From Academic Integrity to Competence Integrity

By Dr Riaan Steenberg

A student submits a polished assignment. The argument is coherent. The sources are plausible. The language is clean. The work meets the formal requirements of the task.

The institution still does not know what the student can do.

That is the problem artificial intelligence has exposed. Universities have spent much of the AI era asking whether submitted work is genuinely the student's work. It is an understandable question, but it is no longer enough. The more important question is whether the institution has sufficient evidence of the student's competence.

Academic integrity is concerned with the honesty of the academic act. Competence integrity is concerned with the validity of the claim that follows from it.

When a university awards a mark, a credit or a qualification, it makes a public claim. It says that this person has demonstrated particular knowledge, skill and judgement to an acceptable standard. If the institution cannot show how it knows that, the problem is larger than cheating. The qualification itself has become weak evidence.

The Finished Product Has Lost Its Meaning

For a long time, assessment relied heavily on the finished product. A student wrote an essay, prepared a report, solved a case study or submitted a piece of code. The lecturer inspected the product and inferred the capability behind it.

That inference was never perfect. Students received help from friends, tutors, editors and the internet. Some memorised without understanding. Some produced strong work under generous conditions and then struggled to reproduce the same quality elsewhere.

Generative AI changes the scale of the problem. It can produce a credible finished product quickly, privately and at low cost. It can improve structure, repair language, generate examples, propose sources, write code and imitate disciplinary forms. The product may be good while the student's understanding remains thin.

This does not mean that every AI-assisted submission is dishonest. A calculator does not invalidate mathematics, and an editor does not invalidate an argument. Tools have always mediated intellectual work. The question is what the assessment was intended to establish and whether the evidence still supports that conclusion.

If the purpose of the task is to assess a student's ability to construct an argument independently, extensive AI generation may invalidate the evidence. If the purpose is to assess how well the student can interrogate, improve and defend an AI-supported analysis, prohibiting AI may make the task less authentic.

The tool is not the standard. The competence is the standard.

Detection Is a Weak Foundation

The instinctive response has been to detect AI use. Institutions have bought detection tools, amended misconduct policies and asked lecturers to identify machine-written prose.

Detection may sometimes produce a useful signal. It cannot carry the weight of the assessment system.

A probability score is not proof of authorship. False positives can harm students, especially those who write in a formal or second-language style. False negatives are inevitable as tools change. Students can edit generated text, mix methods or use AI in ways that leave little visible trace. The institution ends up fighting over the origin of sentences while avoiding the more basic question of what was learned.

The finished artefact is becoming a weaker proxy for the process that produced it.

This is why academic integrity cannot be protected through policing alone. A university may catch more prohibited use and still remain unable to substantiate competence. Conversely, a student may use AI openly and still provide strong evidence that they understand, can perform and can exercise judgement.

The design problem sits upstream of detection.

Assessment Should Produce a Chain of Evidence

Competence is rarely demonstrated in one moment. It appears across a chain of activity.

A student frames a problem. They identify what they know and what they need to find. They make assumptions. They try an approach. They encounter an error. They explain why it failed. They revise the work. They respond to critique. They apply the idea in a different context. They defend the final judgement.

That chain is much harder to fake than a final document. More importantly, it is better evidence of learning.

A competence-integrity model might include:

• an initial problem framing completed under controlled conditions;
• visible drafts and decision points;
• disclosure of tools and assistance used;
• a record of sources, prompts, tests or calculations where relevant;
• short oral explanations or demonstrations;
• feedback followed by revision;
• transfer of the same capability to a new problem;
• direct observation of selected high-risk skills;
• an assessor's recorded judgement against an explicit standard.

Not every assessment needs every component. A first-year factual test and a final-year professional simulation carry different risks. The evidence burden should match the consequence of the claim.

The principle is simple: the more consequential the competence, the stronger and more varied the evidence should be.

Authentic Assessment Is Not Merely More Realistic

The phrase "authentic assessment" is often used to describe tasks that resemble the workplace. That is useful, but realism alone does not solve the evidence problem.

A student can use AI to complete a realistic consulting report just as easily as a conventional essay. A simulated workplace task is only valuable if it reveals the student's reasoning, choices, adaptation and performance.

Authenticity has at least three dimensions.

The task should resemble meaningful practice. The conditions should reveal how the person works. The evidence should justify the competence claim.

This may mean allowing AI because the profession will allow AI. It may also mean creating moments in which the student must work without it because the underlying capability matters. A future accountant may use intelligent tools to analyse transactions, but still needs to recognise when the output conflicts with the evidence. A future teacher may use AI to prepare learning material, but still needs to read a classroom and respond in real time. A future software developer may use an agent to write code, but still needs to diagnose failure and accept responsibility for what is deployed.

The institution should therefore stop asking whether AI is permitted in the abstract. It should specify the role AI may play in this task, the capability that remains attributable to the student, and the evidence required to support that attribution.

Disclosure Must Become Ordinary

Students need a vocabulary for legitimate assistance.

Current rules often force a false choice between declaring that AI was not used and admitting to something that sounds like misconduct. This encourages concealment and leaves lecturers unable to distinguish productive tool use from substitution of thought.

Disclosure should be ordinary, structured and proportionate. A student should be able to state:

• which tool was used;
• what it was used for;
• which parts of the output were accepted, rejected or changed;
• how factual claims were verified;
• what decisions remained the student's own;
• where the use of the tool affected the final result.

This is not a ritual confession. It is part of the evidence trail.

The disclosure itself can reveal competence. A strong student can explain why an AI recommendation was unsuitable, why a source was weak or why the first approach failed. A weak student may present a clean output but remain unable to account for the choices inside it.

The Lecturer's Role Changes

When the finished assignment stops being sufficient evidence, assessment becomes more relational.

The lecturer is no longer only a marker of products. They become an assessor of evidence across time. They decide which moments require observation, which claims require verification, where tool use is appropriate and when a student's explanation is strong enough to support the result.

This is more demanding than running an AI detector. It requires better assessment design, clearer standards and professional judgement. Institutions cannot simply tell lecturers to redesign everything while leaving workloads, class sizes and systems unchanged.

Technology can help. Learning platforms can retain drafts, capture feedback, vary problems, record short explanations and assemble evidence portfolios. AI can generate practice, propose misconceptions and help lecturers inspect patterns. But the system must serve the assessment model. More surveillance is not the same as better evidence.

The institutional question is not which detector to buy. It is which claims the university makes, what evidence each claim requires, and how that evidence can be collected without making education impossible to administer.

Qualifications Are Trust Instruments

A qualification has value because people who were not present during the learning process are willing to trust it.

Employers trust that graduates can perform certain work. Professional bodies trust that specified standards were met. The public trusts that people entering consequential roles have been assessed properly. Students trust that the time and money invested in the qualification will produce a credible signal.

That trust weakens when institutions cannot distinguish the quality of a submitted product from the capability of the person submitting it.

This is not an argument for returning to rows of handwritten examinations as the only legitimate form of assessment. Controlled examinations can establish some things well and others poorly. Nor is it an argument for banning useful tools from education. Graduates should learn to work with the systems they will encounter outside the institution.

It is an argument for evidentiary diversity.

Written work, oral defence, practical demonstration, observed performance, iterative development, peer work and controlled tasks can support one another. No single item needs to carry the entire claim. The institution should build a case for competence, not infer it from one polished artefact.

From Integrity Policy to Evidence Architecture

Most universities already have academic-integrity policies. They define misconduct, prescribe processes and allocate authority. Those policies remain necessary.

They now need an evidence architecture.

An evidence architecture connects learning outcomes to assessment events, permitted tools, required disclosures, verification methods and progression decisions. It identifies which competencies are foundational, which are professionally consequential and which can be demonstrated through AI-supported work. It also defines what happens when the evidence conflicts.

A high mark on a submitted report, a weak oral explanation and no visible development trail should not be averaged into administrative comfort. The conflict is itself evidence. It requires further assessment or a narrower claim.

The same discipline should apply to positive results. Course completion is not competence. Attendance is not competence. Tool use is not competence. A beautiful portfolio is not competence unless the institution can account for how the work demonstrates the standard.

This is where competence integrity becomes a governance issue. Institutions must resist promoting activity into achievement without sufficient evidence.

The Better Question

Artificial intelligence did not create the weakness in assessment. It made the weakness impossible to ignore.

Universities have often relied on convenient proxies: time spent, content covered, tasks submitted and marks accumulated. AI can now satisfy some of those proxies without doing the learning they were meant to represent.

The response should not be a permanent contest between generation and detection. It should be a redesign of the evidence.

For every important assessment, ask:

What competence are we claiming? What evidence would justify that claim? Which part of the performance must belong to the student? Where may tools legitimately extend the student's capability? How will uncertainty or conflicting evidence be handled?

Academic integrity still matters. Honesty still matters. Authorship still matters when authorship is part of the task.

But the public promise of education is larger than honest submission. It is that the person who receives the qualification can do what the qualification says they can do.

That is competence integrity.

Sources

• EDUCAUSE, The Impact of AI on Learning Assessment (2026)
• South Africa, Draft National Artificial Intelligence Policy (2026)
• OECD, AI and Skills (2026)