moving in completely opposite directions.
215
00:20:59.226 --> 00:21:06.630
So it was taking them almost 20% longer to complete a task, but they thought it was taking them 20% less time to complete a task.
216
00:21:07.431 --> 00:21:08.872
And so I think this finding...
217
00:21:11.147 --> 00:21:15.808
shows that AI usage can produce wildly inflated self-assessments.
218
00:21:15.908 --> 00:21:18.409
And that's the problem with a lot of these things, self-assessments.
219
00:21:18.909 --> 00:21:20.889
This is the problem with engagement surveys.
220
00:21:20.929 --> 00:21:22.910
This is the problem with well-being studies.
221
00:21:22.950 --> 00:21:25.110
This is the problem with the four-day workweek studies.
222
00:21:26.010 --> 00:21:33.732
When you do these types of self-assessed studies, the data is not that valuable.
223
00:21:34.690 --> 00:21:46.378
I mean, obviously, if you tell me, for example, hey, Jacob, you can make the exact same amount of money, get the same speaking gigs, the same podcast views, the same everything, and only have to work two, three days a week, I'd say fantastic.
224
00:21:47.579 --> 00:21:49.500
And then, of course, I'm going to say I'm so much more happy.
225
00:21:50.260 --> 00:21:50.961
I'm much happier.
226
00:21:50.981 --> 00:21:51.681
I'm more productive.
227
00:21:52.282 --> 00:21:54.203
But does that mean I'm actually more productive?
228
00:21:54.603 --> 00:21:58.366
Am I actually, you know, are the numbers actually better from a business perspective?
229
00:21:59.287 --> 00:21:59.467
No.
230
00:21:59.487 --> 00:21:59.547
No.
231
00:22:00.458 --> 00:22:02.739
Right, subjectively, I might tell you I feel better.
232
00:22:03.559 --> 00:22:05.600
I think I'm more productive, I'm happier.
233
00:22:07.141 --> 00:22:07.581
But so what?
234
00:22:08.702 --> 00:22:12.243
Those types of self assessments do not mean much.
235
00:22:14.004 --> 00:22:18.967
So this is what I think a lot of organizations are missing.
236
00:22:18.987 --> 00:22:21.548
They're not paying attention to this, right?
237
00:22:21.848 --> 00:22:27.731
So a leaderboard, for example, an 80% adoption target,
238
00:22:28.620 --> 00:22:37.686
where a manager's watching token counts, you know, that creates, I think, the exact condition for Goodhart's law to operate at full force.
239
00:22:38.826 --> 00:22:44.430
Employees are optimizing for a number that was already, I think, a pretty poor proxy for value.
240
00:22:44.890 --> 00:22:47.932
Why does token usage mean that I'm adding more value?
241
00:22:48.592 --> 00:22:53.655
Why does token usage mean that I'm a more productive or a better employee or making more of an impact on the business?
242
00:22:53.695 --> 00:22:53.936
Why?
243
00:22:53.956 --> 00:22:54.296
It doesn't.
244
00:22:56.857 --> 00:22:57.098
So...
245
00:22:58.522 --> 00:23:00.443
I think what's happening is companies are actually making things worse.
246
00:23:01.624 --> 00:23:04.166
And the broader data out there confirms that it's not just Amazon.
247
00:23:05.147 --> 00:23:07.708
Thomson Reuters study from February earlier this year
248
00:23:08.931 --> 00:23:17.994
found that 82% of leaders say their organizations are either not collecting ROI metrics around AI usage or are unsure whether they are.
249
00:23:18.614 --> 00:23:28.757
A KPMG survey from September, 2025, found that 78% of leaders believe traditional business metrics are becoming insufficient to measure AI's impact.
250
00:23:30.218 --> 00:23:37.260
Companies are mandating the adoption of a technology whose value they largely cannot measure or don't know how to measure
251
00:23:38.904 --> 00:23:45.689
which is creating literally the exact environment where Goodhart's law takes over.
252
00:23:46.249 --> 00:23:48.811
The leaderboard is not the answer to the problem.
253
00:23:48.851 --> 00:23:52.374
The leaderboard is the acceleration of the problem.
254
00:23:53.895 --> 00:24:03.803
Token consumption, I think, is a treacherous metric for organizations to be optimizing for because it is invisible to common sense.
255
00:24:05.824 --> 00:24:08.126
A thousand tokens burned on a genuine customer...
256
00:24:10.029 --> 00:24:25.400
analysis or customer project and a thousand tokens burned on a mesh cloth agent that's running overnight, they might look identical on a leaderboard, on a dashboard, right?
257
00:24:25.440 --> 00:24:26.481
I'm burning a thousand tokens.
258
00:24:26.501 --> 00:24:27.662
You're burning a thousand tokens.
259
00:24:27.682 --> 00:24:29.563
You're doing who knows what with it.
260
00:24:29.623 --> 00:24:31.345
I'm actually trying to solve a customer problem.
261
00:24:33.066 --> 00:24:34.107
Is what we're doing the same?
262
00:24:34.127 --> 00:24:34.967
It's not.
263
00:24:35.008 --> 00:24:35.348
It's not.
264
00:24:36.397 --> 00:24:46.585
This is what a lot of researchers call a counterfeit reality, a data set that tells leadership the transformation is proceeding while the underlying behavior is actually degrading.
265
00:24:47.626 --> 00:24:53.070
And so the lesson for leaders is not that AI adoption metrics are necessarily bad or wrong.
266
00:24:53.610 --> 00:24:55.392
Some measurement of adoption is necessary.
267
00:24:55.412 --> 00:24:56.553
It's worth paying attention to.
268
00:24:57.033 --> 00:25:04.559
But the lesson is that activity metrics, the tokens consumed, the tools launched, the sessions logged, the percentage of employees who opened the app,
269
00:25:05.473 --> 00:25:06.854
These things are always going to get gamed.
270
00:25:07.094 --> 00:25:07.395
Always.
271
00:25:08.135 --> 00:25:08.636
They always have.
272
00:25:09.016 --> 00:25:09.577
They always will.
273
00:25:10.778 --> 00:25:16.242
The only metrics that resist Goodhart's Law are outcome metrics.
274
00:25:17.383 --> 00:25:18.985
Did the quality of the decisions improve?
275
00:25:19.946 --> 00:25:22.228
Did the customer problems get solved faster?
276
00:25:23.168 --> 00:25:24.970
Are we able to bring on more customers?
277
00:25:25.851 --> 00:25:27.352
Are we entering new markets?
278
00:25:27.812 --> 00:25:30.355
Did the work product actually get better?
279
00:25:31.359 --> 00:25:44.283
These are obviously much harder things to measure and they are much slower to see, but these are the only numbers that cannot be faked by any of these leaderboards, any of these dashboards like a mesh cloth agent that's burning tokens while an employee sleeps.
280
00:25:45.943 --> 00:25:49.304
So if you want to get serious about it, then get serious about it.
281
00:25:50.785 --> 00:25:52.425
What you measure grows.
282
00:25:53.445 --> 00:25:55.466
If you want to measure tokens, then guess what?
283
00:25:55.906 --> 00:25:59.227
You're going to see your employees burning through tokens like you wouldn't believe.
284
00:26:00.513 --> 00:26:02.756
But at the end of the day, who really cares?
285
00:26:04.198 --> 00:26:12.248
Story three, Princeton ends 133 years of trust-based exams because AI made cheating invisible.
286
00:26:13.392 --> 00:26:19.758
Now, before I get into this one, if you're watching this and you're listening to this, you're a chief human resource or chief people officer.
287
00:26:20.338 --> 00:26:26.323
We now actually have over 40 chief human resource officers at companies around the world who joined us.
288
00:26:26.363 --> 00:26:30.167
It's a big, exciting milestone for us.
289
00:26:32.028 --> 00:26:33.970
So we're very, very excited for this.
290
00:26:35.291 --> 00:26:37.573
Over 40 CHROs from companies around the world.
291
00:26:38.694 --> 00:26:38.914
And...
292
00:26:41.140 --> 00:26:46.943
If you want to learn more, you can check us out, futureofworkleaders.com.
293
00:26:46.963 --> 00:26:50.265
We get together every single month virtually.
294
00:26:50.285 --> 00:26:52.486
We get together one or two times a year in person.
295
00:26:53.207 --> 00:26:59.691
There's no reason for why you should be spending $20,000, $30,000, $40,000, $50,000 for some of these CHRO groups out there.
296
00:27:00.231 --> 00:27:00.791
You join us.
297
00:27:00.831 --> 00:27:03.373
You spend a tiny fraction of that amount.
298
00:27:04.193 --> 00:27:06.534
And I promise you it is well worth your while.
299
00:27:06.554 --> 00:27:09.355
We're specifically focused on the future of work.
300
00:27:09.775 --> 00:27:13.557
We're not talking about compensation and benefits and any of that kind of traditional stuff.
301
00:27:14.017 --> 00:27:16.078
We're talking more about the future of work.
302
00:27:16.098 --> 00:27:20.059
So if you want to learn more, check us out, futureofworkleaders.com.
303
00:27:20.139 --> 00:27:21.200
Or you can also email me.
304
00:27:21.900 --> 00:27:26.002
That is jacobatthefutureorganization.com.
305
00:27:26.642 --> 00:27:28.823
And you can learn a bit more there.
306
00:27:29.740 --> 00:27:31.001
Okay, so last story of the day.
307
00:27:31.521 --> 00:27:40.007
Princeton University faculty voted this week with a single dissenting vote to require proctors at all in-person examinations beginning July 1st.
308
00:27:40.748 --> 00:27:44.570
That ends 133 years of unbroken tradition.
309
00:27:45.211 --> 00:27:50.635
Princeton's honor system was established in 1893, not by administrators, but by students themselves.
310
00:27:51.435 --> 00:27:57.920
And these are people who petitioned the faculty to eliminate proctoring and take personal responsibility for academic integrity,
311
00:27:58.768 --> 00:28:04.271
The pledge, quote, I pledge my honor that I have not violated the honor code during this examination.
312
00:28:05.052 --> 00:28:08.614
This has been signed at the bottom of every Princeton exam since then.
313
00:28:09.254 --> 00:28:17.359
Faculty was explicitly banned from exam rooms under the university's rule for the entirety of that period.
314
00:28:18.140 --> 00:28:18.840
And what broke it?
315
00:28:19.381 --> 00:28:21.122
Well, you guessed it, AI.
316
00:28:22.806 --> 00:28:34.612
The dean of the college, Michael Gordon, wrote that generative AI programs and discrete personal devices have made it far more difficult for students to identify and report classmates suspected of cheating.
317
00:28:35.333 --> 00:28:45.258
The policy proposal stated plainly that AI tools on a small personal device have changed, quote, the external appearance of misconduct during an examination.
318
00:28:46.078 --> 00:28:50.020
end quote, making cheating much harder for other students to observe and hence report.
319
00:28:50.761 --> 00:28:59.426
Princeton's own senior survey found that nearly 30% of seniors reported cheating on an assignment or exam during their time there.
320
00:29:00.126 --> 00:29:05.549
44.6% reported knowing about the honor code violations they chose not to report.
321
00:29:05.649 --> 00:29:07.250
Only 0.4%, 0.4% reported a peer.
322
00:29:10.875 --> 00:29:15.924
One faculty member who supported the change said simply, quote, I think it's a shame, but it's necessary.
323
00:29:16.946 --> 00:29:22.896
Now the futurist lens here, I love that ending quote, I think it's a shame, but it's necessary.
324
00:29:23.957 --> 00:29:27.658
The shame is genuine, I think, and so is the necessity.
325
00:29:28.778 --> 00:29:31.379
So the Princeton honor system was not just a policy.
326
00:29:31.399 --> 00:29:32.619
This was a philosophy.
327
00:29:33.059 --> 00:29:41.300
This was an original idea proposed by the very students themselves at Princeton in 1893.
328
00:29:42.761 --> 00:29:43.161
Crazy.
329
00:29:43.341 --> 00:29:44.321
133 years ago.
330
00:29:45.941 --> 00:29:53.563
This was that a community's shared commitment to honesty could be stronger than any external enforcement mechanism.
331
00:29:54.643 --> 00:30:00.508
So if we all agree to be honorable, then no external force or mechanism is going to be able to break that honor code.
332
00:30:01.930 --> 00:30:06.814
And that trust is now gone.
333
00:30:08.396 --> 00:30:08.636
Right?
334
00:30:08.716 --> 00:30:16.043
So if everyone shared the same values and could see each other's behavior, the system ideally should be able to self-regulate without requiring any form of surveillance.
335
00:30:16.143 --> 00:30:16.463
Right?
336
00:30:16.948 --> 00:30:18.849
Kind of a radical idea, a crazy idea.
337
00:30:19.550 --> 00:30:21.991
Most institutions didn't trust students that way.
338
00:30:22.091 --> 00:30:25.013
Princeton did, and for 133 years, that's the way that things worked.
339
00:30:25.713 --> 00:30:30.997
But what AI has done is not just make cheating easier.
340
00:30:31.737 --> 00:30:33.138
I mean, cheating has always been possible.
341
00:30:35.508 --> 00:30:38.169
you know, in high school, in college.
342
00:30:38.189 --> 00:30:40.711
I mean, we've all seen cheating.
343
00:30:40.851 --> 00:30:43.332
We've all probably have cheated in some form of capacity.
344
00:30:43.352 --> 00:30:46.754
Try to like, you know, sneak at your neighbor's paper and see what's going on there.
345
00:30:48.075 --> 00:30:53.999
But what AI has actually done now is they've, I mean, they've made, it made cheating invisible.
346
00:30:55.559 --> 00:30:56.160
And it matters.
347
00:30:56.180 --> 00:30:56.500
It matters.
348
00:30:57.370 --> 00:31:00.152
A 2026 analysis made a striking observation.
349
00:31:01.052 --> 00:31:06.995
In 2012, 17% of students used phones to text answers during assessments.
350
00:31:07.696 --> 00:31:11.858
In 2026, 18% used AI to submit unedited work.
351
00:31:12.638 --> 00:31:16.680
The overall rate of outright academic dishonesty has barely moved.
352
00:31:17.981 --> 00:31:19.462
So what changed is not the...
353
00:31:21.319 --> 00:31:45.728
ethical disposition of students what changed is the behavioral signature of cheating so previous technologies calculator the internet smartphones these things all left traces all right someone typing on a phone under the desk you could see that someone glancing at a smart watch you could see that somebody taking out a calculator you could see that it was detectable
354
00:31:46.906 --> 00:31:54.713
The University of Reading found that 94% of AI-generated academic work goes undetected under current methods.
355
00:31:55.594 --> 00:32:04.202
Three-quarters of university CTOs now say AI is a moderate or significant risk to academic integrity at their institutions.
356
00:32:04.963 --> 00:32:10.648
And a growing number of universities across multiple continents are returning to paper-based exams.
357
00:32:11.772 --> 00:32:22.180
not out of nostalgia, but because paper is currently the only way, the only medium where you can be reasonably confident the person sitting in the exam is actually doing the work.
358
00:32:23.401 --> 00:32:25.443
I mean, when I was in college, all of my exams were on paper.
359
00:32:26.464 --> 00:32:28.405
Blue books, remember those Scantrons?
360
00:32:28.425 --> 00:32:35.771
All right, I mean, all of my exams when I was doing, I double majored in economics and psychology.
361
00:32:35.791 --> 00:32:37.632
It was business management, economics, and psychology.
362
00:32:37.653 --> 00:32:37.993
All of my
363
00:32:39.599 --> 00:32:43.981
Economics, the calculus, the business calculus, everything was done on paper.
364
00:32:45.222 --> 00:32:48.803
My economic history courses where you had to write essays, they were all done in blue books.
365
00:32:50.224 --> 00:32:53.686
So I never even had the experience of using these digital tools for the exams.
366
00:32:54.766 --> 00:32:58.108
Now, AI in a small device, it looks like thinking.
367
00:32:59.208 --> 00:33:01.369
There's no behavioral signature to observe.
368
00:33:01.770 --> 00:33:08.713
The pure accountability loop, which is the entire mechanism that Princeton was relying on, completely breaks it.
369
00:33:10.106 --> 00:33:11.127
The community still exists.
370
00:33:11.187 --> 00:33:13.429
The shared values kind of mostly still exist.
371
00:33:14.010 --> 00:33:20.297
But the ability to observe cheating and therefore to self-regulate, that's been removed.
372
00:33:22.279 --> 00:33:30.827
And I think Princeton's honor system is an unusually pure and well-documented version of a governance model that most organizations operate in some form.
373
00:33:31.740 --> 00:33:40.223
We set rules, we explain the rules, we explain the values behind them, we hire people who share those values and we trust the community, we trust our employees and our peers to self-govern.
374
00:33:42.144 --> 00:33:52.028
And then we have things like performance management, expense reporting, conflict of interest, disclosures, time tracking, quality, attestation and assessments.
375
00:33:53.568 --> 00:33:59.951
All of these rest on some of the assumptions and assertions of what Princeton built in 1893.
376
00:34:01.858 --> 00:34:11.142
The assumption underlying all these things is that human peers, managers, colleagues, they can observe behavior well enough that self-governance will function.
377
00:34:12.363 --> 00:34:26.610
And AI comes along and it's systematically disrupting that assumption, not because people are suddenly less ethical, although I suppose you could potentially make the argument for that, but because the observability
378
00:34:28.277 --> 00:34:39.121
the ability to see somebody using these tools and cheating and skirting the system is collapsing across every domain that AI touches.
379
00:34:40.841 --> 00:34:51.865
And so when you cannot actually see whether somebody's doing the work or an agent is doing the work, the trust architecture, that self-governance that...
380
00:34:53.445 --> 00:34:56.207
I just talked about at Princeton, it starts to erode.
381
00:34:56.907 --> 00:34:58.948
It erodes slowly at first and then faster.
382
00:34:59.889 --> 00:35:04.092
And so Princeton's response was to add external enforcement proctors.
383
00:35:04.732 --> 00:35:10.275
And that's because the community-based alternative no longer had the informational foundation it needed to function.
384
00:35:11.616 --> 00:35:17.420
So every organization, I think, is going to face a potential version of this decision internally.
385
00:35:18.721 --> 00:35:20.942
And the most important thing to understand is...
386
00:35:24.110 --> 00:35:36.115
It's not a moral failure that's gonna be forcing this, although it will become harder for an employee to act in a more moral way, but it's more of an architectural challenge.
387
00:35:37.436 --> 00:35:43.018
The community doesn't necessarily stop believing in honesty, you just lose the ability to observe it.
388
00:35:43.899 --> 00:35:49.121
And these are different responses that require different problems that require different responses.
389
00:35:49.141 --> 00:35:51.062
I mean, look at what organizations like Meta are doing.
390
00:35:52.302 --> 00:35:53.823
Look at what so many organizations are doing.
391
00:35:53.843 --> 00:35:55.663
They're trying to track employee keystrokes.
392
00:35:56.124 --> 00:36:06.848
They're trying to use AI to see if an employee is using certain words during calls or during drive-throughs.
393
00:36:06.888 --> 00:36:14.292
Like if a customer orders food through a drive-through window, we're using AI to like listen in to see if you're using the right keywords.
394
00:36:15.052 --> 00:36:21.015
So we're creating this kind of surveillance world to track, to monitor, to keep an eye on you.
395
00:36:21.855 --> 00:36:28.820
And so you can see that this really presses in on honesty, on the morality inside the organization.
396
00:36:28.840 --> 00:36:37.705
And I think this is going to lead to some very deep and structural philosophical debates and discussions around how moral can and should you be?
397
00:36:38.286 --> 00:36:40.507
How much surveillance should a company have?
398
00:36:40.627 --> 00:36:42.248
Can you bypass that surveillance?
399
00:36:42.268 --> 00:36:44.950
And you start to get into this situation.
400
00:36:45.816 --> 00:36:51.338
this battle that seems to transcend beyond just tools, beyond just governance.
401
00:36:51.478 --> 00:36:57.521
And it starts to, again, feel a little bit more philosophical in a lot of ways, right?
402
00:36:57.561 --> 00:36:58.581
Questioning morality.
403
00:36:59.781 --> 00:37:02.322
What's true inside your organization?
404
00:37:02.362 --> 00:37:02.783
What's not?
405
00:37:05.264 --> 00:37:08.645
So the next few years with these AI tools is going to be very interesting.
406
00:37:08.665 --> 00:37:13.567
And we already see so much pushback against AI, right?
407
00:37:14.630 --> 00:37:18.455
There was actually a, somebody might be able to fact check this.
408
00:37:19.596 --> 00:37:24.401
I can't remember who it was, but there was, somebody was giving a graduation speech at a university.
409
00:37:24.742 --> 00:37:26.043
And I just randomly saw a clip of it.
410
00:37:26.063 --> 00:37:27.585
I think it was on Instagram or something like that.
411
00:37:28.406 --> 00:37:35.234
And this lady gets up on stage and she is giving the commencement speech, I think a graduation speech, something.
412
00:37:36.566 --> 00:37:38.207
And she starts off her speech.
413
00:37:38.247 --> 00:37:42.468
And then one of the first things she says is that AI is the new industrial revolution.
414
00:37:42.508 --> 00:37:52.472
And you know, she's trying to talk about AI a little bit and you could just hear all the recent college grads start booing her and they're like, boo, boo.
415
00:37:52.532 --> 00:37:53.873
And she turns around, she's like freaking out.
416
00:37:53.893 --> 00:37:54.613
She doesn't know what to do.
417
00:37:54.653 --> 00:37:55.433
She's like, Oh my goodness.
418
00:37:55.493 --> 00:37:58.234
I, I guess I struck a curve, a chord.
419
00:37:58.854 --> 00:38:00.415
And she's like, can I, can I finish my speech?
420
00:38:00.435 --> 00:38:02.676
You know, she's looking behind her like what's going on and what do I do?
421
00:38:05.162 --> 00:38:20.348
So we see a lot of backlash, a lot of pushback, whether it's from unions, whether it's from employees entering the workforce, whether it's protests against data centers, whether it's, you know, throwing a Molotov cocktail in Sam Altman's house, whether it's removing people through elections on city councils.
422
00:38:21.609 --> 00:38:22.409
Massive pushback.
423
00:38:24.750 --> 00:38:30.473
So it's going to be very interesting to see how the next one to five years of this is going to play out.
424
00:38:31.412 --> 00:38:33.533
So those are the top stories of the day.
425
00:38:33.593 --> 00:38:34.874
I hope you enjoyed today's episode.
426
00:38:34.994 --> 00:38:42.698
If so, please consider rating and reviewing the podcast on Apple or on Spotify, whatever your preferred platform or channel is.
427
00:38:43.518 --> 00:38:46.720
It takes just a couple seconds and my team and I greatly appreciate it.
428
00:38:47.380 --> 00:38:50.481
I will be back tomorrow, of course, but I want to leave you with a quote.
429
00:38:50.541 --> 00:38:55.144
This one is from William Bruce Cameron, who's a sociologist or was a sociologist.
430
00:38:56.010 --> 00:39:00.692
And it's a quote that oftentimes gets misattributed to Albert Einstein, but it's actually not his.
431
00:39:01.032 --> 00:39:02.812
It's William Bruce Cameron's.
432
00:39:03.633 --> 00:39:06.494
Quote, not everything that can be counted counts.
433
00:39:07.174 --> 00:39:09.975
And not everything that counts can be counted.
434
00:39:10.915 --> 00:39:11.695
I'll say that one more time.
435
00:39:11.735 --> 00:39:13.436
Not everything that can be counted counts.
436
00:39:14.056 --> 00:39:16.617
And not everything that counts can be counted.
437
00:39:17.277 --> 00:39:19.898
Something for you to ponder and think about on this wonderful Wednesday.
438
00:39:20.578 --> 00:39:21.619
I will see you all tomorrow.
439
00:39:21.919 --> 00:39:22.799
Have a great rest of your day.
