Resilience Starts With Us: Ethical Pathways to Innovation & Trust in AI

Responsible AI requires more than oversight—it demands ethical frameworks, human discernment, transparent governance, and embedding contextual judgment at key decision points. 

Humans provide reasoning, nuance, ethics, and context that machines cannot replicate. Integrating human oversight ensures safer, more reliable, and more trustworthy AI deployment. 

A responsible AI framework typically includes data governance, model validation, continuous monitoring, bias detection, explainability tooling, human-in-the-loop checkpoints, and risk-driven escalation mechanisms. The framework ensures safe deployment, traceability, and accountability across the AI lifecycle. 

Please view video here for a time-stamped transcript

00:12 – 00:15

Our guest has an extraordinary resume.

00:15 – 00:19

uh She is the CEO of Humane Intelligence.

00:19 – 00:26

She was also the United States Science Envoy for Artificial Intelligence, appointed by President Biden.

00:26 – 00:30

And I cannot wait to introduce her and have her join us, Dr.

00:30 – 00:32

Ruman Chowdhary.

00:42 – 00:44

Hi, good morning.

00:44 – 00:45

It’s so good to see you all.

00:45 – 00:47

I hope you enjoyed your pretzels.

00:47 – 00:48

My name is Dr.

00:48 – 00:56

Aman Chowdhury, and it was amazing to hear all the talks and all the conversations and actually see some familiar faces in the audience.

00:56 – 00:57

Hello.

00:58 – 01:03

So I’m here to talk to you today about ethical pathways to innovation and trust.

01:04 – 01:08

So how many people know who this man is?

01:09 – 01:18

I usually find that an audience, maybe one person does, but often I met with a lot of very confused looks when I say that he’s the reason we’re all alive today.

01:21 – 01:22

That’s right.

01:22 – 01:23

OK, so I’m going to explain.

01:23 – 01:25

name is Stanislav Petrov.

01:25 – 01:31

And in 1983, he was what we would today call the human in the loop.

01:31 – 01:38

He was on the receiving end of a technical system that would say whether or not the United States had launched a nuclear weapon.

01:38 – 01:39

He was on the Russian side.

01:39 – 01:51

And his job was to, in his bunker, receive the output that a missile had launched and then push the button for what, if you remember your world history or social studies, what’s

01:51 – 01:51

known as mutual

01:51 – 01:52

assured destruction.

01:52 – 01:56

And we’re all very, very fortunate that he didn’t.

01:56 – 02:02

And I love this story because when we think about human in the loop in technology, we don’t think about it very deeply.

02:02 – 02:05

We just think, there’s a person at the end and they’re just going to dot, dot, dot do something.

02:05 – 02:07

Well, he was a human in the loop.

02:07 – 02:10

And it’s a good thing that he didn’t just go ahead and push the button.

02:10 – 02:17

And he used a very, very unique thing to human beings, which is contextual clues and critical thinking.

02:17 – 02:20

And he sat there and thought, well, the Cold War has been going on for decades.

02:20 – 02:21

And there hasn’t been much

02:21 – 02:25

escalation, why would the United States set off a nuclear weapon?

02:25 – 02:27

And he chose not to push the button.

02:27 – 02:31

And that is the reason why we did not actually end up entering a nuclear war.

02:31 – 02:32

It’s incredibly important.

02:32 – 02:37

And what’s absolutely wild is really very few people know who this man is.

02:37 – 02:48

And I love this story again because it really demonstrates how human critical thinking cannot be replaced by even the most advanced technology.

02:49 – 02:52

So when we talk about artificial intelligence, what do we mean?

02:52 – 02:53

So in his day, Mr.

02:53 – 02:59

Petrov was uh subject to some of the most advanced technology.

02:59 – 03:08

And maybe it seemed intimidating or frightening that a human being could stand up and defy the output of such a sophisticated technological tool.

03:08 – 03:11

And that’s sometimes how we think about artificial intelligence.

03:11 – 03:16

And these are the kinds of headlines you may have opened up and seen in your news app today.

03:16 – 03:19

The solution to our education crisis might be

03:19 – 03:31

AI, are educators next in line for a robot takeover, lawyers as the next profession replaced by computers, and the omnipresent AI is as good as diagnosing disease X as any

03:31 – 03:33

human or doctor.

03:33 – 03:36

And there’s a couple of things I want to note about these headlines.

03:36 – 03:37

So number one.

03:37 – 03:42

These probably sound like they’ve come from the last three years, a generative AI revolution.

03:42 – 03:45

Actually, they all come from 2018 to 2020.

03:45 – 03:49

I’ve been in this field for almost a decade at this point, well, specifically in responsible AI.

03:49 – 03:52

I’ve been in tech for over 15 years.

03:52 – 03:56

And it seems like every new iteration, we see these headlines over and over again.

03:56 – 04:00

And of course, with every iteration, the technology is getting more more sophisticated.

04:00 – 04:04

But there’s a couple of things here that I find quite problematic.

04:04 – 04:07

Number one, there is this anthropomorphizing of the technology.

04:07 – 04:20

If you read these sentences, it sounds like there is some AI person in a machine that really wants to be a doctor or a teacher or a lawyer, and it goes and it drags your kid’s

04:20 – 04:21

teacher out kicking and screaming.

04:21 – 04:24

I mean, a robot takeover, my goodness.

04:24 – 04:25

Dramatic much.

04:25 – 04:29

But importantly, where is the person in this sentence?

04:29 – 04:37

In all of these headlines, and again, these are really common headlines you probably see every single day, the human being is just a passive

04:37 – 04:41

recipient of what the technology has chosen to do.

04:41 – 04:44

And that could not be further from the truth.

04:44 – 04:50

So what happened is what people think AI is is a sort of hoard of robots.

04:50 – 04:51

And instead, it’s not.

04:51 – 04:55

It’s something that looks fine.

04:55 – 04:56

Can I go back, please?

04:56 – 04:56

Sorry.

04:58 – 04:59

see if that happens.

04:59 – 05:00

There you go, thank you.

05:00 – 05:11

What people think AI is these hordes of robots coming in to change our lives and take our jobs and maybe teach our kids, but the reality is if you’re a developer like I am, it

05:11 – 05:13

looks a little bit more like that.

05:13 – 05:22

And actually the reality is those hyperbolic headlines don’t reflect the truth, which is that a lot of human effort and labor for everything from data collection, data curation,

05:22 – 05:29

cleaning, security, uh programming, debugging, and then product launching, scaling, and delivery,

05:30 – 05:49

and deployment, all of that actually requires people.

05:49 – 05:50

So we just don’t go into educators in line for robot takeover.

05:50 – 05:52

there’s a lot of human steps that happened in the meantime.

05:52 – 05:58

we talk things, and when we talk when we human things,

05:58 – 06:04

We are interacting directly with this technology that seems so futuristic and so capable.

06:04 – 06:11

Well, what happens is that we start to push technology forward that’s not actually thinking about what human beings might need.

06:11 – 06:22

And then specifically when the technology is not actually delivering the fantastical narratives uh that might be promised in a pitch deck, then what happens is we actually

06:22 – 06:28

assume that the human beings and that society needs to adjust to the limitations of the technology,

06:28 – 06:34

rather than thinking about how we can build the technology to adjust to what society needs.

06:34 – 06:37

And that’s something I call the retrofit human.

06:37 – 06:44

So when we think about this retrofit human, one of the ways we think about AI technology is, what is it good at?

06:44 – 06:46

What is it capable of?

06:46 – 06:50

And there’s a few things that machine learning and AI is actually quite good at.

06:50 – 06:55

So all of this preamble is not to say AI is not capable or machine learning can’t do things.

06:55 – 06:57

It actually can do quite amazing things.

06:57 – 07:02

But what’s important here is parsing out what people are really good at versus what technology is really good at.

07:02 – 07:03

So what’s technology really good at?

07:03 – 07:06

It’s good at identifying overarching patterns.

07:06 – 07:08

It’s good at parsing through

07:08 – 07:10

and identifying things that are similar to each other.

07:10 – 07:13

That’s basically most of the basis of computer vision.

07:13 – 07:17

And it’s generally consistent, depending on the error rate.

07:17 – 07:27

So when we build AI, and again, I’ve been doing this for quite some time, one kind of shorthand is to say that we collect a lot of data to build for the average person of a

07:27 – 07:29

particular type.

07:29 – 07:32

And we can collect very hyper-specific data.

07:32 – 07:35

And it’s not to say we’re doing it for the average person overall.

07:35 – 07:37

But we define you and me and the

07:37 – 07:41

person sitting next to you as part of a customer profile or community.

07:41 – 07:45

And maybe it’s based on 10, 20, 100 different data points.

07:45 – 07:50

And then you kind of assume that there are some assumptions made about the data that’s missing.

07:50 – 07:50

All right.

07:50 – 07:55

Well, I’m going to tell you a story about what it means to build for the average person.

07:55 – 08:04

And what does the average person even mean, even if you have a hyper-specific subset of people that one might think would be very, very predictable.

08:04 – 08:07

So one thing I love, by the way, is giving anecdotes and stories and a talk

08:07 – 08:07

that.

08:07 – 08:07

that.

08:07 – 08:07

that.

08:07 – 08:11

artificial intelligence, then nothing to do with artificial intelligence.

08:11 – 08:19

One of the biggest mistakes we do is we think that this technology is new and that all the problems we face are things that we’re never going to see, that we’ve never seen before.

08:19 – 08:21

We’ve never asked these questions.

08:21 – 08:31

And while AI is introducing new paradigms and new questions and new issues, the things that we are solving for are not net new questions.

08:31 – 08:32

So this is one of my favorite stories to tell.

08:32 – 08:37

So in 1940s, uh airline pilot crashes were at an all time high.

08:53 – 09:07

And so the US Air Force went back to the airplane seat manufacturers and said, look, we need you to redesign these seats because the average profile of the

09:07 – 09:08

our pilots has changed.

09:08 – 09:13

Of course, if we think about who these pilots are, they’re actually a pretty narrow subset of people.

09:13 – 09:17

They’re young men between probably 18 to maximum 30.

09:19 – 09:20

They’re quite healthy.

09:20 – 09:22

They’re able-bodied, et cetera.

09:22 – 09:24

And in that time, there would be no women pilots.

09:24 – 09:26

It would be pretty standardized.

09:26 – 09:33

So they did in 1940s what they did in the 1920s, which is they took over 4,000 of these pilots.

09:33 – 09:36

They measured them across 10 different data points.

09:36 – 09:37

Does that sound familiar?

09:37 – 09:43

If you’re a data scientist in the room, this sounds like a pretty reasonable way to design a good seat.

09:43 – 09:48

So over 4,000 men, 10 different data points, and they took the average and they made a seat.

09:49 – 09:59

And it served not a single person, not a single one of those 4,000 some odd pilots sat in that seat and said, yep, this is comfortable for me.

09:59 – 10:04

And the reason is that this idea of an average is this 50th percentile person.

10:04 – 10:07

If you know a bell curve, it’s the biggest part of the bell curve.

10:07 – 10:11

curve, uh but that is not necessarily a real person that exists.

10:11 – 10:20

It’s a fantastical person that is an amalgamation of all the other data points you’ve collected about the people like them.

10:20 – 10:22

So they whittled it down to three dimensions.

10:22 – 10:26

So what if we just take three data points, have more course data, right?

10:26 – 10:29

That only fit about 13 % of pilots.

10:29 – 10:34

Well, then the Air Force went back and said, OK, airplane manufacturers, this is not working for us.

10:34 – 10:36

Because of course, this is a life and death situation.

10:36 – 10:41

These seats need to actually be functional and useful, and pilots are in very extreme situations.

10:41 – 10:43

So this is not an afterthought.

10:43 – 10:50

So they went back and they said, actually, rather than fitting for the 50th percentile person, the average person, what I want you to do is actually make

10:50 – 10:53

the seats work for 95 % of all pilots.

10:54 – 10:57

which is a radical departure of how you measure success, right?

10:57 – 11:02

Because so much of what you’re making is how you’re defining what success is and what good looks like.

11:02 – 11:09

So they redefined good to say, rather than making a seat for the average pilot, why don’t we make the seat that fits 95 % of all pilots?

11:09 – 11:13

And again, this is still a small subset of people with not a lot of variation.

11:13 – 11:19

So all the airplane manufacturers came back and said, impossible, can’t do it, never done it before, we’re going to go broke.

11:19 – 11:21

This is an absolutely wild idea.

11:21 – 11:24

And of course, it’s the Air Force, so they can do what they want.

11:24 – 11:27

like, all right, fine, we’ll go find somebody else who’ll do it.

11:27 – 11:28

And what they did.

11:29 – 11:31

is they just made adjustable seats.

11:31 – 11:42

So whenever you get in your car or a bus or an airplane and you can do things like move the armrest and lift the headrest and tilt the seat differently, it’s a reason why someone

11:42 – 11:49

who’s 5’3 like me and the average 5’10 guy can actually get into the same car and drive it.

11:49 – 11:56

It’s because the military pushed back and actually said, we want you to rethink how you design.

11:56 – 11:59

We want you to rethink how you’re making this technological tool.

11:59 – 12:06

again, rather than saying we’re building for an average, we’re going to say we want this to embrace a wider range of people.

12:06 – 12:12

A radical departure that actually, again, has led to better designed and better output products.

12:12 – 12:13

So what are your key takeaways?

12:13 – 12:22

Number one is in order for AI to work, we need to think radically and think differently about how to customize and build for the users that we have.

12:22 – 12:26

The second is clear change of accountability and responsibility.

12:26 – 12:29

I coined this term moral outsourcing, and it ties

12:29 – 12:38

to this anthropomorphizing of artificial intelligence, how we use phrases that make it seem like AI has intent, AI has reason, it actually has none of it.

12:38 – 12:41

It’s human beings that build and deploy technology.

12:41 – 12:49

Rather than saying we’re building AI to replace lawyers, what’s actually happening is somebody is making a company that is legal tech.

12:49 – 12:56

And you can choose to make legal tech in a way that is augmenting a human, or you can choose to make legal tech in a way that’s intentionally trying to replace them.

12:56 – 12:59

And it’s a human-driven decision, a human-driven decision on how

12:59 – 13:00

how technology is being used.

13:00 – 13:08

And then finally, achieving the gains of artificial intelligence requires as much investment in the human side as the technical side.

13:08 – 13:18

So rather than seeing the fact that all of our bodies are different sizes and shapes and heights as a failing, it’s actually an opportunity to design better and to design more

13:18 – 13:19

broadly.

13:20 – 13:22

So how will AI impact our society?

13:22 – 13:33

So I will say the number one talk I’m asked to give, and this started this year, and I’ve done so many speaking engagements, I’ve talked a lot about ethical responsible AI, but the

13:33 – 13:37

future of work in 2025 has really become my number one requested topic.

13:37 – 13:47

ah And what I’m glad to report is that actually there’s a lot of empirical evidence, there’s a lot of really great work that’s happening where people are talking about, know,

13:47 – 13:50

quantitatively they’re measuring and studying the impact of

13:50 – 13:51

AI in the workforce.

13:51 – 13:59

So in one hand, again, what you may see in the news when you open your news reader is something like people calling for universal basic income, people saying that there will be

13:59 – 14:01

no jobs left for anybody.

14:01 – 14:07

And what we’re finding instead is that it’s actually quite, it’s more granular conversation.

14:07 – 14:10

So I’m going to talk to you about three different papers that I really love.

14:10 – 14:13

There’s actually quite a few papers out there, but there’s three papers that I love.

14:13 – 14:18

I want to start with a big picture, like macroeconomic impact across the globe over the next 10 years.

14:18 – 14:20

Medium-sized picture, which is vertical

14:20 – 14:30

by industry and the most granular and experimental study that was done at a specific company to see how employees reacted and how they responded to working with AI and how

14:30 – 14:31

that impacted them.

14:31 – 14:40

I will also start by actually saying that I speak at a lot of universities and actually there is a lot of very justified concern amongst young people.

14:40 – 14:50

And what we’re finding, and again this is the first year we’re really seeing it, that Gen Z entering the market to get jobs, entry-level jobs are having a harder time because

14:50 – 14:53

artificial intelligence is actually able to do some of the basic level jobs.

14:53 – 14:56

And this is not just research and writing.

14:56 – 14:58

It’s actually primarily in programming.

14:58 – 15:02

So what AI is really, really good at actually is junior level programming.

15:02 – 15:07

Now, you still need senior developers to ensure that the code is robust, it’s functional, et cetera.

15:07 – 15:11

But a lot of the junior roles are just not being replaced.

15:11 – 15:17

So as people are aging up and they’re getting promoted, their junior roles are just not getting replaced.

15:17 – 15:20

And it’s worrisome about the future.

15:20 – 15:27

So the first paper I want to talk about is by a labor economist, uh Professor Dara Nesemoglu at MIT.

15:27 – 15:32

he studied, and again, he’s somebody who’s been doing econ for very, long time.

15:32 – 15:37

And he’s really entered the AI space thinking about the impact on work and productivity.

15:37 – 15:45

So he looked at what’s called total factor productivity, which is pretty much if you summed up the GDP of every country, that would be sort of the TFP.

15:45 – 15:50

ah And what he’s calculated is that over the next 10 years, it’s going to be what he calls

15:50 – 16:01

So, about non-trivial but modest impact, so like sub 1%, which may not sound like a lot, but by the way, sub 1 % of the entire GDP of the entire world, so pretty significant.

16:01 – 16:03

So, that’s a pretty significant amount of automation.

16:03 – 16:06

But he dug even deeper to say what is being automated?

16:06 – 16:13

And he points out that what a lot of people have is that it’s actually more white-collar jobs than blue-collar jobs getting automated.

16:13 – 16:14

And this is interesting.

16:14 – 16:19

This really points out one of the things that’s particularly radical about generative AI.

16:32 – 16:33

And point.

16:33 – 16:36

And And And I think that’s important point.

16:36 – 16:36

And point.

16:36 – 16:38

And really important point.

16:38 – 16:50

And I important And really important point.

16:50 – 16:50

And

16:50 – 16:54

really knowledge labor, rudimentary knowledge labor that’s being automated.

16:54 – 16:56

We haven’t really thought about that before.

16:56 – 16:59

We’ve not had to deal with the impact of that.

16:59 – 17:08

And by the way, going back to Gen Z, some of you, if you have kids of that age, might know that your kids or their friends might actually be thinking about getting blue collar jobs.

17:08 – 17:14

Because increasingly, young people don’t know what the value is of college education if AI is actually able to do that work.

17:14 – 17:20

But an AI is never going to be able to mow your lawn or fix your plumbing or lay out a

17:20 – 17:21

roof.

17:21 – 17:32

And what they’re seeing is that, for many reasons, a blue collar job to a lot of them seems a lot more, you know, predictable and reliable than trying to aim really high for

17:32 – 17:40

some sort of a senior level VP role that may or may not manifest itself because you can’t get your first job, you’re never going to get promoted.

17:40 – 17:48

So, you know, he noted that AI is predicted to widen the gap between capital and labor income, but also that there’s going to be this negative societal value that we really have

17:48 – 17:50

to take into account.

17:50 – 18:02

most interesting things I’m seeing is as we continue to measure the value of AI, people are increasingly bringing in the negative externality on society as a factor to consider.

18:02 – 18:09

Because what is the purpose of making a productivity machine if we are all completely incapable at the end of it?

18:09 – 18:14

There is serious concern about things like over-reliance, about people use this phrase, AI psychosis.

18:14 – 18:20

uh we’re seeing there’s education tech coming out, and there’s concern

18:20 – 18:22

about young people being over-reliant on the technology.

18:22 – 18:28

How do we think about and address these considerations while building technology that is still net beneficial?

18:29 – 18:37

So the second paper I like to talk about is actually from 2023, and it’s actually by a team at OpenAI that worked with another group of economists.

18:37 – 18:41

It’s called GPTs or GPTs, and it looks at different verticals.

18:41 – 18:49

And rather than kind of saying blanket, all of these jobs are going to get automated away, it said, well, sector by sector, what is more likely to be automated away?

18:49 – 18:59

And the shorthand to take away is that about 80 % of the workforce would have about 10 % of your tasks automated, and roughly 19 % would have about half of tasks impacted.

18:59 – 19:04

And that pretty much tracks, that was a few years ago, and it kind of tracks the trends that we’re seeing.

19:04 – 19:13

I think a lot of us use AI tools for very basic productivity, Like helping you review your notes or maybe doing meeting summarizations for you.

19:13 – 19:18

It’s not taking over your job, and in fact, it’s probably giving you a little bit of extra time back.

19:18 – 19:22

most of us in this room are probably part of that 80%.

19:22 – 19:29

That 19 % of workers that are seeing 50 % automated away, they pointed at professions like journalism, paralegal.

19:29 – 19:35

more of like a knowledge research role that maybe does not require deep expertise.

19:36 – 19:37

So now the third paper.

19:37 – 19:46

And again, one of the biggest topics that people want to understand is what does artificial intelligence mean in my workforce if I use it, if my employees use it, if my

19:46 – 19:47

company adopts it?

19:47 – 19:56

Well, overwhelmingly, will tell you that the answer is human plus AI is greater than human without AI or AI alone.

19:56 – 19:58

And I love this study because, first of it has a fabulous name.

19:58 – 20:00

It’s called The Cybernetic Teammate.

20:00 – 20:04

And it was a case study done at Harvard Business School working with Procter & Gamble.

20:04 – 20:06

And they set up a randomized controlled experiment.

20:06 – 20:17

experiment, about 800 P &G employees, they had a multi-day hackathon, and they basically put people into groups, it’s either teams or individuals, some got AI and some didn’t.

20:17 – 20:24

And what I like about this paper, often a lot of the papers about uh AI in the workplace is about productivity.

20:24 – 20:30

And I really, really, really hate that measurement, because uh there’s one thing AI can do, it’s just produce.

20:30 – 20:31

That’s all it does.

20:31 – 20:34

And actually, it’s often confidently incorrect.

20:35 – 20:36

It takes the human being to be

20:36 – 20:38

to understand whether this output.

20:38 – 20:43

I mean we have a word for it now we call it slop workforce slop AI slop right.

20:43 – 20:45

is literally the term for it now.

20:45 – 20:50

So just measuring pure production is not really giving you any answers.

20:50 – 20:57

But what they tested for were things like quality of work time to completion and importantly the ability to assist non expert workers.

20:57 – 21:05

Because on the other end the concern for more senior people or at least one of the narratives is that well maybe in the future you won’t need expertise.

21:05 – 21:06

You just need

21:06 – 21:11

an inexperienced human, you slap on some AI, and then they’re the same as a senior developer.

21:11 – 21:13

So they wanted to test all of that.

21:14 – 21:20

So how do they do it so you don’t have to read a lot of complicated uh bar charts?

21:20 – 21:22

the first thing they looked at was average solution quality.

21:22 – 21:25

So remember, they had the individual with no AI as the baseline.

21:25 – 21:27

They have the teams with no AI.

21:27 – 21:29

They have individual plus AI and team plus AI.

21:29 – 21:33

So you’ll see that adding AI was somewhat of a benefit.

21:33 – 21:38

So the individual with no AI clearly did not do as well as any of the other subgroups.

21:38 – 21:41

And if they were given AI, they tended to do better.

21:41 – 21:44

But there’s actually not much of a discernible difference between

21:44 – 21:46

individual plus AI and team plus AI.

21:46 – 21:50

And remember, I’m going to get little statsy with you for a second.

21:50 – 21:53

The bar, like don’t worry about where the bar actually is.

21:53 – 22:00

It’s actually the error term, that line that’s telling you the upper, that is the expected area where the value is going to fall.

22:00 – 22:09

So all of that is to say if the bars are overlapping significantly, which they are, the black lines, they’re likely to be fairly, fairly similar numbers.

22:09 – 22:15

So the second thing they looked at is whether or not it being the core job mattered.

22:15 – 22:19

So what if you took a completely inexperienced person, gave them AI?

22:19 – 22:23

uh Well, the reality is that doesn’t quite work.

22:23 – 22:25

And I think intuitively we all know this.

22:25 – 22:27

Again, I just referred to AI slop.

22:27 – 22:29

AI is very convincingly incorrect.

22:29 – 22:30

It sounds smart-ish.

22:30 – 22:39

It kind of sounds like you gave a kid, if you heard of the term Dunning-Kruger syndrome, really suffers the biggest Dunning-Kruger syndrome.

22:39 – 22:47

It’s synthesizing a bunch of information without context and it’s sort putting it out there, but it’s not necessarily good.

22:47 – 22:49

It’s just volumetric.

22:49 – 22:59

And overwhelmingly what they found was that individuals and teams with no expertise or limited expertise got somewhat of a boost, but really you still needed people in the room

22:59 – 23:05

who understood the output and what was needed in order to really get the gains of this technology.

23:05 – 23:07

And then finally, of course, time saved.

23:07 – 23:09

So this is super interesting, which is that

23:09 – 23:11

And we know this intuitively.

23:11 – 23:15

Teams are maybe less good at getting things done fast.

23:15 – 23:19

is better, but they’re not as good.

23:19 – 23:22

You have to do lot of consensus building, getting people’s inputs and opinions.

23:22 – 23:28

And AI did save some time, but of course, if they were working by themselves with AI, they work the fastest.

23:28 – 23:32

And the most important part, who did the best?

23:32 – 23:39

Well, it turns out that teams using AI, even if they were not necessarily the fastest at producing something, actually

23:39 – 23:43

tended to turn out the most top 10 % solutions.

23:43 – 23:44

So this a really fascinating finding.

23:44 – 23:45

Again, very, very early.

23:45 – 23:49

There’s another really great paper that I don’t have here by Microsoft.

23:49 – 23:54

And this is actually about the use of Copilot, which is great because that was actually a much bigger study.

23:54 – 23:58

I think it was 20, 30 different industries, really looking a lot more deeply at it.

23:58 – 24:00

So there’s a lot of great research coming out of Stanford.

24:00 – 24:04

If there’s something that interests you, there is a significant amount of research happening.

24:04 – 24:08

Eric Ben-Yolson at Stanford is doing a lot of this work.

24:08 – 24:09

And again, lot of the companies are coming

24:09 – 24:12

out with as well, because it is something that we’re all quite curious about.

24:12 – 24:14

How do we integrate this technology?

24:15 – 24:21

Well, and one of the things we have talked about for quite some time is this term, you know, digital literacy.

24:21 – 24:23

What does it mean to be AI literate?

24:23 – 24:28

And I think the earlier, more naive version of it was something like, oh, everyone needs to know how to program.

24:28 – 24:30

Everyone just needs to code.

24:30 – 24:31

But there was no context for it.

24:31 – 24:33

Like, what does that mean?

24:33 – 24:39

What does it mean for a CEO to code or program or a senior VP or a product

24:40 – 24:42

It’s not really relevant, right?

24:42 – 24:49

So one of my favorite books about what digital literacy is, or the concept of digital literacy, is actually written by an English professor.

24:49 – 24:52

And as you probably know by now, I like using older examples.

24:52 – 25:01

His book was written in 2003, and it was written for his English students who are now navigating doing research on this crazy new thing called the internet.

25:01 – 25:09

What did it mean for research and the credibility of your work if you no longer went to a library or interviewed people and had a microfiche

25:09 – 25:19

and these primary sources, what happens now that you’re using a computer to do this work, or the internet, or this new unreliable thing called Wikipedia to help you with your

25:19 – 25:19

research?

25:19 – 25:22

So there’s three levels of literacy that he talks about.

25:22 – 25:24

And it’s so relevant to AI.

25:24 – 25:32

And I really like this rubric because it really helps you get your head around how you should be thinking about being AI literate, no matter where you are in your usage of the

25:32 – 25:33

technology.

25:33 – 25:34

So level one is functional literacy.

25:34 – 25:37

This is for people who are just on the receiving end of a technology.

25:37 – 25:39

And this is your average customer.

25:39 – 25:42

The person interacting with the chatbot, right?

25:42 – 25:46

The person who’s trying to use the AI agent to do shopping for them.

25:46 – 25:48

And really, your job is effective employment.

25:48 – 25:54

Critical literacy are for people who are not the programmers, but they’re the ones who are making the project happen.

25:54 – 25:55

So they’re maybe the product owners.

25:55 – 25:58

Their job is to have informed critique.

25:58 – 26:03

And then third, rhetorical literacy, which is necessary for people who are the developers.

26:03 – 26:09

What’s really fascinating about this is that it gets less and less technical the closer and closer to the code that you are.

26:09 – 26:13

So when you are the programmer, the developer, your job is to have reflexive practice.

26:13 – 26:15

So we’ll talk a little bit about what that means.

26:15 – 26:17

So let’s say you are just the recipient.

26:17 – 26:19

You’re the person just interacting with the chatbot.

26:19 – 26:21

You want to know what your bank balance is.

26:21 – 26:22

You want to transfer some money.

26:22 – 26:25

And there’s this AI agent that’s going to do that for you.

26:25 – 26:29

Well, how should you think about being literate in that world, and if that’s your role?

26:29 – 26:37

The most important thing is that part on the bottom is learning to engage with the technology versus as a result of the technology.

26:37 – 26:38

That’s really, really important.

26:39 – 26:47

you want your users and your customers to be critically literate in understanding the output and how they can take action or whether or not the technology should be taking

26:47 – 26:48

action for them.

26:48 – 26:54

So rather than blindly accepting, they should actually be able to engage with the technology.

26:54 – 27:03

And then that’s actually a much better experience as somebody using anything if you’re engaging with it and you’re actually seeing and understanding the benefits of working with

27:03 – 27:03

it.

27:05 – 27:14

And for the second level of people, the product managers, the people managers, the product owners, the most important thing is to be able to understand when things are going wrong,

27:14 – 27:15

when things are not working well.

27:15 – 27:18

This is even more important, the era of generative AI.

27:18 – 27:20

Remember, it’s confidently incorrect.

27:20 – 27:21

It will say output.

27:21 – 27:28

I think last week, the big debacle was how many lawyers are now using ChatGPT to help write their legal briefs for them.

27:28 – 27:30

And guess what ChatGPT does?

27:30 – 27:31

It hallucinates.

27:31 – 27:35

So it’s absolutely fabricating cases that don’t exist.

27:35 – 27:40

So again, this is about being smart enough to interrogate the technology.

27:40 – 27:42

Look at the sources it’s giving you.

27:42 – 27:43

Double check it.

27:43 – 27:46

uh And being able to challenge the system.

27:47 – 27:52

And then when you’re the developer, the builder, the programmer, the person sticking their hands in the code.

27:52 – 27:54

questions are the most rhetorical.

27:54 – 27:57

Where does the power lie in the system?

27:57 – 28:05

If there was a human in the loop and Stanislav Petrov had to decide whether or not to set off a nuclear weapon, would he be able to actually say no?

28:05 – 28:08

Or would he just have to just follow what the AI said?

28:08 – 28:16

So how do you design a system so that there is the reality of the human in the loop, that people can actually be engaged?

28:16 – 28:22

My way of addressing it, and I know Anna gave the introduction, I run an organization called Humane Intelligence.

28:22 – 28:34

I’m a statistician by background, a social scientist, and what interests me about Gen AI is the fact that every single person gets to have this technology in their hands.

28:34 – 28:42

And during my time at Accenture, really, I thought about this question of what does it mean to have human in the loop, and what does it mean to have AI assurance?

28:42 – 28:50

AI assurance is just the ability to say that this product or this technology is going to act, is going to do what I think, what I have promised it will do.

28:50 – 28:51

And with Gen.

28:51 – 28:52

AI, that’s actually really, really difficult.

28:52 – 28:55

So how do you do this within an organization?

28:55 – 28:58

This is a paper that we wrote in 2020 about it.

28:58 – 29:09

Really it’s about having clear and accountable processes, diffuse accountability, and also being transparent and actually more responsible with what the system can and can’t do.

29:10 – 29:16

During my time at Twitter, I actually did something quite radical, which is I opened up our code for public scrutiny.

29:16 – 29:19

Now, the concept of a bug bounty here in cybersecurity is nothing new.

29:19 – 29:23

But an algorithmic bug bounty was something that never happened before.

29:23 – 29:27

So we opened up our code to the entire world, and we said, find problems, and we’ll give you money.

29:27 – 29:31

I actually ended up making a nonprofit, that’s Human Intelligence, to actually do that work.

29:31 – 29:40

So we’ve worked with a wide range of organizations, governments, uh civil society groups, et cetera, to do these kind of bias bounty programs.

29:41 – 29:44

And the new thing I’m known for is red teaming.

29:44 – 29:48

So again, if you’re from cybersecurity, red teaming is not a new concept.

29:48 – 29:58

But the understanding the use of red teaming, not just to understand malicious actors, but unintended consequences when your system is behaving poorly, but it was not an intentional

29:58 – 30:00

action of your user, is actually quite difficult to do.

30:00 – 30:04

And the way I tackled it is what I called structured public feedback.

30:04 – 30:10

The thing I’m the most known for is this event where we had over 3,000 people at DEF CON, which is the largest cybersecurity

30:10 – 30:18

conference in the world actually hack and analyze uh eight different AI models actually worked with the model owners and the companies.

30:18 – 30:20

And this is against the AI Bill of Rights.

30:20 – 30:26

So whether or not these models would reveal personal data, whether it would have hallucinations.

30:26 – 30:30

And we worked with the US government and the Biden administration to perform this work.

30:30 – 30:32

And we continue to do that today.

30:33 – 30:40

And what’s fascinating is that the average person, while you can have this technology in your hand, we’re more and more removed from

30:40 – 30:43

the ability to give feedback on whether or not it’s working.

30:43 – 30:46

So one of the projects we did was work with the National Institutes of Standards and Technology.

30:46 – 30:50

And we opened up this concept of red teaming to anybody in America.

30:50 – 30:52

And this actually happened last year.

30:52 – 31:00

So anybody in America could actually red team AI models and give feedback to NIST in order to create better standards of how this technology should work.

31:01 – 31:05

And really what I work for, had a TED Talk last year on this concept of a right to repair.

31:05 – 31:12

And red teaming is a starting point, this idea of just building that knowledge base and that reflex for people to say, hey, something’s wrong here.

31:12 – 31:13

I want to raise my hand.

31:13 – 31:15

I want to be able to point that out.

31:15 – 31:20

That’s actually not something we’ve thought about building into the technology that we have.

31:20 – 31:27

So how can you utilize AI to solve problems uh positively and productively?

31:28 – 31:30

And so what’s next?

31:30 – 31:31

So AI is a productivity tool, right?

31:31 – 31:32

So it’s not magic.

31:32 – 31:33

It does not have intent.

31:33 – 31:36

It’s not dragging your kid’s teacher out of the classroom.

31:36 – 31:41

So when you’re thinking about building and deploying AI, how can you think about it being user first, right?

31:41 – 31:49

So how do you think about the pain points of what AI is good at doing and also importantly, what people are good at doing and merging the two?

31:49 – 31:53

And then finally, how can you utilize this system to solve problems?

31:53 – 31:57

And again, when we think about this idea of the human in the loop, it’s not just a phrase.

31:57 – 31:58

It’s not just a term.

31:58 – 32:00

It’s what people are asking for today.

32:01 – 32:05

you know, we’re at an all-time low with trust as it pertains to AI systems.

32:05 – 32:10

During my time in the Biden administration as science envoy, I had the privilege of going all around the world.

32:10 – 32:20

And you’d be amazed at everybody from, you know, tribal elders in Fiji to the most sophisticated young person working with Gen.

32:20 – 32:20

AI.

32:20 – 32:22

They all have the same questions.

32:22 – 32:28

There’s questions about data privacy, security, whether or not algorithms are being, whether or not they’re being manipulated by algorithms.

32:28 – 32:31

whether or it’s being used maliciously against them.

32:31 – 32:33

Everybody has the same questions.

32:33 – 32:39

So if you want to build trust in the systems that you are making, think user first, not technology first.

32:39 – 32:41

Thank you so much for your time.

32:46 – 32:47

Thank you Dr.

32:47 – 32:49

Chaldry, absolutely fantastic.

32:49 – 32:50

We appreciate you being here today.

32:50 – 32:51

Thank you so much.