Latest / The Road to Accountable AI / Katie Fowler (Thomson Reuters Foundation): How 3,000 Companies Approach AI Governance

Transcript

Katie Fowler (Thomson Reuters Foundation): How 3,000 Companies Approach AI Governance · The Road to Accountable AI

0:09
Werbach: Hi, I'm Kevin Warbeck, professor of Legal Studies and
0:11
Business Ethics at the Wharton School of the University of Pennsylvania.
0:15
For decades, I've studied emerging technologies from broadband to blockchain.
0:19
Today, AI is promising to transform our world, but AI needs accountability
0:26
mechanisms to ensure it's developed and deployed in responsible,
0:29
safe, and trustworthy ways.
0:32
On this podcast, I speak
0:33
with the experts leading the charge for accountable ai.
0:40
One of the biggest challenges in responsible AI is the
0:43
lack of sufficient data.
0:45
What are companies actually doing?
0:48
As an aside, this is something that the Wharton Accountable AI lab is working on.
0:51
More on that later, but there are a number of important initiatives happening.
0:56
My guest is Katie Fowler, director of Responsible Business at the
0:59
Thomson Reuters Foundation.
1:01
It has just released a report based on the largest available dataset.
1:06
Of corporate AI adoption, 3000 companies across 13 sectors all around the world.
1:13
We talk about what they found about the gap between conceptualization and
1:19
operationalization of AI governance, other striking findings about what companies
1:24
are and aren't doing, and patterns about different adoption practices, uh, around
1:29
the world and in different contexts.
1:32
And how we can move towards an environment where data helps inform both what
1:37
organizations do internally as well as incentive structures from investors
1:42
and other stakeholders to promote effective AI governance practices.
1:49
Katie, so nice to have you with me on the road to accountable ai.
1:52
Katie: :Thank you so much for having me.
1:55
Werbach: Thomson Reuters is a, a major content and technology firm,
1:59
which people may be familiar with from things like the Reuters News Service
2:02
and Westlaw and, and other services.
2:05
You work for the foundation.
2:07
So talk a little bit about the work that your organization does
2:11
in general first, and how does it relate to the larger for-profit firm?
2:16
Katie: Absolutely.
2:17
Yeah, thanks again so much for having us.
2:19
It's such a great opportunity to talk about the Thomson Reuters Foundation
2:23
which is an independent charity.
2:25
Um, it's the c Corporate Foundation of Thomson Reuters, as you mentioned.
2:29
Um, and our mission is really to strengthen free,
2:32
fair, and informed society.
2:34
We have three key big work streams.
2:37
One is around strengthening the resilience of independent media.
2:41
Of course, we carry the Reuters brand and we take our commitment to the resilience
2:46
of independent media really seriously.
2:48
Second is our access to law work stream, where we actually run
2:52
the world's largest corporate pro bono platform called trust Law.
2:56
So we've been connecting, um, civil society organizations and, and.
3:00
For purpose organizations across the world with pro bono
3:04
legal support for many years.
3:06
Really, really proud of the work that we do in that area.
3:09
And then finally the responsible business work, which is of course led by myself.
3:13
I've got an amazing team of, um, data professionals and private sector
3:17
engagement, um, professionals too.
3:20
And our work is really to.
3:23
Drive responsible business practice through corporate transparency and
3:28
really to advocate for good jobs worldwide and increasingly to advocate
3:33
for the responsible adoption of ai as it pertains to the private sector.
3:37
So a really broad mission in relation to the work that Thomson Reuters does.
3:44
I think the best way to put it and to sort of explain the difference in our
3:47
approaches is that Thomson Reuters is very much powering the world's most informed
3:53
professionals through the services that you mentioned, of course and others.
3:57
So it provides fast, reliable, independent, verified news,
4:01
data and tech that is informing decisions at the highest level.
4:04
And contra contrary to that and complimentary to that and the
4:09
foundation's work is empowering often those same professionals
4:13
actually to use their skills, expertise, and know how to strengthen
4:16
free, fair and informed societies.
4:18
Because every day we are working with those same professionals, lawyers.
4:22
In-house councils, journalists, corporate leaders, and investors,
4:25
to really bolster the resilience of independent media and to strengthen
4:30
access to law, as I mentioned, through the world's largest pro bono platform.
4:35
Mm-hmm.
4:35
And also then to promote responsible business.
4:37
Yes.
4:38
But we're also connecting with civil society organizations and governments.
4:41
So we take a, a really whole ecosystem approach to the
4:44
positive impact that we can have.
4:47
Werbach: We're going to talk mainly about the new initiative
4:49
you have on AI company data.
4:51
How did that, uh, emerge out of the overall responsible
4:55
business work that you do?
4:57
Katie: So it was quite a natural evolution for me, not least, because ignoring
5:02
what is probably the most seismic, um, issue that it's happening in the
5:06
private sector would be slightly odd.
5:08
Um, and of course it's having a huge impact on business behavior currently.
5:12
Um, but we already run, um, the Workforce Disclosure Initiative or
5:16
WDI, which is a world leading plaque.
5:19
Form, which collects survey data from companies on how they treat
5:22
their workforce throughout their direct operations and supply chain.
5:26
And of course, because we already have that foundation and infrastructure, it
5:30
was quite a natural progression to be looking at how the impact of corporate AI
5:35
adoption and in fact data from our 2020.
5:38
Four WGI initiative actually found that only about a quarter of the
5:43
participating companies, which is around 250 of the world's largest
5:47
companies, could actually report on how AI was affecting their workforce.
5:52
Mm-hmm.
5:52
Uh,
5:52
really indicated to us this low awareness of AI's business and workforce impact.
5:58
So we therefore thought that it was a a really huge gap really.
6:03
I wanted to provide a comprehensive framework through which companies could
6:07
really start to to examine and report against this, this growing issue.
6:13
And then of course, secondly, I'd be remiss to not mention that we have a
6:18
longstanding partnership with UNESCO and the opportunity since they.
6:22
Publish their recommendations on the ethics of AI to be able to apply
6:27
that to a corporate setting mm-hmm.
6:29
Was just, open for us.
6:32
It's such a huge comprehensive framework.
6:35
As you probably know, it's been globally ratified by 193
6:39
countries, so very intentionally.
6:41
It is the most globally accepted governance framework,
6:45
so we really couldn't.
6:46
Um, miss the opportunity to be able to translate that into something
6:50
that the corporate sector could actually use as a framework itself.
6:56
Werbach: The report that you recently published, and we'll put a
6:58
link in the show notes covers 3000 companies, lots and lots of data.
7:04
And my understanding is that, uh, you did that partly using AI to analyze
7:08
information that companies put out.
7:10
So, so talk about, um, what that experience was like and you know, how
7:14
you can have confidence in the accuracy of what you're producing in the report.
7:18
Katie: Yeah, absolutely.
7:19
I really, I'm really pleased that you raised this actually, because naturally
7:24
we couldn't go into this huge data collection exercise without acknowledging
7:29
that we also had a responsibility to, to use ai appropriately and effectively
7:34
in order to gather this information.
7:36
So, I first do want to say how much consideration went into that approach?
7:40
Um, you're right in saying we collected public publicly available disclosures
7:44
and we used an LLM with very specific prompts to identify, um, the exact
7:50
relevant information in line with the framework that we'd already
7:52
established, which included then linking outputs back to original source text.
7:58
And I guess my T key takeaways from, um, using that approach was that the LM
8:04
LLMs proved very effective at scale for passing large volumes of public dis.
8:09
Corporate disclosures.
8:10
Ultimately a really big underlying, um, objective for us was to build a
8:15
large data set because of the global discourse that surrounds corporate AI
8:20
adoption and the lack of that sort of volume of data to support a lot of the
8:24
hypotheses that, that we're talking about.
8:27
Um, and this was particularly the case for.
8:30
Identifying explicit evidence-based statements, um, consistently across a
8:35
diverse range of formats, such as reports and policies, and also then web content.
8:40
But the prompt design was absolutely critical.
8:43
So outputs were really highly sensitive to how instructions were framed.
8:47
Mm-hmm.
8:47
So making clarity, specificity, and traceability, and linking extracted
8:53
information back to exact source.
8:56
Text was essential mm-hmm.
8:57
In order to avoid misinterpretation, which is actually really, really
9:01
easy when it comes to, um, a lot of sort of ESG and governance data.
9:06
So humans, human herbicide just remained totally indispensable throughout the
9:10
whole process, and there was a lot of it.
9:12
So corporate disclosures obviously vary very, very widely.
9:16
Mm-hmm.
9:17
In the language and structure.
9:19
A lot of organizations are voluntarily disclosing information, not
9:22
necessarily using explicit frameworks.
9:25
So we had to be really considered mm-hmm.
9:27
In the reach of our of our scraping.
9:30
And, you know, look at that language and structure and modeling risk uh, or risk
9:36
mitigation in terms of the inconsistency and incompleteness of the extraction.
9:40
Mm-hmm.
9:41
Um, and unless that was obviously a guided by human oversight, we would've been yeah.
9:46
Not doing very well in terms of the ness of the, of the dataset.
9:50
Overall we are really, really pleased with the outcome and
9:53
the approach that we deployed.
9:55
Lots of lessons obviously we'll go in to this as as we're sort of
9:59
brandishing this last first annual report, we're intending very much
10:03
to, to release one next year as well.
10:05
And I'm sure that we'll be looking at exacting the process again.
10:09
And of course, looking at the developments that AI get off,
10:11
Werbach: yes, much will happen in the next year.
10:13
No question.
10:14
Yeah.
10:14
Well, it's an interesting, it's an interesting point that you make
10:16
because this is something that, that many companies are struggling with.
10:19
The, the LLMs are incredibly powerful, but we know they have accuracy issues.
10:24
It's sensitive to the prompt hallucinations and so forth.
10:27
So certainly human oversight is essential in this context.
10:32
How do you have confidence, uh, given something at this scale that
10:36
you can't necessarily have humans look at every single data point.
10:39
How do you actually have confidence that the report meets the objectives?
10:43
Are there ways that you are able to measure and assess
10:47
whether you're analyzing these disclosures in the right way?
10:51
Katie: O other than sampling and testing mm-hmm.
10:53
Um, throughout the whole process, we are, we are looking at the coherency mm-hmm.
10:58
Of the data.
10:59
What are the trends telling us how many anomalies exist within the trends?
11:03
There are lots of different tests models that we can deploy, and I
11:06
think as our team and capability also grow in that area, as we become
11:10
a little bit more familiar with the type of information that we're.
11:14
That accuracy will likely build over time.
11:17
Mm-hmm.
11:17
It's very much, I mean, there's a, there's a full kind of section in our
11:21
report that outlines our methodology and our approach and, you know, we
11:24
really want to be on a learning journey in developing this sort of data that's
11:29
useful and usable for the corporate sector as much as they are, able to put
11:33
it out into the, um, into the ether.
11:36
And ultimately I might also.
11:38
Add that our goal is that a large proportion, a much larger proportion
11:42
of the dataset is put together from direct disclosures in the future.
11:47
Mm-hmm.
11:48
So we can deploy LLMs to scrape data, we can pre-populate surveys for a
11:53
really large proportion of companies who want to engage with with us.
11:57
Mm-hmm.
11:57
But then allowing them to do, to also deploy the human oversight,
12:01
not just also is like such a key part of this quality control.
12:06
Werbach: There's a tremendous amount of fascinating data in this report.
12:09
I want to get into it, but let me first just ask you a general question.
12:12
You, you mentioned using the UNESCO AI ethics framework as a starting point.
12:18
Do you have any general insights that you gain from looking at what companies
12:23
are doing versus that framework?
12:26
Katie: I think the headlines for us were just this enormous governance gap.
12:32
So we saw a couple of our key findings really highlighted and emphasized how
12:37
much, how much onus and imperative companies do feel to be talking about
12:42
their AI strategies, but how little that is translating into operate.
12:47
Operationalization and being implemented within their organizations.
12:51
Our overall feeling is that there's a piece here where we're unable to unpick
12:56
the extent to which it's fear and the extent to which actually companies
13:00
are being slightly modest about the extent to which they're actually
13:04
deploying ai because there is such a lack of public information about it.
13:09
They are totally unaware of what their peers and competitors are doing.
13:12
They're un unaware of what.
13:14
Good looks like at an industry level or at a regional level.
13:18
And we really wanted to try and dispel some of those myths
13:20
using that that framework.
13:24
The recommendations on the ethics of AI that UNESCO publish cover
13:27
such a wide breadth of issues?
13:29
Yes.
13:30
We tried to boil it down into governance and oversight, human
13:34
capital and data safety and security.
13:36
And I think even then you can go a lot further, um, in terms of looking
13:41
at how different, particularly.
13:43
Particularly different industries need to look at AI governance more specifically.
13:47
But certainly we, our overall feeling is that using the UNESCO
13:52
recommendations did allow us to appeal to organizations, companies in all regions.
13:59
It gave us a really great solid benchmark and foundation of what,
14:04
theoretically good can look like.
14:06
Mm-hmm.
14:06
And then really helping to build on that.
14:08
And, you know, as our data.
14:11
Increases.
14:12
But the other key kind of standout indicator that we've highlighted
14:17
is just how lacking workforce preparedness is across all sectors.
14:23
It's unanimously being ignored.
14:25
And, um, my feeling sort of.
14:27
Totally is that there's, again, there's a fear and a, and a
14:31
trepidation amongst corporate leaders about the extent to which engage
14:36
their, to engage their workforce.
14:37
Are their workforce actually going to be redundant or are
14:41
they actually a key stakeholder in the success of AI deployment?
14:45
And it's this sort of tension that is being fueled, I think a lot by.
14:50
By sort of media coverage around mm-hmm.
14:52
What AI is going to do for the workforce in the future.
14:55
And then sort of confusion about the lack of evidence that we have in at an
15:00
individual company level about exactly the efficiencies that it is offering.
15:05
Werbach: Right.
15:06
With regard to workforce preparedness, what, what external data might
15:11
we see that is evidence of that?
15:15
Because it, it may well be that companies are doing a lot internally,
15:18
but they're not necessarily putting a formal disclosure out
15:20
about, say, their training policy.
15:22
So what, what would you look for to see indications of
15:24
workforce preparedness for ai?
15:26
Katie: That's really, that's a really good point.
15:28
And ultimately our data is very much corporate disclosure data.
15:34
So there are limitations on the extent to which we are able
15:37
to to make huge assertions.
15:40
But what we can see say is that there are, there is enough evidence sort of out.
15:46
In the broader kind of research landscape to be able to corroborate a lot of the
15:51
trends that we were seeing around human capital and human capital risks whereby.
15:57
Employees themselves are not undergoing training, for example, where, there's an,
16:03
an assertion that training is available or com or employees themselves are not
16:08
aware of specific company policy around AI adoption and, and even ethical,
16:14
considerate considerations of ai.
16:17
And so those sorts of indicators are, are those that we can look for.
16:22
And then just to also highlight that our Workforce disclosure initiative
16:27
As I, as I mentioned at the start, we have seen a consistent lack of reporting
16:32
around the impact of technology on work.
16:34
Course and so just being able to highlight to companies the importance
16:38
of being able to monitor and manage redeployment, upskilling, you know, all
16:43
of the changes that they're hoping to make to their sort of their workforce
16:47
over time is really, really critical.
16:49
Werbach: Mm-hmm.
16:51
I wanna get back to the, the point that you made earlier about the
16:53
governance gap, and I, I believe the numbers in your report were 44%.
16:58
So nearly half of companies have an AI strategy and only 13% have a,
17:04
uh, a formal AI governance framework that, that you were able to see.
17:07
So I'm, I'm curious your take on that, uh, because, you know, one question
17:10
is you know, there's obviously a difference in those two numbers,
17:13
but in some ways both of those are pretty big numbers given that.
17:16
We're really only three years in since chat, GPT, when, when companies broadly
17:21
were having to wrestle with AI across the organization in the way they are now.
17:26
Uh, and in some ways that's, that's a lot.
17:28
Both of those numbers seem like a lot in that short period of time.
17:30
But anyway, I'm interested in your, your thoughts about that.
17:33
Katie: Yeah, I think that's a really a really useful pointer.
17:37
And, and ultimately we are not here to emphasize any negative trends.
17:42
We're, we are really wanting to spell out what we're seeing, um, in the data.
17:46
So ultimately, yeah, you are absolutely right.
17:49
It's absolutely, it's an actual, it's a brilliant number that 44% of organizations
17:54
say that they have an AI strategy.
17:56
I think that.
17:57
Because
17:57
they may be a very good AI strategy.
17:59
Well, indeed, indeed.
18:01
And of course, because only 13% of companies are publicly committing to
18:05
a formal AI governance framework, we are then left to fill in the gaps.
18:08
So what exactly does that tell us?
18:10
And people like me are gonna tell you that.
18:13
Ultimately I, my concern is that there is a, a mismatch
18:18
and the gap is suggesting that.
18:19
Probably most companies, AI strategies are focusing on accelerating
18:24
adoption and extracting value rather than on establishing robust
18:29
governance and ethical commitments, um, or protecting their workers.
18:33
And of course, without these guardrails, AI oversight may not
18:37
actually be working in practice.
18:39
And I also feel that it's important to look at this from a, an, the
18:42
perspective of of other stakeholders.
18:44
So a really large stakeholder group that back, our AI company data
18:49
initiative are our investor signatory group, and transparency is often.
18:54
Used as a proxy for risk management.
18:57
And where there's a transparency gap, they are then left to ask questions.
19:01
So are there reputational, regulatory or other risks that might actually
19:06
undermine all of that accountability?
19:08
So I think that, that, for me, it's, uh, often it's not what you are
19:12
telling, it's what you are not telling.
19:16
Leaves you exposed somehow.
19:17
So really what we're trying to encourage is to identify these gaps
19:22
and help companies appreciate that.
19:25
That without further explanation, other
19:28
External parties are then left wondering and trying to fill
19:31
that gap with an explanation.
19:33
Um, so fibs maybe be adopting ai.
19:36
Practices very responsibly in their day-to-day operations, but without
19:40
transparency to their consumers or to their investors, they're
19:43
actually risking, undermining trust.
19:46
So that's my kinda key takeaway from that gap.
19:49
Werbach: Absolutely let, let me ask about a, a few of the specific findings there.
19:52
I, I think they're interesting and I'm curious how you reacted to them.
19:56
Only about one eighth of the companies.
19:58
Uh, we, we talked, um, a minute ago about human oversight in terms of your
20:01
own work in building this report and, and how important it's to have humans
20:06
in the loop in various ways engaged with ai, but only, only about an eighth
20:09
of the companies that you surveyed.
20:11
Had policies explicitly about human oversight.
20:14
Did that surprise you?
20:15
Katie: Yeah.
20:16
Um, did it surprise me?
20:19
I, I think that it's not a great it's not a great data point.
20:23
And given how much we know about the importance of Hu human oversight,
20:28
it, it is quite alarming to look at on, on a, an aggregate level.
20:32
Mm-hmm.
20:33
And I think that you do have to question whether it is a disclosure problem or
20:36
whether companies do genuinely lack that.
20:39
Mm-hmm.
20:39
Human oversight, of course.
20:41
Um, and is human oversight something that, you know, you would've
20:45
publicly communicated about it?
20:48
Are companies privy to how much an external stakeholder wants to know
20:54
about the level of human oversight?
20:56
I think those are all really key questions to ask.
20:58
Stage that we're at in collecting this data.
21:01
I feel confident that a lot of companies do have internal practices.
21:06
They're just simply not communicating externally.
21:08
Mm-hmm.
21:09
But the reports, broader findings do actually point to more of a
21:12
structural problem, which goes a little bit beyond poor disclosure.
21:16
And I think it is important not to just dismiss it as a disclosure
21:19
problem, because across the dataset where we are seeing consistent
21:23
patterns emerging mm-hmm.
21:24
Around organizations being able to describe governance at a concept.
21:28
To level, but not really being able to demonstrate how it that functions
21:33
on a day-to-day operational level.
21:35
So a lot of companies felt much better at announcing governance intentions,
21:39
perhaps, than building operational infrastructure to deliver them.
21:43
So human oversight is probably squarely falling in that sort
21:48
of infrastructure category.
21:51
Werbach: No.
21:51
Interesting.
21:52
And, and I, I see that a lot when I talk to organizations that it, it's
21:55
fairly easy to, to say, okay, here's our principle, but then they sit
21:59
around looking at each other about what it means to actually implement it.
22:03
And, and to your point before they say what's everyone else doing?
22:05
And, and they don't even know.
22:06
So this, this kind of information you're showing I think will be
22:09
really valuable to help organizations start to, to benchmark in that way.
22:12
And,
22:12
Katie: and also I think another point.
22:15
For me that sort of resoundingly important to express is how,
22:19
where our experts sit in this.
22:21
So often our experts in, you know, the people who see the importance of human
22:26
oversight are not sitting on the boards, they're not sitting in leadership.
22:30
They're sitting deep.
22:31
Within teams of people who've been hired to operationalize ai.
22:35
So they're not necessarily also the people who are responsible for, broad rollout
22:41
of robust governance infrastructure.
22:43
And I think that's a really important point to raise too, and something that
22:47
we've been called on is to kind of demystify amongst leadership, you know.
22:53
Directors, what, what does responsible AI actually look like and mean?
22:57
Because I've been in my career for 30 years and I'm not
23:00
exactly sure how to deploy that.
23:02
So, um, I think that's also another important point to raise and not
23:06
to undermine actually the really amazing work that is happening deep
23:10
within organizations and teams.
23:12
And that's sort of the ethics that exist at that individual level.
23:17
Werbach: Yeah.
23:17
One of the, the antecedents for that responsible AI work is, uh, work on
23:21
data protection and privacy that that many organizations, especially global
23:25
organizations, have had to do with the adoption of GDPR in, in Europe and
23:30
with, broader global, uh, implications.
23:32
And so I was, I was interesting that, you know, your report seemed to.
23:36
Suggest that there were, many more companies that talked about doing
23:40
assessments for privacy with AI than things like human rights and ethics,
23:44
but even with regard to data, um, I believe the number was only about a
23:48
quarter of the companies did analysis of the provenance of the training
23:52
data that they're using for ai.
23:54
Which you would think would be the kind of thing that, you
23:55
know, if they're more attuned to.
23:58
Privacy and data protection, they would start there.
24:00
So there's a couple different pieces there.
24:02
I I, I wonder if you could elaborate more on some of those.
24:05
Katie: Yeah, I mean, just with your, with respect to your point about that policy
24:10
in GDPR, I think that is a really, really interesting perspective and something that
24:15
you could apply to a lot of the findings of the report in lots of different ways.
24:19
Werbach: Mm-hmm.
24:19
Katie: Um, the way that GDPR continues to influence AI governance through
24:23
AI systems that process personal data and must comply with GDPR principles,
24:28
including lawlessness purpose limitation.
24:31
Data minimization, et cetera.
24:33
So that's a, that is an interesting facet to this.
24:36
And I think that the def facto way that this global standard in the way that GDPR.
24:43
Um, sort of applied to data protection at large.
24:45
We are also potentially seeing a little bit of that with regard to the governance
24:49
framework of the EU AI Act as well.
24:52
Mm-hmm.
24:53
Um, so the, because the AI Act is re remains the most advanced.
24:57
An atte advanced attempt to set clear legally binding standards
25:02
for safe, transparent, and accountable AI across all sectors.
25:06
The broader regulatory landscape covering digital technologies and
25:10
shaping AI operations indirectly is sort of a lot more disparate.
25:16
And so the way the fascinating takeaway from the report for me was that over
25:20
half of the companies actually cite.
25:24
Governance site, the eu a act, AI act as their sort of governance framework,
25:30
even those that exist outside of Europe.
25:32
And I think that's a really an interesting reflection for a
25:37
lot of policy makers worldwide.
25:39
In fact, you know, the, the absence of a regulatory foundation actually
25:44
leaves a lot of corporates a little bit paralyzed to know, you know what the.
25:50
The where the goalposts are and.
25:53
You know, the environment in which they're operating.
25:55
So contrary to this sort of rhetoric that regulation stunts innovation and
26:02
growth, actually we might be seeing something that's quite contrary to that,
26:06
where anti-regulation actually stunts you know, the pace of innovation and growth
26:11
that a lot of organizations, um, sort of need that safety net around them in
26:15
order to be able to operate and yeah.
26:17
Werbach: Safely and without constraints.
26:19
So I think that's, that's quite an interesting.
26:21
Perspective.
26:22
Yeah, I, I hear that from a lot of companies too.
26:24
Obviously they don't love regulation and they have particular objectives
26:28
to certain legal rules, but really what they want is clarity.
26:32
They, they keep saying to me, just tell us what the rules are, and we struggle
26:36
with an environment where we don't know how what we do is going to be evaluated.
26:41
Katie: Yeah, absolutely.
26:42
It's it's a really interesting, um, time and we are obviously mapping
26:47
global, regulatory, um, jurisdictions across our framework as well.
26:52
So throughout the survey framework, we're actually able to tell companies which
26:58
data point that they're responding to.
27:00
Mm-hmm.
27:01
What that's relevant for in terms of global regulation.
27:03
We've mapped 23.
27:05
Cybersecurity and AI regulatory frameworks, for example,
27:08
including obviously the EU AI Act.
27:11
But we are seeing, you know, a lot more pace in emerging economies to
27:14
be able to provide that kind of.
27:17
Framework and environment for, for companies to thrive in.
27:20
So, you know, the next couple of years gonna be really fascinating mm-hmm.
27:23
To see what happens.
27:24
Werbach: Yes.
27:25
Did you find any other interesting differences among regions in terms of
27:30
how, uh, people were responding, who got companies from all over the world?
27:34
Katie: Yes, we did.
27:36
We found some quite stark differences, which probably won't
27:39
surprise the readers an awful lot.
27:42
European and UK companies disclose a lot more on AI safeguards for workers than
27:47
those in Asia, Latin America, and Africa.
27:49
Of course, it's fairly reflective on on the.
27:54
Uh, labor standards that exist across all those different regions anyway,
27:58
but also potentially it's reflective of the the categories of workforce,
28:03
which AI is most affecting at this time.
28:06
So it's slightly to suggest, obviously, a more mature regulatory and governance
28:10
expectations around workforce, of course, in the UK and Europe.
28:14
Mm-hmm.
28:15
Um, but there's also greater institutionalization of formal
28:18
reporting channels in that region.
28:19
And of course, the data finds that.
28:22
Although it's still really low, the presence of AI related compliant
28:27
mechanisms are higher in the UK and Europe compared to the rest of the world.
28:31
So that was quite an interesting, trend.
28:33
Werbach: One other finding that that isn't surprising is that
28:36
larger companies were much more likely to have formal AI governance
28:39
policies than, than smaller ones.
28:41
Did.
28:41
Did you get any sense about.
28:44
How small medium enterprises can engage in AI governance in an effective way.
28:50
Katie: Yeah.
28:51
I think this question's really interesting and actually one we
28:53
want to do a lot more work on.
28:55
We really struggled to engage with the the SME kind of sectors and naturally, because
29:03
we did a lot of public data scraping, of course that information is more
29:07
available from, um, from public markets.
29:10
So there's.
29:11
There's a really obvious answer to that to some extent, which
29:14
is reflected in our data.
29:16
However, we do hear anecdotally that, um, small and medium enterprises are
29:20
a lot more driven in order to be able to adopt responsible AI practices,
29:26
and of course, a lot earlier.
29:28
Phase organizations are being established around, AI operationalizing
29:33
their business in the first place.
29:34
So there is a sort of imperative.
29:37
Also what's really interesting is looking at, um, our different
29:40
segments of investors who are you know, really driving for.
29:44
Responsible decision making and practice around AI adoption.
29:48
Um, organizations that are invested in really large established organizations.
29:54
Pension funds and other asset owners are really, really motivated to understand
29:58
what this looks like long-term.
30:00
Whereas, you know, when we're trying to look at.
30:03
Um, the venture capitalist market, private equity, we are seeing a lot
30:08
more more of a splattering of interest.
30:10
Mm-hmm.
30:11
Which is very much steeped in individual ethics rather than, corporate framework.
30:17
'cause of course, fast turnaround investments the longer term risks
30:21
of AI for people's society and the environment aren't so interesting
30:25
for them 'cause it's not really going to impact their, um, their kind of.
30:30
Grace and turnover and, and yields in the short term.
30:33
Mm-hmm.
30:33
So I think that whole piece is really interesting.
30:37
The startup kind of network is, uh, one that we've had a lot of interest from.
30:42
Um, and may well look at the possibility of adapting the AI company data initiative
30:48
survey even further so that it's more accessible for smaller businesses.
30:53
But I have to say that we we made a really considered effort to mm-hmm.
30:57
Ensure that the survey was appropriate to be responded to by any company
31:02
of any size and at any stage of their corporate adoption of ai.
31:07
And so we're, now looking at potentially making some updates to the survey
31:12
so that it's as relevant as it can be for the next cycle of disclosure.
31:16
And we're concerned about too many changes 'cause it may sort of damage the integrity
31:21
of the, ethical framework that we want.
31:24
Companies to be able to deploy.
31:26
So whilst companies might not be able to actually answer all of the
31:29
questions, we still want them to be there as prompts for business leaders
31:33
to be able to take back to their organizations and have a think about
31:36
and maybe respond to the following year.
31:38
Werbach: You mentioned a couple times the, the role of investors and, and
31:42
obviously with regard to things like worker protection or broadly ESG.
31:46
There there's been, you know, tremendous effort around the world to have metrics
31:50
and have ways that investors can assess companies and, and make their decisions.
31:55
Um, the kind of work you're doing obviously can contribute
31:59
to doing that kind of assessment for responsible use of ai.
32:04
How do you see, uh, US going forward in terms of investors being able to
32:08
play a role in incentivizing companies?
32:11
Katie: So, um, I think our long-term goal would be to be able to deploy our
32:17
data into more materiality, um, research.
32:21
Ultimately there needs to be a link to financial materiality.
32:25
We would love to think that we'd be able to contribute to a
32:27
responsible AI index, for example.
32:30
Mm-hmm.
32:30
Where we are actually matching the.
32:33
The maturity and um, and growth of financial metrics
32:38
to responsible AI practice.
32:39
Obviously it's gonna take some years to be able to build a data set that is,
32:44
comparable and established enough to be able to develop that sort of framework.
32:49
But that's going to be a really key lever for investors ultimately, because
32:53
they'll be able to really tangibly request that of their portfolio companies.
32:57
But in the meantime, as I mentioned, a lot of our investors are saying to us.
33:01
That they are, you know, in the absence of any kind of governance framework
33:06
that they can really hang their hat on, they're looking at transparency
33:09
as a key proxy for risk management.
33:11
And so we are able to offer a framework to companies to, um.
33:17
To enhance their transparency around their corporate adoption.
33:20
And what we have seen through our Workforce disclosure initiative is
33:24
that investor buy-in and investor engagement is such a key part mm-hmm.
33:30
Of this theory of change.
33:32
Ultimately, they are the people who are gonna be able to impress upon
33:36
corporations to change their behavior.
33:38
It's not little little.
33:41
You know, civil society organizations and their advocacy,
33:43
you know, the money will talk.
33:45
Mm-hmm.
33:45
And where investors start to feel that they have enough evidence to be able
33:49
to withhold investments on the grams of, responsible practice around, um,
33:54
ai, where there's a materiality link.
33:56
I think that's gonna be the absolute, you know, winning strategy in order to,
34:01
to emphasize the importance of this.
34:04
And of course, as we.
34:05
You know, move forward, we will start to see real impacts and, you know, it will
34:11
become much more, um, of a reputational risk for organizations mm-hmm.
34:15
Rather than an anecdotal one.
34:17
Werbach: Absolutely.
34:18
I, I agree that this is the direction that the things are
34:20
inevitably going to have to go.
34:21
Uh, it's, it's a very useful report and, uh, very much look forward to seeing.
34:26
What you do going forward with this line of work.
34:28
Uh, we need to wrap up but Katie, it's really been a pleasure speaking with you.
34:31
Katie: Thank you so much.
34:32
Appreciate it.
34:35
Werbach: This has been the Road to accountable ai.
34:37
If you like what you're hearing, please give us a good review and
34:40
check out my substack for more insights on AI accountability.
34:44
Thank you for listening.
34:48
If you want to go deeper on AI governance, trust and responsibility with me.
34:53
Another distinguished faculty of the world's top business school.
34:56
Sign up for the next cohort of Wharton's Strategies for Accountable
34:59
AI Online Executive education program, featuring live interaction
35:04
with faculty expert interviews and custom designed asynchronous content.
35:08
Join fellow business leaders to learn valuable skills you can
35:11
put to work in your organization.
35:14
Visit exec ed.warden.upen.edu/acai for full details.
35:21
I hope to see you there.