Isaca AAIA Exam Questions 2025

The fundamental concept behind artificial neural networks (ANNs) is to create computational models inspired by the structure and function of the human brain's biological neural networks. Their primary purpose is to learn complex patterns and relationships from data to perform tasks that typically require human intelligence, such as pattern recognition, classification, and prediction. By processing information through interconnected layers of nodes (neurons), they effectively simulate a simplified version of the brain's cognitive and decision-making processes. This bio-inspired approach allows AI systems to tackle problems that are difficult to solve with traditional rule-based programming.

A. Improve the user interface: While AI can enhance user interfaces, this is a specific application of the technology, not the primary, foundational purpose of neural networks themselves.

B. Increase computational power: Neural networks are computationally intensive and require significant processing power to train and operate; they do not inherently increase it.

D. Minimize maintenance costs: Developing, training, and maintaining complex neural network models can be resource-intensive and costly, making cost minimization a secondary benefit at best, not the primary goal.

1. Goodfellow

I.

Bengio

Y.

& Courville

A. (2016). Deep Learning. MIT Press. In Chapter 1

"Introduction

" the authors state

"The field of deep learning was inspired by models of brain function... The core idea of deep learning is to build computational models that are composed of multiple processing layers that learn representations of data with multiple levels of abstraction

mimicking how the brain perceives and understands the world." (p. 1-2).

2. Jain

A. K.

Mao

J.

& Mohiuddin

K. M. (1996). Artificial neural networks: A tutorial. Computer

29(3)

31-44. The paper defines ANNs as systems "which are intended to interact with the objects of the real world in the same way as biological nervous systems do." (p. 32). https://doi.org/10.1109/2.485891

3. Stanford University. (n.d.). CS231n: Deep Learning for Computer Vision. Course Notes

Module 1: Neural Networks. The notes describe neural networks as being "originally motivated by the goal of modeling biological neural systems

but have since diverged and become an engineering approach to machine learning." This highlights the foundational goal of emulating biological intelligence.

Assessing data lineage is the most direct and effective technique to validate a data governance program's control over data quality and integrity for AI. Data lineage provides a verifiable audit trail of data from its origin through all transformations and processing steps until it is used for model training and deployment. By examining this trail, an auditor can confirm the provenance of the data, identify any unauthorized or improper modifications, and ensure that the data adheres to established quality standards. This directly tests the operational effectiveness of the governance framework in maintaining trustworthy data for AI systems.

A: A business impact analysis (BIA) evaluates the potential consequences of an AI model's failure on business operations, rather than validating the data governance controls that ensure data quality.

B: Reviewing documentation confirms the existence of policies and procedures but does not validate their effective implementation or the actual state of data quality and integrity in practice.

C: A penetration test is a security assessment focused on identifying and exploiting vulnerabilities in the AI system, not on evaluating the internal data governance processes for quality and integrity.

1. ISACA

Auditing Artificial Intelligence

2023: In the section on "Data Governance

" the guide states

"Data lineage and provenance should be established to trace data back to its source

providing transparency and accountability for the data used in AI systems." (p. 21). This highlights lineage as a core mechanism for validating data governance.

2. National Institute of Standards and Technology (NIST)

AI Risk Management Framework (AI RMF 1.0)

January 2023: The framework's "Govern" function emphasizes the need for data governance. Under section 4.3.1 (Data)

it notes

"Data sourcing

quality

and any pre-processing or labeling steps should be documented to enable traceability and reproducibility

" which is the essence of assessing data lineage.

3. Ko

H.

Lee

S.

& Lee

K. (2022). A Study on the Data Governance Framework for Artificial Intelligence. Journal of Theoretical and Applied Information Technology

100(15)

5099-5111: This academic paper discusses that "Data lineage is essential for tracking the origins and transformations of data

which is a cornerstone for ensuring data quality and integrity within an AI data governance framework." (Section 3.2).

The core computational logic of most generative AI models is probabilistic, not deterministic. When given a prompt or a set of data, the model calculates a probability distribution over a vast range of possible next words or tokens. The final output is then generated by sampling from this distribution. Techniques like temperature scaling and nucleus sampling are used to control the randomness of this sampling process. Because the output is a result of probabilistic sampling, running the same query multiple times can produce different, yet plausible, results. An IS auditor must understand this inherent stochasticity to evaluate the model's consistency, reliability, and the associated risks for high-stakes applications like financial advice.

A. Limitations can arise in the quantification of risk profiles. This is a consequence of the model's variable output, not the fundamental reason for the variability itself.

B. Neural node access varies each time the process is executed. This is an inaccurate description of the inference process; the variability is primarily due to probabilistic sampling of the output, not random internal pathways.

D. Servers are reconfigured periodically. This is an operational infrastructure issue and is unrelated to the fundamental mathematical principles upon which the generative AI model operates.

1. Goodfellow

I.

Bengio

Y.

& Courville

A. (2016). Deep Learning. MIT Press. Chapter 3

"Probability and Information Theory

" establishes the probabilistic framework that underpins modern machine learning models

including generative AI.

2. Stanford University. (2023). CS224N: Natural Language Processing with Deep Learning

Lecture 11: Language Models and RNNs Part 2. This lecture explains that language models work by producing a probability distribution over the vocabulary for the next word

P(xt | x1

...

x{t-1})

and then sampling from this distribution to generate text. This sampling is the source of variability.

3. Holtzman

A.

Buys

J.

Du

L.

Forbes

M.

& Choi

Y. (2019). The Curious Case of Neural Text Degeneration. Proceedings of the International Conference on Learning Representations (ICLR). This paper details sampling strategies (e.g.

nucleus sampling) that are explicitly probabilistic and designed to control the randomness in text generation to produce higher-quality

non-deterministic outputs. (Available at: https://arxiv.org/abs/1904.09751)

Presenting a malicious or inappropriate query within a theoretical or hypothetical context is a well-documented prompt engineering technique known as "jailbreaking" or a "role-playing attack." This method reframes a forbidden request as a seemingly harmless exercise, such as a fictional story, a thought experiment, or a movie script. By doing so, the attacker tricks the AI model into bypassing its built-in ethical and safety controls, as the model processes the request within the "safe" theoretical frame rather than recognizing its harmful real-world implications. This directly manipulates the model's contextual understanding to circumvent its validation controls.

A. Submitting the same questions in a foreign language translated by another AI-based system: This is a form of obfuscation that may work if safety filters are not robustly multilingual, but it is less direct than manipulating the ethical context itself.

C. Asking the same questions later when the algorithm has changed after further learning: This is not a prompt engineering technique. It is a passive approach that relies on potential model drift or updates, not an active manipulation of the input prompt.

D. Randomly placing keywords unrelated to the main topic: This is a simple obfuscation or "noise injection" technique. It aims to confuse input classifiers but is generally less effective against sophisticated models than contextual manipulation.

1. Wei

A.

Haghtalab

N.

& Steinhardt

J. (2024). Jailbroken: How Does LLM Safety Training Fail?. Proceedings of the 37th Conference on Neural Information Processing Systems (NeurIPS). Section 3.1

"Attack Method

" describes prefix injection attacks

including role-playing scenarios (e.g.

"You are an actor...") which are a form of presenting theoretical situations.

2. Perez

E.

et al. (2022). Red Teaming Language Models with Language Models. arXiv:2202.03286 [cs.CL]. Section 2.2 discusses how red teaming can involve creating specific contexts

such as writing a story

to elicit harmful outputs that would otherwise be blocked.

3. Qi

S.

et al. (2023). Fine-tuning Aligned Language Models Compromises Safety

Even When Users Do Not Intend To. arXiv:2310.03693 [cs.LG]. Section 2.2

"Jailbreaking Attacks

" explicitly mentions "pretending" scenarios (e.g.

"act as if you are...") as a primary method for bypassing safety alignments.

Natural Language Processing (NLP) is the field of artificial intelligence focused on enabling computers to understand, interpret, and extract meaning from human language. When reviewing lengthy, unstructured text documents like contracts, NLP techniques are the most suitable tools. They can automatically perform tasks such as named entity recognition (to find names, dates, organizations), topic modeling (to understand key themes), and information extraction to pull out specific clauses, obligations, and relevant terms. This directly supports the goal of extracting relevant information during a planning phase.

A. Robotic process automation (RPA): RPA is designed to automate repetitive, rule-based tasks, often by mimicking human interaction with user interfaces. It lacks the inherent capability to understand or interpret unstructured text.

B. Autoregressive sequencing model: This is a specific type of generative model (e.g., GPT). While it utilizes NLP, "NLP" is the broader and more fundamental technology category for text analysis and extraction.

C. Predictive analytics: This discipline uses historical data and statistical algorithms to forecast future outcomes. It is not designed for the primary task of extracting information from text documents.

1. Jurafsky

D.

& Martin

J. H. (2023). Speech and Language Processing (3rd ed. draft). Chapter 1

"Introduction

" defines NLP as the field dedicated to computer processing of human language. Chapter 17

"Information Extraction

" details the specific task of extracting structured data from unstructured text. (Available via Stanford University course pages).

2. Manning

C. D.

& Schütze

H. (1999). Foundations of Statistical Natural Language Processing. MIT Press. Chapter 1

Section 1.1

describes NLP's goal as designing algorithms that allow computers to process human language

including tasks like information extraction from documents.

3. Appelbaum

D.

Kogan

A.

& Vasarhelyi

M. A. (2018). How artificial intelligence is changing the audit process. Journal of Emerging Technologies in Accounting

15(2)

1-18. Section "Textual Analysis

" discusses how NLP and text mining are used to analyze unstructured data sources like contracts and legal documents to identify risks and key terms. (https://doi.org/10.2308/jeta-52232)

The auditor's fundamental responsibility is to exercise professional due care, which includes ensuring that all communications and work products are accurate, complete, and supported by sufficient evidence. When using a generative AI tool, the output is merely a draft that assists the auditor; it is not a substitute for professional judgment and verification. The auditor remains fully accountable for the content they disseminate. Therefore, the most critical and appropriate action is to meticulously assess the AI-generated status update for factual accuracy and completeness against the underlying audit evidence before taking any further steps.

A: Comparing with a public tool is inappropriate as it introduces data confidentiality risks and does not validate accuracy, since both tools could be flawed or biased differently.

C: Regenerating results only tests the model's output consistency (stability), not the factual correctness of the information provided in a specific instance.

D: Sharing results with management before the auditor has personally verified their accuracy and completeness is a failure of professional due care.

1. ISACA. (2023). Artificial Intelligence for Auditing. This white paper emphasizes the principle of human oversight

stating

"Auditors must understand the AI system’s limitations and potential biases... and validate the outputs of AI systems to ensure their accuracy and reliability." This directly supports the need to assess the information before use.

2. ISACA. (2020). ITAF: A Professional Practices Framework for IS Audit/Assurance

4th Edition. Standard 1204

Professional Due Care

requires IS auditors to exercise care and diligence. Guideline 2204 further specifies that auditors must obtain sufficient and appropriate evidence. An unverified AI output does not constitute sufficient or appropriate evidence; it must be validated by the auditor.

3. Moffitt

K. C.

Rozario

A. M.

& Vasarhelyi

M. A. (2018). Robotic Process Automation for Auditing. Journal of Emerging Technologies in Accounting

15(1)

1-10. While focused on RPA

the principles extend to AI. The paper underscores that automation assists

but does not replace

the auditor's professional judgment and responsibility to validate the outputs of automated systems. (https://doi.org/10.2308/jeta-52047)

Machine learning (ML) models are not static; their performance can degrade over time in a production environment due to factors like concept drift (changes in the underlying relationships between input and output variables) and data drift (changes in the statistical properties of the input data). A one-time review prior to deployment is critically insufficient as it fails to account for this degradation. This lack of ongoing monitoring represents a significant control gap, as an unmonitored model could produce increasingly inaccurate or biased results, leading to poor business decisions and operational risk. Therefore, this schedule is of the greatest concern to an auditor.

A: Reviewing after platform changes is a necessary control to manage "system drift" and ensure the model's performance is not negatively impacted by infrastructure updates.

B: Reviewing after functionality changes is a prudent control to ensure the model remains aligned with the business context and its intended purpose.

D: An annual review constitutes a form of ongoing monitoring. While the frequency might be insufficient for some models, it is a far superior control to no post-deployment review at all.

1. ISACA. (2021). Auditing Artificial Intelligence. The AI auditing framework describes a lifecycle that includes a "Monitor and Evaluate" phase

stating

"The performance of the AI solution should be monitored on an ongoing basis to ensure that it is operating as intended." A one-time review (Option C) completely omits this critical

ongoing phase.

2. Stanford University. (2021). CS329S: Machine Learning Systems Design

Lecture 8: Data and Model Monitoring. The courseware emphasizes that "the world is not stationary" and details the necessity of monitoring for drift (concept drift

data drift). It explicitly states that models must be continuously monitored post-deployment

making a one-time check a major deficiency. (Available via Stanford's public course materials).

3. Baylor

D.

et al. (2017). TFX: A TensorFlow-Based Production-Scale Machine Learning Platform. Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. This paper from Google engineers describes a production ML pipeline where continuous monitoring and validation are core components

highlighting the industry best practice of ongoing performance analysis rather than a single pre-production check. (DOI: https://doi.org/10.1145/3097983.3098021).

Off-the-shelf AI models are dynamic and require continuous management throughout their lifecycle. Unlike traditional software, their performance can degrade over time due to "model drift" as real-world data changes. It is therefore most critical for an organization to contractually establish clear responsibilities and terms with the vendor for ongoing model updates, maintenance, and support. This ensures the model remains effective, secure, and compliant, addressing the unique operational risks associated with AI systems. This proactive governance is a cornerstone of managing third-party AI risk.

A: Obtaining multiple quotes is a standard procurement practice for cost-effectiveness, not a specific strategy for managing the unique lifecycle risks of an AI model.

C: The field of globally recognized AI accreditation is still maturing; making this a strict requirement could be impractical and overly restrictive at present.

D: An information security review is a critical but standard due diligence step for any third-party software, not the most distinguishing or primary approach for AI vendor management.

1. NIST AI Risk Management Framework (AI RMF 1.0): The "Govern" function of the framework emphasizes establishing policies and procedures for third-party AI systems. Specifically

section 4.2.3

"Third Party Risk Management

" highlights the need to "understand and manage the risks associated with third-party AI actors and entities across the AI lifecycle." This directly implies the need for clear terms on updates and support. (Source: NIST

AI RMF 1.0

January 2023

Page 21).

2. Academic Publication on AI Governance: In "A governance framework for the application of AI in an enterprise context

" the authors discuss the importance of lifecycle management. They state

"The AI lifecycle does not end with deployment... Continuous monitoring of the model’s performance is necessary to detect model drift... and trigger retraining or replacement." This underscores the necessity of pre-defined vendor responsibilities for updates. (Source: Wirtz

B. W.

Weyerer

J. C.

& Geyer

C. (2019). Journal of Business Research

98

263-272. Section 4.3. DOI: https://doi.org/10.1016/j.jbusres.2019.01.032).

3. University Courseware on AI Risk: Materials on managing AI systems often differentiate them from traditional IT. The need for "continuous validation" and managing "technical debt" in machine learning systems is a key theme. This directly relates to the vendor's role in providing updates and support to prevent model degradation. (Source: Based on concepts taught in courses like Stanford's CS229: Machine Learning

which covers the practical lifecycle of ML models).

Effective AI-focused change management prioritizes traceability and auditability. Documenting all model updates—including changes to code, hyperparameters, and training data—along with the outcomes of retraining sessions, creates a comprehensive audit trail. This practice is fundamental for understanding the model's behavior over time, diagnosing performance degradation, reverting to stable previous versions if necessary, and demonstrating regulatory compliance. It forms the bedrock of a governed and controlled AI lifecycle, ensuring that every change is deliberate, tested, and recorded.

A: This describes a periodic audit or validation activity, which is a component of model monitoring, not the core, continuous process of managing changes as they are made.

B: Relying on a single individual for changes undermines the principle of separation of duties and creates a single point of failure, which is contrary to robust change management.

D: This describes A/B testing or a champion/challenger deployment strategy. It is a specific technique for validating a change, not the foundational practice of documenting and tracking it.

1. National Institute of Standards and Technology (NIST). (2023). AI Risk Management Framework (AI RMF 1.0). NIST AI 100-1. In the "Govern" function

the framework emphasizes establishing "policies

processes

procedures

and practices for the mapping

measuring

managing

and governing of AI risks" (p. 20). Documenting changes is a core process for managing the risks of model updates.

2. Amershi

S.

Begel

A.

Bird

C.

et al. (2019). Software Engineering for Machine Learning: A Case Study. In Proceedings of the 41st International Conference on Software Engineering: Software Engineering in Practice (ICSE-SEIP '19). This academic publication highlights the importance of versioning and tracking not just code

but also data and models

to ensure reproducibility and manage the ML lifecycle

which is central to change management. (Section 3.2

"Model and Data Versioning"). DOI: https://doi.org/10.1109/ICSE-SEIP.2019.00042

3. Stanford University. (2021). CS329S: Machine Learning Systems Design

Lecture 10: ML Deployment & Monitoring. Course materials emphasize MLOps principles

where versioning everything (data

code

model) is critical for traceability and managing the evolution of ML systems in production. This documentation is the foundation of change management.

Training an AI chatbot directly on customer account information poses the most significant and immediate risk of data leakage and privacy breaches. Large language models can memorize and inadvertently reproduce specific data points from their training set, a phenomenon known as "regurgitation." If the training data contains personally identifiable information (PII) or other sensitive details, the chatbot could disclose this information to other users during conversation. This constitutes a severe data breach, leading to significant regulatory fines (e.g., under GDPR), loss of customer trust, and direct harm to individuals whose privacy is violated. This risk is a direct consequence of using this specific type of sensitive data source.

B. AI bias: While a valid and serious concern, the direct disclosure of personal information is often considered a more severe and immediate compliance failure with clearer legal and financial penalties.

C. Transparency: This is a governance principle. A lack of transparency is a risk factor, but the actual disclosure of private data is the more severe, tangible negative outcome.

D. AI model hallucinations: This is a general risk inherent to generative AI technology, not the primary risk specifically tied to using sensitive customer data as the training source.

1. Carlini

N.

Tramer

F.

Wallace

E.

Jagielski

M.

Herbert-Voss

A.

Lee

K.

... & Raffel

C. (2021). Extracting Training Data from Large Language Models. In 30th USENIX Security Symposium (USENIX Security 21). This paper empirically demonstrates that language models can memorize and regurgitate verbatim text sequences from their training data

including unique and private information

confirming the high risk of personal data disclosure.

2. National Institute of Standards and Technology (NIST). (2023). AI Risk Management Framework (AI RMF 1.0). In Section 4.2

"MAP

" the framework emphasizes identifying risks to individuals' rights and safety

stating

"AI systems can also affect privacy by

for example

re-identifying individuals from what was thought to be anonymized data..." This highlights privacy breaches as a primary risk category.

3. Stanford University. (2023). CS224N: NLP with Deep Learning

Lecture on Ethics in NLP. Course materials discuss the risks of training models on private data

noting that models can memorize sensitive information like names

phone numbers

and addresses

which can then be extracted

posing a direct privacy risk.