An instructor asks AI to rewrite a lesson in plain English. The result reads beautifully. It also changes a requirement from “must” to “may.”
Nothing about the formatting reveals the problem. There are no obvious spelling mistakes. A learner who follows the revised version could come away with the wrong understanding.
This fictional example captures an everyday challenge with generative AI: fluent writing can make a result feel finished before anyone has checked it.
Using AI safely means matching the tool to the task, protecting the information you give it, verifying important outputs, and keeping a person responsible for consequential decisions.
For instructors and small teams, the goal is a workable review habit. You should be able to explain what the AI contributed and why the final result is fit to use.
Decide how much a mistake would matter
Start with the consequence of an error.
Brainstorming five possible titles for a workshop has a relatively small downside. Drafting an explanation of tuition refunds, generating an answer key, or deciding whether someone meets a prerequisite has a greater one. Instructions involving clinical care, equipment operation, or personal safety need review by someone qualified in that subject.
Before starting, complete this sentence: “If the answer is wrong, someone could ___.”
That sentence tells you how much checking the task needs. It may also tell you that the decision belongs with a person or an authoritative source rather than a general-purpose chatbot.
NIST’s Generative AI Profile treats risk as dependent on the use case and its consequences. It identifies confidently presented false information—often called hallucination or confabulation—as one of the risks organizations need to manage.
A tool can be useful for a low-stakes draft and unsuitable for making the final decision in a high-stakes case.
Set an information boundary before writing the prompt
A prompt can disclose information just as an email attachment can. Think about what is in the document, who it concerns, and whether your organization has approved that use.
For routine drafting, use fictional examples or the smallest amount of information needed. An AI tool does not need a real student’s name, account balance, identification number, and personal circumstances to help you write a general payment-reminder template.
Removing a name may still leave identifying details. A rare diagnosis, a specific incident, a course date, and an employer name can point to one person. Review the whole context, including screenshots and attachments.
Harvard’s guidance on generative AI use separates approved tools, sensitive information, connected permissions, and output review. Its institutional rules apply to Harvard; other schools need their own approval process. The general lesson is to evaluate the actual account and features being used.
Before allowing real student or customer data into a tool, establish how the service handles retention, access, sharing, and model training, and whether that setup meets your organization’s requirements. A paid account or a privacy-related setting alone does not answer every question.
Give the AI a task you can check
“Make this better” is difficult to review because it leaves the task open-ended. A bounded request makes changes easier to inspect.
For an approved, non-sensitive course outline, you might ask:
Rewrite the explanation for an adult learner encountering this topic for the first time. Preserve the requirements, numbers, and exceptions. Use only the source text below. List any statements that are unclear instead of supplying missing facts.
The prompt is an instruction, not a guarantee. Compare the output with the original even if the tool says it followed every requirement.
A useful division of work is to let AI propose structure or wording while a subject expert checks meaning. Keep the original beside the draft. Pay particular attention to dates, thresholds, exclusions, definitions, and the difference between a recommendation and a requirement.
If a document changes regularly, record which version you used. A careful summary of an outdated source can still be wrong for today’s learner.
Verify the evidence behind a confident answer
Ask for sources when a claim needs support, then open them. Check that the source exists, says what the answer claims, applies to your setting, and is current enough for the question.
A real link can still be attached to an unsupported statement. Two chatbots agreeing with each other does not replace checking the underlying evidence. They may share similar source material or produce similar mistakes.
For a school policy, consult the approved policy. For a course requirement, use the responsible authority’s current materials. For calculations, work through the arithmetic independently.
When generating quiz questions, review more than the answer key. Check whether the wording is unambiguous, whether more than one option could reasonably be correct, and whether the question measures what you actually taught. An explanation can sound convincing while rewarding the wrong answer.
Build review into the time you allocate. A draft produced quickly still needs enough attention to become useful teaching material.
Connected AI tools need a second kind of review
Some tools can read files or take actions in other applications. In those cases, also ask what the tool is allowed to access and change.
Start with the narrowest permissions needed for the job. If the task is summarizing a document, it may not need permission to send messages or modify records. Keep approval before actions such as publishing a policy, sending a message to a class, or changing a financial record.
Harvard’s guidance also highlights prompt injection: instructions hidden in content that attempt to redirect an AI system. A document being summarized can contain text that tries to make a connected tool disclose data or act outside the user’s request.
For a small team, the practical response is to limit access, preview proposed actions, and keep a clear route to report unexpected behavior. “Only follow my instructions” is not a complete security control.
Make the human role clear to learners
Education researchers Fengchun Miao and Wayne Holmes, authors of UNESCO’s guidance on generative AI in education and research, emphasize human agency, privacy, and educational suitability.
Apply that thinking to the learning activity itself. What should the student be able to do after using the tool? How will they demonstrate understanding? What assistance is allowed, and what must they do independently?
For example, an instructor could allow AI to generate a practice scenario, then ask learners to identify its assumptions and justify their own response using course materials. That gives the tool a clear educational purpose and keeps the learner’s reasoning visible.
Explain your expectations before the assignment. Students should not have to guess whether brainstorming, translation, editing, and answer generation are treated the same way. Offer an appropriate alternative when a tool is inaccessible or cannot be used within your school’s requirements.
Keep a short record of the review
You do not need to retain every experiment. For material you will teach, publish, or use in an important decision, a short record can answer the questions that matter later:
- What task did the tool assist with?
- Which approved source or document version was used?
- Who checked the result, and what did they correct?
- Where is the final approved version?
If an error reaches learners, correct it where they received it. Explain what changed and whether they need to take any action. Quietly replacing a file may leave people relying on the earlier version.
If sensitive information was entered into an unauthorized tool, notify the person responsible for privacy or security promptly. Follow the tool’s deletion options and your organization’s response process, while recognizing that removing a visible chat may not establish that every retained copy has been deleted.
Begin with one repeatable use
Choose a modest task such as rewriting a public course description. Keep private information out of the input, preserve the original, and have someone check each important detail before publication.
Notice where the tool helped and where it created extra work. Use that experience to decide whether to repeat the process, change the prompt, or choose a different task.
For Steams Online’s community of training providers, AI literacy should make instruction and administration more dependable. A useful result is one your team can explain, verify, and stand behind.
Sources and further reading
- NIST: Artificial Intelligence Risk Management Framework — Generative Artificial Intelligence Profile, 2024.
- Harvard Information Security and Data Privacy: Best Practices — Using Generative AI.
- Fengchun Miao and Wayne Holmes / UNESCO: Guidance for generative AI in education and research, 2023; publication page updated in 2026.
