Artificial intelligence in accounting assessment: Seeking a balance
by Dr Neil Dunne, Trinity Business School, Ireland
Artificial intelligence (AI) has exploded in recent years, in both its power and its accessibility. Like many accounting academics, I have grappled with its impact on assessment. If we assume that our in-person exams required for exemption-bearing modules are shielded from AI’s effects, then our dilemma mainly concerns ‘take-home’ assessment. Should we allow students to use AI in such assessment, and if so, how?
Drawing on my own experience as an accounting academic, thus mobilizing an autoethnographic perspective (Dunne, 2025), I submit that confusion and tension dominate this matter. Specifically, I observe two conflicting and, in my estimation, unsatisfactory stances, namely (1) an outright prohibition on using AI in assessment; and (2) a laissez-faire approach, allowing unhindered use of AI. I will first detail these two perspectives, setting out my concerns with each. I will then propose a middle ground that I have found to work well in practice.
Approach 1: Outright prohibition
This approach potentially supports assessment of learning outcomes associated with knowledge accumulation and critical thinking. In theory, if a student is unencumbered by the imperfect contribution of AI, then educators can more comprehensively identify and assess assurance of learning. However, banning AI also creates a policing problem; to ensure compliance, busy academics will have to rely on subjective judgment and/or flawed detection tools (Hadra et al., 2026). Furthermore, even if all students faithfully refrained from using AI, one might problematize the authenticity of such an approach; is the assessment preparing students for a ‘real world’ accounting context whereby they will need AI for work-related tasks (Abbas, 2026)?
Approach 2: Laissez faire
In allowing unhindered use of AI, this approach mirrors the ‘real world’ and liberates academics from tedious ‘policing’ roles. However, although we may intuit that students would prefer this approach, they can also find this ‘freedom’ overwhelming (Klimova and Pikhart, 2025) and worry that the assignment becomes less a test of accounting skills (which, after all, is what they have signed up for) and more about AI proficiency. This liberal approach may also result in learning outcomes surrounding critical thinking and technical knowledge going unassessed. More broadly, over-reliance on AI deprives students of the inherent pleasure associated with reading, writing, and learning independently (Hasan, 2025).
Seeking a balance
In September 2025, cognizant of the two above approaches and not entirely happy with either, I modified the group project for students on my MBA module, ‘Accounting and Financial Management’. In previous years, I had required students to conduct a ratio analysis on a given annual report. I had taken a somewhat ‘liberal’ approach to student use of AI, reasoning that the nature of the ‘deliverable’ (a recorded video presentation, rather than an ‘essay’), and the fact that I required students to explicitly connect their analysis to the lecture material, would mitigate against over-reliance on AI. Although this assessment generated lively presentations and a reasonable mark distribution, class reps subsequently informed me that they were never entirely sure what use of AI was acceptable and what wasn’t. If there’s one thing that accountants don’t like, it’s uncertainty, and I knew that in 2025 I would need to take a more definitive approach regarding AI use in the assessment. To do this, I explicitly advised in the 2025 assignment where AI could (laissez faire) and could not (prohibition) be used, as shown below:
1. Upload the annual report to a generative AI of your choice (e.g., ChatGPT) and prompt it to analyse the financial health of the company in 1,000 words. Please keep your prompt to one sentence and keep a record of it.
2. Then, critique the summary provided by the AI. Specifically, what does it include/exclude, relative to the ratio analysis we discuss in Session 3? (see more details below for what a critique might look like).
3. Deliverable = PowerPoint presentation (max 15 minutes/12 slides).
As can be seen, Part 1 explicitly permitted (required!) students to use AI to analyse the company. However, students had not previously been using AI in-class, and I wanted to ensure that the accounting focus of the module did not become diluted by a parallel need for AI mastery. I thus kept the AI requirement in Part 1 non-technical, merely asking students to devise and apply a prompt. Part 2 required a critique of the AI summary produced in Part 1. I hoped that, in ‘isolating’ the AI aspect of the project to Part 1, students would approach the Part 2 critique in good faith and provide an independent and self-authored critique of the AI analysis, deploying the ratio material that they learnt in class. To increase the likelihood of this desired outcome (and the chances of detecting its breach), I required students to present their findings (a recorded Zoom presentation), and allocated 25% of marks to ‘engagement’, which I defined for students as the “ability to capture audience's attention in a structured, creative and compelling manner. It is optimal to speak naturally rather than read from slides. Stick to the slide/time limits, and the slides should be as professionally presented as possible.” Merely reading from slides would not score well; rather, I wanted students to talk ‘around’ the slides, in a way that evidenced (their own) comfort and mastery of the material being presented. I discussed the assignment with my Teaching Assistant and then released it on Canvas, keen (anxious!) to see how the 32 students (split into eight groups) would approach it.
Student approach
For Part 1 (where students asked AI to analyse the annual report), the most popular AI used was ChatGPT (three groups) followed by Google Gemini (two groups). Microsoft Copilot and Perplexity Pro were each used by one group, whilst one group did not disclose the AI used. The prompts used were relatively homogenous, although some were more specific than others. For instance, whilst one group asked the AI to “analyse the financial health of this company in 1,000 words”, another utilized a slightly longer prompt, namely “acting as a financial analyst, summarize the financial health of [the company] in no more than 1,000 words referencing only the attached report.”
Groups approached Part 2 (critique of the AI analysis) in a more heterogeneous manner. One group declared that the “AI assessment of ‘strong liquidity position’ is misleading”, basing their verdict on the company’s low current and quick ratios, whilst another concluded that the AI provided “no mention of capital invested, equity, or efficiency in the profitability summary.” Students thus judged the AI misleading and incomplete, although not technically ‘incorrect’. However, groups also raised some noteworthy benefits of the AI analysis relative to the ratio analysis. In particular, AI “quickly simplifies complex concepts for non-financial audiences”, “connects the numbers to the story behind them”, and explained the “why” of certain numbers. In other words, groups identified that AI provided speed, simplicity and context.
The video presentations in general were enjoyable and well presented. Most groups ‘engaged’ as required, presenting the material in a natural and convincing manner, and were transparent in describing their approach. I graded all eight, returned the grades with written feedback, and then, once student queries on marks had been addressed (only one group), I took time to reflect on the experience.
Post-experience reflections
How do we contemplate the ‘success’ of a particular assessment? If we take student ‘enjoyment’ as a metric, one positive indicator was the rise in score (out of five) for the evaluation question ‘how would you rate this module overall?’ from 4.5 (N = 18) in 2024 to 4.6 (N = 20) in 2025. Of course, this is a composite mark that takes into account more than just assessment. Of closer concern to the assessment, and its utility to the ‘real world’, may be the response to the question ‘rate your ability to apply your learning in a work or other practice situation.’ The responses here increased from 3.7 to 4.4, tentatively indicating that the assessment, which accounts for 70% of their overall mark, bears a meaningful connection to their work context. Given that this is a post-experiential program for MBA students that are generally in employment, this development is encouraging. Students value clarity, and explicitly stating where AI could (i.e., Part 1) and could not (i.e., Part 2) seems to have been well-received. It is also worth mentioning that I genuinely enjoyed grading the assessment. Almost twenty years lecturing tells me that my experience of a module is generally positively correlated with that of students’, and I am optimistic that my own experience was ‘contagious’.
Implications for accounting education
I perceive the new assessment to have managed the balance between ‘prohibition’ and ‘laissez faire’ in a way that mobilized the best parts of each approach, i.e., latitude was granted to students regarding AI, but in a controlled manner that provided specificity and certainty. I would urge fellow accounting educators to find an approach that balances ‘real world’ tasks, such as the use of AI and presentations, with opportunities for critical thinking (for example, the requirement I introduced for students to critique the AI finding with reference to the lecture material). This ‘hybrid’ approach replicates the professional contexts that students will find themselves in, where blending AI and critical thinking will be the norm.
Of course, this is only one possible approach. These are still early days in AI, and no one individual has all the answers. However, I would advise readers to approach with care the false binary of ‘prohibition’ versus ‘laissez faire’. There are other paths, one of which I have set out in this article. More broadly, and regardless of which path you take, it seems to me that if we follow an overarching principle of balancing the needs of the workplace with a requirement for critical thinking, we will serve our students well. AI is here to stay, but it must support rather than replace our duty of care to students.
Continue the conversation: https://www.linkedin.com/in/neil-dunne-aa245419/
Email: nedunne@tcd.ie