# How to evaluate students when they use artificial intelligence

> Learn how to evaluate students who use artificial intelligence through process evidence, rubrics, verification, oral defense and clear rules.

- Site: Genialoh (https://genialoh.org)
- Language: en
- Category: Assessment
- Reading time: 16 min
- HTML version: https://genialoh.org/#/en/blog/evaluate-students-artificial-intelligence

---
Generative artificial intelligence can produce essays, solve problems, create code, summarize documents, propose designs and answer questions in seconds. This capability requires educational institutions to review how they evaluate learning.

When an activity is graded only by its final product, it can be difficult to know what the student understood, what decisions they made, which part was generated by a tool and what evidence exists of their reasoning.

The solution is not just to ban artificial intelligence or to use automatic detectors. Proper assessment must answer more important questions:

- What learning does the student need to demonstrate?
- What use of artificial intelligence is allowed?
- What part of the process must they do personally?
- How did they verify the generated information?
- Can they explain and defend the submitted work?
- What decisions did they make and what errors did they identify?
- What evidence do they keep of their process?
- Does the activity distinguish assistance from substitution?
- Were the assessment criteria communicated in advance?

> **Shift the focus**
> 
> Assessing in the AI era means observing not only what the student submits, but also how they research, verify, decide, apply and explain.

## The problem of assessing only the final product

A traditional activity may ask: "Write a 2,000-word essay on the effects of artificial intelligence on employment." The instructor receives a finished document and evaluates it based on structure, clarity, argument and references.

However, if a tool could generate almost all the content, the final product offers little information about the student's understanding, research ability, source selection, argument development, quality of decisions and command of the topic.

This does not mean essays are no longer useful. It means they may need additional elements that make the process and the student's intellectual participation visible.

## The central question: what evidence demonstrates learning?

Before deciding how to control AI use, the instructor must define what to assess: conceptual understanding, application, problem solving, data interpretation, argument, critical thinking, design, communication, decision making, professional judgment, research, reflection, creativity or collaboration.

The required evidence changes with the objective. To assess spontaneous writing, a supervised activity without AI may fit. To assess critical thinking, the tool may generate a response for the student to examine, verify and improve. To assess a professional competency, the student may be required to use AI and also to justify each decision and recognize risks.

## Four assessment modalities in the AI era

| Modality | Description | Typical formats |
| --- | --- | --- |
| No AI | The tool is not allowed. | In-person exam, oral interview, in-class problem solving, practical demonstration. |
| Limited use | AI in specific stages, not in the core. | Generate study questions, preliminary outline, prior practice. |
| Allowed and disclosed | AI as process support, with transparency. | Projects, essays with disclosure, analysis with documented AI use. |
| AI required | The tool is part of the assessed competency. | Prompt design, comparison of results, bias analysis. |

In the last modality, assessment measures not only the ability to obtain an answer. It measures the ability to direct, review and responsibly use the technology.

## Principle 1. Assess the process, not only the result

An assessment can request evidence of the work stages: initial question, plan, draft, sources consulted, prompts used, relevant AI responses, verification table, changes made, final product, reflection and oral defense. Evidence must be proportional to the activity's value and risk.

### Example of staged submission

| Stage | Deliverable |
| --- | --- |
| First submission | Research question, objective, initial sources and work plan. |
| Second submission | Draft, disclosed AI use, issues found and verification evidence. |
| Final submission | Finished product, table of changes, reflection and presentation or defense. |

## Principle 2. Ask the student to explain their decisions

A student can produce correct work without understanding it. Assessment must include questions like: Why did you choose this approach? What alternatives did you consider? What suggestions did you reject? What information did you have to verify? What was the hardest part? What limitations does your answer have? How would you apply the result to another case?

- Written reflection or logbook.
- Comments within the document.
- Presentation or brief interview.
- Oral defense or explanatory video.
- Random questions or peer review.
- Practical demonstration.

## Principle 3. Include verification as an assessment criterion

A generated response may contain incorrect data, nonexistent references, invented citations, calculation errors, faulty code, generalizations, biases or outdated information. The student must show they can check it.

### Verification table

| Generated claim | Source or test | Result | Decision |
| --- | --- | --- | --- |
| First claim | Official document | Confirmed | Kept |
| Second claim | Academic article | Partially correct | Modified |
| Third claim | No evidence found | Not verifiable | Removed |
| Fourth claim | Independent calculation | Incorrect | Corrected |

The quality of this verification can carry specific weight in the grade.

## Principle 4. Use oral defenses

A brief oral defense checks whether the student understands what was submitted. It does not need to become a long exam: a five-minute conversation can include a problem summary, approach explanation, justification of a decision, review of a source, concept explanation, solving a variant, description of AI use and identification of a limitation.

> **Sample defense questions**
> 
> - Explain the main conclusion in your own words.
> - What would change if this data were different?
> - Which is the most important source?
> - What AI recommendation did you reject?
> - Where is there greater uncertainty?
> - What mistake could someone make when applying your proposal?

## Principle 5. Redesign easy-to-delegate activities

Some activities let a tool produce almost all the work: general essays, summaries, definitions, lists, descriptive reports, standard problems, generic presentations or simple code without explanation. They can be strengthened by incorporating local cases, student-collected data, progressive submissions, justified decisions, application to a context, oral defense and process evidence.

| Original activity | Redesigned activity |
| --- | --- |
| Summarize three leadership theories. | Select three theories and ask a tool for an initial comparison. |
|  | Verify each description with sources and identify simplifications. |
|  | Analyze a real or local case and apply the theories. |
|  | Defend which one best explains the situation and document AI use. |
|  | Explain which parts of the generated answer you modified. |

## Principle 6. Adapt the rubric

A rubric for AI activities must include criteria on process and responsibility: understanding, quality of analysis, verification, source use, error identification, decision justification, original production, transparency, reflection, defense capacity, rule compliance and information protection.

| Criterion | 4: Outstanding | 3: Adequate | 2: Developing | 1: Insufficient |
| --- | --- | --- | --- | --- |
| Understanding | Deeply explains and applies concepts | Shows correct understanding | Shows significant gaps | Does not show understanding |
| Verification | Rigorously checks claims and sources | Verifies main elements | Verification is partial | Accepts results without checking |
| Decisions | Justifies what was accepted, changed and rejected | Explains main decisions | Offers limited justification | Does not explain decisions |
| Original product | Integrates evidence and develops a solid response | Product shows personal elaboration | Depends excessively on generated result | Presents generated response as own work |
| Responsible use | Discloses accurately and respects rules | Meets essential requirements | Disclosure is incomplete | Omits or contradicts rules |
| Defense | Explains, responds and applies to new situations | Defends main elements | Has difficulty explaining | Cannot explain the content |

The rubric must be shared before the activity begins.

## Principle 7. Request a use disclosure

> **Short template**
> 
> I used [tool] for [purpose]. I used it during [stage]. I verified the result through [sources or procedure]. I modified or rejected the following parts: [explanation]. I take responsibility for the final content.

The extended version adds tool used, purpose, stage, relevant prompts, result used, changes, verification, errors found, decisions and confirmation of responsibility. The disclosure can be assessed as evidence of transparency and judgment.

## Principle 8. Combine assessment with and without AI

An institution does not have to choose between banning the technology in every activity or always allowing it. It can combine formats: individual quiz without AI, project with allowed AI, oral presentation, practical exercise, written reflection, peer review, portfolio, application exam and professional demonstration. A student may use AI in a project but must individually show that they understand the concepts and can apply them without assistance.

## Principle 9. Use personalized questions or variants

Personalization must not depend on sensitive data. It can consist of assigning different cases, conditions, data, sectors, perspectives, constraints, audiences or professional scenarios. After submission, the instructor can change one condition and ask the student to adapt their answer. The ability to respond shows understanding beyond the prepared document.

## Principle 10. Do not rely solely on AI detectors

An automatic detector cannot on its own prove which tool was used, which parts were generated, whether an infraction occurred, who wrote the content, whether the use was allowed or what the student understood. It can produce false positives and false negatives.

> **Fair review**
> 
> A grade or sanction should not depend solely on an automatic score. Review must consider instructions, allowed use level, disclosure, drafts, change history, sources, process evidence, ability to explain and conversation with the student.

## Principle 11. Protect information

Students should not enter into unauthorized tools full names, student IDs, records, grades, medical or financial information, family data, passwords, confidential assessments, question banks, unpublished research, organizational data or unauthorized materials. Instructors should also avoid uploading identifiable work when the tool is not approved for it. Information protection can be part of the rubric and disclosure.

## How to evaluate different types of products

| Product type | Request | Assess |
| --- | --- | --- |
| Essays and academic texts | Own question, verified sources, outline, draft, disclosure, table of changes, oral defense. | Argument, evidence, interpretation, originality, verification, explanation. |
| Code | Runnable code, comments, tests, edge cases, explanation, error log, practical defense. | Functionality, understanding, efficiency, security, debugging, responsible use. |
| Presentations | Sources, script, notes, disclosure, adaptation to audience, response to a scenario change. | Command, organization, evidence, communication, coherence. |
| Designs and creative products | Brief, sketches, iterations, references, prompts to AI, reflection on authorship. | Concept, decisions, process, coherence, originality, justification. |
| Problem solving | Statement, procedure, assumptions, verification, variant without assistance, oral explanation. | Reasoning, method, accuracy, understanding, adaptation. |
| Research | Question, methodology, sources, data, analysis, disclosure, ethical limits. | Methodological coherence, source quality, analysis, transparency, interpretation. |

## 10 assessment formats suited to the AI era

- Progressive portfolio gathering drafts, decisions and final product.
- Brief oral defense to check understanding and application.
- Practical assessment under defined conditions.
- Contextualized case study that requires applying concepts.
- Decision logbook documenting what the student did and why.
- Critical comparison of responses, sources or methods.
- Staged assessment with intermediate feedback.
- Product plus reflection on the process.
- Review and correction of a generated result.
- Combination of individual and collaborative assessment.

## Complete assessment example

Title: Critical evaluation of an AI-generated recommendation. Objective: apply disciplinary criteria to assess a recommendation, verify its evidence and develop an original proposal. Use of artificial intelligence: required.

### Instructions

- Analyze the provided case and formulate a prompt to request a recommendation.
- Save the response and identify its five main claims.
- Verify each claim and classify errors and omissions.
- Write an original recommendation and explain what elements you kept.
- Submit a disclosure and defend your work for five minutes.

### Suggested distribution

| Criterion | Percentage |
| --- | --- |
| Case understanding | 15% |
| Verification | 20% |
| Critical analysis | 20% |
| Original proposal | 20% |
| Justification | 10% |
| Transparency in AI use | 5% |
| Oral defense | 10% |

The percentages are an example and should be adjusted to the course's purpose.

## Institutional process to update assessment

| Step | Action |
| --- | --- |
| 1 | Identify vulnerable activities that can be almost fully completed with AI. |
| 2 | Prioritize important assessments: capstones, high-stakes and core competencies. |
| 3 | Define use categories: prohibited, limited, allowed with disclosure or required. |
| 4 | Create templates for instructions, disclosures, logbooks and rubrics. |
| 5 | Train instructors with real examples from each discipline. |
| 6 | Pilot the changes in two manageable programs. |
| 7 | Collect evidence on understanding, faculty workload, clarity and incidents. |
| 8 | Adjust instructions, formats and criteria. |
| 9 | Create a shared repository of activities and rubrics. |
| 10 | Review periodically as tools and practices evolve. |

## Common mistakes when assessing in the AI era

| Mistake | Consequence |
| --- | --- |
| Keeping easy-to-delegate activities. | The final product no longer represents learning. |
| Prohibiting without explanation. | Students do not understand which functions are included. |
| Allowing without setting limits. | Support and substitution are not distinguished. |
| Assessing only style. | Well-written text may contain weak reasoning. |
| Not requiring verification. | Generated references and claims may be incorrect. |
| Requesting every conversation. | Disproportionate burden and unnecessary information. |
| Trusting only detectors. | A score does not prove an infraction. |
| Not offering a chance to respond. | Review loses fairness and considers partial evidence. |
| Applying the same rules to every activity. | Each learning outcome needs different conditions. |
| Ignoring faculty workload. | New formats must be sustainable and supported. |

## Frequently asked questions

### Can work supported by AI be assessed?

Yes. Clear rules must be set and understanding, analysis, verification, decisions, original production and transparency must be evaluated.

### How do I know the student really learned?

Through process evidence, follow-up questions, application to variants, oral defense, supervised activities and combined formats.

### Should AI be banned in exams?

It depends on the objective. Some assessments must be done without assistance; others can assess precisely the ability to use AI critically.

### Is it mandatory to submit every prompt used?

Not always. Evidence must be proportional. Main prompts and an explanation of how they influenced the work may be enough.

### Can a detector be used to grade?

It should not be used as sole evidence or as a direct grading criterion.

### How to assess originality?

It can be observed in decisions, adaptation to context, justification, selection of evidence and ability to defend the result.

### What to do when a student did not disclose use?

Apply institutional policy considering instructions, importance of the omission, intent, impact and repetition.

### Can AI help the instructor to grade?

It can support rubric creation, initial organization of comments or identification of items requiring review, but the academic decision must remain a human responsibility.

### Should every assessment change?

Not necessarily. Prioritize those in which AI can easily substitute the learning being observed.

### How can an institutional platform help?

It can provide contextualized tutors, common guidelines, instructor support, training, analytics and adoption tracking within an experience defined by the institution.

## Assess what really matters

Artificial intelligence can produce increasingly complete answers. That is why assessment must focus on capacities that still require student participation: understanding, verifying, applying, comparing, deciding, justifying, correcting, creating, communicating and reflecting.

An institution does not need to respond only with more surveillance. It needs to design better evidence of learning, train its instructors and set common rules.

> **Genialoh for your institution**
> 
> - Institutional name, logo and colors.
> - Context based on syllabi and programs.
> - Tutors configured by degree, level or program.
> - AI support for instructors and personalized learning.
> - Student and group reports, and institutional analytics.
> - Faculty training, implementation, support and adoption tracking.

Book a first 30-minute meeting. If your institution decides to move forward, it will select two degrees or programs and provide the required authorized academic documentation. Genialoh will prepare a functional demo with the institution's identity and academic context in less than 14 days from receiving the complete documentation. After the presentation, up to five authorized people can explore it for 30 calendar days.