Latest / The Road to Accountable AI / Cameron Kerry: From Gridlock to Governance?

Transcript

Cameron Kerry: From Gridlock to Governance? · The Road to Accountable AI

• 0:01
  Cam, always great to talk to you.
• 0:03
  Well, Kevin, it's great to see you and nice to be here.
• 0:07
  Thanks for the invitation.
• 0:09
  So let me start by asking how in general would you describe the current moment we're in for, well, let's start with AI policy.
• 0:17
  um Yeah, well, that encompasses an awful lot of uncertainty, I guess.
• 0:24
  But let's focus on AI.
• 0:28
  I like a lot of other things in some turmoil, but I think with a little bit more consensus, to some extent internationally, think certainly so domestically.
• 0:43
  It's one of the few...
• 0:46
  relatively bipartisan issues.
• 0:51
  And, you know, there are certainly notes of difference, but I think on the substance, a lot in common.
• 0:58
  uh If you look at what the Trump administration has done, there's a lot of points in common with the Biden administration and the Biden administration built on things that the
• 1:12
  first Trump administration did.
• 1:15
  Where would you say those points of agreement are?
• 1:17
  I would tend to agree with you, but obviously rhetorically there's a lot of difference and you hear a lot of people in the current administration saying what they're doing is
• 1:24
  totally different.
• 1:25
  Yeah, I think that's exactly right, Kevin.
• 1:32
  think there's a lot of rhetoric that's performative, like talking about woke AI, the approach of innovation in unleashing things.
• 1:55
  I think there certainly are elements in that.
• 1:57
  think uh with the involvement of some of the Silicon Valley voices that are particularly anti-regulation and libertarian, there are elements of that in the policy.
• 2:16
  But there's also, I think, a lot of substance uh to like.
• 2:24
  uh
• 2:25
  Some of the rhetoric as I said is performative some is is is unhelpful, but you know I think you have You know a lot of substantive elements when it comes to to the measures to
• 2:41
  to promote uh adoption uh in the federal government and otherwise open source and uh Improving AI literacy uh
• 2:54
  It talks about worker first AI.
• 2:58
  mean, that to me sounds an awful lot like the Biden administration in some respects.
• 3:06
  talks about an evaluation ecosystem, and that's built in the Office of Management and Budget guidance that came out earlier in the year and all of that, notwithstanding JD
• 3:23
  Vance's.
• 3:24
  sneer about hand wringing uh about AI safety among Europeans.
• 3:35
  You know, there are certainly contradictory notes as well.
• 3:40
  mean, yeah, there's talks about the executive order talks about the Trump executive order, the AI action plan uh talks about combating synthetic media.
• 3:54
  Same time, know, voices saying, you know, no, we're basically we're, you know, in favor of, you know, deep thanks.
• 4:04
  uh
• 4:05
  I'm...
• 4:09
  Similarly, I think to my mind, the biggest contradiction is on science policy.
• 4:19
  It talks about investing in science, theoretical, computational, and experimental uh science.
• 4:31
  Hard to have that when you are cutting out so much of the funding for
• 4:39
  Excuse me.
• 4:43
  It's hard to have that when you are cutting out so much of the investment base in scientific research and so much of the academic base as well.
• 5:01
  Do think we need legislation on AI?
• 5:04
  then if so, do think we'll get it?
• 5:08
  Well, Ted Cruz was just yesterday saying, uh you know, we ought have some kind of legislation on AI, mostly, I think, to preempt uh state laws.
• 5:25
  And, you know, look, I was at the Department of Commerce.
• 5:30
  Its overarching function is to promote
• 5:36
  the foreign and interstate commerce of the United States.
• 5:39
  So I do believe in having, particularly in areas uh like information technology and communications, which are intrinsically interstate and international, we should have
• 6:01
  consistent national laws, whether that's privacy.
• 6:06
  or in AI.
• 6:10
  But I don't think we should completely preempt the field, which may be the intention here.
• 6:21
  uh But I think there are ways to get at this to do some baselines that I think can be very helpful to.
• 6:36
  thoughtful uh deployment, development of AI uh without impairing innovation and without uh heavy regulatory hands.
• 6:52
  What about globally?
• 6:54
  Where do you see the key fault lines right now?
• 6:58
  Well, it's interesting, Kevin.
• 7:00
  We're talking, well, I'm actually in New York, uh where a couple, three years ago in the wake of chat GPT, there was an enormous push towards global uh governance on AI.
• 7:21
  I think subsequently we've backed
• 7:26
  a little bit away from that, but there are certainly strains of that out there.
• 7:32
  My view is you need to deal with AI in ways that can be a lot more agile.
• 7:47
  This is a global phenomenon.
• 7:56
  research and development.
• 7:57
  A lot of the research teams are across institutions, across borders.
• 8:05
  Certainly, the data, the technology move quickly.
• 8:12
  we need some alignment.
• 8:17
  But we need a system that uh operates uh
• 8:23
  a lot like the information technology networks that it runs on.
• 8:29
  That is networked, decentralized, with a lot of different nodes and points of contact.
• 8:43
  And I think we need that, Kevin, because there's so much that is unknown when it comes to AI.
• 8:53
  It's going to take experimentation.
• 8:57
  It's going to take a lot of careful and thoughtful inquiry.
• 9:07
  And we are also seeing that AI is really not one thing, notwithstanding the major, the large language models and a lot of focus on that area.
• 9:23
  uh It's a lot of domains, a lot of functions.
• 9:27
  Those are multiplying.
• 9:30
  And even the major uh large language model developers like OpenAI, like Anthropic, tailoring the models to very specific purposes.
• 9:46
  And they have different functions.
• 9:47
  We just saw this week that people are seeing that Claude
• 9:53
  And open AI have different evolving uses and strengths.
• 10:01
  Claude gets used a lot more for coding.
• 10:04
  Open AI much more for interactions and search and generative language.
• 10:17
  What does that mean for policy and regulation though, that there is this differentiation?
• 10:22
  Yeah.
• 10:23
  No, I think that is the key question.
• 10:26
  uh Look, I think it needs to be extraordinarily uh agile uh and also extraordinarily alert.
• 10:37
  mean, look, there are going to be problems and dysfunction uh with AI.
• 10:44
  We know enough about it to know that's clear.
• 10:48
  What exactly some of those are going to be, we don't know.
• 10:51
  So I think at both the organizational level as well as the governmental level, it needs to be done with a lot of intentionality and not driven by FOMO, but carefully applying best
• 11:18
  practices, benchmarks.
• 11:21
  measurements and monitoring.
• 11:24
  And what exactly does agile regulation mean in practice?
• 11:29
  ah So my thought is, look, I think in practice, um organizations that are looking at this, privacy professionals and lots of uh trade associations.
• 11:57
  other organizations, as well as leading developers, think have developed fairly well-established best practices and benchmarks and measurement tools.
• 12:12
  So to have some baseline environments uh to assess uh AI uh models.
• 12:26
  deployments going in and before uh release, before deployment uh is part of it.
• 12:38
  But then uh to monitor after uh deployment and to continue uh to do that.
• 12:51
  I say that in carpentry, the maximum
• 12:58
  is measure twice, cut once.
• 13:02
  When it comes to AI development and employment, it's measure, measure, and measure again.
• 13:12
  And I think to have some basic legal obligations to do that is given the state of the industry, given the state of best practices, is not.
• 13:25
  I have me left.
• 13:27
  So do you think that the practice and the science around measurement for AI is good enough at this point to drive those kinds of decisions?
• 13:38
  Look, I think there's a lot of work to be done to develop a lot more measures.
• 13:45
  you know, so exactly what they are, we don't know, know, but that can evolve, I think.
• 13:58
  Here, standards development is going to be hugely important.
• 14:05
  There are, I think.
• 14:06
  process standards in place.
• 14:08
  ISO IEC uh has a risk management standard that sets some basic uh policy or practices in place.
• 14:23
  I was in a discussion at the International Telecommunications Union this summer with a number of
• 14:34
  people and organizations involved in standards.
• 14:39
  The consensus at the table was that process standards are easy, but there's so much in terms of the range of applications, um the domains, the functions that are different that
• 14:58
  product standards, we got a long way to go.
• 15:03
  You've done a great deal of work for many years on privacy policy and famously or infamously, we still don't have comprehensive federal privacy legislation in the United
• 15:13
  States.
• 15:14
  Should that give us pause in terms of thinking about the capacity to rise to the AI moment?
• 15:22
  Look, I think that is a great question.
• 15:25
  And I think we've got some of the same dynamics, I think, in terms of complexity, in terms of some of the industry interests that have ended up blocking things in the privacy area
• 15:46
  over the years.
• 15:51
  But I think in both of those cases, there's certainly for this polarized era, uh a lot of bipartisanship.
• 16:07
  uh There was a congressional task force uh in the last Congress.
• 16:13
  I think it's good work on outlining some of the basic issues for privacy legislation.
• 16:21
  which I think is much along the lines of what I said about what could be in legislation.
• 16:30
  So I'm hopeful if we go forward and adopt some national legislation, it can be along those lines.
• 16:40
  I think that will do a tremendous amount to help inform where we go in the future with regulation and develop some standards that
• 16:51
  then set a baseline for industry expectations.
• 16:58
  Yeah, that was going to be my next question is whether you think that the all these AI developments might reshape or change the debate around privacy.
• 17:09
  I certainly hope so.
• 17:11
  mean, look, data is key here.
• 17:14
  And it's clear that AI...
• 17:23
  exponentially increases the privacy risks.
• 17:28
  the privacy debate, mean, 15 years ago, it was moving from focusing on what identifies personally identifiable information, oh information like uh identifiers, social security
• 17:48
  numbers, other unique
• 17:52
  identifying information.
• 17:56
  But at the time, the realization that given big data and analytics, it was possible to identify people uniquely based on data that didn't fit that personally identifiable data
• 18:18
  category.
• 18:19
  uh
• 18:21
  um And so the privacy debate has moved to, in general, in state laws and federal drafts, to talk about information that is linked, personally identifiable information, or linkable
• 18:40
  to a specific individual or a device.
• 18:45
  Now, with AI, uh
• 18:51
  What is linkable is almost anything.
• 18:56
  And that really puts a premium on, I think, particularly developing privacy enhancing technologies, ways to uh share information, to manage uh information, for example, in the
• 19:17
  corpus of a training data.
• 19:19
  em So that you get the benefits oh of the training, but not expose individuals.
• 19:31
  clearly, if you're doing, scraping everything on the internet, you're sweeping a lot more individuals into that training data.
• 19:49
  absolutely.
• 19:49
  No, and that's what I was going to ask you about that, you you talked about the, you know, you know, political issues and, you know, corporate incentives and so forth that made it
• 20:01
  so difficult on privacy.
• 20:04
  Now it seems like we've ramped that up even more.
• 20:07
  We're now, as you say, it's talking about all data and we're talking about it.
• 20:10
  It's so central to these frontier models that they have access to all of this data, you know.
• 20:17
  does the pressure come from to appropriately address these privacy concerns?
• 20:22
  I'm...
• 20:26
  You know, that's been a challenge because people are talking about the privacy paradox, the notion that people express concerns about privacy but don't act on it.
• 20:40
  I think part of that is it's just so damn hard.
• 20:50
  it's far more than any of us could do, no matter.
• 20:55
  sophisticated you are about technology or managing your privacy.
• 20:59
  um I don't know any privacy experts who can do all of the things that in theory we ought to do to protect our privacy.
• 21:14
  I certainly don't.
• 21:18
  And we rely on a certain level of trust.
• 21:25
  legal basis on how data is used to really give a foundation to that trust.
• 21:34
  think that will do, I think, in the long run, a tremendous amount to enable the free flow and use of data.
• 21:48
  I firmly believe that information
• 21:54
  is a public good, that getting uh access to uh data, personal data, can be enormously beneficial, both to individuals and to society and to the store of human knowledge.
• 22:13
  But we need ways to ensure that that's done appropriately.
• 22:20
  Yeah, I'm curious on your thoughts on this.
• 22:22
  One of the things that I've noticed in talking to lots of enterprises about AI governance is in many cases to the extent that they feel like they have some foundation, it starts
• 22:35
  with privacy and it starts with frankly GDPR implementation in many cases, even if it's an American company, you know, the ones that say, we actually had to ask these questions
• 22:45
  about what are we collecting and what's the provenance.
• 22:48
  ah So it would seem like there's potentially an incentive there for companies to say, we don't like regulation, but we know we've got to understand our data just for AI to deploy.
• 22:58
  So do you see that happening too?
• 23:02
  Yeah, I have seen that.
• 23:05
  certainly, after I left the Obama administration in 2013, about a year or so later, I affiliated with a law firm and was involved in counseling boards on cybersecurity,
• 23:24
  advising clients on privacy, including
• 23:28
  implementation of the GDPR, implementation of initial California privacy law.
• 23:38
  And what I would say to boards is...
• 23:47
  This is all about governing your data.
• 23:54
  And for both privacy and particularly for cyber security where they clearly have interest in protecting information, if you don't know what data you're collecting, how it's being
• 24:09
  used, you can't protect it.
• 24:12
  And more broadly,
• 24:16
  that in this digital era, that is also what you need to have a digital strategy.
• 24:25
  And that applies exponentially to AI.
• 24:37
  So much of AI in the enterprise context is taking the tools, taking the models.
• 24:46
  using that to analyze uh the data that exists within the enterprise, whether it's uh marketing communications, the body of corporate emails.
• 25:08
  It's really uh giving a lot of life to the business intelligence component.
• 25:18
  And really, at long last, enabling, I think, enterprises to make sense of a lot of that, to take what's been unstructured data and make sense of it using, essentially, language.
• 25:46
  But knowing where what that data is uh and how it flows is at the baseline of having that kind of strategy.
• 26:02
  What do you think companies most need in order to be effective in that effort?
• 26:10
  Either uh in terms of just incentives or knowledge or where public policy could actually help.
• 26:19
  I'm...
• 26:22
  You know, I think some of the baselines that we've talked about around assessment can help.
• 26:36
  I don't know that they can get to the specificity that any particular company would have to oppose.
• 26:52
  apply that in ways that help them.
• 26:58
  I think you need to look at the tools that are out there.
• 27:04
  now the model developers are putting out assessments of the usage of their models and how that plays out.
• 27:23
  And a lot of the companies and platforms uh that are doing this are putting out uh information on what some of their tools, their key performance indicators, their
• 27:40
  benchmarks are.
• 27:42
  And I think there's that, and create sort of uh a body of resources that people can look at and say, all right, what do we need?
• 27:51
  business environment, what can we adopt?
• 27:57
  And then just going forward, what do you see over the next, say, one or two years, whether AI generally, privacy policy, uh globally, what should people be most focused on in terms
• 28:14
  of understanding how all this is going to?
• 28:17
  I'm...
• 28:20
  Boy, Kevin, I'm...
• 28:24
  Yeah, really.
• 28:25
  You know, I've been the optimist about getting privacy legislation done because I saw a pathway and saw that there was the possibility that people were actually working on
• 28:46
  something that could get there by...
• 28:50
  by that pathway.
• 28:52
  I've uh occasionally been called, as a result, a ridiculous optimist.
• 29:02
  It's harder for me to sustain optimism given uh the experiences that we've had.
• 29:11
  came very close to privacy legislation in 2002.
• 29:17
  uh
• 29:18
  I mean, in 2022.
• 29:22
  But we had basically two members of Congress who blocked it for less because of the industry, just for political reasons of their own.
• 29:42
  that was Nancy Pelosi, who would not bring in the bill that was passed out of committee.
• 29:50
  the heavy vote in favor, wouldn't bring it to the floor.
• 29:57
  And in the Senate, the chair of the Commerce Committee, Maria Cantwell, didn't join the three committee leaders in sponsoring the bill.
• 30:12
  think had either of them done differently, that could very well have passed.
• 30:20
  I'm so.
• 30:21
  oh
• 30:24
  And now there's not the same sort of momentum that we had then.
• 30:31
  um So I think the answers are really going to have to come bubbling up from underneath, uh from the process that I've talked about, the assessment, the deployment, the experience,
• 30:51
  the monitoring.
• 30:54
  That's what is going to teach organizations and ultimately I think what's going to teach societies and governments.
• 31:05
  That is, I think, the sort of networked decision making that I've talked about.
• 31:13
  Great, Cam, thanks so much.
• 31:14
  Great to talk to you.
• 31:16
  well thank you Kevin.
• 31:17
  I enjoyed talking to you.