AI Ethics in Internal Audit

00:00:02 The IIA

The Institute of Internal Auditors presents All Things Internal Audit Tech.

00:00:06 The IIA

In this episode, Antonio Cachaputi and Alessandro Casarari unpack the ethical challenges of artificial intelligence in internal audit and anti-financial crime.

00:00:16 The IIA

They discuss AI hallucinations as a risk to be governed, not eliminated, and examine why governance, accountability, and human judgment are central to ethical AI.

00:00:28 Antonio Cacciapuoti

How do you think internal audit can evaluate a system that by definition can hallucinate?

00:00:34 Alessandro Casarotti

For internal audit, we need to change a little bit the question, because the point is not if an AI system can hallucinate, because by design, of course, it can and it will.

00:00:48 Alessandro Casarotti

The real question here is if the organization is ready to live with that, because from

00:00:55 Alessandro Casarotti

an internal audit perspective, AI cannot be assessed in the same way as a deterministic system.

00:01:02 Alessandro Casarotti

So the point is not to expect that the result will be correct every single time.

00:01:09 Alessandro Casarotti

For me, what really is important is understanding what the organization has put in place

00:01:15 Alessandro Casarotti

to manage and to mitigate situation when the model can be wrong.

00:01:20 Alessandro Casarotti

So in my opinion, the internal auditors should focus on three important things.

00:01:26 Alessandro Casarotti

The first is the governance.

00:01:27 Alessandro Casarotti

So who owns the AI and who is accountable for the use of AI.

00:01:33 Alessandro Casarotti

The second is the control around the usage and not the model itself.

00:01:38 Alessandro Casarotti

I mean that auditors should try to audit the algorithm line by line, but they should assess if there are human in the loop, if there are controls, especially in high-risk use cases.

00:01:52 Alessandro Casarotti

And the last but not least, the user knowledge and transparencies, because

00:01:59 Alessandro Casarotti

it's important to understand if people understand that the system can hallucinate or if the system is perceived as an authoritative source.

00:02:09 Alessandro Casarotti

Because here the distinction is very important because it changes the risk profile and the type of control we can put around it.

00:02:19 Alessandro Casarotti

For internal audit, the hallucinations are not a technical failure, but it's just a risk, a normal risk to be mitigated.

00:02:28 Alessandro Casarotti

So

00:02:29 Alessandro Casarotti

So if you want to put an ethical point on this, I would say that an ethical AI system, it's not a system that never hallucinates, but it's a system where hallucinations are expected and safely assorted by the organization.

00:02:46 Alessandro Casarotti

So this is my view on your question, Alessandro.

00:02:50 Alessandro Casarotti

And now I would like to shift the focus to your experience in the financial crime space.

00:02:57 Alessandro Casarotti

And I'm really

00:02:58 Alessandro Casarotti

curious about your answer.

00:03:01 Alessandro Casarotti

So based on what you have seen so far, of course, how are the main AI providers building controls to make sure that decisions remain ethical and that these tools are not used for fraud?

00:03:15 Alessandro Casarotti

And please, if you have any concrete examples, just give it to us.

00:03:20 Antonio Cacciapuoti

Remain ethical for big AI players is a big point of concern.

00:03:26 Antonio Cacciapuoti

Lately, I think we had a big statement from Sema Altman of OpenAI, but also from the CEO of Anthropic in regards to where really they are looking for also ethical regulation that gives a bit of grounds on how to address

00:03:42 Antonio Cacciapuoti

those topics and the development that is taking because it's going faster and faster.

00:03:47 Antonio Cacciapuoti

Because big players have taken different levels of safeguards on ethical perspective, which can be a combination of technical safeguards, policy frameworks, active monitoring,

00:04:02 Antonio Cacciapuoti

to try to, of course, limit the risk also that fraudster can exploit.

00:04:08 Antonio Cacciapuoti

But I think, at least in my perspective, the most interesting one is from Anthropic, because they're building their own model and AI constitution which evaluates the answer on a model perspective.

00:04:22 Antonio Cacciapuoti

And I think that's quite interesting that it's a private institution that builds

00:04:28 Antonio Cacciapuoti

in its own model, an AI constitution to evaluate the ethical perspective.

00:04:34 Antonio Cacciapuoti

And there is no public push for having this as a consensus.

00:04:41 Antonio Cacciapuoti

Actually, we have also very different big AI system, which has been criticized actually for the opposite, to not have ethical safeguards.

00:04:49 Antonio Cacciapuoti

So I think as a first approach, it's very interesting where Anthropic went and the fact that it's a private institution pushing for that

00:04:58 Antonio Cacciapuoti

and not a public one, even though we know that regulation, at least in Europe, they are trying to ensure that some safeguards are put into place.

00:05:09 Antonio Cacciapuoti

But then if we look more overall on all other safeguards, I mean,

00:05:14 Antonio Cacciapuoti

There are several things that have been built in the model, such, I mean, classification on wireless detection, toxicity, bias detection, that can run to ensure that the model blocks certain outputs.

00:05:30 Antonio Cacciapuoti

But that goes a bit also in line with what you were saying before, I mean, and it's how the model was validated, because the model was validated, trained by human.

00:05:40 Antonio Cacciapuoti

And so the famous human judgments needs to stay in somehow.

00:05:44 Antonio Cacciapuoti

And it's the famous human in the loop that also the UAI act is pushing for.

00:05:50 Antonio Cacciapuoti

I mean, it's important in a gray area and really subjective area such ethics to have human judgments that is let somehow also controlled by the human in the end, whether with human feedback, whether with some specific testing, which can be

00:06:10 Antonio Cacciapuoti

also done by prompts that are studied for that.

00:06:14 Antonio Cacciapuoti

I mean, there is a specific team testing also that goes a bit in that direction also to ensure that social engineering is not used for fraud models or so to phishing or other potential manipulation.

00:06:28 Antonio Cacciapuoti

So all in all, I think what is important is that we keep in mind that fraudsters are always the one step ahead.

00:06:38 Antonio Cacciapuoti

And I mentioned a bit earlier, I mean, if they know how to prompt, they can always rewrite the question to move around the ethical safeguards through to looks legit.

00:06:50 Antonio Cacciapuoti

And that's why human in the loop is important to ensure that the model.

00:06:57 Alessandro Casarotti

So we need to act as criminal to stay one step ahead.

00:07:04 Alessandro Casarotti

And.

00:07:05 Antonio Cacciapuoti

Usually we used to say as a motto that you need to think like a criminal to catch such criminals.

00:07:12 Alessandro Casarotti

And maybe we will talk later about it.

00:07:14 Alessandro Casarotti

But I think also in your opinion, the human are the main

00:07:20 Alessandro Casarotti

characters, I mean, first human and then machine.

00:07:24 Antonio Cacciapuoti

Yes, I mean, I mentioned anthropic, but anthropic and then AI constitution made by humans.

00:07:29 Antonio Cacciapuoti

That's very interesting.

00:07:31 Antonio Cacciapuoti

I mean, in the end, the human is the one creating the machine.

00:07:36 Antonio Cacciapuoti

So being the one that creating a machine is also sure the ethics in the machine, because it's the ultimate responsible.

00:07:43 Antonio Cacciapuoti

I said, the human has this creativity,

00:07:46 Antonio Cacciapuoti

that can help to circumvent the safeguards.

00:07:49 Antonio Cacciapuoti

So that's why fraudsters are always one step ahead also of artificial intelligence actually.

00:07:54 Antonio Cacciapuoti

Right.

00:07:55 Antonio Cacciapuoti

Because if they know how to prompt.

00:07:57 Antonio Cacciapuoti

They know also how to circumvent those safeguards.

00:08:03 Antonio Cacciapuoti

And if you want to stay a bit on this philosophical part, at least what I find interesting is that you need to be able also to, as you mentioned, I mean, you don't need to know how to code and to read all the code, but you need to be able to explain a bit the process to ensure you can audit it.

00:08:19 Antonio Cacciapuoti

So how ethical is for you to automate a process if you're not able to explain it?

00:08:26 Alessandro Casarotti

I love philosophy, and here I don't think ethics is just about if we can fully explain a system, but it's more about if we understand the consequences of using that system.

00:08:40 Alessandro Casarotti

Because many processes were never fully explainable, even before AI.

00:08:45 Alessandro Casarotti

Think about some example linked to my

00:08:48 Alessandro Casarotti

asset management work, for example, the complex credit model, so the investment decision made by the portfolio managers.

00:08:56 Alessandro Casarotti

We often can't explain every step, but we still have accountability for them.

00:09:02 Alessandro Casarotti

So the ethical problem for me starts when automatation removes responsibility.

00:09:09 Alessandro Casarotti

If nobody can explain what the system does, why it does it, and especially who is accountable when something goes wrong, then automatation becomes ethically weak, not because it's so back, but just because it's not well governed.

00:09:26 Alessandro Casarotti

So if also here, because we stay in an internal audit perspective, if we jump in the internal audit world, the key question for me is not if we can explain, as I said before, the algorithm line by line, but if we can explain when the system should be used and when not, and especially if we can explain who is responsible for the outcome produced.

00:09:55 Alessandro Casarotti

So if you cannot explain those things, you as a company, you shouldn't start to automate the process.

00:10:04 Alessandro Casarotti

So ethically speaking, for me, automatation without the possibility to explain what is acceptable, the automatation without possibility to explain, it's only acceptable when the impact is low.

00:10:20 Alessandro Casarotti

So when the risk is low, as you were saying before, the humans remain in the control.

00:10:25 Alessandro Casarotti

And if there are clear control put in place by the organization when things go wrong, because things, of course, can be wrong.

00:10:36 Alessandro Casarotti

So otherwise, if you combine just opacity with automation, we are not talking about ethic.

00:10:43 Alessandro Casarotti

It's not ethic.

00:10:45 Alessandro Casarotti

It's just risk moving faster, which is not responsible.

00:10:49 Antonio Cacciapuoti

And that's interesting because what you said, I mean, mainly speaking about the risk perspective, it's a bit where there, at least in Europe, the regulation is trying to give a bit of safeguards when we say, okay, depending on certain type of activities can be defined whether it is low, medium, high, or very high risk, and therefore certain level of controls is expected in the activities.

00:11:11 Alessandro Casarotti

Yeah, because then the perception of risk

00:11:15 Alessandro Casarotti

can change and also the control that every organization can put in place to mitigate this risk.

00:11:21 Alessandro Casarotti

But, Alessandro, today the world is more connected than ever, but especially in our sector and in yours, I imagine that you often have to navigate local regulatory barriers from clients operating under different legislation.

00:11:39 Alessandro Casarotti

So with that in mind,

00:11:41 Alessandro Casarotti

However, AI players approach ethical decision when operating under such a different regulation.

00:11:48 Alessandro Casarotti

For example, if we can mention Europe.

00:11:52 Alessandro Casarotti

Asia or United States?

00:11:54 Antonio Cacciapuoti

I mean, if we take these three macro regions, I think we can see already three quite different approach on AI and also mainly regulation.

00:12:04 Antonio Cacciapuoti

I mean, we mentioned already Europe and we know that also Europe is somehow also teased about over-regulating, but for sure gave us quite clear ground

00:12:16 Antonio Cacciapuoti

to build up some governance on what is EU AI with the EU AI Act, and also therefore to ensure that there are some parameters embedded for having safeguards that are brought forward.

00:12:31 Antonio Cacciapuoti

On that perspective, it's way different from United States and Asia, which have more a pragmatic approach also for

00:12:39 Antonio Cacciapuoti

constant development.

00:12:41 Antonio Cacciapuoti

Because if we think on an overall perspective, I mean, we mentioned so far the EUAI Act, but at least here in Europe, any big players needs to ensure multiple regulatory compliance.

00:12:56 Antonio Cacciapuoti

Not only the EU Act, EU Act is a foundation, but there is also other consideration to do.

00:13:00 Antonio Cacciapuoti

For instance, the GDPR regulation.

00:13:02 Antonio Cacciapuoti

I mean, all the data that you are inputting, where are they going, how they are treated, because also this may have impacts.

00:13:09 Antonio Cacciapuoti

on how the model is learning and therefore also about also some ethical or bias discrimination.

00:13:18 Antonio Cacciapuoti

And if for you in your sector, in the financial sector, you have even other regulation that goes on top.

00:13:24 Antonio Cacciapuoti

So the Dora one, the Cloud one, the outsourcing one, all together, I mean, it becomes very complex to put a system in place that really fits all activities.

00:13:36 Antonio Cacciapuoti

But at the same time, I think it helps

00:13:39 Antonio Cacciapuoti

to perform a reflection in the development of the model and of the need that AI is required to answer to ensure that all potential safeguards and risks are actually assessed.

00:13:54 Antonio Cacciapuoti

Because when you have to consider all these regulatory compliance, I mean,

00:13:58 Antonio Cacciapuoti

the technology is moving way faster.

00:14:00 Antonio Cacciapuoti

And that need of having a reflection on compliance on multiple regulation help the governance behind AI to ensure that the model is strong enough to assess the risk and the safeguards required.

00:14:16 Antonio Cacciapuoti

Because if we think all this regulation was potentially written or drafted even before ChatGPT was publicly released,

00:14:24 Antonio Cacciapuoti

which means that, I mean, usually we say technology is five years forward, the regulation, or regulation is five years behind the technology.

00:14:33 Antonio Cacciapuoti

So if you consider how much is going fast now, the AI development is regulation with the risk to stay a bit behind more and more.

00:14:43 Antonio Cacciapuoti

But at the same time, it helps to ensure that when we build a model,

00:14:49 Antonio Cacciapuoti

we have some reflection on risk and safeguards.

00:14:52 Antonio Cacciapuoti

That's how we see the things.

00:14:54 Antonio Cacciapuoti

And that's where the human still clear, having a very important part, since ethics is very subjective, a judgmental topic.

00:15:01 Antonio Cacciapuoti

So when you build, when each company builds its own safeguards, I mean, it's managed and built by humans.

00:15:10 Antonio Cacciapuoti

So this is not a thing that cannot be delegated to a machine.

00:15:13 Alessandro Casarotti

I will just jump in because the fact that here in Europe, we have a strong

00:15:19 Alessandro Casarotti

So we are maybe more regulated than other countries.

00:15:24 Alessandro Casarotti

For you to represent a barrier for technology development or is just for you an advantage in terms of a more ethical AI compared to, I mean, Asia and the United States?

00:15:39 Antonio Cacciapuoti

For sure, give us some time to recover the development that they may have in USA and

00:15:47 Antonio Cacciapuoti

Asia, because they have potentially less regulatory barrier, but at the same time gives, I think, comfort and ground, mainly when we go on ethical safeguards, because as I mentioned before, the reflection that you need to give and make when you build the model and the governance on risk and barrier and safeguards that you're implementing.

00:16:09 Antonio Cacciapuoti

And since this is mainly done by humans, I mean, it helps really to have this human in the loop, also in the build of the governance, the model, and manage such a subjective, a judgmental topic like ethics that...

00:16:25 Antonio Cacciapuoti

can be handled not by a machine, or at least at the beginning, cannot be dedicated to a machine.

00:16:31 Antonio Cacciapuoti

Then we go in the famous human in the loop, because it's also part of the regulation that is expected here.

00:16:38 Antonio Cacciapuoti

I mean, the human in the loop is sufficient for you, Antonio, to ensure an ethical AI, or it's just a slogan?

00:16:48 Alessandro Casarotti

Alessandro, to be honest with you and of course with all the audience, human in the loop is necessary, but it's not enough.

00:16:59 Alessandro Casarotti

Yes, I have to admit, sometimes it's just a slogan, because having a human in the loop only works if that human has knowledge, has time and authority

00:17:14 Alessandro Casarotti

and of course, more knowledge to operate.

00:17:16 Alessandro Casarotti

But if that person is just clicking approve because the system is very complex or is fast, in that case, the human in the loop is just theory.

00:17:29 Alessandro Casarotti

But I mean, it's just theory and you cannot apply that in practice.

00:17:33 Alessandro Casarotti

From an ethical perspective, here the risk is the, as I said before, is the automation bias.

00:17:41 Alessandro Casarotti

Because when AI results, when AI

00:17:44 Alessandro Casarotti

high outcomes looks too confident, too good, people tend to trust them even when they shouldn't.

00:17:51 Alessandro Casarotti

So in those cases, the human becomes just a blind approval and not a safeguard.

00:17:58 Alessandro Casarotti

And internal audit, in my opinion, should understand if the human can understand why the system produced that kind, that specific output.

00:18:10 Alessandro Casarotti

and especially if they allow to override it without any justification or any penalties.

00:18:17 Alessandro Casarotti

Because for me, of course, ethic AI requires real human in the control, not just a symbolic control.

00:18:25 Alessandro Casarotti

And that for me means that you need to, if I can give you

00:18:30 Alessandro Casarotti

example in what we do in my organization, in the internal audit function, we create a strong training program, we create escalation paths, we create clear decision limit, and especially a good communication with all the assurance providers.

00:18:52 Alessandro Casarotti

Because we need to know if humans

00:18:55 Alessandro Casarotti

really intervene, really go in when something looks wrong.

00:19:00 Alessandro Casarotti

So to conclude, yes, human in the loop, it's a good starting point, but without governance, it's not a control, it's just a checkbox.

00:19:10 Alessandro Casarotti

And as you know, ethics built on checkboxes will never survive with the reality.

00:19:16 Antonio Cacciapuoti

Indeed, indeed.

00:19:17 Antonio Cacciapuoti

You cannot put boxes for something subjective.

00:19:20 Alessandro Casarotti

Of course, you need, as we were saying before, you need to stay one step ahead of a criminal.

00:19:26 Alessandro Casarotti

And for doing that, at least what we are doing is you need to enhance your risk assessment framework.

00:19:35 Alessandro Casarotti

especially for this kind of AI fraud.

00:19:39 Alessandro Casarotti

You need to maybe enhance your detection system.

00:19:43 Alessandro Casarotti

You need to anticipate risk.

00:19:45 Alessandro Casarotti

You need to move faster in order to apply the best control you can.

00:19:52 Alessandro Casarotti

You need to increase training and the employees

00:19:56 Alessandro Casarotti

knowledge about this kind of new risk and also collaborate, as I was saying before, with all the assurance providers, especially IT, compliance, risk management, the other second level control function in order to respond faster to this kind of risk and to put in place a good control.

00:20:18 Alessandro Casarotti

Now, I suggest to stay in the famous real human in the loop.

00:20:23 Alessandro Casarotti

I would like to know how did you implement a real human in the loop process in insensitive domains like financial crime?

00:20:33 Antonio Cacciapuoti

We said that ethics is something subjective and very grey because it's a grey area when there is a red thin line between what it is, ethical and unethical, and it's very subjective.

00:20:44 Antonio Cacciapuoti

And anti-financial crime too is a very grey area because there are some areas that are not well defined, also for law, for regulation, or I mean, as we said, since frauds are always one step ahead, I mean, they're always trying to find

00:21:01 Antonio Cacciapuoti

the right loophole or the new technology to jump in.

00:21:04 Antonio Cacciapuoti

That's the complexities because, I mean, we have some great advantages in using AI for anti-financial crime because it can help you completely transform the activity.

00:21:17 Antonio Cacciapuoti

I mean, if we take one of the most common controls that you have in anti-financial crime, such that your customer,

00:21:25 Antonio Cacciapuoti

So far is always seen as an administrative activity.

00:21:28 Antonio Cacciapuoti

You collect info, you collect documents, you need to check it off, it's more a checklist exercise.

00:21:34 Antonio Cacciapuoti

And the AI, it helps to change completely the perspective because it will help you to automate this part as much as possible or to identify also patterns.

00:21:45 Antonio Cacciapuoti

So I'm to anticipate certain thing that of course if you repeat just the activities of checking documents and info you may lose.

00:21:53 Antonio Cacciapuoti

To really have a human

00:21:55 Antonio Cacciapuoti

in the loop that performs the investigative process.

00:21:58 Antonio Cacciapuoti

But since it's very sensitive, a human involvement is there from the beginning.

00:22:05 Antonio Cacciapuoti

So that's how we implemented, for instance, when we implemented some AI for KYC, human in the loop was at the beginning.

00:22:12 Antonio Cacciapuoti

So because the model required to be tested, trained, and so validated and trained data and scenarios by human to ensure that all the different scenarios can be at least considered.

00:22:24 Antonio Cacciapuoti

But

00:22:25 Antonio Cacciapuoti

Another point is also human group needs to stay after, because as we said, the fraudster are always ahead.

00:22:31 Antonio Cacciapuoti

But one thing is having more in an investigative approach, so to identify patterns, learning, so that human in the group just assess, okay, yeah, this is risky or not risky.

00:22:41 Antonio Cacciapuoti

And it drives already to him with a predefined assessment, but can have the sensibility and creativity that the machine doesn't have to ensure that there was no safeguards that are delayed.

00:22:54 Antonio Cacciapuoti

But there is also a very important starting point there that is the data quality.

00:22:59 Antonio Cacciapuoti

I mean, and that's why a human involvement is the first safeguard there, because at this moment of transition, we are building everything on data that we have.

00:23:10 Antonio Cacciapuoti

But we know very well also that the data quality is essential to train the model.

00:23:14 Antonio Cacciapuoti

And that's why the human needs to be at the very first stage also in the model training, because it depends on which model you train.

00:23:21 Antonio Cacciapuoti

and which data you use to train the model.

00:23:26 Antonio Cacciapuoti

And therefore, the data may also have some biases and you may create biases in the model even without that you really want.

00:23:34 Antonio Cacciapuoti

So that's why the human needs to be set as a first safeguard at the very beginning and also at the end of the process, mainly when we have these

00:23:47 Antonio Cacciapuoti

risky areas such as anti-financial crime.

00:23:49 Antonio Cacciapuoti

Of course, there are potentially some parts that can be fully automated when you see really that there is no risk, low risk, and there is sufficiently confidence in the data also behind, because of course data will improve along the time.

00:24:05 Antonio Cacciapuoti

But as a starting point, cannot be a machine generating everything mainly in such domain.

00:24:12 Antonio Cacciapuoti

So we cannot let the machine decide for ethical points.

00:24:16 Antonio Cacciapuoti

I mean, such a subjective line, machine can help to assess and the human therefore can focus on the really the subjective assessment and able to refocus his activity on the least red thin line that is on ethical perspective.

00:24:34 Antonio Cacciapuoti

and also to ensure that bias or hallucination checked in advance.

00:24:39 Alessandro Casarotti

Yeah, but Alessandro, we, of course, we need to keep in mind that the risk cannot be eliminated.

00:24:45 Alessandro Casarotti

We, our work as control function is just to reduce us, you know, the risk and to bring it within the organizational risk appetite.

00:24:57 Alessandro Casarotti

And here we talked about safeguard and ethical frameworks.

00:25:02 Alessandro Casarotti

So

00:25:04 Alessandro Casarotti

from the other side of this table, how do fraudsters try to bypass those protections to achieve their goals?

00:25:13 Antonio Cacciapuoti

Indeed, that's a very good question, Antonio.

00:25:15 Antonio Cacciapuoti

I mean, so far we always look on maybe on how we implement safeguards, but not how fraudsters try to make the safeguards derailed.

00:25:23 Antonio Cacciapuoti

So as we said, and again, I think we go a bit on the human aspect of ethics.

00:25:30 Antonio Cacciapuoti

the creativity is essential and fraudsters developed.

00:25:34 Antonio Cacciapuoti

And so if you know already how to prompt, you can find very good turnarounds.

00:25:41 Antonio Cacciapuoti

Because if you ask, for instance, to an IM model how to cook books, even those, let's say, with less barriers, but then I think they will block you.

00:25:50 Antonio Cacciapuoti

I mean, there is Dark GPT for sure that if we find help, but let's say all others, and then they should

00:25:59 Antonio Cacciapuoti

block you.

00:25:59 Antonio Cacciapuoti

However, if you ask how I can explain to my audience that is a classroom how to not cook books, then you start to, for instance, look for turnarounds of the ethical safeguards that you may have put in place.

00:26:17 Antonio Cacciapuoti

So that's where really the first

00:26:21 Antonio Cacciapuoti

type of activity that fraudsters do, what is so-called prompt obfuscation.

00:26:27 Antonio Cacciapuoti

So we're trying to rephrase to try to bypass the filters that AI system has, like in the financial institution.

00:26:37 Antonio Cacciapuoti

I mean, they know very well the controls you have to try to circumvent the controls, like with AI, they know very well the safeguards that they have.

00:26:48 Antonio Cacciapuoti

to try to turn around them.

00:26:50 Antonio Cacciapuoti

Then, of course, they can also exploit biases.

00:26:53 Antonio Cacciapuoti

For instance, if they know that AI may have some certain biases, they may act with some profiles.

00:27:02 Antonio Cacciapuoti

that they know that they will be protected due to these biases and they will get further information, or they can create scientific identities to bypass AI because maybe more tolerant to avoid discrimination.

00:27:17 Antonio Cacciapuoti

So that's also the interesting parts, because depending on the different models, there may be

00:27:24 Antonio Cacciapuoti

different approach, but for sure what I see most is that fraudsters have really great creativity.

00:27:31 Antonio Cacciapuoti

They're always one step ahead.

00:27:33 Antonio Cacciapuoti

And that's why, again, being such very thin line ethics and the safeguards, you need

00:27:42 Antonio Cacciapuoti

the humans still in the loop.

00:27:44 Alessandro Casarotti

And for me, I don't see another way, Alexandro.

00:27:47 Antonio Cacciapuoti

Still, at least so far, as I said, also based on anyway, the limited data and knowledge that we have now, because of me, what I see also on how the models were beat so far, I think for sure there are safeguards, but I'm always conscious that

00:28:03 Antonio Cacciapuoti

The data quality is the key first components and how the data quality were input, how the trained model were trained, how the human validated the data quality and the model.

00:28:14 Antonio Cacciapuoti

And so the output generator is the essential part on how the AI is beat and learn.

00:28:20 Antonio Cacciapuoti

So that's why the human aspect is still very relevant, mainly in such a topic like in such gray areas, like ethics, like financial crime, like fraud, where a human interaction creativity still needs to be put in to assess a subjective domain.

00:28:38 Antonio Cacciapuoti

And that's why I think, anyway, human still will be essential even

00:28:44 Antonio Cacciapuoti

in development of AI.

00:28:47 Antonio Cacciapuoti

And maybe also to close, I mean, a final question to you.

00:28:50 Antonio Cacciapuoti

How do you think the internal audit works for us of the future will interact with AI?

00:28:57 Alessandro Casarotti

Well, I love this question and thanks for it.

00:29:00 Alessandro Casarotti

And when we talk about the future of internal audit and the work in general, we also have to talk about the new generations.

00:29:10 Alessandro Casarotti

For me,

00:29:11 Alessandro Casarotti

As I always say, AI is an incredible accelerator, but acceleration without any foundation is just dangerous.

00:29:20 Alessandro Casarotti

And here the real risk is that people stop learning how to do things just because AI will do it anyway.

00:29:30 Alessandro Casarotti

If you don't understand what you are doing, what you are looking for, I mean, you cannot

00:29:35 Alessandro Casarotti

evaluate if the AI output makes sense.

00:29:39 Alessandro Casarotti

I mean, you cannot understand if it's an AI hallucination.

00:29:44 Alessandro Casarotti

And that's how the tool start controlling people and instead of the other way around.

00:29:51 Alessandro Casarotti

No, you were saying before, it's people creating AI, but if we don't know what we are using, it will be the other way around.

00:30:01 Alessandro Casarotti

So just prove me wrong, but you learn to walk before drive a car.

00:30:06 Alessandro Casarotti

So you learn to arithmetic before using a calculator and you learn to read a map maybe before using a GPS.

00:30:16 Alessandro Casarotti

So it's the same thing.

00:30:18 Alessandro Casarotti

And in the audit, you must understand the risk, governance, and you must understand controls and judgment before asking AI to help you assess them.

00:30:30 Alessandro Casarotti

And as the godfather of AI, Geoffrey Hinton said, I'm not afraid of AI, but I'm afraid of human laziness.

00:30:39 Alessandro Casarotti

And I totally agree with that.

00:30:42 Alessandro Casarotti

And because here the key word is use and it's not outsourcing thinking that we have not to outsource

00:30:51 Alessandro Casarotti

thinking that we are not to outsource training and knowledge.

00:30:55 Alessandro Casarotti

And because for me, AI should come after the competence and not instead of it.

00:31:01 Alessandro Casarotti

Otherwise, the next generation will not be improved, will not be augmented by AI, but they will be just depending on it.

00:31:10 Alessandro Casarotti

And the dependence is not the future of internal audit.

00:31:14 Alessandro Casarotti

Judgment is and the knowledge is.

00:31:17 Alessandro Casarotti

So

00:31:17 Alessandro Casarotti

Alessandro, this was a really great conversation about a very hot topic and it's also very hard, I mean, to talk about ethics in this kind of tool, because ethic on AI, it's a very...

00:31:32 Alessandro Casarotti

difficult topic.

00:31:33 Alessandro Casarotti

So thank you very much for this discussion.

00:31:37 Alessandro Casarotti

I really enjoyed and I also want to thank you, the Institute of Internal Auditors, Grazie Miller from my side.

00:31:45 Alessandro Casarotti

And until next time.

00:31:46 Antonio Cacciapuoti

Thank you, Antonio, and thanks for having me again to speak with you.

00:31:50 Antonio Cacciapuoti

Always a pleasure and always a pleasure to be part of this podcast.

00:31:56 Antonio Cacciapuoti

Thank you, everyone.

00:31:57 Alessandro Casarotti

Thank you.

00:31:59 The IIA

Ready to strengthen your audit work with analytics, automation, and AI?

00:32:03 The IIA

Well, join the IIA's AAAI virtual conference on April 7th.

00:32:09 The IIA

You'll hear practical strategies, real-world insights, and earn up to 9 CPE credits, plus a bonus AI nano course included with registration.

00:32:18 The IIA

Learn more at the iia.org.

00:32:22 The IIA

If you like this podcast, please subscribe and rate us.

00:32:25 The IIA

You can subscribe wherever you get your podcasts.

00:32:27 The IIA

You can also catch other episodes on YouTube or at the iia.org.

00:32:31 The IIA

That's T-H-E-I-I-A.org.