Certification in the Age of AI: Designing Proof of Skill When Answers Are Cheap
When AI can produce passable answers cheaply, certification must prove human capability through multiple forms of evidence, authentic performance, explanation, and transfer.
AI breaks the old shortcut from submitted artifact to inferred human skill
TEQSA’s assessment-reform guidance argues that generative AI intensifies existing assessment challenges because a submitted product may no longer provide trustworthy evidence of the learner’s personal attainment.[1] This is the core certification problem. Essays, reports, code, slide decks, and analyses can all be useful work products, but their quality no longer automatically proves who performed the underlying reasoning. A credential must therefore specify what capability it claims and collect evidence that connects the person to that capability. The answer is not simply more surveillance; it is better assessment design.
Certification should begin with a claim, not an exam format
“Can independently diagnose and fix a production bug” is a capability claim. “Passed a 60-question test” is evidence that may or may not support it.
Trustworthy judgments require multiple, contextualized forms of assessment
TEQSA’s 2025 implementation guidance emphasizes that assuring learning in a generative-AI environment requires multiple, inclusive, contextualized approaches rather than dependence on a single assessment type.[2] A certification can combine a knowledge check, live performance, oral defense, portfolio, AI-assisted task, and transfer challenge. Different evidence channels cover one another’s weaknesses. A candidate who submits excellent code can be asked to explain tradeoffs, modify it under new constraints, and debug a fault introduced after submission.
Triangulation is stronger than trying to make one task AI-proof
General-purpose AI will continue improving. Assessment resilience comes from combining evidence that is difficult to fake consistently, not from endlessly redesigning one take-home assignment.
AI-assisted performance and human learning need to be separated explicitly
The OECD Digital Education Outlook 2026 warns that successful performance with generative AI does not necessarily imply learning, while also recognizing that AI is becoming part of authentic work.[3] Certification should therefore state whether a task measures unaided capability, capability with approved tools, or the ability to supervise AI. All three can be legitimate, but they mean different things. A modern software credential might test core debugging without AI, then separately assess effective use of an AI coding assistant, verification, and responsibility for the final result.
Tool policy should be part of the construct definition
Instead of asking “Is AI allowed?” after the assessment is designed, define which human capability remains important when AI is available and design evidence around that boundary.
Quality agencies are increasingly framing the problem as assurance that assessment shows what was learned
QAA’s 2026 work on GenAI and standards identifies the central question as whether assessment genuinely demonstrates what a student has learned, because the value of qualifications depends on trust in that link.[4] That framing is useful far beyond universities. Professional certification, corporate badges, bootcamps, and online course certificates all face the same problem. A credential should make its evidentiary logic legible to employers and learners: what was demonstrated, under what conditions, with which tools, and how recently.
A badge needs an evidence trail
Digital credentials can expose the kinds of tasks completed and conditions of assessment without publishing private work. This makes the claim more interpretable than a generic completion badge.
AI competency frameworks show that evaluating AI itself is becoming part of skill
UNESCO’s AI competency framework for students emphasizes human-centered judgment, ethical understanding, application, and creation with AI.[5] Future certification therefore should not assume competence means working without AI in every context. In many fields, skill includes knowing when to use AI, how to verify it, how to protect data, and when to reject its output. The challenge is to prevent AI use from obscuring foundational competence while still testing the tool-augmented practice people will actually use.
Authentic performance should include variation so memorized demonstrations cannot carry the entire credential
A candidate can rehearse one polished project or memorize an oral defense. Strong certification changes the case while preserving the competence being tested. Ask the analyst to interpret a new dataset, the manager to handle a different interpersonal conflict, the programmer to modify unfamiliar code, or the designer to explain tradeoffs under a changed constraint. AI can help generate variants, but domain experts should define the invariants that make tasks comparable.
Process evidence can complement outcomes when the process itself is part of expertise
For some capabilities, the route matters. A security analyst should show how evidence was gathered; a researcher should distinguish source claims from inference; an AI user should document prompts, tool outputs, verification, and corrections. Process logs should not become performative surveillance, but selective trace evidence can reveal judgment that a final artifact hides. The candidate can be asked to annotate key decisions and defend them, turning the assessment into a demonstration of reasoning rather than a contest in artifact polish.
The strongest credential describes what the person can still do when the answer is cheap
Generative AI lowers the price of many outputs. That makes certification more valuable when it focuses on capabilities that remain costly: framing the right problem, judging evidence, diagnosing errors, transferring knowledge, making tradeoffs, acting responsibly, and performing under novel conditions. The future credential may contain fewer conventional answers and more demonstrations, defenses, simulations, and tool-mediated tasks. The goal is not to prove the candidate never used AI. It is to prove that there is a reliable human capability underneath the tools.
Works Cited
- 01
- 02
- 03
- 04
- 05
CodeHistory is a living archive. Citations document the evidence used for this edition; later evidence may refine the account.