>> People worry about AI for all kinds of
00:11
reasons. It's polarizing, it spreads
00:14
misinformation, it's coming for our
00:16
jobs. And those are all good reasons to
00:18
But what really keeps me up at night is
00:22
something else. Something that feels
00:24
more personal and perhaps even more
00:26
concerning for the human experience
00:28
I worry that AI will make us boring.
00:31
You, me, all of us.
00:34
That's because the more we outsource our
00:37
decisions to AI, the more we surrender
00:39
something deeply human.
00:41
Our capacity to explore, to take risks,
00:43
to stumble into the unknown, and
00:47
sometimes surprise even ourselves.
00:48
I worry without that without the
00:51
potential for discovery, for
00:53
serendipity, we risk becoming more
00:54
shallow and unidimensional versions of
00:57
who we are and who we could be.
00:59
Not just as humans, as individuals, but
01:02
Now, let me make this idea a bit more
01:06
concrete and take you somewhere perhaps
01:08
unexpected, but also totally relatable.
01:10
Imagine walking into a Baskin Robbins.
01:14
What you'll see is an illustrious
01:17
assortment of 31 different ice cream
01:18
flavors. There's chocolate chip,
01:20
pistachio almond, lemon sorbet, and many
01:22
You're now facing a difficult choice.
01:26
Are you going to go with an old
01:28
favorite, chocolate maybe, or are you
01:29
going to take a risk on something funky
01:31
and new like wild and reckless sherbet?
01:33
And yes, it's a real flavor.
01:37
The choice you're facing is a classic
01:40
human dilemma scientists call the
01:42
exploitation exploration trade-off.
01:44
Are you going to play it safe and
01:47
capitalize on what you know you like, or
01:49
are you going to take a risk with the
01:51
hope of finding something even better?
01:53
And our lives are full of these
01:56
trade-offs. Stick with your favorite
01:57
restaurant or try to fancy new spot on
01:59
the corner. Keep your usual haircut or
02:00
ask for something new and edgy. Stay in
02:03
your stable job or finally try to become
02:05
a pickleball champion.
02:08
We all differ in our appetite for
02:12
exploitation versus exploration, but
02:14
we're also all hardwired by evolution to
02:17
strike a certain balance between the
02:19
Because for our ancestors on the African
02:22
savanna, the trade-off was actually
02:23
pretty simple. Stick too closely to
02:25
what's safe and you risk starvation when
02:27
the grove runs dry. Stray too far into
02:29
the unknown and you might get poisoned
02:32
Luckily for you and me, the trade-offs
02:36
we face today are far less about
02:37
But our choices are still governed by
02:41
this instinct to balance caution with
02:42
If all you ever did in life was play it
02:46
safe and exploit, you'd never get
02:49
disappointed. But you'd also never get a
02:51
chance to advance and grow.
02:53
If all you did instead was discover and
02:56
explore, you'd collect endless
02:58
experiences, but you'd also never get a
03:00
chance to capitalize on your learnings.
03:02
So it's really this balance between
03:05
exploitation and exploration that fuels
03:07
Not just as a species, but also as
03:11
But that's exactly where AI throws a
03:15
wrench into the works.
03:17
Because AI hates risk.
03:19
And that's not because of its
03:22
algorithmic DNA, it's because the
03:23
systems we rely on to navigate our world
03:26
from Spotify to Netflix to ChatGPT are
03:28
overwhelmingly trained to optimize for
03:31
Or more specifically for short-term
03:35
engagement and satisfaction.
03:36
Did you click the link, watch the video,
03:39
like the song? If yes, AI gets a clap on
03:40
the shoulder. If no, gets a slap on the
03:44
Risk, discovery, and exploration are
03:47
simply not part of their programming.
03:49
Think back to Baskin Robbins. If 60% of
03:52
people prefer the flavor Pralines 'n
03:55
Cream, which is their best-selling
03:57
flavor, then that's what the AI is going
03:59
It's not that AI lacks imagination, it
04:02
Safe bets protect against
04:07
disappointment, which in turn reduces
04:09
So instead of taking this risky gamble
04:13
that helps you explore, companies tend
04:15
to err on the side of exploitation when
04:17
training their AI systems.
04:19
And to be clear, there's nothing
04:23
inherently wrong with that. I've spent
04:24
the last 15 years as a computational
04:26
social scientist working at the
04:28
intersection of psychology, computer
04:30
science, and business. And I know that
04:32
rewarding AI that way can be extremely
04:34
Algorithms in a way only become so good
04:38
at figuring out what you might enjoy or
04:41
currently be looking for because they
04:43
focus on exploitation.
04:44
And that's not just good for business,
04:47
it's also good and sometimes even
04:49
necessary for us as consumers.
04:51
For me, the mere thought of having to
04:54
choose between 31 different ice cream
04:56
flavors makes me dizzy.
04:57
And yet, most of the choices that we
05:00
face day-to-day are far more complex
05:01
Netflix, for example, has over 5,500
05:05
movies to choose from.
05:07
Spotify, over 100 million songs.
05:09
So the truth is that without a little
05:12
help from our exploitation-loving AI
05:14
friends, we simply don't stand a chance
05:17
of navigating modern life.
05:19
But I've also become increasingly
05:22
concerned that this upside comes at a
05:23
somewhat existential cost.
05:26
And to give you a first flavor of what I
05:28
mean by that, let's go back to Baskin
05:30
Before coming here, I ran a simple
05:33
experiment. I asked ChatGPT to recommend
05:35
one of their ice cream flavors. I did
05:38
that 100 times, each round pretending to
05:40
96 times, it recommended one of their
05:45
two most popular flavors, Pralines 'n
05:47
Cream and Mint Chocolate Chip.
05:50
Which, I realize and I'm sure you'd
05:53
agree, is not exactly a very diverse set
05:55
And it could mean that we soon find
05:59
ourselves in a world where Baskin
06:01
Robbins offers only those two flavors.
06:02
If no one ever picks the other 29, why
06:05
bother offering them in the first place?
06:07
Now, this might sound trivial and not
06:10
really existential at all.
06:13
But that's only until you realize that
06:15
this flattening of the human experience
06:17
happens across every aspect of life.
06:19
In studies with my collaborators and
06:22
students, we've shown that when people
06:24
use AI for guidance,
06:26
their preferences become more normative
06:28
and less diverse. Their creative output
06:30
becomes less unique. And their choice of
06:32
the most important scientists, athletes,
06:35
and historical figures becomes the same
06:37
as everyone else's.
06:39
So in a nutshell, AI turns this infinite
06:42
diversity of human opinions, beliefs,
06:44
and preferences into statistically safe
06:47
And that's only part of the problem
06:53
because even when AI learns your quirks,
06:55
say it figures out that you like nutty
06:57
coconut better than Pralines 'n Cream,
06:59
it will still play it safe within your
07:02
preferences. Meaning it is going to
07:04
optimize for what you are most likely to
07:06
I went back to ChatGPT for a second
07:10
experiment and this time told it about
07:11
my personal ice cream preferences. I
07:13
said that during my last 100 visits, I
07:15
picked nutty coconut 70% of the time.
07:18
The remaining 30% I split evenly among
07:21
three of my other favorites.
07:24
I then asked it to make the next 100
07:26
choices on my behalf.
07:28
And guess what it did.
07:30
It picked nutty coconut every single
07:32
are my adventures into occasional
07:38
chocolate fudge cookie dough or mango.
07:40
All I am at that point is literally a
07:43
But forget about ice cream. My students
07:50
and I have repeatedly shown the same
07:52
patterns when studying people's actual
07:54
AI narrows your taste, it flattens your
07:57
personality, and it scrubs away the
08:00
edges that make you interesting and keep
08:02
And the insidious part of all of this is
08:06
that the impact of AI on human
08:08
complexity is subtle. You won't notice
08:10
it happening overnight.
08:12
Rather, it's a death by a thousand
08:14
algorithmic recommendations. One
08:15
slightly safer movie, one slightly more
08:18
popular book, one slightly more
08:20
mainstream vacation at a time.
08:21
With every decision you outsource to AI,
08:24
you become more narrow. Then AI learns
08:27
from that shallower version of you and
08:29
narrows its recommendations even more.
08:31
The New York Times reporter Kashmir Hill
08:34
captured this dilemma perfectly in one
08:37
After outsourcing her decisions to AI
08:41
for a week, she complained about its
08:42
hidden agenda to turn her into a basic
08:44
And the funny part is it's the word
08:51
basic here that's the biggest insult.
08:52
Because being basic means being
08:57
unoriginal, unexceptional, and
08:59
It's offensive not because it suggests
09:02
failure, but because it implies
09:05
An absence of complexity.
09:08
And who really wants that for
09:12
Nobody. We all like the feeling of being
09:15
unique, of being special, of being
09:18
someone who doesn't easily fit into a
09:19
I for one don't want to become exactly
09:23
like everybody else. And I also don't
09:25
want to become a singular version of
09:27
So what can we do to course correct and
09:31
avoid destination basic
09:33
We can't hit the reset button on AI and
09:37
we shouldn't. As I said before, these
09:39
tools are incredibly valuable for
09:41
navigating modern life and they make us
09:43
better off in many ways.
09:45
But we need to reclaim our ability to
09:47
take risks and discover.
09:50
We need to find a way to rebalance
09:52
exploitation and exploration.
09:54
And ironically, AI could actually help
09:58
us accomplish just that.
10:00
But only if we ask it the right
10:02
questions and reward it for the right
10:03
So, instead of asking it to help us find
10:07
something we like, we could start asking
10:09
it to help us find something new.
10:11
Something we're likely to love even
10:13
though we've never tried it before.
10:15
And by that, I don't necessarily mean
10:18
you typing that question into ChatGPT
10:19
Although you should certainly try.
10:23
What I mean is getting companies to
10:24
harness AI superpower. It's ability to
10:27
detect patterns in vast amounts of human
10:30
data for exploration purposes.
10:32
Because AI has seen the entire universe
10:36
of preferences, it knows not only what
10:38
you currently like, but also what lies
10:40
just beyond the boundaries of your
10:42
typical preferences.
10:44
Which means that it can help you explore
10:46
smartly. To take curated risks when you
10:48
Now, here's what this could look like.
10:52
Imagine a dial on your Netflix account
10:54
or your Google search bar that lets you
10:56
decide how far from your typical
10:59
preferences you want to stray at any
11:01
given point in time.
11:03
On a regular day, you'll probably keep
11:05
the dial close to the spot-on setting
11:07
because you want to get the most
11:09
relevant hits right away.
11:10
But then on other days, you might feel
11:13
adventurous enough to push it a little
11:14
closer to the me with a twist setting
11:16
and ask for content that's a little
11:19
outside of your comfort zone, but still
11:21
And then on other days, for those rare
11:25
moments when you feel like you're really
11:26
ready to take on the world, you might
11:28
push the dial all the way to the wild
11:30
card setting and ask it to help you
11:32
discover something entirely new.
11:34
I for one would love to have such a
11:37
dial. It doesn't force me to leave my
11:39
comfort zone. I can always leave it on
11:42
the spot-on setting and benefit from the
11:43
convenience that comes with it.
11:46
But it gives me the choice and the
11:48
agency to break out of my little bubble
11:50
whenever I want to or I feel like I have
11:52
But here's the catch. For AI to toggle
11:56
its superpower between exploitation and
11:58
exploration, we need to incentivize it
12:01
Merely asking it to be more creative or
12:05
help you become more adventurous isn't
12:07
going to do the trick.
12:09
In most cases, it will still default to
12:11
the tried and tested output because it
12:13
craves that clap on the shoulder.
12:15
So, instead of punishing AI every time
12:17
it takes a swing and misses, we need to
12:20
start rewarding it for taking smart
12:22
For making bets that are a little bold,
12:25
a little unusual perhaps, but still
12:27
grounded in what it knows about us.
12:29
AI doesn't have to throw darts in the
12:32
dark. It can take informed risks. It can
12:34
help us optimize our exploration.
12:37
And when it does, we should treat it as
12:41
a success, not a failure.
12:42
Now, this isn't about desserts or
12:46
playlists. This is about preserving what
12:48
makes us uniquely, messily, and
12:50
It's your random passion for ET
12:54
memorabilia, your weird detour into
12:56
geocaching, or your stubborn preference
12:58
for a stick shift car.
13:00
It's about all the contradictions that
13:03
AI can't quite explain and the beauty of
13:05
us being just a little weird once in a
13:07
And if we believe that human complexity
13:12
is worth preserving, then the time to
13:14
Because we're at this inflection point
13:17
where AI is no longer just recommending,
13:19
but starting to act on our behalf.
13:21
It's choosing, not suggesting.
13:24
And that's where the stakes become
13:28
really existential.
13:29
So, next time AI offers you pralines and
13:31
cream, maybe say no and ask it to help
13:34
you go wild and reckless instead.
13:37