Pestal guide·Assessment
AI-Resistant Assessments: How to Measure Learning When Students Have AI
The goal is not to defeat AI. The goal is to preserve the quality of the evidence.
John Segreto, Founder of Pestal
Published September 2026 · Reviewed September 2026 · 21 min read
On this page
Generative AI has made it possible for students to produce polished essays, correct solutions, working code, detailed reports, presentations, and other high-quality work with substantial assistance.
That creates an assessment problem.
If a student submits excellent work, what can we reasonably conclude that the student understands?
This is the question behind AI-resistant assessment.
At Pestal, we use the term cautiously.
No assessment should be assumed to remain permanently resistant to increasingly capable AI. And preventing AI use is not always the educational objective.
A more useful definition is:
An AI-resistant assessment is an assessment designed so that an educator can still obtain credible evidence of what a learner understands or can do, even when AI is readily available.
The goal is not to defeat AI. The goal is to preserve the quality of the evidence.
Assessment is an inference problem
Assessment has always required educators to make inferences.
We cannot directly observe "understanding."
Instead, we observe something a learner does.
They might:
- answer a question;
- solve a problem;
- write an essay;
- explain an idea;
- build something;
- perform a procedure;
- analyze a source;
- design an experiment;
- write a program;
- or respond to an unfamiliar situation.
We then use that performance to make a claim about what the learner knows or can do.
In other words:
Assessment inference
- Observed performance
- Evidence
- Inference about the learner
That relationship is fundamental to assessment.
Generative AI matters because it can change the relationship between the performance we observe and the learner capability we are trying to infer.
A 2026 Stanford Accelerator for Learning and ETS initiative on responsible assessment makes a closely related point: as AI becomes embedded in learning, assessments based primarily on final products may no longer capture the knowledge and skills they were intended to measure. The initiative argues for richer systems of evidence that can capture learning across contexts and over time.
This means the challenge created by AI is not simply:
Can AI complete this assignment?
The deeper question is:
Does this assessment still give us the evidence we need to make the inference we want to make?
That is the problem AI-resistant assessment should try to solve.
AI has weakened an assumption behind many traditional assessments
For a long time, educators have often made a reasonable inference from student work:
Quality of student work → evidence of student understanding
That inference was never perfect.
Students have always received help from teachers, tutors, parents, classmates, calculators, search engines, reference materials, and other resources.
Generative AI changes the conditions under which many artifacts can be produced.
A learner can receive substantial assistance with reasoning, writing, coding, problem solving, editing, research, explanation, and revision while still producing an artifact that appears highly accomplished.
This concern is increasingly visible in assessment research. UNESCO has argued that AI's ability to produce essays, solve problems, write code, and generate other outputs challenges the use of those outputs as straightforward indicators of individual understanding. Educational-measurement researchers have likewise identified validity, reliability, fairness, transparency, equity, and test security as important issues when generative AI intersects with assessment.
Consider a student who submits an excellent essay.
The essay may reflect the student's:
- understanding of the subject;
- ability to construct an argument;
- writing ability;
- research ability;
- ability to use AI;
- ability to evaluate AI output;
- editing ability;
- or some combination of these.
The artifact itself may not tell us which.
That does not make the essay worthless.
It means we have to be more careful about the inference we make from it.
Student work and evidence of learning are not necessarily the same thing
This distinction is increasingly important.
Artifact
What the learner produces or performs.
Evidence
What that artifact and its production conditions allow us to conclude.
A student's essay, lab report, solution set, program, presentation, model, or project is an artifact.
Evidence is what that artifact—and the conditions under which it was produced—allows us to reasonably conclude about the learner.
Those are different things.
- A beautiful engineering design does not necessarily demonstrate that the student can explain the physics behind it.
- A correct set of homework answers does not necessarily demonstrate that the student can solve similar problems independently.
- A polished historical essay does not necessarily demonstrate that the student can evaluate an unfamiliar primary source.
- A working computer program does not necessarily demonstrate that the student understands why the program works.
What claim about this learner does this evidence actually support?
That is a more useful question than simply asking whether AI was used.
The Pestal Evidence Test
At Pestal, we have developed a five-question framework for thinking about assessment when learners have access to AI.
We call it the Pestal Evidence Test.
The framework is not an empirical measurement instrument or a claim that every assessment must follow one format. It is a practical design framework intended to help educators examine the relationship between a learning claim and the evidence an assessment actually produces.
Before redesigning an assignment, ask five questions.
- 1
What are we trying to know?
Start with the claim you want to make about the learner.
What knowledge, skill, reasoning, judgment, or capability are you actually trying to assess?
Too vague: The student understands projectile motion.
More useful: The student can construct an appropriate projectile-motion model from an unfamiliar physical situation, identify relevant assumptions, and use the model to make and justify predictions.
The clearer the intended claim, the easier it becomes to decide what evidence is necessary.
- 2
What evidence would demonstrate it?
Ask what a learner would need to do for you to reasonably conclude that the intended learning occurred.
Sometimes a final artifact may provide enough evidence for the instructional decision being made.
Sometimes it does not.
Evidence might include:
- a solution;
- an explanation;
- an observed performance;
- a design decision;
- intermediate reasoning;
- an independent demonstration;
- a revision;
- application to a new situation;
- or performance across several assessments.
The assessment should be designed backward from the evidence needed to support the intended claim.
- 3
Could AI produce the artifact without the learner possessing that understanding?
This is the new vulnerability question.
Imagine that a capable AI system produces or substantially assists with the submitted artifact.
Would the resulting work still allow you to make the same conclusion about the learner?
If the answer is no, there is an evidence gap.
That does not necessarily mean AI needs to be prohibited.
It means the assessment needs another source of evidence if the educator still wants to make that particular claim.
- 4
What additional evidence would strengthen the inference?
Often the solution is surprisingly small.
You may not need to replace the assignment.
You may need to supplement it.
Ask the student to explain one important decision.
Give them a related problem they have not seen before.
Ask them to revise a portion of their work in response to new information.
Observe part of the performance independently.
Ask why one approach works and another does not.
Compare assisted practice with an independent demonstration.
The objective is to close the evidence gap without unnecessarily turning every assignment into a high-security examination.
- 5
How much evidence is enough?
This question matters.
Not every assignment needs to prove learning beyond all doubt.
A homework problem may exist primarily for practice.
A draft may exist to generate feedback.
A low-stakes quiz may provide a quick instructional signal.
A final examination or culminating performance may support a stronger conclusion about what the learner can independently demonstrate.
Assessment should therefore be proportional to the importance of the claim being made and the decision that will be based on it.
The goal is not maximum security. The goal is sufficient evidence.
The Pestal Evidence Test in one question
The five questions can ultimately be condensed into one:
Given how this work was produced, what am I justified in concluding about this learner?
That question changes the way we think about assessment.
Instead of starting with:
How do I stop students from using AI?
we can start with:
What do I need to know about the student, and what evidence would allow me to know it?
AI use then becomes one part of the assessment context.
Evidence becomes stronger when we can see learning from multiple angles
A final submission is one evidence channel.
It does not have to be the only one.
Different parts of a learning experience can reveal different things about the learner.
Evidence channels and what each can help reveal
Evidence channelFinal artifact
What it can help revealQuality of the submitted result
Evidence channelProcess
What it can help revealHow the learner developed the work
Evidence channelExplanation
What it can help revealWhether the learner can explain reasoning and choices
Evidence channelIndependent performance
What it can help revealWhat the learner can do without assistance
Evidence channelTransfer
What it can help revealWhether understanding can be applied in a new situation
Evidence channelRevision
What it can help revealWhether the learner can respond intelligently to feedback
Evidence channelLongitudinal evidence
What it can help revealWhether understanding develops or persists over time
| Evidence channel | What it can help reveal |
|---|---|
| Final artifact | Quality of the submitted result |
| Process | How the learner developed the work |
| Explanation | Whether the learner can explain reasoning and choices |
| Independent performance | What the learner can do without assistance |
| Transfer | Whether understanding can be applied in a new situation |
| Revision | Whether the learner can respond intelligently to feedback |
| Longitudinal evidence | Whether understanding develops or persists over time |
None of these channels is universally superior.
A teacher does not need to collect all of them for every assignment.
The important principle is that different evidence supports different claims.
If the claim is important and one evidence channel has become insufficient, another channel may strengthen the inference.
This idea is consistent with the direction of recent assessment work. The Stanford/ETS responsible-assessment initiative argues for moving beyond isolated testing events toward richer evidence of learning, including portfolios, formative feedback, conversation-based assessments, authentic performance tasks, and competency demonstrations.
Assistance is part of the assessment context
There is another distinction that becomes important in an AI-rich learning environment:
What a learner can do with assistance and what a learner can demonstrate independently are both useful information—but they are not the same information.
With support
What can the learner accomplish with support?
Independently
What can the learner demonstrate without that support?
Suppose two students submit identical correct solutions.
One student solved the problem independently.
The other initially could not solve it, asked an AI tutor for an explanation, worked through two examples, and then completed the problem.
From a learning perspective, the second interaction may be productive.
The student sought help and may have learned from it.
But if the teacher is trying to determine what each student can currently demonstrate independently, the identical final answers do not tell the whole story.
This is why the context around an assessment matters.
Ideally, we want to know something about:
- What was taught?
- What did the learner attempt?
- What assistance did the learner receive?
- What could the learner ultimately demonstrate?
The presence of assistance does not automatically invalidate learning.
It changes what the evidence means.
Research also gives us reason not to treat AI assistance and learning as mutually exclusive. A 2025 systematic review and meta-analysis of 57 higher-education studies found positive effects of generative-AI interventions across several measured learning outcomes, although effects varied substantially by context and outcome. That evidence does not mean AI assistance is always beneficial, but it does reinforce the importance of distinguishing using assistance from failing to learn.
Practice and assessment do not have to operate under the same rules
One consequence of this framework is that educators do not necessarily need the same AI policy everywhere.
During learning, assistance can be valuable.
Students ask teachers questions.
They consult examples.
They work with classmates.
They look things up.
They receive hints.
They use tutoring systems.
And increasingly, they may use AI.
The purpose of practice is often to develop capability.
The purpose of some assessments is to measure capability.
Those goals are related, but they are not identical.
A student might therefore be allowed to use substantial AI assistance while practicing and later be asked to demonstrate an important capability independently.
This produces two useful signals:
What can the learner accomplish with support?
and
What can the learner demonstrate without that support?
The difference between those signals can itself be informative.
What kinds of assessments are more resistant to AI?
There is no single AI-proof assessment format.
Instead, educators can combine assessment patterns depending on what they are trying to measure.
The point is not to choose the format that is hardest for AI to complete.
It is to choose evidence that matches the intended learning claim.
Independent demonstrationAsk students to demonstrate an important capability under conditions where assistance is limited or known.
Examples include:
- short in-class writing;
- problem solving;
- whiteboarding;
- laboratory work;
- practical demonstrations;
- supervised assessments;
- or other independent performances.
Independent work is particularly useful when the claim being made is specifically about independent capability.
It is not automatically the best format for every learning goal.
Explain and defendAllow substantial work, then ask the student to explain important parts of it.
For example:
- Why did you choose this method?
- What assumption does your model depend on?
- Which piece of evidence is most important to your argument?
- Why wouldn't this alternative solution work?
The purpose is not to catch the student.
It is to gather another kind of evidence.
Novel transferAsk the learner to apply the same underlying idea in a situation they have not rehearsed.
Change a condition.
Introduce an unfamiliar source.
Modify a variable.
Present a new case.
Ask for a prediction before calculation.
Transfer tasks can provide evidence about whether a learner can use knowledge beyond the exact form in which it was originally practiced.
Process-visible workCollect meaningful intermediate evidence rather than only the final submission.
Depending on the task, that might include:
- drafts;
- calculations;
- design decisions;
- checkpoints;
- experimental results;
- revisions;
- prototypes;
- or reflections on important decisions.
The goal is not to create surveillance of every keystroke.
Collect process evidence when the process itself helps answer the learning question.
There is emerging research supporting the idea that AI interactions themselves can sometimes become useful process evidence. A 2025 study of 70 graded essays in STEM writing-intensive courses analyzed 1,450 annotated GenAI interactions and found different patterns between passive task delegation, information retrieval, text refinement, and more substantive idea co-development. The authors proposed that process-focused evaluation of student-AI interactions can reveal information not available from the final artifact alone.
That does not mean every AI conversation should be collected or graded.
It means the interaction process can sometimes contain educationally relevant evidence.
Authentic performanceSome capabilities are best assessed by having students actually perform them.
Students might:
- conduct an experiment;
- build and test a device;
- give a presentation;
- participate in a debate;
- analyze a case;
- perform a procedure;
- troubleshoot a system;
- or respond to a realistic scenario.
AI may still be useful during preparation.
The performance provides an additional evidence channel.
Recent assessment discussions have similarly emphasized authentic performance and competency demonstrations as ways of gathering richer evidence than a single final product or score.
AI-inclusive assessmentSometimes the ability to use AI well is itself part of the intended capability.
In that case, banning AI may make the assessment less authentic.
Students might instead be asked to:
- critique an AI-generated response;
- identify errors;
- verify claims;
- compare alternative AI solutions;
- improve an AI-generated draft;
- document important decisions;
- or defend why they accepted or rejected AI recommendations.
The assessment then measures the learner's judgment in an AI-assisted environment.
Research has begun exploring these designs directly. A 2024 study that redesigned writing assessments to incorporate generative AI and other online tools found that students made purposeful choices among tools and argued for assessment practices that allow appropriate flexibility in AI use. Other work has similarly suggested that assessment design can focus on the quality of student interaction with AI rather than treating any AI involvement as inherently invalid.
Again, the starting question is the same:
What capability are we trying to measure?
Examples
History essayTraditional task: Explain the major causes of the American Civil War.A polished final essay may provide weaker evidence of historical reasoning.Explore example
Consider a traditional take-home prompt:
Explain the major causes of the American Civil War.
If the goal is to assess historical reasoning, a polished final essay alone may provide weaker evidence than it once did because a capable AI system can substantially assist with producing that artifact.
The assignment does not necessarily need to disappear.
Instead, the evidence architecture can change.
Redesigned assessment
Students develop an argument about the relative importance of several causes using a defined set of primary and secondary sources.
They submit:
- their final argument;
- an evidence map showing how sources support the argument; and
- a brief explanation of an important interpretive choice.
Then, in class, students receive an unfamiliar source and explain whether it strengthens, weakens, or changes their argument.
The final essay still matters.
But it is no longer carrying the entire evidentiary burden.
The teacher can observe the final product, source reasoning, explanation, and transfer to a new source.
Physics problem solvingTraditional task: Complete ten projectile-motion problems at home.The homework may be useful practice while providing limited evidence of independent problem solving.Explore example
Consider:
Complete these ten projectile-motion problems at home.
Those problems may be excellent practice.
But if students can obtain complete solutions from AI, the homework alone may provide limited evidence of independent problem-solving ability.
That does not mean the homework needs to become AI-free.
Students could be allowed to use notes, examples, classmates, teachers, or AI while practicing.
Later, they receive an unfamiliar physical situation and independently:
- decide how to model it;
- identify relevant assumptions;
- predict what should happen;
- construct and solve the model; and
- explain how changing one variable would affect the system.
Now the practice and assessment serve different purposes.
The homework helped develop capability.
The independent task helps measure it.
Engineering designTask: Design a device that satisfies engineering constraints.Preventing all AI use may be neither realistic nor educationally useful.Explore example
Imagine students are designing a device to satisfy a set of engineering constraints.
Trying to prevent all AI use may be neither realistic nor desirable.
Students might use AI to brainstorm approaches, investigate components, debug code, analyze data, or consider alternatives.
The assessment can instead collect evidence from several places:
- Final artifact: Does the prototype work?
- Design history: How did the solution develop?
- Engineering decisions: Why were particular tradeoffs made?
- Calculations: Can the student justify important design parameters?
- Testing: How was performance evaluated?
- Revision: How did the design respond to failure?
- Defense: Can the student explain why the final design behaves as it does?
AI use does not automatically destroy the validity of the project.
The assessment becomes stronger because the grade is not based solely on the polish of the final artifact.
Why AI detection does not solve the assessment problem
AI detection and assessment answer different questions.
Detection asks
Was AI probably involved in producing this work?
Assessment asks
What does this learner understand or know how to do?
An AI detector attempts to answer something like the first question.
An assessment should help answer the second.
Even a hypothetical perfect AI detector would not make those questions equivalent.
Knowing that AI was involved does not tell us whether the student understood the material.
Knowing that AI was not involved does not automatically prove that the student understood it either.
Detection may be relevant to enforcing a clearly communicated AI-use policy.
But it should not substitute for assessment design.
Concerns about relying on detection are not merely theoretical. UNESCO's assessment analysis notes problems with the reliability of AI-detection approaches and argues that the response to generative AI should focus more fundamentally on what educators need to measure.
A stronger assessment strategy is therefore to generate useful evidence in the first place rather than relying primarily on forensic analysis of the finished artifact.
Don't mistake inconvenience for rigor
There is a danger in responding to AI by simply making assessments harder to complete with AI.
That can produce assignments that are inconvenient without producing better evidence.
Avoid assuming that:
- every assessment should become an oral examination;
- handwritten work is automatically more valid;
- every student should produce extensive process documentation;
- obscure prompts are better because AI has seen fewer examples;
- AI should be banned even when its use is compatible with the learning objective;
- or a difficult-to-outsource assignment is necessarily a good assessment.
Making an assessment harder to outsource is not automatically the same as making it a better assessment.
The purpose is not to make students prove that they are human.
The purpose is to understand what they have learned.
This is also why assessment redesign should consider accessibility, fairness, workload, and context rather than adopting a single supposedly AI-proof format for every learner and every objective. Validity and fairness remain core assessment concerns in an AI environment.
The Pestal AI Assessment Audit
The Pestal Evidence Test provides the underlying framework.
The Pestal AI Assessment Audit turns that framework into a practical way to examine an existing assignment.
It is intended as a design aid, not as a validated psychometric instrument.
Take an assessment you already use and work through the following questions.
- 1
Learning goal
What should this assessment tell me about the learner?
Write the intended claim as precisely as possible.
Avoid beginning with the assignment format.
Begin with what you want to know.
- 2
Artifact
What does the learner submit or perform?
Identify what you actually observe.
It might be an essay, solution, presentation, program, prototype, performance, lab report, discussion, examination, or something else.
- 3
AI capability
How much of that artifact could a capable AI system produce or substantially assist with?
Do not assume that AI involvement automatically makes the task invalid.
Simply identify the vulnerability.
- 4
Evidence gap
If AI produced much of the artifact, what could I no longer conclude confidently about the learner?
This is often the most important question in the audit.
It identifies the difference between the artifact you receive and the claim you want the artifact to support.
- 5
Additional evidence
What is the smallest additional source of evidence that would strengthen the assessment?
Consider:
- explanation;
- process;
- independent performance;
- transfer;
- revision;
- observation;
- or another evidence channel.
The goal is not to add requirements unnecessarily.
It is to close the evidence gap.
- 6
AI policy
For this particular assessment, is AI prohibited, limited, permitted, or expected?
The policy should follow from the learning objective rather than habit.
If independent writing is the capability being measured, AI writing assistance may undermine the intended evidence.
If evaluating and improving AI output is the capability being measured, AI access may be essential.
- 7
Verification
Does anything need to be demonstrated independently?
If independent capability matters, identify where it will actually be observed.
Not everything needs independent verification.
The question is whether the learning claim requires it.
- 8
Transfer
Does the learner ever need to apply the underlying learning somewhere new?
If transfer matters to the learning goal, assess it.
A student who can reproduce a familiar solution and a student who can recognize when and how to apply the underlying idea in an unfamiliar situation may provide different evidence of understanding.
- 9
Decision
Given all of this evidence, what conclusion about this learner am I willing to make?
That is the assessment decision.
From individual assessments to evidence of learning
AI-resistant assessment points toward a larger problem.
Most educational systems fragment the information needed to interpret learning.
The lesson may exist in one system.
Homework exists somewhere else.
Tutoring happens somewhere else.
AI assistance may happen in another application entirely.
Assessment produces a score.
And then the teacher is expected to determine what the learner actually understands.
But evidence becomes more meaningful when we understand the learning experience that produced it.
At Pestal, we think about that experience as a connected loop:
Pestal learning system
- Goal
- Instruction
- Activity
- Assistance
- Assessment
- Evidence
- Adaptation
- Goal
- What is the learner trying to accomplish?
- Instruction
- What was the learner taught?
- Activity
- What did the learner actually do?
- Assistance
- What help did the learner use?
- Assessment
- What was the learner asked to demonstrate?
- Evidence
- What does the resulting performance tell us?
- Adaptation
- What should happen next?
This is why we do not think the future of assessment is simply better AI detection or more secure assignments.
The larger opportunity is connected evidence.
This direction is consistent with the broader movement in assessment toward collecting evidence across learning experiences rather than relying exclusively on isolated final products. The Stanford/ETS work, for example, argues that learning is continuous and contextual and that assessment systems should increasingly reflect that reality.
For Pestal, that idea goes one step further.
When we can interpret what a learner demonstrates in the context of what they were taught, what they practiced, and what assistance they received, we can form a better-informed estimate of the learner's current state.
And a better estimate of learner state can support a better next instructional decision.
That is the broader system Pestal is being built toward.
The purpose of assessment has not changed
Generative AI has changed what students can produce with assistance.
It has not changed the fundamental reason we assess.
Teachers still need to make decisions.
- Does this student understand the concept?
- Are they ready to move forward?
- Where are they struggling?
- What support do they need?
- Can they transfer this knowledge somewhere new?
- What should we teach next?
AI-resistant assessment is therefore not primarily about defending old assignments against new technology.
It is about designing evidence that remains useful for making those decisions.
The question we should keep asking is not:
Can AI do this assignment?
It is:
What will this assessment allow me to know about the learner?
That is a more durable foundation for assessment in the age of AI.
Frequently asked questions
What is an AI-resistant assessment?
An AI-resistant assessment is designed to preserve credible evidence of a learner's knowledge or capability even when AI is readily available. It does not necessarily prohibit AI. Instead, it aims to ensure that the evidence collected still supports the conclusion the educator wants to make about the learner.
Is "AI-resistant" the same as "AI-proof"?
No. "AI-proof" implies designing a task that AI cannot complete. That is unlikely to be a durable goal as AI capabilities continue to change. AI-resistant assessment focuses instead on preserving useful evidence of learning. An AI system may be able to contribute to an assignment while the overall assessment still provides meaningful evidence through explanation, independent performance, process, transfer, revision, or other channels.
Is any assessment completely AI-proof?
It is safer not to assume that any assessment format will remain permanently impervious to increasingly capable AI. The more durable strategy is to design assessments around the evidence needed to demonstrate learning rather than trying to invent assignments AI cannot complete.
Should teachers ban AI from assessments?
Sometimes. If the purpose of an assessment is to determine what a student can independently demonstrate, restricting AI may be appropriate. In other assessments, AI use may be permitted or even expected. The policy should follow from the capability being assessed.
Are in-class assessments more AI-resistant?
They can provide stronger evidence of independent performance when assistance is controlled or known, but simply moving an assignment into class does not automatically make it a better assessment. The task still needs to produce evidence relevant to the intended learning goal.
Are oral assessments AI-resistant?
Oral explanation can provide an additional evidence channel because students can be asked to explain reasoning and respond to follow-up questions. It does not need to replace every other assessment format, and educators should consider accessibility, fairness, time, and practical constraints when deciding how to use it.
Can students use AI and still demonstrate learning?
Yes. AI assistance and learning are not mutually exclusive. A student may learn through AI-supported explanations, feedback, examples, practice, critique, or revision. The assessment question is what the student can ultimately demonstrate and whether that demonstration supports the intended learning claim.
Why isn't AI detection enough?
AI detection focuses on whether AI may have contributed to an artifact. Assessment focuses on what the learner knows or can do. Those are different questions. Even knowing exactly how much AI was used would not, by itself, establish what the student understands.
What is the difference between an artifact and evidence of learning?
An artifact is something the learner produces, such as an essay, solution, project, program, or presentation. Evidence is what that artifact, together with the conditions under which it was produced and any other relevant observations, allows an educator to reasonably conclude about the learner.
How can I make an existing assignment more AI-resistant?
Do not automatically replace the assignment. First identify the learning claim and determine what evidence AI use may have weakened. Then add the smallest useful evidence channel—such as an explanation, independent demonstration, process checkpoint, revision, or novel transfer task—that strengthens the inference you need to make.
Sources and further reading
This guide combines Pestal's first-party assessment framework with research and institutional work on assessment in the age of generative AI.
Stanford Accelerator for Learning and ETS — Responsible Assessment in the AI Era (2026).
A convening and white paper examining how AI changes what is assessed, how evidence is collected, and why assessment may need to move beyond isolated final products toward richer and more continuous evidence of learning.
UNESCO — "What's worth measuring? The future of assessment in the AI age" (2025).
A discussion of how generative AI challenges traditional assessment outputs and why assessment may need to place greater emphasis on higher-order thinking, creativity, ethical reasoning, and authentic demonstrations of learning.
Jiangang Hao, Alina A. von Davier, Victoria Yaneva, Susan Lottridge, Matthias von Davier, and Deborah J. Harris — "Transforming Assessment: The Impacts and Implications of Large Language Models and Generative AI" (2024).
An educational-measurement perspective on the opportunities and challenges generative AI creates for validity, reliability, transparency, fairness, equity, and assessment security.
"Assessing students' DRIVE: A framework to evaluate learning through interactions with generative AI" — Computers and Education: Artificial Intelligence (2025).
Research examining whether students' interactions with generative AI can themselves provide evidence about learning and distinguishing patterns such as passive task delegation from more substantive idea-development interactions.
"Reinventing assessments with ChatGPT and other online tools: Opportunities for GenAI-empowered assessment practices" — Computers and Education: Artificial Intelligence (2024).
A study of redesigned writing assessments that incorporated generative AI and other tools and examined how students selected and used them.
Chen and Cheung — "Effect of generative artificial intelligence on university students learning outcomes: A systematic review and meta-analysis" (2025).
A synthesis of 57 studies examining measured learning outcomes associated with generative-AI interventions in higher education and showing substantial variation by context and outcome.