AI familiarity does not prove that students can think independently

A first-year accounting student submits a polished analysis of an unfamiliar business problem. The structure is clear, the terminology sounds credible and the recommendation is confidently expressed. Yet when asked why one assumption was used, the student cannot explain it.

Was the difficulty weak accounting knowledge, uncritical reliance on AI or simply an inability to articulate the reasoning under pressure? The finished work alone may not tell us.

This is the useful question raised by a QS feature on the first supposedly “AI-native” university cohort. Its contributors discuss students who may have used generative AI before entering higher education and propose responses ranging from clearer course rules to oral examinations and ambitious real-world projects.

The feature is expert commentary, not an empirical study. It includes no direct investigation of students, courses or learning outcomes, and its contributors disagree about whether the AI-native label is justified at all. There is also no accounting-specific evidence. Its proposals should therefore be treated as prompts for course design and local investigation, not proven solutions.

Within those boundaries, the feature makes a distinction worth carrying into accounting education: experience with AI is not the same as expertise, and neither necessarily demonstrates independent competence.

Start by finding out what students can actually do

It is tempting to infer capability from familiarity. A student who uses a chatbot frequently may appear fluent because they can produce prompts quickly and obtain polished responses. That does not show whether they can assess the response’s accuracy, identify missing evidence or recognise when a technically plausible answer rests on the wrong jurisdiction, reporting period or assumptions.

Nor should lecturers assume that all students arrive with similar experience. Some may use several paid tools regularly. Others may have tried only free versions, relied on a shared device or avoided generative AI altogether. Confidence may also conceal substantial differences in how students check outputs.

A short diagnostic activity could provide better local information than the AI-native label. For example, students might receive an AI-generated response to a deliberately bounded accounting question, together with the relevant course materials. They could be asked to identify:

• the assumptions made by the response;
• statements supported by the supplied materials;
• claims that require further evidence;
• calculations that should be checked;
• information missing from the question;
• the conclusion they would retain, revise or reject.

The purpose would not be to catch students out or establish a definitive measure of AI literacy. It would help the lecturer see how students approach verification and where teaching may be needed. Asking students separately about tool access and previous use would also help distinguish technical familiarity from critical evaluation.

Decide which capability each outcome requires

The QS contributors offer different views about how extensively AI should be integrated. Some favour complex projects in which students begin with AI and build more ambitious outputs. Others argue that students still need opportunities to struggle with a problem, develop their own reasoning and write without assistance.

Accounting courses do not need one answer for every task. A more useful starting point is to examine each learning outcome and decide what kind of performance students should demonstrate.

Some outcomes may require unaided capability. Students might need to perform a calculation, explain a basic accounting relationship or construct a coherent argument without AI support. This provides evidence about knowledge and reasoning they can call on independently.

Other outcomes may require AI-assisted capability. In professional practice, graduates may be expected to use digital tools while remaining responsible for the resulting work. Students could use AI to generate possible explanations or approaches, then check the output and document their revisions.

A third category is evaluative capability. Here, the intended skill is judging a machine-generated response. Students might have to locate unsupported assumptions, reconcile calculations, question sources or explain why apparently precise language does not settle a matter requiring professional judgement.

These categories can overlap. A student might first complete a short technical task unaided, then use AI on a related case and finally explain what the tool changed and why. The aim is not to create an elaborate sequence for every assessment. It is to avoid treating a polished AI-supported submission as evidence of every underlying capability.

This curriculum mapping also makes course rules easier to explain. Instead of telling students vaguely to use AI “responsibly”, a module guide can connect permission to purpose:

• AI is excluded here because independent recall or reasoning is being assessed.
• AI is permitted here, but its contribution must be documented and checked.
• AI is expected here because evaluating and improving its output is part of the task.

Programme-level coordination may help identify and reduce contradictory expectations, although the source does not evaluate whether it does so. Consistency should not mean identical rules in every module. An introductory bookkeeping task, an audit judgement case and a data-analysis project may require different forms of evidence.

Make reasoning visible, then evaluate the method

Several contributors to the feature recommend oral examinations, live problem-solving, presentations, simulations and reflective accounts. These formats may help a lecturer question students about assumptions, evidence and decisions that are difficult to observe in a finished submission.

They do not automatically provide valid or fair evidence of understanding. The source does not test whether these approaches improve learning or reliably distinguish independent thinking from AI-supported performance. Larger-scale use would also raise questions about staff time, consistency, accessibility and student anxiety.

A limited pilot is more defensible than wholesale redesign. A lecturer might add a brief, structured explanation to an existing assignment. Students could be asked to describe one important decision, identify the evidence supporting it and respond to a follow-up question. If AI was permitted, they could also explain one output they rejected or revised.

The questions should be linked closely to the stated learning outcomes and applied consistently. Lecturers could then examine whether the added component reveals useful information that was not visible in the written submission. They should also collect feedback on accessibility, workload and whether different students received comparably demanding questions.

Reflection requires similar care. Asking students merely to state that they “used AI for ideas” is unlikely to reveal much. More focused prompts could ask:

1. What did the tool contribute to your work?
2. Which part of its response required the most checking?
3. What evidence did you use to verify or reject that part?
4. What accounting decision remained yours?

Even then, a reflective account is evidence to interpret, not proof that the reported process occurred exactly as described.

Replace the cohort label with better questions

The central risk in calling students AI-native is not simply that the phrase may be inaccurate. It can encourage course teams to design for an imagined uniform student who is technically fluent but critically weak. The feature supplies no evidence that this describes accounting students, and its own contributors offer conflicting views about the depth of incoming students’ experience.

Accounting educators can ask more productive questions. What tools have our students used? Can they detect unsupported claims and missing assumptions? Which outcomes require independent performance? Where should AI use form part of the capability being assessed? Do students have equitable access to the tools a task requires?

These questions will not produce one permanent policy. Tools, professional practice and student experience will continue to change. They can, however, lead to clearer assessment decisions than either assuming that students are AI experts or responding as if none of them has used the technology.

The immediate task is not to redesign an entire accounting programme around a generational label. It is to specify the capability that matters, gather evidence from the students actually enrolled and test modest changes before relying on them.