Introduction

Thanks for the introduction.

I'm going to talk about controlling AI.

So it basically means

using AI in a safe way.

I think most of you are here because you're interested in AI,

because you realize that AI is an extraordinary new technology we have.

But because it's so powerful, it's also important that we're able to use it in a safe way.

Especially

because of the nature of AI which I'll talk a bit more about later.

So I hope I

can in this presentation give you a bit of an idea of how organizations are

working on making AI safe in their workspace.

So also I have my QR code

right here which is for my personal profile because I'm quite active on

LinkedIn and if people are interested in AI they can add me there if they're

interest in that.

Speaker background and why this topic matters

Okay so Andrea has already talked a bit about this.

Yeah so how did I end

up in AI governance?

I have a bit of a different path I would say than most people working in

governance because I have a more technical background.

I started with studying mathematics

in Delft.

Then I took a year of studying to be a software developer at Hydrogen Race Team

and then I did my master's in math but I realized that now that I've worked on

something which is actually like tangible it was quite difficult to get

back into abstract mathematics so I decided to start Lexfriend and stop

with my master and when working with Lexfriend I realized when talking with

clients talking about AI that they were quite hesitant to use AI mainly because of possible

risk so information security risks they don't know where their information is going they don't know

whether the output will correct will be correct so I realized quite quickly that it's very important

that you can give them some sort of assurances or some sort of rules or frameworks so they're able

able to trust AI because you know using it personally is one thing but if you

have client data that's sensitive you don't want that to just disappear

somewhere so what I did is I did a lot of things to get certified in AI and I

would change my role from a software developer which I used to be completely

specific now to AI governance and information security also at Lex friend

where I work we help organizations with AI in trainings and consultancies and starting from

today I also give talks about AI.

Why AI governance is necessary

Real-world examples of what can go wrong

Yes so the first question you may have is okay it sounds

interesting but why is it necessary to have some sort of governance?

Well I picked three real world

examples that show what can go wrong if you don't have a good structure in place.

So for instance

the first one is with Amazon.

They had used AI for hiring and it was trained on data where

generally speaking where women had a lot less opportunity to be hired than men so what the

AI did is it basically penalized women so made sure it was men were privileged in that way

which is obviously not what they intended to do but they saw that happening in real time which

is just a form of discrimination.

Then the second one is from Air Canada and people were interacting

with their chatbot and it made up a policy so they said you know as a client you are able to

claim this and that but that was completely made up so they had to financially repay those clients.

And the last one is about Samsung, it's more about information security but also AI where sensitive

data got leaked in ChatGPT.

And what's interesting here is that each one of them has a different

thing.

So the first one is about bias in the training data itself.

The second one really has

to do with hallucinations, which is also a big issue now.

And the third one, as I mentioned,

is more about information security.

So where is your data going when you're using it in AI?

So there's lots more that can go wrong.

I'll touch upon that a little bit later.

What an AI system is (and isn’t)

Building a practical definition of AI

and yeah when I tell people that I govern AI I always say that the first

thing that you need to do is know exactly what an AI system is and that

may sound a bit strange to start there but when it comes to AI I see a lot of

people need to have a sort of idea of what it is I mean most of us use it

every day we can sort of in some way express what it is but to really make it

specific what it is to have a definition to determine what it is and exact also what it isn't

that's actually quite difficult to do now people at the european union thought about that for a

long time and they came up with something so what i want to do is basically um build up that

definition in a way that i hope you guys will sort of understand and follow and in that way

understand better what ai is and as i already said what it really isn't

Two core principles: machine-based systems with objectives

so it starts with basically two principles the first one is that it's machine -based

and that sounds a bit strange but what it means is that it's not like human intelligence or animal

intelligence it's something that's software and hardware and then you have another one which might

seem a bit more far -fetched but that's that it has a specific objective so how AI works simply put

is that from the training data it has it learns and in that way it's able to

pursue a certain goal so if you take away the goal there is no learning so it

has to have a form of objective or it works towards because otherwise there is

no intelligence so these are the two core principles and if we build upon

that we get the following so this is next to the thing I just showed and that

Inputs, inference (the “black box”), and outputs

That is, you put something in.

So for instance, when you use ChatGPT,

you put some text in, you put an image in,

that's the input, and something comes out.

That's their answer, so that's the output.

Can be predictions, I don't know, recommendations, whatever.

And the real thing that makes AI special

is obviously what happens in between.

The inference part, or well,

that's what they call it, inference.

And yeah, that's the black box also,

what people call it, and that's where the patterns

or the logic is to infer an answer to your inputs.

So that's also where if you ask the same question,

you don't get the same answer necessarily twice,

which in that sense makes it a very interesting new piece of technology.

Yes, and then we have the last one.

So the last layer on top.

And these are things that I also think will make sense to you.

What makes AI different: autonomy, adaptiveness, influence

That's first of all, the autonomy part.

So we're used to having software

where you tell it what to do and it does that.

But because AI can sort of reason in a way,

it can actually do more on its own.

So what really qualifies or what really is unique about AI

is the fact that it can actually have humans

out of the loop.

So it can continue by itself without human intervention,

like agents, for instance.

And then we have adaptiveness,

meaning it's even after it's in use.

So when you're talking to ChatGPT,

it can still learn and improve even as you're using it that's also new because traditional

software just what you see is what you get it doesn't really change after you've used it

and then there's influence and that's more of a i think more of a philosophical point

and that's that because ai can give answers in the form of recommendations of decisions you

should take it has a real impact on the way we perceive it so it has much more of a direct

influence than a traditional software has.

So this all put together gives you

this overview.

These are all the things we just saw put together in the sort of

one big map and here you see how it basically interacts with this bit of a

simplification but this is the legal definition that we use now to categorize

AI and I'm showing you this because when people think about controlling AI they

they generally only think about controlling the output.

So we think, okay, something comes out

and the only thing we can do to control AI

is making sure that that output is correct.

But the problem is that the part before that,

the inference part, that's the black box.

There's things happening that we don't understand.

So controlling the output is in that sense,

extremely difficult or even impossible.

So what I'm trying to do and show you this

is that the way AI is governed

is by controlling everything around the output in order to assure that the output itself

can be sort of put into boundaries, has some sort of guardrails on it.

So instead of saying, okay, the output has to be correct, we can, for instance, say,

hey, we have to look at the autonomy.

The AI shouldn't be able to do that much on its own.

Or when we look at input, we could say it's only allowed to use this sort of input,

so no sensitive information gets in.

Or we can look at the objectives.

Is the AI actually doing the right thing?

or is it going in a different direction than we want so governance is very much

about changing all these things around the output and the inference to make

sure that you can control it now that's a bit fake but I'm trying to I will

explain how we can actually do that yes so we know what an AI system is and now

How organizations control AI in practice

Frameworks: the EU AI Act and AI management standards

let's look at how we are able to control it yeah this was the difficult word

Andreas mentioned so I just took two frameworks as an example you don't have

to remember the names but I just put them there for yeah so it's like

complete we have the EU AI Act that's the European Union's law about AI

systems and then we have another management system that's like the world

standards for managing AI and those are two frameworks that in our organization

we have implemented so I'll basically walk you through how that works and

their approaches to managing AI because I think it gives you an idea of how you

yourself can handle AI better in that way.

The EU AI Act’s risk-based categorization

So this is the AIX approach, it's

risk -based and what it basically says is you look at your AI system

whatever you have, whether it's a chat or whether it's something that predicts

certain results and you only look at the intended goal, so the purpose.

And

It falls in a risk group.

So it has a list for instance for unacceptable risks

So when you're thinking of you know, manipulative AI

That's just simply not allowed.

So if your use case falls in that group, you can't do it

then you have the high risk and here you have things like

Medical devices the whole list and if your AI system falls in that you're categorized at high risk and you have a lot of obligations

So you have to do a lot of logging you have to do a risk analysis.

You have to do a lot of stuff

and when it's in the limited risk that's for the majority of AI there's very

little regulation actually in the Europe so those people have the idea that

there's a lot of regulation but for most use cases there really isn't and yeah a

very other big group even has no specific rules for instance think about

video game AI things that aren't really affecting people's lives yeah so what

does it mean when you're falling a certain risk group well I took care the

Governance controls: inventory, risk assessment, roles, policies, literacy

other framework as an example so as you can see here what I was trying to make

clear earlier you're not really trying to make the output better in any way

you're just trying to make sure that everything around it is better

organized so first of all what's a very important point is actually making sure

that when you're using AI you know exactly when you're using AI and that

sounds very obvious but what we see is that a lot of people don't really

you realize they're interacting with AI anymore so like using a chat GPT is obvious but a lot of

applications now have AI and a lot of those applications also deal with your sensitive

information so the first step is always looking where am I using it and then we have the risk

assessment so what you do is you look at okay I'm using this AI system can I see where there's

possible risks so what can go wrong and what can what what's yeah really something that can have a

negative influence and you try to mitigate those risks you want to have roles and responsibilities

so if something goes wrong who's responsible for what you need to have some sort of

management in place for people who are for instance going to let your clients know that

something is wrong or people that go contact the organization where the AI is you really need to

have a good structure there and the most important part is policies so those are basically the rules

within your organization where you would say things like we're only using AI from

registered developers if you want to use a new tool you have to contact the IT

person where you say we don't allow sensitive data in AI that's basically

the most important part and it's about continuous improvement so a lot of

people think okay I have a policy in place where I have some sort of rules

tools, but AI adapts.

That's like one of the key properties and organizations change too.

So it's really important that you make a system that's actually able to change in a sort of

natural way.

So it doesn't get stuck right away when some small things change.

And the

last part is very important, AI literacy.

And that means that people in your organization

know what AI is and what its risks are, which is also often overlooked.

Because what you

you see maybe also today is that the because it's so new the level of experience between people can

vary a lot so i'm maybe assuming someone's very good with ai but they've never heard of it and

also the other way around so we see that there's a lot of difference in literacy

so yeah now that you basically hopefully have some sort of idea of how to manage ai in some way

How we implemented AI governance at Lexfriend

Yeah, I want to get from our own account how we did it.

So as I mentioned before, the inventorizing is a very important part.

So I'm also basically asking you tonight to think about how and in which ways you're using

AI already.

And it's probably a lot more than you think, because there's already small applications

like spam filters, like Google search, a lot more things that already use AI, where you

You might also put sensitive information into or interact with other ways that can actually

be harmful.

So that's important.

Yes, then what's important is that you're able to categorize the systems in a way.

I mean, if you want to in your organization or for yourself, want to have an overview

of okay, I know my systems, then you should put them in some sort of framework like an

Excel or any other thing you like and then go through all the steps like okay, is it

using sensitive data are they do they have some sort of certification about how they use their

data do they explain how they use their data do you know that it's ai what sort of ai those are

all the things that you're going to put in an overview what's also important in your team is

that everyone is on board so what we see is a lot that the management for instance of organizations

are not really like invested in AI usually they're yeah a bit more senior

people of course and so AI is not something that they may have used

naturally so it's very important that those people also understand or have

some sort of better idea of where the risks may lie and then of course

something that's often overlooked the people that actually make AI so in our

organization we have people making AI systems we involve those heavily in

making our policy because yeah they know how it works then the AI literacy is

important part I think we're doing that tonight all together making sure people

better understand AI and also understanding its risks also in your

organization if you already have things in place for compliance for instance

with information security or gdpr then a lot of it can be reused and then i have my last

Using AI to help govern AI

point which is quite maybe a controversial point but what we actually found really useful was to

use ai to govern ai so what we did is now let's say we're looking at making certain rules to bound

ai what we could do is then write out some rules we have and then give it to some sort of chat

interface and say okay is this like exhaustive is this really handling all cases and then it says no

because these things you're not covering yet and we're all okay um what we could say oh the ai

access we have to do this but what does it mean exactly and it says oh yeah you're an ai act it

says this and this and this so it's actually quite a helpful tool to govern itself in some weird way

but that really helped us it saved us a lot of time and obviously you have to make sure you do

that in the in the way you're building it here like governing it but that was extremely useful

and made our whole progress a lot faster too so the most important thing i actually like the

Key takeaways: governance builds confidence (and prevents “shadow AI”)

takeaways i want to give today is that this is something a lot of organizations are working on

ai governance but since it's so new there's a sort of at least the way i see it a sort of

gap in experience maybe because when we look at older things like information security GDPR

there's like tons of people who already doing that for years they know a lot about that

and with AI it's very new and you see organizations constantly trying to figure out how to do it best

and one of the most common reactions is just to not do anything so we see shadow AI a lot where

organizations ban the use of AI or extremely limited which means that people are going to

to use it anyway but secretly which is much worse because then there's no policy at all

or they don't want to use any AI.

Now I think AI is going to be maybe the biggest thing in my

lifetime at least so I don't think that's a good strategy either.

So my whole point here today is

that when we look at our own organization we really feel in control and confident about using

our AI because we follow these rules.

So there is a way to deal with AI in a useful way and also

make sure that it's very effective and can really help you yes so that brings me to the end of my

Audience Q&A: adoption, tools, compliance, and regulation

talk um yeah in case anyone is interested if they're thinking about ai in their organization

or they want to learn more about the things i said we also have a website which you could check out

but also i'm very curious to hear your own experiences or your questions if anybody has them

if you have a question

please try to speak up

and speak slow

yes please

do you also feel like

everyone is just trying something

and investing

in AI and then see

what works and what doesn't work

and in the end what they start with can change

completely

yeah

good to

summarize the question

even though you speak up

it's it's better to summarize and then yeah so you basically said if I say

correctly that a lot of projects are starting with AI and they might end up

completely differently than where you started but everybody's just trying

things out yeah we see that a lot so we actually had a talk here a couple of

weeks earlier but someone whose job it was to go to organizations and write all

those ideas out and basically select the projects that are interesting and but we

see that yeah so we have people that don't want to do anything with AI but

but also people want to do a lot,

but then they don't really know the use cases.

And that brings them very weird solutions

that don't really work in the organizations.

So it's actually quite difficult, I think,

or more difficult, but you really have to know

a bit about AI and about your organization

to actually apply it in a good way, I would say.

So we do see a lot of organizations piloting new projects

and then they don't really work.

Yeah, that happens.

Yes.

I'm wondering, so in your industry, do you meet more people who don't know how to use

AI or do you meet more people who are misusing them?

So you ask, do I meet more people that misuse AI or don't really know how to use AI?

More people that don't use AI.

so surprisingly enough especially also in in fields where it would be super

useful because people are for instance working a lot with text like in legal

fields we see that a lot of well a big part of the people we speak to don't

really use AI that much and the problem is that because they don't they don't

see the potential that it has so it's very difficult to explain to them that

that it's interesting because they don't care yeah that's something we do see yes

Yes.

Yeah, there are quite some different AI tools.

And I know in my organization, we started with CoPilot.

Yeah.

And you said goodbye to CoPilot and now JetGPT.

Do you feel like you need to know all the different tools

or can you choose one and use that in the post?

Yeah.

Yeah, so you're asking, there's a lot of different tools.

How do you know which one to use?

Yeah, so that's a good question.

You have specific tools, AI tools that might be very good

at specific things, but you also have these bigger ones

like cloud or co -pilot that especially now a bit later have significantly improved that i would say

are somewhat similar in the way they work i do think some can be a lot better but the way they

work they're quite similar so for instance i use cloud a lot i prefer that over the other ones

but i also use the other ones because i sometimes want to check one answer with another ai

so to see if that actually makes sense but what you do see is that the

beginning had lots of different AI things so one AI product did one

specific thing and now you see the bigger companies doing a lot more so the

the industry of different AI products is sort of decreasing in that way so I do

expect well what it seems like is that there's going to be a few big

big organizations that are going to be very good at a lot of things.

You do depend on the tools that the company chooses to work with.

Yeah.

Yeah.

I have a question too.

Come to you then.

Yes.

The different standards of, let's say, security or the questions you have to ask yourself.

It looks like quite a big program.

Yeah.

And even me personally, when I look at all of these steps I should be aware of, I know

i would skip half of them probably because either i don't want to or it takes too much time or it's

too confusing let alone a company that has to instruct its people to go through all these steps

so how probable is it that this happens yeah that's that's a good point so what what usually

occasionally people say as they draw the comparison with GDPR so that's the

the AVG in the Netherlands like the information security law that came into

practice like a while back and then all of a sudden organizations also had to do a lot of

stuff that they didn't have to do before so the problem of course is communicating that with

with everyone in your organization.

And in a certain sense, I don't think it's,

I mean, you can obviously,

what they do is they give trainings,

they tell them about the policy, where they can find it,

but in practice, it's gonna be difficult.

So what generally happens is

if people don't know what to do,

they have like someone they can contact

and they'll tell them what they can do.

They really know what they shouldn't do.

And yeah, in that way, you try to make it work,

but it's difficult, especially with larger organizations.

So you're basically asking, is the AI act sufficient in really, you know, creating ethical

ethical AI yeah exactly yeah yeah that's a good point so there's a quite a lot of

debate about that so what's important to know is the AI act is coming into

effect and sort of phases so it's not fully in effect now but in there keep

postponing it but in a while the high -risk systems the high -risk group

which is the most regulated group will go into effect right now there's no

regulation at all but yeah there's a lot of debate about that because as I said

you're a high risk only if your use case is on a specific list but you can

imagine that that list yeah I mean it's only a finite list so there may be there

are specific use cases that aren't on there that can be very harmful and also

So the AI Act is quite vague about things having to do

with the algorithm or the AI itself.

So what it does is it can talk a lot about,

do logging, do risk analysis, that kind of stuff.

But it also says, yeah, there shouldn't be any bias

in your data, but yeah, I mean, how are you gonna do that?

And a lot of those fields, it can be quite abstract

and that makes it very open to interpretation

and that's not good.

but also yeah how how could you fix it it's it's such a an intangible thing in that way that it's

very difficult to regulate so yeah it's expected that there will be going to be some amendments

to the laws to make sure that it will work in practice thank you where you had your ban high

risk yeah they had simple like for example chatbots and stuff like that yeah what about

about when agents come in?

Because that's a much more multi -field, I feel like.

Because you start with one use case

and then it keeps building on itself.

And it might move to a different one by itself.

Yeah, that's one of the biggest issues now.

So it's a very good question,

because as you properly said,

it's a use case that matters.

So as your use case dynamically changes

or naturally changes,

you can move into different risk groups.

risk groups but yeah I mean how can you manage that in that way so that is a

serious problem I mean at the time that the AI Act was written wasn't really

expected that these type of things would happen anytime anytime soon so I guess

now if I would think about how I would do it you would probably have to sort of

assess beforehand what could happen and then make sure that that that you can

assess it but i mean i spoke to you earlier you said it's very unclear where it will go so

yeah that's a very interesting point difficult to do so and honestly i just don't think companies

would do it now because it's so uh yeah all over the place yeah so what he's talking about is that

you have this um new ai technology basically well i don't know anyway there's this new thing

and that's much more autonomous than before so you can ask it to do things like for instance

present you the news every morning and then it does that on its own so it goes to different

websites it can log in for you it can decide oh maybe i want to go there too and it does all this

type of stuff but it can also do things on its own that was not intentional at all for instance it

could you know log into a news site and all of a sudden place a comment for you on a website that's

like super politically charged that was something that actually happened with someone so the problem

there is that you can't beforehand really assess the risk that AI can have

because you don't know where it's going to go so that's a very interesting thing

now.

There was one last question I think yeah yeah that's a good question so how

How do other countries or regions deal with regulation?

It varies a lot.

So actually I was working on making a sort of world map to assess it because every day

I see new things like Ireland has something, Taiwan has something, you know, you see it

all over the place.

But it's really sort of individual.

So there's no real worldwide network that does this.

it's each on their own and a lot of countries don't have anything at all so

the Europeans are one of the first to actually push this and what generally

happens which also happens with if you look at for instance and the Apple

charging that's the maybe a good example because the EU says they have to use

USB -C they'll do it everywhere because it's easier for them that way so

generally the European standards will also influence the rest of the world

I have one question though.

Since we have the EU AI Act, do you think in the future

if any individual or organisation use AI for bad purposes, do you think they are going

to face more legal consequences?

Yeah, so you're basically asking if an organization does something bad, will they face consequences

in that way?

Yes, because I think when the Cambridge Analytica scandal happened, there wasn't too many people

concerned.

No, exactly.

And at that time there wasn't really regulation in place.

So what people expect that will happen is the same what happened when GDPR, the information

security laws went into effect is that they'll probably find big companies a

lot at the start as a sort of deterrent and then people will quickly see oh I

don't want this to happen to me so they'll make sure that it gets

updated it's important to know that you can now make an AI system in any way or

not in any way one but in the high -risk category and don't do any sort of

control because it's not in effect yet anyone else okay we wrap this up yes

Thank you, Lauter, give him another round of applause.

Thank you.