Assessing Knowledge in the ChatGPT Era: What Instructors Need to Change

Only a few years ago, instructors could reasonably assume that students had independently done much of the work when submitting essays, solving problems, or preparing analytical reports. Generative artificial intelligence changed that logic. Today, well-structured text, code, presentations, and even reasoned analysis can be created in minutes.

This does not mean traditional assessment has become meaningless. It has, however, become unreliable where only the final product is assessed, rather than the path to it. If a system can produce an output resembling student work, instructors need to define more precisely what they want to measure: memory, understanding, application of knowledge, reasoning, creativity, professional decisions, or the ability to explain one's choices.

By 2026, this had become a systemic issue. EDUCAUSE surveyed 438 higher education instructors and staff directly involved in designing or delivering assessment. A key finding was that instructors and students increasingly use AI, but rules remain inconsistent and participants lack clear guidance on when AI is acceptable and when it is not.

The main question today is therefore not 'how do we ban ChatGPT?' but 'how do we build assessment that once again reveals people's real knowledge and skills?'

Why ChatGPT Changed the Logic of Assessment

Conventional systems often rely on an artifact: an essay, homework, a report, presentation, test, or code. Instructors assess the product and infer the student's level from it.

Generative AI disrupts the direct link between product quality and the author's competence. A struggling student can obtain a strong text with a well-formulated prompt. Conversely, a strong student may produce shorter or less polished work while understanding the material much better.

In Digital Education Outlook 2026, the OECD highlights this distinction between performance and learning. Generative AI can improve task performance without learning necessarily taking place. In some studies, AI users' advantage disappeared when access to the tool was removed for a subsequent assessment.

This is a fundamental signal for assessment. Measuring only final output quality may assess the quality of a student–AI combination, rather than the student's knowledge.

Modern assessment should therefore gather more evidence of learning: not only 'what was created,' but also 'how was it created,' 'why was this approach chosen,' and 'can the student independently repeat the key action?'

Not Every Assignment Needs Saving

The initial response to ChatGPT often sought to protect old assignments: ban AI, check texts with detectors, or return to handwritten work.

Yet AI raised a more useful question: did every old assignment measure learning well in the first place?

If students can earn high grades by retelling a few sources, the problem may extend beyond ChatGPT. Such tasks may already have measured rapid information gathering better than deep understanding.

This idea appears in EDUCAUSE's 2026 discussions: AI has exposed weaknesses in assessment models that for years relied on scalable artifacts—essays, tests, and standard problems—as indirect signals of knowledge.

Before defending a particular format, instructors should ask which competence the task should reveal. If the answer is unclear, the task probably needs redesign regardless of AI.

Start With the Learning Outcome

Strong assessment in the AI era begins with a clear learning outcome, rather than the question of whether to allow ChatGPT.

For example, a short test can remain appropriate for checking key terminology. If the aim is to assess the ability to argue a position, a finished essay alone is no longer enough. If a course teaches professional decision-making, students should demonstrate that decision-making process.

Breaking competence into components helps.

What should students know without external assistance?

What should they be able to do with access to tools?

What should they be able to verify in AI output?

What should they take personal responsibility for?

This distinction replaces an abstract battle against technology with different assessment modes for different objectives.

Assess the Process as Well as the Result

One of the strongest changes is moving from one final file to a sequence of stages.

Instead of assigning 'a 2,000-word essay,' you can assess the research question, initial thesis, source selection, argument structure, draft, final text, and brief reflection.

This does not mean instructors must award seven separate grades. Some stages can serve as checkpoints showing how the work develops.

The approach has two advantages. First, it reveals more about students' thinking. Second, it reduces the value of an entirely automatically generated final product.

If students must explain why they changed a thesis after examining sources, show which arguments they rejected, and defend the final decision, simply copying an AI answer is no longer enough.

In a digital course, short intermediate tasks can record this process. In WeStudy, for example, dividing a major assessment into several checkpoints makes sense, revealing both the final score and the trajectory of performance.

Oral Defenses Restore Authorship

A short conversation often reveals understanding better than several pages of polished text.

Instructors can ask students to explain the main conclusion, identify the weakest point in their work, justify a source choice, or apply the same principle to a new situation.

This does not require a full hour-long oral exam. Sometimes five minutes suffice.

Particularly useful questions cannot simply be memorized from the text: 'Why did you reject the first option?' 'Which argument was hardest?' 'What changes if one condition in the case differs?'

In its 2026 assessment discussion, EDUCAUSE highlights dialogic assessment: dialogue about how work was created and why particular decisions were made.

This format is harder to organize for large groups, but even selective oral defenses substantially improve assessment reliability.

Authentic Tasks Become More Important

The closer a task is to a real professional situation, the harder it is to reduce it to generating a template answer.

Give future marketers data from a specific campaign and ask them to justify a decision. Offer engineers several design options with constraints. Ask managers to analyze a conflict with incomplete information, or instructors to design part of a lesson for a particular audience.

AI can help with these tasks too, but students must evaluate alternatives, work with context, and take responsibility for their choices.

Strong authentic assessment often permits several answers. What matters is demonstrating reasoning, criteria, and decision quality, rather than guessing the 'right' answer.

This makes it more resilient to generative AI: a system can suggest options, but students must show why a particular option suits the context.

ChatGPT Can Become Part of the Task

One particularly interesting approach makes working with AI the subject of assessment, rather than keeping it outside assessment.

For example, students receive ChatGPT's answer to a professional question and must identify errors, verify sources, and produce a better version.

Another option is comparing two AI answers and explaining which is stronger against defined criteria.

You can also ask students to use AI for brainstorming and then show which ideas they rejected and why.

These tasks assess the ability to work professionally alongside AI, rather than the ability to exist in a world without it.

This particularly matters in subjects whose graduates will almost certainly use AI at work. A complete ban throughout a course can create an environment as artificial as having no restrictions at all.

Foundational Skills Still Need Assessment Without AI

Integrating AI does not mean every assessment should permit it.

People need foundational competence to oversee automated outputs. Doctors must understand fundamental principles even when using decision support systems. Programmers must spot code errors even when AI helps write functions. Finance specialists should understand calculations rather than merely copy a ready-made figure.

Courses therefore benefit from 'closed' stages: short oral checks, tests without AI, basic practical exercises, or other formats where students demonstrate foundational skills independently.

The OECD recommends preserving the development of foundational knowledge and independent thinking alongside educational and general-purpose GenAI.

This is a way to ensure people can evaluate automated results instead of depending on them blindly, rather than a return to the past.

AI Usage Rules Should Be Specific

Unclear policies are one of students' biggest problems.

'AI may be used responsibly' explains almost nothing. Can it outline an essay? Translate a paragraph? Check grammar? Generate code? Suggest counterarguments?

Among its key 2026 findings, EDUCAUSE identifies a need for clearer student policies and guidance on AI use in assessment.

A practical approach defines rules separately for each major assignment.

For example, AI may be prohibited in an individual test. Another assignment permits it for ideas, but requires students' own final argument. A third allows language editing with disclosure. A fourth requires AI because it assesses professional use of a generative system.

The more specific the rule, the less room for misunderstanding.

AI Disclosure Is Better Than AI Hunting

If generative tools are partly permitted, teaching students to document their use makes sense.

This can be a short note at the end: which tool was used, for which tasks, and what the student changed independently.

For more complex projects, request a few key prompts or brief reflection on where AI helped and where its results were weak.

The purpose is transparency and professional responsibility, rather than bureaucracy.

Professionals also need to explain which parts of a workplace decision were automated, which data were used, and who checked the result. Transparency in AI use therefore becomes an important competence itself.

Why AI Detectors Do Not Solve the Problem

Automatically identifying AI text sounds appealing: students submit work, a service shows a percentage, and instructors decide.

But a probabilistic detector does not prove authorship.

Human text can be misidentified as machine-generated, while edited AI text can pass. Language style also varies with proficiency, subject, and genre, leaving a risk of misinterpretation.

More importantly, a detector does not answer the educational question. Even if AI use is certain, we still need to know whether it was allowed, which work it performed, and whether the student developed the required competence.

Rather than relying on one technical indicator, gather several forms of evidence: intermediate versions, sources, defenses, practical application, and explanations of the process.

Tests Also Need Review

A conventional multiple-choice test may seem automatically protected from AI in a controlled environment. But another question remains: does it measure what really matters?

Simple factual recall questions remain useful for checking foundational knowledge. However, an assessment system consisting only of them provides weak evidence of the ability to apply knowledge.

AI encourages greater use of scenario questions, error analysis, comparisons of approaches, and tasks requiring reasons for choices.

An automated test can also be combined with a short follow-up: explain one answer, apply a principle to a new case, or find an error in a supplied solution.

The test then becomes one layer of learning evidence, rather than the endpoint.

Assessing Group Work Becomes More Complex

AI adds another layer to the familiar group-project problem: who actually did the work?

A team may include four students and several AI tools. The final presentation can look professional while individual contributions remain hard to see.

Assess an individual component alongside the group product, such as brief reflection, a personal defense of one's section, a decision log, or answers about one's project contribution.

Also assess AI interaction as part of teamwork: who verified generated data, who was responsible for sources, and which automated recommendations the team rejected.

Group work then comes closer to real professional practice, where responsibility is also distributed among people and technologies.

Rubrics Should Assess Thinking as Well as Presentation

If criteria mainly reward fluent text, structure, and error-free language, generative AI has a natural advantage.

Rubrics therefore need review.

More weight can go to justified decisions, evidence quality, source use, recognition of limitations, explanations of choices, and transfer of knowledge to new contexts.

Writing quality still matters. However, distinguish communication form from subject competence.

In a business case, for example, a beautifully presented recommendation should not earn a high score if it rests on weak assumptions.

Such criteria are harder to satisfy with one successful prompt because they require connections between evidence, context, and one's own decision.

Feedback May Change Too

AI affects both how students complete assignments and how instructors review them.

Generative systems can help group common errors, draft initial comments, or compare work with a rubric. This can potentially save time.

However, the final educational judgment should not be entirely delegated to a system.

Grades often depend on context: previous progress, task features, ambiguous reasoning, and what the instructor intended to assess.

The best scenario uses AI for routine work while people remain responsible for substantive decisions.

In particular, do not transfer student work containing personal or confidential data to an external service without a clear information-processing policy.

What About Large Groups?

Recommendations for oral defenses and personal feedback sound good until a group contains 200 students.

A new assessment system therefore needs to remain scalable.

One option combines automated and human elements. Tests assess foundational knowledge. More complex assignments pass through several short checkpoints. Oral defenses can be selective or reserved for key projects.

Another approach uses peer assessment with clear criteria: students analyze each other's work. This reduces workload while developing the ability to evaluate quality.

On WeStudy, this approach is best organized as a system of different assessment formats rather than one large final test. It gives instructors more progress signals and distributes review throughout the course.

Portfolios and Cumulative Assessment Restore Context

Portfolios are another format gaining value in the generative AI era. Instead of one final artifact, students gradually collect evidence of development: early attempts, drafts, revisions after feedback, practical work, brief reflections, and the final result.

Portfolios show changes in competence over time. Instructors see more than 'strong work at the end': they see how students learned to frame problems, verify sources, correct errors, and explain decisions.

AI need not be banned here. Instead, require students to mark where it was used and briefly explain the system's recommendation, what they accepted or rejected, and why.

This format is particularly useful for project-based and professional courses, where competence develops gradually. It also reduces the risk that one brilliant final file conceals weak understanding.

The Ability to Work With AI Also Needs Assessment

In many professions, refusing AI entirely no longer reflects the real workplace.

Alongside tasks without AI, include tasks in which students demonstrate competent AI use.

For example: formulate a good prompt, verify an answer, identify an error, compare alternatives, protect confidential data, or explain an automated recommendation's limitations.

Assessment then checks AI literacy—the ability to use a tool critically and responsibly—rather than simply whether students can press a button.

EDUCAUSE explicitly identifies AI literacy, including understanding when AI should not be used, as an important competence for students, instructors, and institutional staff.

A Practical Model: Three Assessment Modes

Rather than one rule for an entire course, use three modes.

The first is without AI. Use it to assess foundational individual competence. Students should demonstrate that they can perform a key action independently.

The second is AI permitted as an aid, for example for brainstorming, language editing, or counterarguments. Use is disclosed, and assessment criteria focus on the student's own decision.

The third is AI as part of the task. Students must work professionally with the model: verify, edit, compare, identify risks, and explain their choices.

This model is far clearer than a universal 'allowed' or 'not allowed.'

It also allows autonomy to increase gradually. Early in a course, more foundational tasks are completed without AI. Later, students learn to use the tool while retaining control.

How to Redesign an Existing Assignment in 15 Minutes

Not every course needs rewriting from scratch.

For a conventional essay, add a brief oral defense and a requirement to show two key sources.

For an analytical report, add an 'alternatives I considered' section and a question about the recommendation's limitations.

For a problem, change one condition after the solution and ask students to explain what changes.

For a presentation, ask one unexpected question afterward that requires applying the material.

If AI is permitted, add a brief disclosure of its role.

Small changes often substantially improve learning evidence without sharply increasing instructor workload.

How to Tell Whether New Assessment Works

A strong assessment system should answer a few simple questions.

Can students explain their own results?

Can they apply knowledge in a slightly different situation?

Is the work's development process visible?

Is it clear where AI use is permitted?

Does the grade avoid depending solely on attractive presentation?

Can assessment be delivered under the group's real conditions?

Does it assess the competence stated in the course outcomes?

If most answers are 'yes,' the next AI model should not completely dismantle the system.

Technology will change. Reliable assessment should center on what remains stable: understanding, reasoning, application, responsibility, and people's ability to explain their decisions.

Assessment After ChatGPT Becomes More Honest, Even if More Demanding

Generative AI has created an uncomfortable situation for education. Many assignments considered normal for years no longer provide enough information about students' real knowledge.

But this also creates an opportunity.

Instead of mainly measuring the ability to produce the right artifact, education can pay more attention to thinking processes, practical application, oral explanation, evidence use, and responsibility for decisions.

This is harder than checking a file and awarding a score. Yet it restores assessment's main function: showing what people actually know and can do.

ChatGPT does not abolish assessment. It forces a more precise answer to a question that should always have been central: what do we want student work to reveal—a polished result or real learning?

Sources

OECD Digital Education Outlook 2026: Exploring Effective Uses of Generative AI in Education, OECD, January 19, 2026.

The Impact of AI on Learning Assessment, EDUCAUSE, June 1, 2026.

The Assessment Debate in an AI World, EDUCAUSE, June 22, 2026.

The Current State of Play: AI in Higher Education and the Road Ahead, EDUCAUSE, June 16, 2026.

Seeing Learning Differently: What Generative AI Reveals About Human Capability Development, EDUCAUSE, August 17, 2026.