What a generative AI tool does
A generative AI tool produces new content—such as text, images or code—from patterns learned during training. A language model processes text in units called tokens and estimates what might come next. It repeats that process to build a response. It does not think or understand like a person: it generates a likely sequence based on the context and patterns learned during training.
Fluency does not show that a response is true. A system can give a convincing explanation and still invent a fact, a reference or a mathematical step. Nor is it necessarily a search engine: generating a response and retrieving documents are different operations. Some tools consult webpages or databases and add context to the generation process. That mechanism can make checking easier, but it does not eliminate errors, omissions or misinterpreted citations.
Models can also reproduce stereotypes, omit perspectives or favor particular language varieties. A formal register or fluent writing does not make a response neutral. In class, compare fictional examples and examine who is represented, which assumptions are taken for granted and whose voices are missing. Do not use students’ real identities for these tests. One comparison can help discuss a particular result, but it does not measure every bias in a tool.
For example, a fictional answer might confidently claim that “photosynthesis mainly takes place in the roots.” Its tone does not turn that claim into fact: students would need to check it against reliable biology materials. A tool may give different answers to the same question; repeating or rephrasing it is not independent verification.
It helps to distinguish four functions. A generator drafts or transforms content; a search engine locates documents; an information-retrieval tool selects passages from a collection to provide context; and a person checks whether the result is supported. One product may combine some of these functions, but teachers should identify what the tool actually does before designing an activity.
Functions that should not be confused
| Function | What it does | What it does not show |
|---|---|---|
| Generation | Produces text, images, audio or code from a prompt and context. | That the content is true or that the sources exist. |
| Search | Finds webpages or documents in response to a query. | That a document is reliable or relevant. |
| Information retrieval | Selects material from a collection to provide context for a response. | That the model has interpreted that material correctly. |
| Verification | Checks a claim against appropriate documents, data or evidence. | That one check resolves every uncertainty. |
Choose uses that serve learning
The main question is not whether students can use AI in general, but what they need to demonstrate in each assignment. If the objective is to practise a basic skill, delegating that practice may obscure what a student can do. If the objective is to analyse, revise or critically evaluate a response, AI can become an explicit part of the activity, provided the rules are clear.
Cornell University suggests considering three options: prohibit AI when it interferes with foundational skills; allow it with disclosure when it can help but students remain responsible for their work; or actively integrate it when it serves the learning objectives. These are design choices, not a universal policy. One course may need different rules for a maths test, a brainstorming exercise and a final project.
UNESCO’s recommendations propose a human-centred approach and pedagogical and ethical validation of tools. Its student competency framework covers a human-centred mindset, AI ethics, AI techniques and applications, and AI system design, with progression from understanding to creating. These frameworks can help define objectives: students should be able to explain, evaluate and decide, not simply obtain an answer.
The practical test is to connect each permission or restriction to a specific skill. If an assignment assesses the ability to write an argument, allowing a tool to write it prevents the teacher from directly observing that ability. If the assignment assesses critical review of arguments, analysing an AI response may be relevant. The same use can therefore be appropriate at one stage of a course and unsuitable at another.
Design choices for three assignments
| Assignment | Learning objective | Possible AI use | Evidence of learning |
|---|---|---|---|
| Secondary-school essay on a novel | Interpret passages and support an independent reading. | Allow initial ideas if disclosed; do not allow the tool to write the analysis being assessed. | A selected passage, a draft, an explanation of changes and a short reflection. |
| Problem-solving in an introductory course | Show the method and understand foundational concepts. | Limit AI in exercises that assess skills students have not yet acquired; allow it later to compare methods. | Working steps, a rationale for the method and a check of the result. |
| University research project | Develop a question, evaluate sources and build an argument. | Allow limited help with organising ideas, provided sources and claims are verified. | Sources consulted, justified decisions, critical review and a discussion of limitations. |
Redesign assessment to make the process visible
An assignment can ask for more than a final product without turning into constant surveillance. Teachers can request relevant evidence of reasoning, such as an outline, an initial version, an explanation of revisions and a brief reflection on key decisions. They do not need to demand a complete history of private conversations with a tool: a specific example of how a student checked or changed a response may be more useful and less intrusive.
For an essay, students might select a claim generated by a tool, compare it with an assigned text and explain what they keep, correct or reject. In maths, they can show their own working, compare the result with another strategy and identify where a suggested solution goes wrong. In a university project, they can annotate why a source is suitable and what uncertainty remains.
A short oral explanation or an in-class activity can provide information about learning, but should not be the only assessment format for everyone. Accessibility needs, language barriers and individual circumstances matter. Offer an equivalent alternative when an in-person, oral or handwritten format creates a barrier unrelated to the assessed objective.
A rubric should assess the intended learning, not the appearance of a piece of writing or a suspicion about who wrote it. It can include the quality of an argument or method, appropriate use of sources, the explanation of decisions and the ability to revise a response in light of new information. For a foundational skill, teachers can also assess an individual demonstration without AI, if that skill is an explicit objective and students have an accessible way to demonstrate it.
Asking for a draft or reflection does not, by itself, guarantee that work is authentic. Those materials are useful when they form a normal part of the assignment and help assess decisions tied to the objective. Explain in advance what is required, why it is required and how it will be assessed, without demanding unnecessary data or private records.
A short rubric centred on learning
| Criterion | What the student can show | What should not replace it |
|---|---|---|
| Reasoning | Explains how they reached a conclusion or chose a method. | A correct final answer without an explanation when the process is part of the objective. |
| Use of sources | Connects claims to relevant documents and checks their content. | An automatically generated bibliography that has not been reviewed. |
| Critical revision | Identifies errors, omissions or limits and explains specific changes. | The claim that a response “looks right.” |
| Authorship and decisions | Discloses the assistance used and identifies their own decisions. | A text detector used as a verdict on authorship. |
Activities for learning about generative AI
AI literacy activities can teach students to evaluate outputs without requiring personal accounts or real personal data. UNESCO proposes that students develop critical judgement and the knowledge to participate responsibly. The suggestions below are adaptable teaching examples, not guarantees that a particular activity will improve grades.
In secondary school, a class can analyse a short AI-generated explanation of a science question and check each claim against the textbook. The resulting table could have three columns: claim, check and decision. Another activity is to compare two explanations of the same concept and ask students which is clearer, what it leaves out and which source they would consult before accepting it.
At university, a group can compare an answer about a disciplinary concept with primary sources or course materials and flag unsupported claims. In a coding activity, students can review a suggested solution in a safe environment, explain what each part does and test edge cases without running unknown code on a device containing personal data.
One adaptable activity is to review a paragraph for generalisations or biased language. The class should justify its changes using explicit criteria and avoid sharing personal information. Another option is to discuss which tasks in a discipline require practice without AI and which might benefit from its use, connecting the decision to learning objectives.
In every case, success is defined by an observable product: a checked claim, an explained error, a justified revision or a defended rule for use. If no approved tool is available, a teacher can prepare a fictional sample or write one themselves. Students do not need to create an account to critically examine a response.
Activity: check an answer about a course topic
- 01The teacher poses a question linked to a specific objective and provides a short answer, generated or prepared as an example.
- 02Each student separates the answer into checkable claims and notes what source would be suitable for each one.
- 03Students consult course materials or relevant primary sources and record whether each claim is supported, corrected or unresolved.
- 04Each student submits a checking table and a reflection explaining what they accepted, changed and why.
Use policies and student disclosure
A useful policy explains the purpose of the rule and specifies how it applies to each assignment. It should say which uses are allowed, which require disclosure and which conflict with the learning being assessed. It is also helpful to clarify which tools the institution has approved, what data must not be entered, how to request an alternative and how questions will be handled. A clear rule is more useful than an ambiguous ban that every student interprets differently.
The following models are illustrative examples, not official policies. Adapt them to course objectives, individual assignments, approved tools, and local rules and procedures. In all three cases, do not enter personal, sensitive or confidential data, or another person’s work, into a tool unless the institution has expressly authorised that processing. The person submitting the work remains responsible for reviewing it: check results and original sources, and do not present a claim as verified if you have not been able to confirm it.
Broad permission. Purpose: to learn how to use AI with judgement and as support without replacing the learning being assessed. You may use AI tools to explore ideas, organise your work, ask for explanations, get feedback or review a draft, unless an assignment’s instructions set a specific limit. You may not present generated answers as your own or submit unchecked claims or references. In a note accompanying your work, identify the tool you used, what you used it for and which contributions you kept or rejected; cite reproduced outputs and original sources in the style specified for the course. If you lack access, need an accessible alternative or have questions about a use, contact the teacher through [course contact method]. Undisclosed uses or uses that break these rules will be handled under the institution’s academic integrity rules and applicable procedures; a detector alone will not determine the outcome.
Uses limited by assignment. Purpose: to practise without assistance the skills targeted by particular activities, and to learn to evaluate AI when its use is part of learning. For [assignment or stage], you may use AI for [permitted uses, such as generating study questions], but not to write, solve or revise [the part that demonstrates the assessed skill]. For [another assignment or stage], you may use it for [permitted uses] and must disclose the tool, purpose and contributions you kept; cite any reproduced output and original sources as directed for the course. Do not upload personal, sensitive or confidential data, or someone else’s work, to services not approved for that use. Check results and references: you are responsible for what you submit. For questions, accessible alternatives or access problems, contact [teacher or course contact method]. Violations will be addressed under the institution’s applicable rules and procedures, without automatic penalties based solely on a detector.
Prohibition of specified uses. Purpose: to directly assess individual learning of [skill or knowledge]. For [identified assignments], do not use generative AI to produce, rewrite, translate or solve the assessed part, and do not present generated content as your own work. The prohibition applies only to those uses and assignments; any classroom activity involving AI must be expressly identified by the teacher. If an authorised activity includes a tool, disclose which one you used and how; cite reproduced outputs and original sources as directed for the course. Do not enter personal, sensitive or confidential data, or another person’s work, into unauthorised tools. Check the results and sources of any material you are allowed to use: you remain responsible for the work you submit. For clarification, contact [teacher or course contact method]. If the prohibition or assignment format creates an access barrier, request an accessible and equivalent alternative before the deadline. Any breach will be handled under the institution’s academic integrity rules and applicable procedures.
A short disclosure should explain the process, not become a fictional bibliography. For example: “I used an AI tool to suggest possible titles and review the clarity of two sentences. I checked the claim about the topic in the assigned chapter, rejected an unsupported suggestion and wrote the final argument. The AI did not write the analysis.” Students should still cite the original sources they used under the course rules; disclosing a tool does not replace citing a book or article.
The policy should match actual assessment practices. If an assignment prohibits a use, explain which learning requires completing it without that assistance and what alternatives are available when access is a barrier. If an assignment allows or encourages AI, specify how students should disclose assistance and which parts they need to check. Consequences for a violation depend on the institution’s rules and procedures; do not promise automatic penalties or replace the institutional process with a detector score.
Privacy, access and protecting minors
Before using a tool, the institution should assess who provides it, what data it receives, how long it retains that data, whether it can use the data for training and what processing terms apply. Do not enter identifiable student data—such as names, grades, health or disability information, images or another person’s work—into services the institution has not approved for that purpose. Removing a name does not always prevent someone from being identified through context.
The rules depend on the country, type of institution, purpose and service. FERPA governs certain education records in the United States and is not a worldwide law. In Spain, the Spanish Data Protection Agency explains that schools and education authorities have responsibilities for the technology used in educational activities. That guidance on online classes and exams does not by itself authorise uploading student records to any AI tool. The institution should consult its data protection officer and review the provider’s terms.
There is no universal age at which every service can be used in every context. Conditions may depend on the platform, jurisdiction and institutional rules. Do not require students to create personal accounts or share data to complete an assignment; prepare an accessible, equivalent alternative. An education plan or a commercial promise does not, on its own, show that data processing is appropriate.
Accessibility is also part of the decision. An activity should not require a paid account or personal device if those requirements are not necessary for the learning being assessed. The institution can provide authorised access, prepare an example to analyse in class or allow an equivalent alternative. Consider cost, support needs, age and whether students can participate without handing over private information.
Practical checks before using a tool
| Question | What to check | Prudent response if unclear |
|---|---|---|
| Is it approved? | Institutional authorisation and the defined purpose of use. | Do not enter student data until the institution provides guidance. |
| What data does it collect? | Types of information, retention, deletion and use for training. | Use fictional data or an alternative that requires no data sharing. |
| Who is responsible? | The roles of the institution, provider and data protection officer. | Consult the responsible officer before processing identifiable information. |
| Can everyone participate? | Access, cost, age, accessibility and whether a personal account can be avoided. | Offer an equivalent activity that needs no account or private-data upload. |
AI detectors: an uncertain signal, not proof of authorship
Detectors classify text using statistical patterns and can produce false positives and false negatives. A score does not establish who wrote a text or prove a violation. Detecting similarities to existing texts is also different from detecting authorship: each method answers a different question.
A 2023 study by Liang and colleagues evaluated seven detectors using specific samples of English essays. The team observed frequent false positives in the TOEFL essays they studied and showed that linguistic changes could affect classification. The result describes those detectors, samples and research conditions; it does not measure the performance of every current tool or predict a rate for a different group or language.
The University of Nebraska–Lincoln advises against using available detectors on their own to validate student work and highlights the risk of errors. If a concern arises, teachers should follow institutional procedures, listen to the student and consider relevant evidence, such as drafts that are a normal part of the assignment, sources used or an explanation of decisions. Do not upload work to an unapproved service, and do not impose an automatic penalty based on a score.
A respectful conversation gives a student the chance to explain how they developed the work and respond to a specific concern. The teacher should distinguish a statistical signal from relevant evidence and follow the institution’s academic procedures. This caution supports both academic integrity and fairness: an automated classification alone does not prove authorship or intent.
Verify claims, sources and results
Students and teachers can check a response with a straightforward process: separate out verifiable claims, locate the original source, confirm its authorship and date, and check that the document supports exactly what is being claimed. For scientific information, review the study or responsible institution; for a rule, check the official text and applicable jurisdiction. Even a complete-looking reference may be invented: confirm its title, author and identifier in the original record.
When a decision depends on a calculation, redo it using an appropriate method. If a tool suggests code, review and test it in a safe environment, using test data and without running unknown instructions on personal files. If the response uses search or document retrieval, open the cited documents and check that the wording reflects their content and scope.
The number of sources needed depends on the claim’s importance and the quality of the primary source. A genuinely independent second source may help check a consequential fact, but several webpages repeating the same claim do not count as independent checks. If documents do not resolve an issue, retain the uncertainty rather than filling it with a convincing answer.
Verification should also fit the kind of claim. Check a course definition against the relevant teaching material; a scientific claim against the study or responsible institution; and a rule against the official text that applies in the relevant jurisdiction. Checking only a source’s title is not enough: confirm that its content supports the full claim and its conditions.
A checking routine for an important claim
- 01Write down the specific claim to check; do not treat a whole answer as if every part has the same support.
- 02Find the original document or a competent institutional source and confirm authorship, date, context, and applicable market or jurisdiction.
- 03Check that the source supports the full claim, including relevant conditions and limitations.
- 04Repeat calculations or technical tests using an appropriate and safe method.
- 05Record what is confirmed, what is corrected and what remains unresolved.
Checklist before assigning work
Before publishing an assignment, teachers can check whether the learning objective is clearly stated; whether AI use is defined for that task; whether students have an accessible alternative; and whether the institution has approved the tools and data handling. They should also make sure the instructions explain how to disclose use, which sources to check and how the work will be assessed.
Next, check that the rubric measures the reasoning or skill the course intends to teach and does not rely on a detector score. If students must provide evidence of process, it should be relevant, proportionate and explained in advance. A policy should also provide a way to resolve questions without assuming bad faith.
To get started, a teacher can redesign one assignment and see whether the requested evidence gives a clearer picture of what students can do. Review the quality of the observed learning, access, privacy and any problems that arise; no tool replaces pedagogical judgement.
Students can also use the policy as a guide before submitting work: check which uses are allowed, disclose any required assistance, review claims and cite original sources. If an instruction is unclear or a format creates an access barrier, ask the teacher in advance and request an equivalent alternative.
Final assignment-preparation checklist
| Check | Yes / no |
|---|---|
| The learning objective is clear and demonstrable. | |
| The policy specifies which AI uses are allowed or limited for this assignment. | |
| An accessible alternative is available without a personal account or private-data upload. | |
| The institution has approved the tools and data processing. | |
| The rubric assesses learning, sources, reasoning and critical revision. | |
| Students know how to disclose use and ask for help. | |
| A detector is not used as sole evidence or as a verdict. |
Keep exploring
Sources consulted
Corrections and transparency
If you spot incorrect or outdated information, send us a correction with the page and source we should review.
Submit a correction