Latest / The Road to Accountable AI / Dragos Tudorache: Next Steps for the World's First Comprehensive AI Law

Transcript

Dragos Tudorache: Next Steps for the World's First Comprehensive AI Law · The Road to Accountable AI

0:01
Welcome and thank you for joining me on the Road to Accountable AI podcast.
0:07
Well, thank you very much for having me.
0:10
My pleasure.
0:11
The AI sector is developing incredibly rapidly.
0:15
Why is this the right time to adopt a comprehensive AI regulation?
0:22
Well, as you said it yourself, AI is advancing very rapidly and I remain
0:27
convinced that this is a technology that is going to drive the future with enormous
0:32
possibilities, enormous benefits, enormous potential, but at the same time with some
0:37
risks that by now we have already started seeing materializing.
0:42
And I'm not only talking about the big threats to humanity and stuff that we've.
0:46
seen on the television, but mostly day-to-day risks that basically involve
0:52
you and me.
0:54
And that was the basic thinking when, at the European Union level, we started
0:59
saying, OK, so if these risks are real, and at the same time if this potential is
1:04
real, can we just remain as we are and as we have been for quite some time in the
1:11
digital sector?
1:12
with absolutely no rules with the companies governing themselves and
1:16
deciding what is right and what is wrong?
1:19
Or do we want to actually put some guardrails in place to signpost in a way
1:23
the evolution of this technology?
1:25
And the answer to this question was we have to take on the responsibility of
1:30
creating those standards ourselves, of saying and trying to develop those
1:36
standards, to develop those guardrails in a way that brings trust ultimately.
1:41
It's all about trust.
1:42
trust in the technology and the way it interacts with our rights and with the
1:46
interests of society.
1:49
In Europe though, it's not the case that there's no regulation of digital activity.
1:53
You have GDPR, you have the Digital Services Act, Digital Markets Act, various
1:57
other regulations.
1:58
So what exactly is the gap that the AI Act is designed to address?
2:04
Well, it is designed to address exactly the applications of AI, which were not
2:07
necessarily covered, many of them not at all covered by the GDPR.
2:11
GDPR was only focusing on the use of personal data on privacy and what that
2:17
means for the digital realm.
2:19
And also all the other legislation that basically was adopted in this mandate, the
2:25
Digital Markets Act, the Digital Services Act, the data, they all are part of the
2:28
same vision back in 2019.
2:31
when basically this European Commission with this European Parliament and Council
2:36
set out for these five years to actually bring all of these tools in place
2:40
recognizing that this was time from our point of view to bring in some rules in
2:45
the digital market and they cover different aspects.
2:49
The market sector covers the markets, the anti-trust issues, the accumulation of
2:56
market dominance with some players.
2:59
The Data Services Act
3:00
is basically the responsibility of platform for the content that they host,
3:06
whereas the AI Act actually looks at the AI itself, but it does not regulate the
3:10
technology of artificial intelligence.
3:12
It looks at the uses of AI and then tries to identify what are those uses, those
3:17
applications that actually brings risks and how to mitigate those risks.
3:21
And as I said earlier, it puts clear rules in place only for those that actually
3:27
bring those risks.
3:28
And then the rest of the AI, and it's important to note that, the rest of the AI
3:32
out there, which is estimated to be in fact somewhere between 80 to 90 percent,
3:37
is going to remain unregulated even after the AI Act comes into force.
3:42
Isn't there a danger though, given how flexible the technology is that leaving so
3:47
much out of regulation then creates an opportunity for services to evolve or be
3:52
used in ways that still raise some of those risks?
3:56
It's a very valid question.
3:57
Is this regulation, and it's a question that many ask, is this regulation actually
4:03
future proof?
4:04
Is it able to basically remain relevant, to remain adapt to the evolution of
4:09
technology, to whatever, change GPT 6.7 or to any other application of AI?
4:15
The answer is, well, of course, the reality will tell whether we're able to
4:20
make it adapt or not.
4:22
But we have addressed that in two ways.
4:24
First of all,
4:25
We've tried to identify those or to express the obligations in a way that are
4:32
technology neutral.
4:34
So when we formulated the rules around transparency, we used wording that would
4:39
ensure that obligation will sound and be just as relevant five years or ten years
4:44
from now, no matter how sophisticated or changed the technology will be.
4:49
At the same time, we also left open
4:52
the areas that are identified right now as being, for example, high risk,
4:57
understanding that the market will evolve, technology will evolve.
5:01
There will be new areas that might prove in the future, one, two, three, five years
5:05
from now, to actually be raising risks and therefore the regulation can be easily
5:10
adapted.
5:10
We have a tool in the EU toolbox which is called Commitology, which allows in
5:16
something between three to six months.
5:19
to adapt a piece of legislation as long as the mandate for adaptation exists in the
5:24
text of the law.
5:24
And we have deliberately included these mandates for adaptation in the text for
5:29
all of those parts where we thought technology will continue to grow, will
5:32
continue to evolve, and therefore those parts will need to be constantly in sync
5:38
with the reality of that evolution.
5:41
It's a balancing act though, right?
5:42
You want something to be clear and defined, but also flexible.
5:48
And as you talked about, this legislation was a particular political compromise, a
5:53
different set of interests.
5:55
So what gives you optimism going forward that this is the right approach as the
5:59
industry and the technology evolves?
6:02
Well, as you said it yourself, it was a very delicate balancing act and it's the
6:06
result of literally hundreds and hundreds, if not more, of hours of work, both
6:12
separately, each of us as co-legislators in our quarters and then afterwards
6:17
through, again, hours and hours of negotiations.
6:20
While that may sound very complex and heavy, it actually does bring something
6:25
good out of those hours of negotiations because it allows it to distill what is
6:29
most important to put in a piece of legislation.
6:31
And I think that we managed to find that balance.
6:34
On one hand, what we try to do is to find the balance between the benefits and the
6:39
risks.
6:40
Because as I said at the beginning, we don't want to put this label of risk on
6:45
AI.
6:46
AI, again, is a fantastic technology.
6:48
And I'm convinced.
6:49
will help us as humanity to solve cancer, it will help us to go to Mars, it will
6:55
help us to do amazing things in almost every walk of life, every walk of human
6:59
activity.
7:00
So on one hand we wanted to make sure that we also put the right enablers in place
7:04
through this legislation and create the space for innovation to continue to grow
7:08
and actually to invite for even further innovation in the market, but at the same
7:13
time to give the confidence that we are addressing properly those risks.
7:16
And we're not ignoring them, we're not sending the signal to society that, okay,
7:20
well, we're leaving you out there to face this technology alone without any
7:26
safeguards.
7:27
And again, one balancing act was to find that equilibrium between the protection
7:32
and the innovation.
7:33
And I think, again, we've managed to solve that equation.
7:37
And at the same time, that balance that you refer to, which is how much do we put
7:41
in the text right now?
7:42
And again, how do we make sure that we...
7:45
are looking sufficiently ahead as to how the technology will evolve to make sure
7:51
that we remain relevant with this piece of text.
7:55
And we've also introduced something that is not known or common in the EU rulebook,
8:00
which is what I like to call flexible enforcement, and also flexible and very
8:06
inclusive governance, which means that the decision makers of the future...
8:12
those that will be called upon based on the mandates that we have introduced right
8:15
now in the text, as I said earlier, that will be called upon to adapt the
8:18
legislation.
8:19
They will not be doing it in absence of being constantly in sync with the reality,
8:25
because we've introduced an advisory group, a scientific board that will be
8:30
constantly feeding what happens in the industry to the decision makers in the
8:36
future AI office and the repeal commission, so that those adaptations will
8:39
actually be mirroring
8:41
as accurately as possible what is happening at a given time in the
8:47
technology itself.
8:50
So one area in particular that relates to this issue of future proofing is
8:54
generative AI.
8:54
The original text of the act didn't really talk about generative AI and then chat TPT
9:00
came along and obviously it was important.
9:02
And so there was something of a scramble to put those provisions in about general
9:06
purpose AI.
9:08
Is it a concern that there was not enough consideration about how to address these
9:12
issues around foundation models and so forth?
9:16
Well, it may sound arrogant if I say it, but we haven't really scrambled doing
9:20
that.
9:21
If one, for historical reasons, if not anything else, if one would want to look
9:25
at my original amendments from back in 2021, I had introduced there the concept
9:31
of AI authors.
9:33
This is how, back then, in my mind, I was thinking of that.
9:37
Of course, ChargPD was not yet on the market, but through the work that we had
9:41
done, and this is again something that I don't know how much is known, but we had
9:44
really...
9:45
taken our time to prepare for the moment when we started this legislation.
9:48
We had a special committee on artificial intelligence in parliament, which I had
9:52
the honor of chairing, where for almost two years, we have looked into every
9:57
aspect of AI.
9:58
We have listened to all possible stakeholders from every walk of life, from
10:02
any perspective looking at the same issue, that allowed us to educate ourselves, to
10:10
grow our understanding of the technology and what it was that we were embarking to
10:14
regulate.
10:15
And during those discussions, I heard for the first time, coming from some of the
10:21
recognized thinkers, let's say, on AI, I heard the concept of general purpose AI,
10:26
as it was called back then.
10:27
And this technology that was bubbling over the surface, which then later popped up to
10:33
be the large language models, the LLMs of today.
10:36
And I was hearing these.
10:41
thinkers is also saying, listen, there is something to this technology that needs to
10:45
be well understood and also cannot leave regulators indifferent, even if they're in
10:51
themselves don't have a purpose that it can easily pigeonhole as the rest of the
10:57
regulation does with other applications.
11:00
So we've realized rather quickly that we need to think of a different kind of a
11:04
regime, of a different approach to this type of AI, let's say, this species of AI,
11:09
the gene AI.
11:11
And already prior to Chad GPD coming to the market, in the team of the parliament,
11:16
we were already discussing how to actually formulate this regime of rights and
11:23
obligations and what we would actually be putting into that regime.
11:27
And then what Chad GPD did, and then the rest of the debate rolling very quickly on
11:34
generative AI, in a way, it gave us the confidence that we were on the right
11:37
track.
11:38
And that we were not just ourselves being paranoid about it, there actually was
11:42
something in this technology that required some rulemaking of its own.
11:47
And that's what we've done, and this is what then came to be the special chapter
11:52
on what we now call foundation models in the AI.
11:57
And at the very end of that process, though, there was significant pushback
12:01
from France and Germany in particular, that segment was overly regulatory.
12:05
So what did it take to eventually resolve that dispute and get enough consensus for
12:11
the legislation?
12:13
There was indeed pushback, not only from France and Germany, but there was also a
12:17
considerable pushback from the industry itself, which basically argued at the
12:21
beginning in two directions.
12:23
One direction saying, we don't need regulations at all for us because what do
12:26
you know?
12:27
We are just language models.
12:29
We are processing text at the end of the day.
12:31
So what can be harmful in that?
12:33
Then they abandoned that line of argument because the public debate and there were
12:38
already evidences of actually what you could do with these models.
12:43
And then there was another line of argument from them initially that was
12:46
saying, listen, the responsibility is not with us, it's further down the line in the
12:49
value chain with those that would be taking our models and doing all sorts of
12:53
things with them.
12:53
So we have no responsibility, the responsibility should be with those that
12:57
are taking our models and using them for a given harmful purpose.
13:02
But as we were discussing and as we were finding counter arguments to those, then
13:07
the discussion pretty much focused, and this is where also, as you mentioned,
13:12
France and Germany came in the debate, in terms of finding a more balanced way,
13:17
speaking of balance, a more balanced way to actually deal with the obligations
13:22
towards these models, understanding how fluid the technology still was, how
13:26
nascent, and how much we still didn't know, not in terms of what we wanted these
13:36
models to do or to prove, but how to actually measure that.
13:41
how to test, how to evaluate, how to actually make sure that you can enforce
13:45
those obligations.
13:46
So this is where we accepted the argument that came from France and Germany, for
13:50
example, for a more flexible enforcement, at least at the beginning, as standards
13:56
and benchmarks would actually start developing in the market.
14:00
And therefore we've landed with this solution that is now the AI Act with cause
14:06
of practices that would
14:09
basically establish how the obligations are set in stone in the regulation, but
14:13
how you actually measure and enforce the obligations is the result or will be the
14:19
result of a code of practice that will be developed and co-developed in fact between
14:24
the between the developers and the regulator and that will plug the gap until
14:29
standards will be available and then standards will be taking over and will be
14:34
representing the measuring stick with which
14:37
developers will be judged by the regulator.
14:41
So I think that, again, with this construction, we've managed to find the
14:45
balance between those that fear that we may be going too fast, too far, and those,
14:51
including myself, that wanted to make sure that at least some responsibility is
14:56
clearly fixed in the regulation.
14:58
There's a general issue though, related to what you just talked about in terms of the
15:03
degree of specificity, not just in the foundation model section, but there are
15:08
many requirements about conformity assessments and explainability and so on
15:11
and so forth.
15:12
The companies are going to have to meet where the details are not yet in place.
15:17
And so, how do companies get started on compliance?
15:23
I understand there's a period until the effective dates, but how do they know
15:27
enough
15:28
exactly they're going to be required to do to be able to comply when some of these
15:32
details are not yet in place.
15:34
Yeah, clearly there's a risk whenever you're doing pioneering legislation such
15:38
as this one, this risk exists, but we have tried to learn a very, I think, valuable
15:44
lesson from GDPR, which came with the market with no technical standards, which
15:50
made it very difficult for companies to figure out, to understand what it was that
15:54
they were expected to do.
15:56
And in fact, without going and calling a lawyer, it was almost impossible,
16:01
particularly if you were a small startup, to
16:04
to make the determination yourself.
16:06
So this is why there is a mandate for technical standards.
16:12
So technical standards will have to be developed by the time the regulation will
16:17
come into effect.
16:18
And that will help in understanding, actually, what it is, what explainability
16:22
is, how data accuracy needs to be understood, so on and so forth.
16:27
So all of those things that right now may sound a bit too generic for some of the
16:31
readers
16:34
via technical standards, and very importantly, these standards will not come
16:37
from heaven, they will not be bestowed upon the world by the European Commission.
16:41
We deliberately change the logic of that by giving this mandate to standard setting
16:46
bodies, those bodies where in fact the industry itself is present, so that they
16:51
can, with their knowledge, with their expertise, they can actually develop and
16:56
work on those standards themselves so that at the end of this process, again, two
17:01
years from now, when the law will come into effect.
17:04
there will be a set of written text that will help an engineer understand what it
17:10
is that they have to do.
17:11
Second, in many parts of the text, if one takes a look, you'll see there is again a
17:16
mandate for guidelines that is supposed to come either at national level from the
17:21
national regulators or at European level from the European Commission and the
17:24
future AI office that are also meant to translate some parts of the text or the
17:29
requirements which we also thought would be too legalistic.
17:33
or too generic, and which require that sort of translation or further
17:38
explanation.
17:39
And those are also mandated to be done before the entry into force of the
17:44
regulation.
17:45
So all this to say that basically by the time that companies will have to comply,
17:50
and including to help them self-assess, which is the basis of this regulation,
17:57
self-assessment, they will have the tools to do that self-assessment, to know.
18:01
where they find themselves on the pyramid of risk, whether they have to ignore the
18:05
text altogether because they're not concerned at all, or if they are
18:08
concerned, what they need to do.
18:12
What's your perspective right now on the Biden administration's approach to AI
18:17
policy in the US?
18:21
Well, I think the Biden administration did what they could do in the absence of
18:26
action from the Hill.
18:28
I think for a while, so first and foremost, the Biden administration renewed
18:34
the dialogue with European Union and created the Trade and Technology Council
18:39
framework, which allowed the two executives, the two administrations, EU
18:43
and the US, to start discussing very concretely, among other things, AI.
18:48
And not only in terms of very generic political narratives and alignment of
18:52
values, also very important, but also in terms of putting together the NIST on the
18:59
US side with the European Commission and then having experts talk to themselves,
19:03
discuss standards, discuss how you measure different things, how do you measure bias
19:06
in an algorithm, so on and so on.
19:08
So very concrete things.
19:10
But beyond that, again, the administration looked at the landscape of norms and said,
19:15
listen, if we
19:17
from a political point of view, if we agree that this technology, apart from the
19:20
huge benefits, comes with certain risks, what do we do about them?
19:23
And again, if Congress does not act legislatively, what else could we do?
19:27
And I think the first response was the Bill of Rights, which was a first
19:33
statement of intent, I think, politically by the administration as to what they
19:38
believed needed to be done or signaled towards the industry working with these
19:44
technologies.
19:45
And then the executive order.
19:47
which I see it as a measure that again is meant to start determining certain
19:55
behavior, starting with the federal administration in terms of how to regard
20:00
AI, to give the impetus to start analyzing sector by sector what could be not only
20:05
the benefits but also to see the potential risks and start thinking and addressing
20:10
these risks, each of them in their sector, which is a from my
20:17
a good approach as well.
20:18
We had also considered at EU level whether to go horizontally as we eventually did or
20:24
to go sector by sector, which shows horizontal for certain reasons but it
20:27
doesn't mean that sectoral approach is the wrong one.
20:30
So in fact if you unpack the logic of the executive order of last autumn, that logic
20:36
is not very different from what we ended up doing with the AI Act, only that the
20:40
Act is a law.
20:42
Whereas the executive order is not.
20:44
But again, I think the administration did the most that it could with the tools that
20:47
it had.
20:50
How do you see the global coordination going forward?
20:53
It's of course not just the US, countries like UK and China and others are very
20:57
active in this area.
20:59
How much does there need to be consensus globally on these regulatory areas?
21:05
From my point of view, it's a must.
21:07
It's not nice to have.
21:09
It's not a luxury.
21:11
It is a priority because, again, AI is the same for everyone.
21:16
And AI will be driving the future.
21:19
So one way or another, there has to be a global conversation about AI.
21:25
And ideally, there has to be an alignment on how you deal with the technology.
21:30
In a way,
21:32
If there was one huge benefit with ChetGPT and the LLMs was that it triggered the
21:36
conversation.
21:38
There were attempts prior to that to actually stimulate that level of interest
21:43
in all jurisdictions out there in various forums, in various formats.
21:50
There was some lip service paid to it, but you couldn't really feel the sense of
21:54
urgency or the sense of purpose in those conversations until last year.
21:59
when all of a sudden the world leadership woke up to the reality of AI, helped also
22:04
by some of the public debate and some of the key figures in the public debate,
22:08
whether in the industry or in the academia saying, listen, folks, start looking at
22:14
this seriously because there are some even essential threats linked to this
22:18
technology.
22:18
So that actually stimulated the attention.
22:20
And then we saw the whole plethora of summits and also some bilateral or
22:25
multilateral.
22:26
for our discussions on AI, which is a good thing.
22:28
Now, what we need to do is to start landing all of these different
22:33
conversations and formats and start, it's not going to be simple, but start trying
22:40
to converge towards, ideally, a single framework in which this conversation takes
22:45
place.
22:47
I think that not only the democracies around the world, certainly the
22:50
democracies around the world need to get themselves aligned because they share
22:53
values and that helps.
22:56
but also going past the democracies of the world, even with those that understand
23:00
technology in a different way and the interaction between state and technology
23:03
in a different way, China and others, I think even with those that conversation
23:08
needs to take place.
23:10
Now, there are again various formats in which these discussions are taking place,
23:14
G7, OECD, GPIE, G20, so on and so forth, bilateral frameworks such as the TTC for
23:20
the EU and the US, there are other TTCs...
23:24
and other bilateral formats going on with India, with Japan.
23:27
But again, I think there needs to be slowly but steadily and surely, there
23:34
needs to be a convergence of all of these conversations so that there would be
23:40
ideally at some point one framework where we discuss it.
23:43
And then the second thing is that we now start seeing different embryos of
23:48
governance starting to emerge in different jurisdictions.
23:51
We have the...
23:52
act with a office which is now taking shape.
23:56
We have a safety institute in the UK following the Bletchley Park summit.
24:00
We have a safety institute based on the executive order in the US.
24:04
I had conversations with the Japanese government recently.
24:07
There is now a safety office in Japan with the leadership already elected.
24:11
I understand there will be someone in Canada.
24:13
And I guess that we will be seeing this trend continuing in many other corners of
24:19
the world.
24:19
I have been speaking quite a lot with
24:22
parliaments and governments in Latin America.
24:24
There is a lot of energy there as well.
24:27
There are actually bills on AI being worked on right now in the Argentinian and
24:32
Brazilian governments, parliaments in Chile, in Peru, in Uruguay.
24:37
So as these governances in every jurisdictions are starting to emerge, I
24:44
think there is a, again, it's not a luxury, it's a must that they have to
24:48
start networking because they will all be
24:52
developing different methodologies, different benchmarks on how they evaluate,
24:56
on how they measure ultimately the same models, because the same models are
25:02
popping up in all of these jurisdictions.
25:03
And it would be totally irrational from my point of view to actually develop these
25:08
methodologies in separation from each other.
25:11
On the contrary, I think there'll be enormous benefits if this networking is
25:16
done as diligently and efficiently as possible so that again
25:20
commonalities are being sought out of that convergence.
25:25
Yeah, I mean, it sounds like what you're saying is that there is a great deal of
25:28
harmony in terms of general viewpoint around the world, which is consistent with
25:33
what I see.
25:33
Maybe that's too strong a word, but yes, at least a similar, I guess what I'm
25:39
thinking of is, is at least from my perspective with regard to data
25:42
protection, there really has been a conceptual difference between the European
25:46
human rights based approach and the US vision that sees this really as more
25:50
commercial transactional and so forth.
25:54
Whereas here, it seems like everyone sees the same broad sets of issues, but a bit
26:01
of duplication.
26:01
Okay, so that's optimistic.
26:04
We are in a way in a different logic than we were in the GDPR, where clearly I think
26:09
the concept of privacy was understood differently in different jurisdictions,
26:13
whereas now I think most jurisdictions are looking at AI in the same way.
26:18
So let's say I'm an executive at a company that is operating in Europe or has
26:23
customers in Europe, it would be subject to the AI Act.
26:27
What should I be doing right now?
26:30
Well, I'll be first of all seriously looking into the AI Act, try to understand
26:35
it.
26:35
I'll also look at my business and see, am I concerned by it or not?
26:39
Because I stress this point again.
26:42
AI Act is not for all of AI.
26:45
There's 80%, maybe even more of AI out there that will remain unregulated.
26:51
Because if your AI deals with any optimization of an industrial process
26:58
optimizing transport or energy flows or whatever it is that they do that does not
27:03
intersect with my rights as an individual or with broader values in society then
27:09
your AI is not concerned by the act and therefore you should not worry you should
27:12
continue to do the business that you do but if you do find yourself In one of
27:16
those business areas where your AI is interacting with my rights, whether you're
27:21
doing AI in recruitment
27:22
or in education, or in health, or in banking, or insurance, or any of the issue
27:28
of the contexts that are listed in the AI Act, then you have to ask yourselves the
27:32
question, OK, so then my AI might come under the Act.
27:37
So then I need to start preparing for that.
27:39
I need to start documenting how my AI works, the data sets that I use to train
27:43
the algorithm.
27:44
So prepare already now in advance of 2026 when the rules will come into force.
27:52
to make sure that by the time I am sufficiently prepared.
27:57
What comes next?
27:59
For you, I assume you could use a lot of sleep and also there will be elections in
28:05
Europe and so forth.
28:06
But substantively, what is still on the agenda or next on the agenda in this area?
28:12
Well, now that we've done the...
28:14
So, all of the legislation, and that's important to note, all of the legislation,
28:19
it's not only AI Act, that was put on the table, this mandate is now done.
28:25
So now, evidently we're looking into the implementation of this legislation, and I
28:29
would venture a guess for the next political cycle at European and Euro level
28:33
in the next five years that we won't be seeing a lot.
28:37
From my point of view, we shouldn't be seeing any new legislation in this area
28:40
for the next five years, but even if we will see, it will be limited in scope.
28:45
So I think in a way, this is it in terms of the package.
28:48
And now for the next mandate, it will all be about landing all this, landing the
28:52
plane, implementing, preparing.
28:56
So if you ask what the next challenges will be for the coming five years, there
28:59
will be a challenge for the regulators, which will now have to shape up, get the
29:03
expertise they need.
29:04
bring in the tools they need to do their jobs, prepare for that moment when they
29:09
will be acting not only as enforcers, because as I was saying earlier, there is
29:13
a balance in this text that is not only about enforcing the law, but also about
29:17
facilitating, about guiding, about working together with the industry to achieve the
29:23
objectives that are in fact shared.
29:26
So there's a lot of shaping up to be done at the level of the future regulators and
29:30
they are starting to do that.
29:31
I had...
29:32
Two days ago, a discussion with the national regulators.
29:34
They are starting to get in shape.
29:35
The AI office will be shaping up soon as well.
29:38
So that is, let's say, objective number one.
29:41
Objective number two for us as policymakers, very importantly, we will
29:44
have to start explaining.
29:46
Because we've moved and we've focused and we spent a lot of energy during these
29:50
years negotiating, adopting these texts.
29:53
Now they have to be explained.
29:54
We'll have to do a lot of pedagogy.
29:56
Not only with the companies that are going to be subject to the law, but we are going
29:59
to have to start explaining AI to our citizens.
30:03
Because if society does not embrace this transformation, if they don't understand
30:07
what is happening, if they don't actually gain that trust in that interaction with
30:12
technology, I think that there'll be no uptake and everyone then loses, everyone
30:16
suffers.
30:16
And also AI is going to have a transformation effect for a lot of things
30:22
that are having to do with people's everyday lives.
30:24
From the way education is going to be done for kids to the way our jobs are going to
30:28
look like two, three, four, five years from now.
30:31
So all that needs to be unpacked for the general public so that again society
30:36
understands what's going on.
30:38
We'll have to again look very seriously as though education methods who are going to
30:42
be revolutionized by LLMs and whatever comes next after LLMs.
30:47
We will have to prepare our labor force and reskill and upgrade our labor force so
30:52
that they start absorbing the reality of AI, the tools that will come and yes,
30:57
replace some parts.
31:00
parts of their jobs, not their jobs entirely.
31:02
Some jobs actually will be maybe even rendered obsolete.
31:05
We have to prepare for that.
31:06
There will be other new jobs being created that we cannot even think of right now.
31:09
All that transformation of the labor market also needs to be understood,
31:13
debated, unpacked so that no one feels left behind in this transformation.
31:18
Because if that happens, again, then the debate about AI becomes ideological,
31:24
becomes partisan.
31:26
And there is the last thing we need.
31:28
for a technology that is again so impactful for our future and if we let it
31:33
again fall into the trap of partisanship and ideological debates and fear and
31:40
scaremongering, then again that whole benefit and potential is going to be lost.
31:47
So quite a lot is ahead.
31:49
quite a lot of work ahead, but quite a lot of opportunity ahead.
31:51
And I share your enthusiasm.
31:55
Absolutely.
31:55
Drago, thank you so much for your time today.
31:57
You're very welcome.
32:01
Okay.