Introduction

One important topic is responsible AI. Who has heard of the term? I know who has an idea, probably most of you.

To give you a little bit of context, I did a couple of weeks ago now in Saudi Arabia a three days training for a group of about 25 people on this topic. three days on this topic now we have 15 minutes so we're pressing now not three days into 15 minutes don't worry yeah but you can see the importance and

relevance of that topic the more the topic of AI evolves the more critical responsible AI becomes and what I want to do here in this 15 minutes just give you a little bit of a perspective. It's an initial overview.

If you want to go deeper, reach out to me. But to give you an initial perspective on this topic.

It's such a complex, large topic and with that you get a little bit of an idea. Principles of Responsible AI, we talk about risks, we talk about a very interesting phenomenon, how automation bias

bias impacts us. And we'll talk briefly about six levels of governance. Don't worry, there's no test afterwards.

So different to the, of course, different to the trainings, there's actually a test at the end. So we are not doing any tests here.

So it's for you, it's really just an overview.

Why Responsible AI Needs Governance

So I think what is important is that governance, which is part of course of responsible AI is not the speed limit there's a lot of sayings is okay governance hampers the advancement of AI who has heard that most of you probably yeah you a I act a lot of discussions around that yeah but as a

matter of fact without governance there's no trust without trust nothing works yeah and this is not we are not talking now about let's say what you do do with ChatGPT and Cloud and whatever.

This is what you do in an enterprise environment for your organization. If you don't have trust in AI solutions, nothing is going to happen. This is why kind of governance is the prerequisite.

The Core Principles of Responsible AI

Let's start with the eight principles of responsible AI.

Feel free to take photos if you want. This is actually quite nice and a little bit bit of a guide when you go into a project.

Anyone here is doing consulting for companies? Probably quite a few. So this is for you.

So of course, you might know already some of that. Incorporate that into your consulting gigs.

Who is kind of doing projects in organizations, not as an external but an internal? Also quite a few. Again, probably applies most of that.

You see here, and you see here kind of the top ones, ones, the four, are more on the personal individual level and this one is on the systems level. These are the eight principles.

We're not going through all the principles. It just gives you a perspective of what is relevant.

Any questions on that? We do have some question time afterwards as well.

This is really helpful in terms of entering a project and thinking thinking about those principles. How do I make sure the solution is fair? Sounds like

an easy question, right? It's actually not easy to answer. It's really difficult to answer.

Accountability the same. Humans remain responsible. All these, of course, the bottom part for those who develop systems, security, robustness, security, we just had this kind of interesting

interesting incident.

Understanding AI Risk Categories

When we look at risks, there are three categories, operational risks, ethical risks, and legal risks.

So you see a very broad kind of risk profile in what's happening and what can go wrong with AI. And you have to cover all of them. And that's the categories.

We have a bit of a more breakdown.

Eight Common AI Risks in Practice

Here are eight AI risks. And one of the things I did is actually it's really nice, is I have these risk cards, yeah, so here's each risk is actually in detail, with more detail, if one wants to have a look later on, yeah, come and see me, this is kind of

in training space, what we use, yeah, we can actually go through which one is actually more important, is it data leaks, is that a problem, yeah, is it automation bias, yeah, is it model total drift. Each of them is a major risk for a project.

Some of them are more critical in the beginning, some are at a later stage. Make sense?

Again, if you have questions, happy to answer those later on.

Discrimination, hallucination, we all know. Who knows what what hallucination is? Let's say, who doesn't know what hallucination is? Everyone knows, great. So, that's now known.

The other question I usually ask and like asking is, is hallucination a feature or a bug? Feature or a bug. Okay, so who is feature? Okay, who is bug? Who doesn't know? Both. It is a feature, yeah?

So when you look at how the systems are built, Hallucination is a feature. It's not a bug. It's a problem and a risk, but it's definitely not a bug. It's therefore a feature. And when you use another term for hallucination, creativity, which means something is produced which doesn't exist, then actually you want that. I hope that makes sense.

Hallucination isn't. I mean, it's a problem, of course, especially in business environments when you do that. but in general is exactly what you want you want new things yeah that's why you

have the temperature who knows what the temperature is in this context quite a few so that's nice who doesn't know so for those of you who don't know it's basically kind of a number which you can kind of put in which is the level of creativity to simplify that yeah so it's kind of you can actually dial in you You can't do that in chat GPT, don't worry. But in systems, you can actually set the temperature at a certain level. And the higher it is, the more kind of creativity there is.

Yes? You can? Okay. Okay. Then so be it. Good. Thank you.

So that's kind of the eight risks. Make sense?

Automation Bias: The Human Factor

This one is an interesting one. And this is probably one of the biggest challenges we see right now. Well, there's quite a few.

So the incident, of course, we had with open AI and hugging face is a huge challenge. But this one is big automation bias because this is about how we behave. So what's behind that?

Basically, and this is the example of radiologists, but it also applies to legal, financial, operational, everything. The problem is as follows. kind of radiologists using AI assist tools detect fewer cancers than those working without AI even kind of when they kind of don't use that anymore it's

like they have this problem so it goes away kind of that knowledge and experience what we also see and basically we did a kind of big case study in the three days is like that humans tend to say yes after a while more than initially.

Make sense? Who has ever kind of signed, kind of ticked the box on T's and C's when you bought a software? Okay, now I ask you the second

question which is a more important one. Who ever read that? One person, two persons.

Okay, that's a lot. I mean, does it matter? Well, I read that but I I just bought the software, so I mean, I can't give it back, yeah, so it's like, it's anyhow

ridiculous, yeah, but it's kind of 10 pages of T's and C's. If I disagree, what should I do then? That's another example for here, the automation bias.

So, the system proposes something, for example, here, radiologists, or in recruitment, said, okay, this is a great candidate, it and initially humans are actually looking kind of is it okay or not after a while they say oh the AI does that right make sense huge problem massive problem it's also an unsolved problem

because it's actually human behavior it's not about hallucination it's not about models it's nothing of technology it's us how we behave make sense good I think it's a personal judgment that you can trust it or not trust it.

It sometimes gives you a push. It's not always false. In my newsletter on LinkedIn, Humanity in the Exponential Age, I've written about kind of laziness versus support. And that's exactly kind of that balance.

Yeah, kind of where you're going and we'd all tend towards laziness. That's just human Yeah, yes Yes

Yes, yeah that we all tend towards laziness, especially when we have a tool which does the thing better than we I Mean it's of course say. Oh, that's better than I can do it. So I just go with it Yeah, big problem.

Explainability and “Code as Law” Challenges

This one is an interesting one. Who knows the kind of the term code as law? What's behind that a few?

Yeah, so of course with code so far You could go into a courtroom and you have a coder and that goes line by line and say okay this system does that Yeah, because it's deterministic

Systems in AI are no longer deterministic So this becomes a real problem, so now it's kind of you can't use that in courtroom anymore Yeah, big challenge, you know, because you have to find new ways now.

I'm just showing you a perspective right now in terms of kind of what is in this responsible AI. Yeah, so here's intent versus learned behavior. I mean, even the experts, even the people and anthropic don't know exactly why the model does what it does. Yeah, so that's a big problem.

Yeah, you can't use that in court anymore. You can still use the models, but you can't use it in court because you can't prove exactly what happened

Now nobody actually I mean nobody knew that this thing finds a vulnerability and actually escapes Yeah, and does something Okay, so that's a big one

Turning Responsible AI into a Competitive Advantage

But the question is always then how can you turn the responsible AI to an advantage? If you are involved in a project ask these questions How can we make this as an advantage?

You have limitations, but it's clear that if you do that in the right way, you build trust. You build trust with your clients, or you build trust with your internal clients, if you do internal projects.

Make sense? So it's kind of really asking this question, how do you turn that around so it becomes a competitive advantage? That's an interesting one.

I'm getting closer to the end.

Where Bias Enters the AI Lifecycle

So where does bias enter? Everyone knows what bias is? Who does not know? I think that's always a better question.

So bias is in the data. It shows kind of a tendency towards something. So you have a bias between male, female, white, colored, kind of these things, kind of West, other countries. You know, so that the models, let's say, favor certain parts of everyone, okay? And you have that still.

Every model is biased. There is no non -biased model. Yes? They can be treated actually to be biased for something or nationality or... Yes, yeah. There's all kind of biases.

The problem is it can be here at every stage of the life cycle of an AI project. There is no, here's a bias, if I solve that, then it's solved. You can see it here from training data. That's a big one, of course.

Model design, how do you design that? How do you deploy that? How do you work with outputs? What does a human do?

The automation bias is a bias, is a form of bias, but it's a behavioral bias. That we tend to say yes to what we're seeing. Okay? Okay, so again, it's not a one -off, you know, bias is quite often,

I remember initially bias was very easy to find in chat GPT. So, yeah, so especially the male -female bias was like very obvious. And the models get better and better, or the harness gets better, yeah? So therefore there's less obvious bias, but it's not gone.

And you have to be very careful when you do a proper project, either with your clients or internally, this can actually really hurt, yeah?

Yes, yeah, I mean here you kind of, it depends on which kind of features are you doing. Are you, for example, kind of picking kind of male, female? Is that relevant or not? It's like at every stage, it's kind of, the elements you're picking have a bias. Because everything has a bias, make sense? Yeah, so it enters everything.

This one we talked about, automation bias. That's kind of humans decide wrong. or kind of tend to kind of move in a dirt in a direction yeah so you had every stage you have different then you have here kind of you create outputs if you use this again for training again then you kind of embed more bias make sense

Privacy and Data-Use Risks in Real Projects

good question thank you good and you see here privacy risks again across all lifecycle I really want you point to out if you do AI projects and get work talk talk here just to clarify.

It's not you using ChatGPT and Cloud. Most of that doesn't really apply here, of course. But if you do projects, solutions, then these things do apply.

And you see that. This one is a big one, by the way. I share that in trainings.

Providers default to using your inputs for retraining. How many of you have turned that off?

How How many don't know that it's actually on? It is on. Go into your ChatGPT in settings, privacy, it is turned on.

Everyone in the training is like, wow, I didn't know that. So all you shared so far, all your stuff, all your personal stuff, all your personal problems are used, everyone not shared personal problems with ChatGPT, I'm the only one. it's on

which means is that a super big problem no, why one second because it's used for training purposes but it's not

Joey I ask kind of your account number even if you put the account number in chat GPT for whatever reasons

it's not a database which stores Joey account number is that that.

It's used for training, overall training purposes, but it can happen that it kind of slips somehow out basically in certain settings. Make sense? I still recommend turning it off.

I did this error even with anthropic encode. Even on an Abo, in a paid version, it's actually on okay still I'm sorry I can't prove all this process but from what I know even if you

if you are personal account and you do turn this off it's still only a legal way legal protection you get if you get a business account and you get only legal protection yeah there's a yes yeah but

But at the end, look, I mean, it's business. You have to, I mean, you're trusting with Office 365 and Google, you're also trusting the organizations. So I think this is a very different conversation

in terms of levels of trust. But in general, if you turn it off, it's off. You have to assume that because that's kind of what basically

a multi -billion company says. But of course, you don't have to. I mean, you just use your own model.

Six Levels of AI Governance

As a last one, I wanted to show this one, and again, this is more for, welcome, this is more for kind of governance, so there's six levels of governance. Again, it's more for if you have a project, and it starts really from the bottom.

So first is policy and standards. So it's kind of defining everything.

Then it's roles and accountabilities, who does certain things, yeah? For a certain period, yeah.

Yeah, let's not go there. So kind of let's focus on here. No, no, because let's discuss that later on. Yeah, so over a pizza, because it's a contentious topic.

So again, it's also about belief and trust and things like that. So yeah, and we can't prove that anyhow.

So the policy and standards is one. If I do that, then next I define who does what. Yeah, that's roads and respond.

Then I can say, okay, now I do a risk assessment. Now you have a better understanding about all the risks.

Then I define, well, who is, what's the human actually doing? What's the oversight, the review?

After that, I decide how do I disclose that, for example, with my clients, with my business partners? So I have a solution, and how do I disclose that?

And finally, I do an incident response. So if something happens, what's the response then?

Hardly an organization is that far to actually cover all these six levels of governance. Good.

Conclusion and Next Steps

I think with that, I close. If you wanted to stay in touch, that's kind of the LinkedIn QR code. Thank you very much.