5G, AI, and Quantum Computing- Mr. Curt Levey

Parliamentary Intelligence-Security Forum
3 followers
1 Video View·Mar 6, 2026
23rd Parliamentary Intelligence-Security Forum
Washington, D.C🇺🇸- December 6-7, 2022
-Mr. Curt Levey
President, The Committee for Justice
Can everyone hear me all right?
Great.
I am the president of the Committee for Justice.
We're a small nonprofit founded 20 years ago by someone
who you actually might be familiar with,
C. Boyden Gray.
Most Americans know him, actually,
as White House counsel under the first President Bush.
But I think in this room, he might be better known
as ambassador to the EU.
And I've been doing this now for almost 20 years.
But before that, I was in the AI field.
And that's what I'm going to talk about.
And one reason that I'm particularly interested in AI
is it sort of combines my two careers, the lawyer career
that I've had for the last 20-plus years,
and also my previous work before law school,
building AI models for a small high-tech startup
in San Diego.
And when I left the AI world and went to law school,
I didn't think I'd probably ever work on AI again.
But in recent years, there's been a lot of interest.
And it's been a pleasant surprise to get back to it.
One of the hottest topics in AI and national security
these days is competition with China
for leadership in the field.
And I decided not to talk about that,
just because at the end of two full days
here, I figured there probably wouldn't be much I could add
about competing with China on technology.
If I had more time, I would just talk generally about AI,
the types of applications that might
apply to national security, the risks, what regulation
might help in that regard.
But given limited time, I decided
to focus on one particular application of AI that's
very relevant to national security,
and that's facial recognition, probably the most important
type of biometric that we have right now.
And one of the risks that certainly
applies to facial recognition that I'll also focus on
is AI and bias, or fairness, as some people call it.
It's a problem I've studied.
And I, again, have sort of a joint perspective,
my former AI scientist perspective.
And also, as a lawyer, I've worked a lot
on discrimination issues.
So I bring both of those to the AI and bias problem.
And the good news is that I think
the problem of AI and bias in general
is probably less severe than some of the hype
would make us think.
But before I get to bias, let me tell you
a little bit about how facial recognition works.
There's two main types.
One, the easier type, I won't say easy, but easier type,
is just comparing two faces, verification.
Is this who the person says they are?
And then the harder type is identification,
where you're basically matching that face against what
could be a huge database of faces, potentially millions
or, at some point in the future, billions.
And you start off with a digital image or a video frame.
And you don't just feed the raw pixels into the system
any more than the human brain works that way.
The human brain extracts features,
and that's what you need to do in a facial recognition system.
So those features, we don't 100% know
what they are in the human brain.
But we can control what they are in a facial recognition system.
Just a few examples would be the distance between your eyes,
or the distance between your eyes and your nose,
or the distance between your nose and the tip of your chin,
things like that.
And it's very important that this feature extraction
be invariant to various transformations.
You wouldn't want the features to be different
because you turned your head, or because it was dark in the room,
or because there was uneven shading.
Humans are very good at that type of invariant feature
extraction, and more and more machines are good at that too.
At some point, though, you're going
to have to compare features between the target and, again,
one other face or a bunch of faces
and compute some type of matching score.
For verification, you would typically apply a threshold.
You are who you say you are if it matches
above a certain threshold.
Again, there's always a trade-off between false positives
and false negatives, but you can set the threshold
where you want.
And then for ID, IDing people against a large database,
typically it'll return the closest matches.
Now, the most common type of technology used,
because it's the best at learning,
is deep learning, or what we used
to call, back when I was in the industry, neural networks,
basically the same thing.
There are other more statistical methods,
but they're not as flexible when it comes to learning.
And the accuracy of your model is
going to depend on, obviously, the technology, how
many examples you have to train on, how good your data is,
and then, ultimately, if you're doing the identification
function, the size of your database.
Obviously, if everyone in the world was in your database,
that might be computationally expensive,
but theoretically, then, you could always
match the person in front of you.
Now, humans are very good at the recognition thing.
If I met you before, I can probably, well,
I'm not that good at it, but most people, I think,
are good at recognizing faces they've seen before.
And so we're lucky if machines can do as good a job.
Where machines do have an advantage is, again,
we can only remember so many faces,
1,000, something around that, whereas there's
almost no limit on how much a facial recognition
system can recognize.
Humans also aren't very good at dealing with unfamiliar faces.
Again, it makes sense if you just
think of it from an evolutionary standpoint
throughout much of our evolutionary history,
there weren't that many people.
There weren't a lot of unfamiliar faces.
So there are many uses for facial recognition.
I'll just quickly focus on a few that involve security,
starting with national security.
Most obvious example is ports of entry.
The US Custom and Border Protection
has tested and deployed AI facial recognition
systems at land, sea, airports.
I think it's about three dozen airports where
it's been used on Americans leaving the country.
Virtually every airport it's used
on people entering the country.
So it's used to search the ID function
to search for known bad guys.
It's used for confirming identities, the verification
function.
And that fits in with creating an entry exit record
for everyone who comes here.
Domestic security wise, it's used more and more
in criminal investigations in much the same way
that fingerprints or DNA evidence are used.
And then it's used in personal security.
I mean, when you look at your phone, it'll unlock it.
Or you log on to your bank through showing it
your face, things such as that.
It's really become very widespread,
despite being a fairly new technology.
You hear a lot of talk about countries banning it.
But the only two that actually ban it now,
and even those have exceptions, are Belgium and Luxembourg.
Was it Morocco?
Yeah, I think Morocco had a moratorium until about two
years ago, which is now.
Lifted.
There was an interesting study that
rated countries around the world based
on how extensive their use of facial recognition is.
So they looked at not just use by the government,
but use in schools, workplace, transportation systems.
A few interesting tidbits from that.
More than, they looked at roughly 100 countries
for which they had data.
More than half of them used it in airports.
About half of them, half of that number
used it in train, subway systems.
About 40% use it in workplaces.
And then they ranked countries based on those factors.
And I said I would avoid the China issue,
but lo and behold, the country that
got the worst score, or the best score,
if you think facial recognition should be used more widely,
was China, which had a score of 5 out of 40,
where 0 is, again, the most extensive possible use
of facial recognition.
Somewhat surprising to me, at least, Russia was second.
And I say that just because I don't
think of them as necessarily having the best technology.
The UAE was third, 10 out of 40.
Japan, India, and Chile were next, at 12 out of 40.
Australia and Brazil, 13 out of 40.
And then if you skip a bit, US was
sort of in the middle of the pack, around 18 out of 40.
And then I promised I would talk about at least
one of the risks, which is bias.
Again, my view is that the threat
is less than some might contend, and that AI will actually
lead to less discrimination, because the alternative isn't
angels making decisions.
It's humans.
And people come to the task with plenty
of irrational prejudices, whereas an AI model starts out
as a blank slate without any biases.
Everything the model knows is from the examples
that are presented to it.
Typically, thousands, tens of thousands of examples
where the correct output is known, so you're teaching it.
So for example, it could be binary.
Are these two faces the same or not?
Or it could be, given the identification function,
given these two faces, here's how well they match.
But the point is it's only going to learn correlations that
rationally help it to reduce error.
It's not going to learn irrational correlations.
Its only goal is to minimize its error rate.
And there is no room for irrational prejudices,
short of a rogue programmer.
I mean, if you had a programmer who
was trying to make it irrational,
then that would be a different story.
They're also much less attuned to race and gender
than human beings are.
The AI system isn't going to know upfront
what your race or gender or any other identity group is
unless you make it specifically an input variable.
And even if, say, it knew your race or gender,
it wouldn't care unless it could use it to lower the error rate.
To use the facial recognition example again,
an irrational, or at least somewhat irrational, bias
might be all Chinese people look alike.
The system isn't going to learn that unless, to some extent,
in your database the Chinese faces do look alike.
So again, I think AI is as close as we're
going to come in the real world to colorblind and blind
to various other things.
And you have to realize when we throw around the term AI
bias or discrimination that the purpose of decision making
is to discriminate, discriminate between the right answer
and the wrong answer.
For example, let's say you have a system that's
trying to predict based on various data who's
going to be a criminal risk once they enter the country.
Well, that model will probably learn
to discriminate against people who have criminal histories.
But is that a bad thing?
Some would say yes.
Some will say that's not fair.
It'll have a disparate impact on people
whose criminal histories might be in their past.
Maybe they'll never commit another crime.
And why should they have to carry around that burden?
But the correlation between past and future criminal record
is not 100%, but it's real.
And if you sacrifice accuracy by forbidding the system
from considering past criminal record,
then there's a trade-off.
Real people will suffer when you admit bad guys who you might
not otherwise have admitted.
Another example, let's say you specifically
decide to make gender an input.
The system may very well improve its accuracy
by learning that men pose a higher risk of criminal
behavior, at least violent criminal behavior.
Is that OK?
There's no really easy answer.
I mean, it's not clear what the fair outcome is.
I mean, again, maybe in an ideal fantasy system,
you wouldn't want it to know that there's any difference
between men and women.
But if your goal, again, is to protect people,
maybe you do want it to know.
So again, I don't know that there's a right answer there.
One last example, let me give you,
you've probably all heard that facial recognition systems
developed in the West often have more trouble identifying
Asian faces, black faces, for example.
Not surprisingly, there are studies
have shown that facial recognition systems developed
in East Asia have trouble on Caucasian faces.
Some would say that's discrimination
against Caucasians.
Some would say that's what you want.
In a domain operating in East Asia,
you want it to be particularly good at East Asian faces.
Still others would say, you know what?
If it comes to at least government surveillance,
I'm kind of happy to be a race that it doesn't do a good job on.
I'm just as happy that the government doesn't recognize me.
So again, it's hard to say what the right answer is.
I'll just quickly, I see I only have a minute,
but I'll wrap up by getting to regulation.
Probably when you hear calling for regulation of AI,
the two top things are bias and privacy.
I don't think, at least in the US,
that we really do need special laws
to protect against AI bias.
I think it can be addressed with the same anti-discrimination
laws that govern human decision makers and the tools they use.
For example, in the US, if you use an employment test
that has a disparate impact, you have to justify
why you needed to use it.
And that shouldn't matter whether the employment test
uses AI or not.
So for now, AI is just another tool.
And I don't think we need new laws.
Another example, if TSA is using a screening tool that
has irrational discrimination in it, then again,
do you care whether it's a dumb tool or an AI tool?
Finally, I'll just say there may come
a day in which we truly have autonomous AI where they really
are functioning not as a tool, but as the equivalent
of humans.
And then maybe we'll need special laws for AI and bias.
But I don't know that that's going to happen in our lifetime.
So I'll wrap up there.
