Rethinking learning outcomes in the age of generative AI

Most of this chapter is aimed at program directors and others responsible for a program's final qualifications: which qualifications can still be assessed reliably, and which need updating because the profession or research domain has itself changed because of AI. If you teach a single course, that curriculum-level material is useful background, but the course-level section is where the question becomes concrete for your own learning outcomes: what students should demonstrate independently, and where they should be coached to use AI productively and responsibly.

#5.1Three levels, two drivers, two lanes

Generative AI is changing both what students need to learn and how we can confirm they have learned it. For learning outcomes, two things are simultaneously true. First, some of what we want students to demonstrate has become much harder to verify because AI can produce competent-looking work that obscures whether the student has actually internalized the skill. Second, the professions and research domains our students are preparing for are themselves changing. They now include productive, responsible use of AI as part of normal practice. Both shifts have consequences for learning outcomes at different levels of abstraction.

This chapter works at three levels. Nothing essential changes at the Dublin Descriptor level (the descriptions of what graduates should be able to do, such as "apply knowledge", "make judgments", "communicate", "continue learning"). Generative AI does not change what it means for a graduate to apply knowledge or to communicate clearly. That goal post is stable. At the program final qualifications level, things do move. Some qualifications need defending against AI-induced erosion of evidence. Others need updating because the profession or research domain itself has changed. At the course level, each learning outcome must be designed with two faces: what students should demonstrate independently without AI, and where they should be coached to use AI productively and responsibly.

This chapter is organized accordingly. It starts with the curriculum-level perspective for program leaders, then moves to the course level, where instructors design specific learning outcomes and their assessments. It closes with worked examples and a practical four-step procedure for program leaders.

#5.1.1Dublin Descriptors: the ceiling that stays put

A useful starting point for program leaders is to recognize that the most abstract level of learning outcomes does not need rewriting. The Dublin Descriptors, the European reference points for what bachelor's and master's graduates should be able to do, describe broad competencies. These cover knowledge and understanding, applying knowledge, making informed judgments, communicating conclusions, and continuing to learn (Joint Quality Initiative, 2004). The emergence of generative AI changes none of these in their general form. A graduate must still be able to apply knowledge in original ways, still make defensible judgments, still communicate clearly.

This stability is itself useful. Program leaders don't need to worry that their entire qualifications architecture is obsolete. The Dublin Descriptors will continue to do what they were designed to do, namely to signal at a high level of abstraction what a higher education diploma stands for, and to provide a reference point for European comparability. Movement happens at the level immediately below.

#5.1.2Program final qualifications: two drivers of change

Program-specific final qualifications, called eindkwalificaties in Dutch, translate the Dublin Descriptors into a discipline. This is where the consequences of generative AI become visible. Two forces push for change here.

The first force is defensive. Some final qualifications were historically assessed through products like essays, code, analyses, and reports. Students can now produce these with AI assistance being undetectable. The question for program leaders is which qualifications still have trustworthy assessment evidence. Where evidence is thin, and AI assistance is untraceable, the qualification needs either rephrasing, redistribution across more assessments, or assessment redesign. A practical instrument here is the step-by-step plan developed by the VU Education Lab, which helps programs systematically identify which assessments are vulnerable and where curriculum coverage of final qualifications is at risk.

The second force is adaptive. Professions and research domains are themselves changing because they incorporate generative AI into normal practice. A junior lawyer is now expected to use AI for drafting and research, while remaining accountable for the accuracy of every output. A research scientist now navigates AI-assisted literature review and AI-suggested code. A graphic designer collaborates with image generators as part of standard workflow. If a university program doesn't update its final qualifications to reflect this, it is preparing students for a profession that no longer exists in that form. Adaptive updates expand the qualifications. They do not change the abstract goal of "applying knowledge in original ways", but they specify that this now includes the responsible and critical use of generative AI as part of disciplinary practice.

These two drivers often point in the same direction. A historically essay-assessed final qualification ("students will construct historical arguments from primary sources") may need both defensive measures (some assessment without AI) and adaptive expansion (some assessment of how the student handles AI-generated historical summaries). They can also point in opposite directions, and recognizing which driver is in play helps avoid confused redesign. Defensive updates are about protecting an existing competence. Adaptive updates are about adding a new one. A program that confuses the two ends up either over-restricting AI in places where the profession demands it, or over-integrating AI in places where independent mastery is the whole point of the qualification.

#5.2Course level: two faces per learning outcome

At the course level, the abstract distinction between defending and adapting, translates into a necessary design choice the instructor makes for each learning outcome. The choice is whether the student should demonstrate the outcome independently, or whether the student should demonstrate productive AI use. This choice isn't new with the rise of AI-use, and has been a challenge for educators with the introduction of earlier tools as well. Such as historically the writing system, computers, the internet, and Google becoming widely available,

was first articulated in the educational technology literature by Salomon (Salomon et al., 1991) as the distinction between:

Effects of a tool:what students learn that persists when the tool is taken away. Or what they know and produce independently.

Effects with a tool: what students can do while the tool is available. Doing so, productively, responsibly, coached and oriented towards the process.

For generative AI, this distinction has direct practical consequences, strongly related to the two-lane approach for assessment.

#5.2.1Effects-of: independent mastery, AI-free assessment

Some competences must be internalized and cannot be delegated to tools or technology. A computer science program needs students who genuinely understand control flow, data structures, and algorithmic reasoning, not students who can prompt their way through a coding task without retaining anything. A law program needs students who have actually digested the foundational principles of their legal system, not students who can paraphrase a generated summary on demand. For these competences, the instructor's design choice is clear. Students should learn the skill in conditions where they can be expected to internalize it, and they should demonstrate the skill in conditions where AI cannot substitute for that internalization.

In practice, effects-of competences are typically assessed under controlled conditions. These include monitored written exams, in-class oral examinations, and observed practical work. A creative alternative used in some computer science programs, is to let students prepare code at home. They can do this with full freedom to use any tools, and then at the final exam you ask them to manually modify part of their code without help. The unsupervised homework serves the learning. The constrained exam serves the verification. The student cannot simply outsource the home assignment because the exam requires them to operate fluently inside their previous work.

#5.2.2Effects-with: productive AI use, coached and process-oriented

Other competences belong to a working life that includes AI as a normal tool. For a graduate, the relevant skill is often not writing Python from scratch with no assistance, but rather: direct, supervise, and verify AI-generated code that does what I want it to do, and recognize when it is wrong. Effects-with learning outcomes specify this directly. The student should learn to use AI productively, critically, and responsibly within the discipline.

Coaching does not mean generic AI literacy here. It is a discipline-specific quality control. A law student preparing case work with AI must learn to fact-check generated case references because fabricated citations have professional and legal consequences. A research methods student using AI for literature review must learn to verify that sources exist, citations are accurate, and important works are not missed. A medical or paramedical student using AI for clinical decision-support must learn where the model is reliable and where it must be overridden. Each discipline shapes its version of effects-with coaching, and assessment correspondingly shifts from the product to the process. The question becomes how the student worked, what the student verified, and how the student justified accepting or rejecting AI suggestions.

#5.2.3Hybrid models: AI in preparation, restricted in performance

Not every learning outcome falls cleanly into effects-of or effects-with. A common third pattern, particularly visible in law and other case-based disciplines, is a scaffolded hybrid. Students may use AI to prepare for a contact moment, but during the contact moment itself (a working group, an oral defense, a clinical role-play) they must perform without AI. This pattern is pedagogically powerful because it keeps the cognitive workload of preparation realistic for modern professional practice, while preserving a verification moment where the student must articulate, defend, and apply the prepared material in real time.

Hybrid models also make something explicit that pure effects-of or effects-with descriptions can obscure. Given the two-lane approach, the lane placement for a single learning outcome can vary across phases of the curriculum. In a law program, foundational legal principles are typically learned and assessed effects-of in the first year. They may be taught in a scaffolded hybrid during case-based working groups. They shift toward effects-with later, as students learn to coach themselves on AI-assisted legal practice. The lane is not a static label glued to a learning outcome. It is a developmental choice that can evolve with the student.

#5.2.4Handoff to the assessment chapter

The distinction we make here at the learning-outcomes level, between what students should demonstrate independently and where productive AI use should be coached, directly anticipates the Two-lane approach to assessment design developed in the next chapter. By the time program leaders and instructors arrive at assessment design, the underlying decision should already have been made. An effects-of outcome belongs in the AI-free lane. An effects-with outcome belongs in the AI-integrated lane. A hybrid outcome assesses in stages. The Two-lane approach is the assessment-side mirror of the design choice this chapter makes at the learning-outcomes side.

#5.3Examples of revised learning outcomes

The framework is best understood through examples that show how the same skill area can land in different lanes depending on program context, and how a single program can place a skill in different lanes across its phases. The illustrations below are not prescriptive. Each program should derive its learning outcomes through the procedure described in the next section. The illustrations below are intended to show the framework in action.

#5.3.1Same skill, different lane: example coding

Coding is a learning outcome in both computer science and theoretical physics programs, but its role differs, and the lane choice follows from that role. The illustration below shows a possible approach for each program, based on conversations with VU Amsterdam colleagues. It is not a description of documented practice in a specific program, but a worked-out example of how the framework could be applied.

In a computer science program, coding is a core competence. The field itself is, in significant part, about building software with disciplined attention to structure, correctness, and efficiency. Students must internalize the basic concepts of programming and be able to demonstrate this mastery without AI assistance. The lane is firmly effects-of. A possible assessment design lets students do home assignments with full freedom to use any tools, then at a final exam asks them to manually modify part of their code without assistance. The homework serves the learning, and the exam serves the verification. In a theoretical physics program, coding is a supporting skill. The actual learning outcome is the ability to apply computational methods to solve physics problems, not the depth of programming expertise across many languages. Here the lane shifts toward effects-with. Earlier phases of the program can already integrate AI-assisted coding into the curriculum, with the emphasis on whether the resulting computation is appropriate to the physics problem rather than on whether the student could have written the code unaided. From this perspective, demonstrating coding without AI matters less than demonstrating that the student can direct AI to produce code that correctly implements the intended physical model.

Same skill, two programs, two lanes. The lane placement could be driven, by what the skill is for, in the qualification architecture of the program.

#5.3.2Same program, evolving lane: example law

A complementary illustration comes from looking inside a single program. The case below also stems from conversations with VU Amsterdam colleagues rather than from documented practice; it shows how the framework could be applied across the phases of a single program. In a law program, the lane assignment for a single skill area, namely legal reasoning supported by sources, could shift as students progress through the curriculum. Early on, students could be expected to learn the basic principles of their legal system and demonstrate this mastery without generative AI, with assessment conducted under controlled conditions; here the lane is effects-of. In case-based working groups, AI could productively support preparation while remaining excluded from the group session itself, a hybrid pattern where the contact moment verifies that the prepared material has been internalized. In later specialization phases, the lane could shift toward effects-with, with students coached in the productive use of generative AI per legal domain. The emphasis then falls on the quality-control habits that distinguish professional from amateur AI use: fact-checking generated references, verifying that cited case law exists and applies, and justifying each AI-assisted output before it is released as professional work. Assessment in this phase could concern process at least as much as product, examining how the student curated, verified, and justified the AI-supported work.

#5.3.3Discipline-specific illustrations

Beyond these two extended cases, every discipline can incorporate concrete learning outcomes that operationalize the framework. The illustrations below are organized by the kind of capability they target.

Evaluating AI output within the discipline

Most fields can incorporate the skill of evaluating AI output as a discipline-specific learning outcome. A first-year writing course might add a learning outcome: that students will identify factual errors, logical inconsistencies, and inappropriate sources in AI-generated text. With assessment that provides AI-generated essays containing deliberate errors for correction. An introductory psychology course could add that students will explain how AI training data introduces biases into system outputs, assessed by having students generate AI descriptions of demographic groups and analyze the outputs for stereotypes. Parallel formulations work in:

Statistics - evaluating AI-generated statistical interpretations against underlying data.

Journalism - verifying facts in AI-generated drafts via primary sources.

Mechanical engineering - validating AI-generated stress analysis by checking boundary conditions.

Molecular biology - evaluating AI-suggested primer sequences against established criteria.

Economics - assessing whether AI-selected econometric models meet underlying assumptions.

History - identifying limitations in AI-generated historical summaries through primary-source evidence.

Learning with AI as a study tool

Some outcomes use AI as a learning tool rather than as a target of evaluation. An introductory chemistry course might include that students will use AI-generated explanations to identify gaps in their own understanding. Students request AI explanations after studying a topic, compare them to their own knowledge, and identify confusion areas to seek targeted help. A calculus course could add that students will generate practice problems using AI and verify solution strategies independently. Students create additional textbook-style problems, attempt solutions, and then use AI to check their work while focusing on differences between their approach and the correct solution.

Professional practice

Many graduates will use AI extensively in their work. An accounting course might revise a learning outcome like "students will prepare financial statements" to "students will review AI-generated financial statements for compliance with accounting standards and calculation accuracy". A data science course could change the learning outcome "students will write data analysis code" to "students will verify that AI-generated code correctly implements intended statistical procedures". An experimental psychology course could include using AI to critique experimental designs and then evaluating whether the suggested confounds are methodologically sound.

Research practice

A research methods course might add a learning outcome that students will conduct literature reviews using AI assistance while ensuring comprehensive coverage and identifying citation errors. Students use AI to surface candidate literature, then verify that papers exist, that citations are accurate, and that important works are not missed. A scientific writing course could add that students will ensure that an AI-assisted manuscript text accurately represents research findings and appropriately credits sources. Students may use AI for drafting methods or improving clarity, and then verify technical accuracy and proper attribution.

Personal and civic life

AI use extends beyond professional practice. A mentoring course might add a learning outcome that students will identify when AI-generated health information requires professional medical consultation. A consumer economics course might add that students will identify when AI-generated financial advice requires professional consultation. An information literacy course could incorporate that students will verify AI-generated information about civic issues by consulting authoritative sources.

#5.4How to derive these for your program

The examples above are illustrations, not templates. To derive these for your program and revise learning outcomes you can follow the four-step procedure below.

1. Map the final qualifications against the two drivers. For each final qualification, ask both questions:

-The defensive question is: where is the current evidence used to certify this qualification thinner than is comfortable under realistic AI use?

The adaptive question is: is the profession or research domain this qualification points to changing because of AI? And if so: is the qualification still describing what graduates need?

Both forces can affect the same qualification, and recognizing which one (or both) is at play helps the program avoid confused redesign.

2. Decide the lane for each affected learning outcome. Move down from the final qualification to the underlying learning outcomes across courses. For each one, decide whether it is effects-of (must be demonstrated independently), effects-with (must be demonstrated through productive AI use), or hybrid (AI in preparation, restricted in performance). The same skill area can occupy different lanes across phases of the program. The same skill can also occupy different lanes in different programs, depending on its role in the qualification architecture.

3. Align learning activities and assessments with the lane decision. Effects-of outcomes need controlled assessment conditions, although their learning activities can still allow AI use as a private study aid. Effects-with outcomes require process-oriented assessment, with explicit coaching on the discipline-specific quality-control habits that make AI use professional rather than amateur. Hybrid outcomes need a clear structural boundary between the AI-permitted preparation phase and the AI-restricted performance phase, and students must understand which phase they are in.

4. Verify coherence at the program level. Once individual outcomes are placed, check that the program as a whole still covers each final qualification through enough independent and AI-restricted assessments to defend the diploma. A program dominated by effects-with assessments may produce graduates who can supervise AI but cannot perform without it. A program dominated by effects-of assessments may produce graduates who can perform without AI but have not learned to supervise it. The balance is itself a curriculum-level design choice.