# AI Ethics in Internal Audit

An episode of [All Things Internal Audit](https://invisiblenetwork.ai/podcast/1758983201), 10 Mar 2026, 33 min.

Audio: https://traffic.libsyn.com/secure/1984d534-f92f-4772-a722-6ac47400e67e/AI_Ethics_V6_MP3.mp3?dest-id=3677289

The Institute of Internal Auditors Presents: All Things Internal Audit Tech In this episode, Antonio Cacciapuoti and Alessandro Casarotti unpack the ethical challenges of artificial intelligence in internal audit and anti-financial crime. 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. • HOST: • Antonio Cacciapuoti, CIA, CAMS Head of Internal Audit, Eurizon Capital Vice President, IIA Luxembourg • GUEST: • Alessandro Casarotti Director, Forensic & Anti-Financial Crime, PwC Luxembourg KEY POINTS: • AI Hallucinations as a Governance Risk (00:00:28 – 00:02:46) • How AI Developers Build Ethical Safeguards (00:02:50 – 00:06:10) • Human Judgment and Ethical Decision-Making in AI (00:06:14 – 00:09:26) • Automation, Accountability, and Ethical Responsibility (00:09:26 – 00:10:49) • Global Approaches to AI Regulation (00:10:49 – 00:14:52) • Why "Human in the Loop" Is Necessary but Not Sufficient (00:14:52 – 00:19:10) • Strengthening Governance and Risk Controls for AI (00:19:10 – 00:21:01) • Using AI in Anti-Financial Crime Investigations (00:21:01 – 00:24:12) • How Fraudsters Attempt to Bypass AI Safeguards (00:24:12 – 00:27:44) • The Future Relationship Between Internal Auditors and AI (00:27:44 – 00:31:17) Visit The IIA's website or YouTube channel for related topics and more. IIA RELATED CONTENT: Interested in this topic? Visit the links below for more resources: • Global Internal Audit Standards • Domain II: Ethics and Professionalism • Knowledge Centers: Artificial Intelligence • Vison 2035 • Become a Certified Internal Auditor (CIA) • 2026 Analytics, Automation and AI Virtual Conference Follow All Things Internal Audit: • Apple Podcasts • Spotify • Libsyn • Deezer

## Transcript

[0:01] The Institute of Internal Auditors presents all
[0:04] things internal audit tech.
[0:06] In this episode, Antonio Kati
[0:08] and Alessandro Kadi unpack the ethical challenges
[0:11] of artificial intelligence in internal audit in
[0:14] anti financianal crime.
[0:16] They discuss AI hallucinations as a risk to be governed,
[0:19] not eliminated, and examine why governance, accountability,
[0:23] and human judgment are central to ethical ai.
[0:28] How do you think internal audit can evaluate a system that
[0:32] by definition, can hallucinate for internal audit?
[0:36] We need to change little bit the question,
[0:38] because the point is not if an AI system can hallucinate
[0:43] because by the sign, of course it can and it will.
[0:48] The real question here is if the organization is ready
[0:52] to live with that, because from
[0:55] an internal audit perspective,
[0:58] AI cannot be assessed in the same way
[1:00] as a deterministic system.
[1:02] So the point is not to expect that the result, uh,
[1:07] will be correct every single time.
[1:09] For me, what really is important is understanding
[1:12] what the organization has put in place to manage
[1:16] and to mitigate situation when the model can be wrong.
[1:20] So in my opinion,
[1:22] the internal auditors should focus on
[1:24] three important things.
[1:26] The first, it's the governance, so who owns the AI
[1:29] and who is accountable for the use of ai.
[1:32] The second is the control around the usage
[1:36] and not the model itself.
[1:38] I mean, that auditors shouldn't try to,
[1:40] to audit the algorithm mal line by line,
[1:42] but they, they should assess if the there are human in the
[1:47] loops, if there are controls,
[1:49] especially in ag risk use cases.
[1:52] And the last but not least, um, the user knowledge
[1:56] and transparencies
[1:57] because it's important to understand if people
[2:02] understand that the system can hallucinate
[2:05] or if the system is perceived as an alterative source.
[2:09] Because here the distinction is very important
[2:12] because it change the risk profile
[2:15] and the type of control we can put around it
[2:19] for internal audit.
[2:20] The hallucination are not a technical failure,
[2:24] but it's just a risk, a normal risk to be mitigated.
[2:28] So if you want to put an ethical point on this, I would say
[2:33] that, um, an ethical AI system, it's not a system
[2:37] that never hallucinates,
[2:39] but it's a system where hallucination are expected
[2:43] and safely sort by the organization.
[2:46] So this is my view on, um, your question, Alessandro.
[2:49] And now I would like to shift the focus
[2:54] to your experience in the financial crime space.
[2:57] And I'm really curious about, uh, your answer.
[3:01] So based on what you have seen so far, of course,
[3:05] how are the main AI providers building controls to make sure
[3:09] that the ion remain ethical
[3:12] and that this tool are not used for fraud?
[3:14] And please, if you have any concrete
[3:18] examples, just give it to us.
[3:20] Remain ethical for big AI players is a big
[3:24] point of concern.
[3:26] Uh, lately I think we had a big statement from, uh, Altman
[3:29] of open ai, but also from, uh, the CEO of, uh,
[3:32] tropic regards, where really they are looking
[3:36] for also ethical regulation that gives a bit of grounds on,
[3:40] uh, how to address, uh, those topics
[3:43] and the development that is taking,
[3:45] because it's going faster and faster
[3:47] because big players have taken different levels of, uh,
[3:53] safeguards on ethical perspective,
[3:56] which can be a combination of technical safeguards,
[3:59] policy frameworks, active monitoring to try to
[4:03] of course limit, uh, the risk also
[4:05] that fraudster can, uh, exploit.
[4:08] But I, I think, at least in my perspective,
[4:10] the most interesting one is from a tropic
[4:13] because they beat in their own model
[4:16] and AI constitution, which
[4:19] evaluates the answer on a model perspective.
[4:22] And I think that's quite interesting.
[4:24] That is a private institution that beats in its own model
[4:30] and AI constitution to evaluate the ethical perspective.
[4:34] And there is no public push
[4:38] for having this as a consensus.
[4:41] Actually, we have also very different big IV system,
[4:44] which has been criticized actually for the opposite,
[4:46] to not have ethical safeguards.
[4:49] So I think as, as a, as a first approach,
[4:52] it is very interesting where Tropic went
[4:55] and the fact that it's a private institution pushing for
[4:58] that and not a public one, even though we know, uh,
[5:01] that regulation, at least in Europe, they are trying
[5:05] to ensure that some safeguards are, are put into place.
[5:09] But then if we look more overall on all, uh,
[5:11] other safeguards, I mean, there are several things
[5:14] that have been built in the model such, I mean,
[5:18] classification on wireless detection, toxicity,
[5:21] bias detection that can run to ensure
[5:25] that the model blocks certain, uh, a full outputs.
[5:30] But that goes a bit also in line with
[5:32] what you were saying before.
[5:34] I mean, and it's how the model was validated
[5:37] because model was validated trained by human,
[5:40] and so the firm was human judgments needs
[5:43] to stay in somehow, and it, the firm was human in the loop.
[5:47] That also the you act is pushing for, I mean,
[5:51] it's important in that gray area
[5:53] and really subjective area, such ethics
[5:56] to have human judgements that is led somehow
[6:01] also controlled by the human in the end, whether
[6:03] with human feedback, whether
[6:06] with some specific testing which can be, uh, also done
[6:11] by prompts that are studied for that.
[6:14] I mean, there is a specific right team testing also
[6:16] that goes, uh, a bit in that direction also to ensure that,
[6:20] uh, social engineering is not used for, uh,
[6:23] fraud models also to phishing, uh,
[6:26] or other, um, potential manipulation.
[6:28] So all in all, I think what is important is
[6:32] that we keep in mind
[6:35] that fraudster are always the one step hired
[6:38] and we mentioned, uh, a bit earlier.
[6:40] I mean, if they know how to prompt,
[6:43] they can always be rewrite the question to move
[6:47] around the ethical safeguards through to looks legit.
[6:50] And that's why human in the loop is important to ensure
[6:55] that the model, uh, blocks any not attempt.
[6:59] So, so we need to act as criminal to pay Yeah, you
[7:03] to stay one step by it.
[7:04] And you know, usually we use to say as a model that you need
[7:08] to think like a criminal to catch a criminal criminals.
[7:12] And maybe we, we will talk later about it,
[7:14] but I think also in your opinion, the human
[7:18] are the main characters.
[7:21] I mean, first human and then machine. Yes.
[7:24] I mean, I mentioned tropic, but tropic
[7:26] and then AI constitution made by humans,
[7:29] that's very interesting.
[7:31] I mean, in the end, the human
[7:33] is the one creating the machine.
[7:36] So being the one that created a machine is also ensure the
[7:39] ethics in the machine because it's the ultimate responsible.
[7:43] I said the human has this creativity that can help
[7:47] to circumvent the safeguards.
[7:49] So that's why frauds that are always one step ahead also of
[7:53] artificial intelligence actually, right?
[7:55] Because if they know how to prompt, they know also how
[7:58] to those Yeah, circumvent those, uh, those safeguards.
[8:02] And if you want to stay a bit on,
[8:04] on this philosophical part, at least, what,
[8:06] what I find interesting is that you need to be able also to,
[8:10] as you mentioned, I mean, you don't need to know how, how
[8:13] to code and to read all the code,
[8:15] but you need to be able to explain a bit the process
[8:18] to ensure you can audit it.
[8:19] So how ethical is for you to, how automate a process
[8:24] if you're not able to explain it?
[8:26] I love philosophy, uh,
[8:28] and here I don't think ethics is just about if we can fully
[8:32] explain a system,
[8:34] but it's more about if we understand the consequences
[8:38] of using that system.
[8:40] Uh, because many processes were never fully explainable
[8:43] even before ai.
[8:45] Think about some example linked
[8:47] to my asset management world, for example,
[8:50] the complex credit model.
[8:52] So the investment decision made by the portfolio managers,
[8:56] we often can't explain every step,
[8:59] but we still have accountability for them.
[9:02] So the ethical problem for me starts when
[9:06] automation remove responsibility.
[9:08] If nobody can explain what the system does, why does it,
[9:13] and especially who is accountable.
[9:16] When, when something goes wrong, then automation becomes
[9:20] equally weak, not because it's sopac, but just
[9:24] because it's not well governed.
[9:26] So if also here, um,
[9:29] because we stay in a internal audit perspective, if we jump
[9:34] in internal audit world, the key question
[9:37] for me is not if we can explain, as I said
[9:41] before, the algorithm line by line,
[9:44] but if we can explain when the system should be used
[9:48] and when not, and especially if we can explain
[9:52] who is responsible for the outcome produced.
[9:55] So if you cannot explain those things, uh, you as company,
[10:00] you shouldn't, uh, start to automate the process.
[10:04] So ethically speaking for me,
[10:06] automation without the possibility to explain
[10:10] what is acceptable, the automation without possibility
[10:14] to explain, not, uh, it's only acceptable
[10:18] when the impact is low.
[10:19] So when the risk is low, as you were saying
[10:22] before, the humans remain in the control
[10:25] and if there are clear control put in place
[10:30] by the organization when things go wrong,
[10:33] because things of course can be wrong.
[10:36] So otherwise, if you combine just opacity
[10:40] with automation, we are not talking about ethics.
[10:43] It's not ethic, it's just risk moving faster,
[10:47] which is not responsible.
[10:49] And that's interesting
[10:50] because what you said, I mean,
[10:52] mainly speaking about the risk perspective, it's a bit
[10:55] where the, at least in Europe, the regulation is trying
[10:58] to give a bit of safeguards when we say, okay,
[11:01] depending on certain type of activities can be defined,
[11:04] whether it is low, medium, high, or very high risk, and
[11:06] therefore certain level
[11:08] of controls is expected in the activities.
[11:11] Yeah, because then the perception of risk can change
[11:15] and, uh, also the control
[11:17] that every organization can put in place
[11:19] to mitigate this risk.
[11:21] But you know, Alessandro,
[11:22] today the world is more connected than ever,
[11:25] but especially in our sector
[11:27] and in yours, I imagine that you often have
[11:31] to navigate local, uh,
[11:33] regulatory barriers from clients operating under different,
[11:38] uh, legislation.
[11:39] So with that in mind, how have AI players
[11:43] approach ethical decision when operating under such a
[11:47] different regulation?
[11:48] For example, if we can mention Europe, Asia,
[11:52] or United States?
[11:54] Yeah, I mean, if we take these three macro regions, I,
[11:57] I think we can see already three quite different approach
[12:01] on, uh, AI and also major regulation.
[12:04] I mean, we mentioned already Europe,
[12:06] and we know that also Europe is somehow also this about over
[12:11] overregulating, but for sure gave, uh, a quite clear ground
[12:16] to build up some governance on what is eu, uh, AI
[12:20] with the UI act, and also
[12:22] therefore to ensure that are some parameters embedded
[12:26] for having safeguards that are broad, uh, forward.
[12:31] On that perspective, it's way different from United States,
[12:35] uh, and Asia, which have more a, a pragmatic approach.
[12:38] Also far constant, uh, development.
[12:41] Because I, if we think on a overall, uh, perspective,
[12:45] I mean, we mentioned so far the I act,
[12:48] but at least here in Europe, any big players needs
[12:52] to ensure multiple regulatory compliance.
[12:56] Not only the U act, U ACT is a foundation,
[12:58] but there is also other consideration to do, for instance,
[13:01] the GDPR regulation.
[13:02] I mean, all the data that you're inputting,
[13:04] where are they going, how they're treated, uh,
[13:06] because also this may have impacts on, uh, how the,
[13:10] the model is learning, and
[13:12] therefore also about also some ethical
[13:15] or bias, uh, discrimination.
[13:18] And if for you in your sector, in the financial sector,
[13:21] you have even other regulation that goes on top.
[13:24] So the DORA one, the cloud one,
[13:26] the outsourcing one all together.
[13:28] I mean, it becomes very complex to put a system in place
[13:33] that really fits all activities.
[13:36] But at the same time, I think it helps
[13:39] to perform a reflection in the development of the model
[13:43] and of the need that AI is required to, uh,
[13:48] answer to ensure that all potential safeguards
[13:51] and risk are actually assessed.
[13:54] Because when you have
[13:55] to consider all these regulatory compliance, I mean,
[13:58] the technologies moving way faster,
[14:00] and that need
[14:02] of having a reflection on compliance,
[14:06] on multiple regulation help the governance behind AI
[14:10] to ensure that the model is strong enough to assess the risk
[14:13] and, uh, the safeguards, uh, required.
[14:16] Because if we think all these regulation was potentially
[14:19] written or drafted even
[14:20] before, uh, JGPT was publicly released, which means that,
[14:25] I mean, usually we say technology is five years forward,
[14:29] the regulation regulation is five years
[14:31] behind the, this technology.
[14:33] So if you consider how much is going fast now,
[14:37] the AI development, this regulation with the risk
[14:40] to stay a bit behind more and more.
[14:43] But at the same time, it helps to ensure that
[14:47] when we build a model, we have some reflection on risk
[14:51] and, and safeguards.
[14:52] That's how, how we see the things.
[14:54] And that's where the human till clear having a very
[14:57] important part since ethics is very
[15:00] subjective, a judgmental topic.
[15:01] So when you build, which when each company builds its own,
[15:05] uh, safeguards, I mean, is managed and build by humans.
[15:10] So this is not a thing
[15:11] that cannot be delegated, uh, to a machine.
[15:13] I'll just jump in
[15:14] because, um, the fact that, uh,
[15:16] here in Europe we have a strong regulation, so, uh,
[15:20] we are maybe more regulated than other countries, um,
[15:24] for you to represent a barrier for, uh, for, um, technology
[15:29] development or is just for you an advantage in terms of, uh,
[15:34] a more ethical AI compared to, I mean, Asia
[15:38] and the United States for sure, give us some time
[15:42] to recover the development that they may have in, uh, USA
[15:46] and, uh, Asia
[15:48] because they have potentially less regulatory, uh, barrier,
[15:53] but at the same time give, I think comfort
[15:56] and ground mainly when we go on, uh, ethical safeguards.
[15:59] Because as I mentioned
[16:00] before, the reflection that you need to give
[16:03] and make when you build the modern, the governance on risk
[16:06] and, uh, barrier and safeguards that you're implementing.
[16:09] And since this is mainly done by humans, I mean,
[16:12] it helps really to have this human in the loop,
[16:15] also in the build of the governance, the model,
[16:18] and manage such a subjective, uh, judgmental topic like, uh,
[16:23] ethics that can be handled not by a machine
[16:27] or at least at the beginning,
[16:28] cannot be delegated to a machine.
[16:31] Then, then we go in the famous human in the loop.
[16:33] Uh, because it's also part of the,
[16:36] the regulation that is expected here.
[16:38] I mean, the human in the loop is sufficient for you, uh,
[16:42] Antonio, to ensure an ethical ai
[16:45] or it's just as slogan well, um, Alessandro to, uh, be, uh,
[16:51] honest with you.
[16:52] And, uh, of course with all the audience, uh,
[16:55] human in the loop is necessary.
[16:57] Uh, but it's not enough.
[16:59] Yes, I have to admit, sometimes it's just as slogan, uh,
[17:03] because having a human in the loop only, uh, works if, um,
[17:08] uh, that human has, uh, knowledge as time and authority
[17:13] and of course more knowledge to operate.
[17:16] But if that person is just clicking approve
[17:20] because the system is very complex
[17:23] or is fast, in that case,
[17:25] the human in the loop is, is just theory.
[17:28] And, um, but I mean, it's just theory.
[17:31] And you cannot apply
[17:32] that in practice from an medical perspective.
[17:35] Uh, here the risk is the, as I said
[17:38] before, is the automation bias,
[17:40] because when AI result, when AI
[17:44] outcomes looks too confident, too good, people tend
[17:48] to trust them even when they shouldn't.
[17:51] So in those cases, the human becomes just a blind approval
[17:56] and not a safeguard.
[17:58] And, um, internal audit, in my opinion, shouldn't,
[18:01] should understand if the human understand why the system
[18:06] produced that kind, that specific output,
[18:10] and especially if they allow
[18:12] to override it without any justification or any penalties.
[18:17] Because for me, course ethic AI
[18:21] requires real human in the control,
[18:23] not just a symbolic control.
[18:25] And that for me means that you need to,
[18:28] if I can give you example in what we do in the,
[18:33] in my organization, in the internal audit function, we,
[18:37] we create a strong, uh, training program.
[18:40] We create escalation path, we create clear
[18:44] decision limit, and, uh, especially a good communication
[18:50] with all the assurance providers
[18:52] because we need to know if humans really intervene,
[18:57] really go in when somethings looks wrong.
[19:00] So to conclude, yes, human in the loop,
[19:03] it's a good starting point,
[19:05] but without governance, it's not a control,
[19:08] it's just a checkbox.
[19:10] And as you know, ethics built on check boxes
[19:14] will never survive with the reality.
[19:16] Indeed, indeed, you cannot put boxes
[19:18] for something subjective.
[19:20] Of course, you need, as we were saying
[19:22] before, you need to stay one step ahead of a criminal.
[19:26] And, um, for doing that, at least what we are doing,
[19:30] it's you need to announce your risk assessment framework,
[19:35] especially for this kind of AI fraud.
[19:38] Uh, you need to maybe, um, announce your detection system.
[19:43] You need to anticipate risk, you need to move faster, uh,
[19:47] in order to apply, uh, the best control you can.
[19:52] Um, you need to increase training
[19:54] and the employee's knowledge about this kind of, uh,
[19:58] new risk and also collaborate, as I was saying
[20:02] before, with all the assurance provider,
[20:04] especially it compliance, risk management,
[20:07] the other second level control function in order
[20:11] to respond faster to the, this kind of risk
[20:14] and to put in place a good control.
[20:18] Now, just to, to stay in the famous real, uh,
[20:21] human in the loop, I would like to, uh, to know
[20:25] how did you implement a real human in the loop process
[20:29] in unsensitive domains like financial crime?
[20:32] Yeah, we said that ethics something, uh, subjective
[20:36] and very gray because it's a gray area.
[20:38] You know, there is a, a red thin line between
[20:41] what it is ethical and analytical, and it's very subjective.
[20:44] And anti financianal crime too is very gray area
[20:48] because there are some areas that are not well defined also
[20:52] for law, for regulation
[20:54] or, uh, I mean, as we said, since fraud are always stay, uh,
[20:57] one step ahead, I mean, they're always trying
[21:00] to find the right, uh, loophole
[21:02] or the new technology to jump in.
[21:04] That's the, the complexities,
[21:06] because I mean, we have some great advantages in using
[21:11] AI for non-financial crime
[21:13] because it can help you completely transform the activity.
[21:17] I mean, if we take one of the, the most common controls
[21:21] that you have in non-financial crime,
[21:23] such the know your customer so far is always seen
[21:26] as a administrative activity.
[21:28] You collect info, you collect documents, you need
[21:31] to check the office more, a checklist exercise,
[21:34] and the ai, it helps to change completely the perspective
[21:38] because it'll help you to automate this part as much
[21:41] as possible or to identify also patterns.
[21:44] Uh, so to anticipate certain thing that of course,
[21:47] if you repeat just the activities of checking documents
[21:51] and info, you may, you may lose to really have a human
[21:55] in the loop that performs the investigative process.
[21:58] But since it's very sensitive, a human involvement
[22:03] is there from the beginning.
[22:05] So that's how we implemented, for instance,
[22:07] when we implemented some AI for KYC,
[22:10] human in the loop was at the beginning.
[22:12] So because the model required to be tested, trained,
[22:15] and so validated and trained data
[22:17] and scenarios by, by, by human to ensure
[22:20] that all the different scenario can be at least considered.
[22:24] But another point is also human group needs to stay after,
[22:28] because as we said, the fraudster are always ahead.
[22:31] But one thing is having more an investigative approach, so
[22:35] to identifying patterns, learning so that human,
[22:37] human group just assess, okay, yeah, this is risky
[22:40] or not risky, and it drives already to him
[22:43] with a predefined assessment, but can have the sensibility
[22:47] and creativity that the machine doesn't have to, uh, ensure
[22:50] that there was no, uh, safeguards that degrade.
[22:54] But there is also a very important starting point there
[22:58] that is the data quality.
[22:59] I mean, and that's why a human involvement is the first,
[23:03] uh, safeguard there.
[23:05] Because at this moment of transition,
[23:07] we are building everything on data that we have,
[23:10] but we know very well, so
[23:11] that the data quality is essential to train the model.
[23:14] And that's why the human needs
[23:16] to be at the very first stage also in the model training,
[23:18] because it depends on which model you train
[23:21] and which data you use to train the model.
[23:26] And therefore the data may also have some biases,
[23:29] and you may create biases in the model even without that,
[23:33] you really want.
[23:34] So that's why the human needs to be set as a first
[23:39] safeguards at the very beginning
[23:41] and also at the end of the process,
[23:44] mainly when we have these, uh, risky areas,
[23:48] such unfinancial crime, of course,
[23:50] there are potentially some parts that can be fully automate,
[23:53] uh, when you see really that is no risk, low risk,
[23:56] and there is sufficiently confidence in the data also
[24:00] behind, because of course, data will improve, uh,
[24:03] around time, but as a starting point, cannot be a machine,
[24:08] uh, generating everything mainly in such domain.
[24:12] So, uh, we cannot let the machine decide
[24:15] for political points.
[24:16] I mean, such a subjective line machine can help to assess,
[24:21] and the human
[24:22] therefore can focus on the really the subjective assessment
[24:26] and able to refocus his activity on the release red thin
[24:30] line that is, uh, on ethical, uh, perspective,
[24:34] and also to ensure that bias
[24:36] or hallucination checked in advance.
[24:39] Yeah, but aand we of course, we, uh, need to keep in mind
[24:43] that the risk cannot be eliminated.
[24:45] We, our work as control function is just to reduce,
[24:51] as you know, the risk
[24:52] and to bring it within the organizational risk appetite.
[24:57] And here, uh, we talked about safeguard
[25:00] and, uh, ethical, uh, frameworks.
[25:02] So from the other side of, uh, this table,
[25:06] how do fraudster try
[25:08] to bypass those protection to achieve their goals?
[25:12] Indeed, that's a very good question, Antonio.
[25:14] I mean, uh, so far we always look on maybe on
[25:17] how we implement, uh, safeguards,
[25:19] but not how fraudster try to make the safeguards derail.
[25:23] So as we said,
[25:25] and again, I think we go a bit on the human aspect
[25:28] of, uh, ethics.
[25:30] The creativity is essential and fraudster developed.
[25:33] And so if you know already how to prompt,
[25:38] you can find very good turnarounds,
[25:40] because if you ask for instance to an I model how
[25:44] to cook books, even those, let's say with less barriers,
[25:49] but then I think they will block you.
[25:50] I mean, there is dark GPT for sure,
[25:52] they will if we find help,
[25:54] but, uh, let's say all, all others,
[25:57] and then they should, uh, block you.
[25:59] However, if you ask how I can explain to, uh,
[26:03] my audience that is a classroom how to not cook books,
[26:08] then you start to, for instance, look for turnarounds
[26:13] of the, uh, ethical safeguards
[26:15] that you may have put in place.
[26:17] So that's where really the, the, the first type of
[26:22] activity that fraudster do really
[26:24] what is so-called prompt obfuscation.
[26:27] So we trying to rephrase, to try to bypass, uh,
[26:32] the filters that AI system has,
[26:35] like in the financial institution.
[26:37] I mean, they know very well definitely the controls you have
[26:40] to try to, uh, circumvent the controls.
[26:44] Like with ai, they know very well the safeguards
[26:46] that they have to try to turn around them.
[26:50] Then of course, they can also exploit biases.
[26:53] For instance, if they know that
[26:56] AI may have some certain biases, they may act
[27:00] with some profiles that they know
[27:03] that they will be protected due to these, uh, biases,
[27:06] and they will get further information,
[27:09] or they can create scientific identities to bypass AI
[27:13] because maybe more tolerant to avoid, uh, discrimination.
[27:17] So that's also the interesting parts
[27:20] because depending on the different models,
[27:23] there may be a different approach.
[27:25] But for sure, what I see most is
[27:28] that fraudster have really great creativity.
[27:31] They're always one step ahead.
[27:33] And that's why, again, being such
[27:37] very thin line ethics
[27:39] and the safeguards, you need the,
[27:42] the humans still in, in the loop.
[27:44] And for me, I don't see another way, Alexandra still,
[27:48] at least so far, as I said, also based on, anyway, the,
[27:51] the limited data knowledge that we have now
[27:53] because of it me, what I see also on
[27:57] how the models were beat so far, I think
[27:59] for sure there are safeguards,
[28:00] but I'm always conscious that the data quality
[28:05] is the key first components
[28:06] and how the data quality were input,
[28:09] how the train model were trained,
[28:11] how the human validated the data quality and the model.
[28:14] And so the who generator is the, the essential part on
[28:17] how the AI is beat and learn.
[28:20] So that,
[28:21] that's why the human aspect still very relevant mainly in
[28:25] such a topic like in such gray areas like ethics,
[28:29] like financial crime, like fraud, where a human interaction,
[28:33] creativity still needs to be put in to assess
[28:37] a subjective domain.
[28:38] And that's why I, I think anyway,
[28:40] human still will be essential even in development
[28:45] of, uh, ai.
[28:46] And maybe also to close, I mean, a final question to you.
[28:50] How do you think the internal audit works force
[28:53] of the future will interact with ai?
[28:57] Well, I, I love this question and thanks for it.
[29:00] And when we, we talk about the future of internal audit
[29:04] and the work in general, we also have to, um,
[29:07] talk about the, the new generations.
[29:09] Um, for me, as I,
[29:12] I always say AI is an incredible accelerator,
[29:16] but acceleration without any foundation is just dangerous.
[29:20] And, um, here the real risk is that people, um,
[29:24] stop learning how to do things just
[29:26] because AI will do it anyway.
[29:29] If you don't understand what you are doing,
[29:32] what you are looking for, I mean, you cannot
[29:35] evaluate if the AI output makes sense.
[29:39] I mean, you cannot understand if it's an AI hallucination.
[29:44] And, uh, that's, uh, how the tool start controlling people.
[29:48] And, uh, instead of the other way around, no, you, you,
[29:51] you were saying before, it's people creating, uh, ai,
[29:55] but if we, uh, don't know what we are using, uh,
[29:58] will be the other way around.
[30:00] Uh, so just prove me wrong.
[30:02] But, um, you learn to walk, uh, before drive a car.
[30:06] So you learn to automatic before using a calculator
[30:11] and, uh, you learn to read a map maybe
[30:14] before using a GPS.
[30:16] So it's the same thing.
[30:18] And in the audit, you must understand the risk governance
[30:22] and you must understand controls and judgment
[30:26] before asking AI to help you assess them.
[30:30] And as the godfather of ai, uh, geography terms said,
[30:35] I'm not afraid of ai, but I am afraid of human ness.
[30:39] And I totally agree with that.
[30:42] And, uh, because here the key word is use,
[30:46] and it's not outsourcing, thinking that we have, we have not
[30:49] to outsource thinking that we are not to outsource
[30:53] training and knowledge.
[30:54] And because for me, AI should come after competence and,
[30:59] and not instead of it otherwise.
[31:02] The next generation will not be, uh, improved,
[31:05] will not be augmented by ai,
[31:07] but they will be just dependent on it.
[31:10] And, uh, dependence is not the future of internal audit
[31:14] judgment is and knowledges.
[31:17] So Alessandro,
[31:18] this was a really great conversation about a very odd topic.
[31:22] And, uh, it's also very hard, I mean,
[31:25] to talk about ethics in this kind of, uh, tool
[31:29] 'cause ethic on ai.
[31:30] It's a very difficult topic.
[31:33] So thank you very much, uh, for, uh, this discussion.
[31:37] Uh, I really enjoyed.
[31:39] And I also want to thank you the, um,
[31:42] internal auditors grad Miller from my side.
[31:44] And until next time, thank you Antonio,
[31:46] and thanks, uh, for having me again, uh, to speak with you.
[31:50] Always a pleasure and always a pleasure to be, uh, part of,
[31:54] uh, this, uh, podcast.
[31:56] Thank you everyone. Thank you.
[31:59] Ready to strengthen your audit work with analytics,
[32:01] automation, and ai.
[32:03] Well join the iias A A A I virtual
[32:07] conference on April 7th.
[32:09] You'll hear practical strategies, real world insights,
[32:12] and earn up to nine CPE credits,
[32:14] plus a bonus AI NAL course included with registration.
[32:18] Learn more@theia.org.
[32:21] If you like this podcast, please subscribe and rate us.
[32:24] You can subscribe wherever you get your podcasts.
[32:27] You can also catch other episodes on YouTube or@theiiaa.org.
[32:31] That's THEI a.org.

Shows are listed from Apple Podcasts' charts and directory, and Podcast Index where it is used. Every show and episode belongs to its publisher and plays from the publisher's own feed; this site hosts none of them.