The Key is Cultivating Professional Skepticism
I have previously blogged about Guidelines for Ethical AI in Higher Education. In this blog I focus attention on the teaching of AI in the classroom. Educators should start with a review of the normative principles an AI system should respect, such as autonomy, beneficence, non-maleficence, justice and explicability.
AI ethics should be guided by three underlying principles that create the foundation of AI systems and its operation.
Responsible AI defines the operating model. It translates those principles into practices such as impact assessments, bias audits, red-teaming, documentation and human review.
AI governance defines the structure. It assigns ownership, policies, controls, approvals and monitoring processes so responsible AI practices are applied consistently.
AI compliance defines the obligation set. It maps AI systems, data uses and operational processes to applicable laws, regulations, standards and internal policies, then produces the evidence needed to show those requirements are being met.
AI ethics gives organizations a principled basis for deciding what “good” means in a specific AI context, then governance and compliance operating practices determine how decisions are enforced. Students should understand the inner workings of AI systems in order to determine how best to interact with AI systems in the decision-making process. AI systems are tools that are reliant on the guidance provided by those who map the systems and use it as the foundation for critical thinking and, ultimately, decision making.
Modern AI-specific Ethical Principles
AI systems draw from the foundational principles above, but they translate them into controls that technical, legal, risk and business teams can apply. The following are principles widely cited in recent writings on AI ethics.
Fairness
Fairness means AI outcomes should not discriminate against people. Teams often evaluate fairness with measures such as demographic parity, equal opportunity, equalized odds and calibration, then decide which metric fits the system’s purpose and risk profile.
Transparency
Transparency means people know when AI is involved, what the system is intended to do, what data or signals it uses and where its limitations are. For generative AI systems, transparency may include disclosure that content is AI-generated, documentation of retrieval sources, model cards, evaluation results and limitations around accuracy or completeness.
Accountability
Accountability means a human or organization is answerable for the system’s behavior. A model can generate an output, but it cannot approve its own deployment, accept regulatory risk, explain an incident to a customer or decide whether a harmful pattern should be remediated. Accountability requires named owners, escalation paths and authority to pause or change the system.
Privacy
Privacy means minimizing personal data use, protecting sensitive attributes, respecting consent and limiting secondary use. In AI systems, privacy may involve data minimization, access controls, retention limits, anonymization, synthetic data or differential privacy techniques that reduce the risk of exposing information about specific people.
Safety and robustness
AI safety and robustness mean systems perform reliably under expected and unexpected conditions. A model should be tested for failure modes, adversarial inputs, drift, unsafe outputs and behavior outside its intended use.
Contestability
Contestability means people affected by AI systems can challenge decisions and seek remedy. This principle matters most when AI influences access to employment, credit, housing, healthcare, education, public benefits or other consequential outcomes. A contestable system gives people a path to correct data, request human review and understand what evidence shaped the outcome.
The National Institute of Standards and Technology (NIST) describes trustworthy AI characteristics that include validity and reliability; safety, security and resilience; accountability and transparency; explainability and interpretability; privacy enhancement; and fairness with harmful bias managed.
Implementing AI in Higher Education
As previously mentioned in my blog, implementing AI in higher education engages various stakeholders, each with unique perspectives and priorities. For instance, students might focus primarily on how AI influences their learning experiences, privacy, and potential employment opportunities. Faculty researchers are likely to focus on how AI can revolutionize their research methods and enhance their findings while also considering the ethical implications of incorporating AI into their work. And a chief academic officer may view AI through the lens of institutional strategy, resource allocation, and the impact on overall educational outcomes.
MacPherson and Gaule, elaborate that “given these diverse perspectives, establishing a shared ethical framework using these guidelines is crucial. This will facilitate a common understanding of AI’s role and promote effective collaboration across the institution. By adopting these principles—either entirely or partially—or by adapting them to their individual contexts, institutional leaders affirm their dedication to the responsible and ethical use of AI for the benefit of all stakeholders. This commitment may manifest in various forms across different institutions, reflecting their distinct missions, values, and community needs. Some may integrate these principles thoroughly into their policies and practices, whereas others may use them as foundational elements for developing customized frameworks. Regardless of the method employed, this commitment is essential for maintaining trust and integrity and for ensuring the well-being and success of students, faculty, staff, and community partners amid rapid technological shifts.”
Professional Skepticism in Accounting
AI poses a significant risk to accounting by eroding professional skepticism, not merely through automation. While firms invest billions in AI for efficiency, the technology automates routine tasks crucial for junior accountants to learn how to identify errors. This could lead to a generation trusting AI output without critical judgment. Professional skepticism, defined as a questioning mind and critical assessment, is traditionally developed through hands-on experience, which AI is now replacing. Studies show less experienced accountants exhibit automation bias, accepting AI’s uncertain or inaccurate outputs. Educators are now implementing AI-powered simulations where the AI can be wrong, forcing students to verify and question. The profession has a narrow window to deliberately teach skepticism before AI-native graduates fill supervisory roles, preventing a critical gap in judgment essential for financial reporting integrity.
The biggest risk AI poses to accounting is not automation. It’s eroding the skill that keeps the system honest. As firms deploy AI to handle reconciliations, classifications and routine testing, they automate the work that once taught junior accountants how to recognize when something is wrong. Without deliberate training, the next generation may trust the machine more than their own judgment.
I asked AI about the risks posed to professional skepticism of using AI systems for decision making. Here is its response: “AI significantly threatens the accounting profession by eroding professional skepticism. As artificial intelligence increasingly automates routine accounting tasks, the essential skill of professional skepticism—critical for ensuring audit reliability—faces decline. This erosion can lead to over-reliance on AI outputs, potentially compromising the quality of audits and financial reporting.” It went on…
Impact on Auditing
- AI enhances efficiency and risk detection but raises concerns about trust and accountability.
- The reliance on AI tools may diminish the human judgment necessary for thorough audits.
Need for Training
- Educators and firms must actively train professionals to maintain skepticism in the face of AI advancements.
- Developing a balanced approach to integrating AI while preserving critical thinking skills is essential.
Professional skepticism is the skill that separates an accountant from an algorithm. And right now, the profession is attempting to automate the very work that was central to practicing that skill.
Concerns About Making Professional Judgments
A study from Stanford and MIT shows that in a controlled experiment with 99 professional accountants, AI assistance improved accuracy by nearly 18 percentage points. However, when the AI flagged its own uncertainty, experienced accountants dug deeper. Less experienced ones didn’t. They accepted uncertain output at face value.
The implications were even more concerning. When the AI offered an unusual classification, participants followed it, even though those suggestions were less accurate. The tool’s authority overrode their own judgment. Researchers call this automation bias, and it’s not new. What is new: 88% of audit professionals surveyed now say AI risks undermining professional judgment. Former PCAOB Acting Chair George Botic warned that while AI can relieve auditors of administrative tasks, “it may also compromise reasoning and judgment,” and he called on auditors to grapple with how to uphold professional skepticism as AI reshapes their work. This affects those who rely on audited financial statements to invest, lend or run a business.
Conclusions
Teaching skepticism in a deliberate, structured way in academic settings can be challenging. More often, skepticism develops through hands-on experience by working through repetitive, detail-oriented tasks that build pattern recognition over time. Activities such as reconciliations, transaction testing and tracing numbers cultivate an instinct for when something doesn’t look correct. While the classroom provides a strong technical foundation, as researchers have discovered, the workplace is where professional judgment and true skepticism develop more fully.
AI is automating the job part. The grunt work that turned junior accountants into seasoned professionals increasingly falls to machines. So, if skepticism was largely learned by doing, what happens when the doing disappears?
Blog posted by Steven Mintz, PhD, professor emeritus from Cal Poly San Luis Obispo, on September 24, 2026. You can contact Dr. Mintz at: smintz@calpoly.edu. Visit Steve’s website to find out more about his activities.