Should teachers worry about students using AI text generators?

A growing number of writing apps now run on large language models, programs trained on huge amounts of text to predict the next word. OpenAI reported in March that more than 300 applications were already using GPT-3 through its API, in categories that include education, productivity and creative writing. Some of these tools can turn a one-line topic into several paragraphs in seconds. It is reasonable for a teacher to ask what that means for the essay assigned on Friday.

The short answer is that the concern is real but manageable. The tools are imperfect, often wrong on facts, and still not available to everyone. The longer answer is that they change which assignments measure learning well, and that is worth thinking through now, while the tools are still new.

What the tools can and cannot do

GPT-3 writes by continuing a prompt with the words that seem most likely to come next. That makes it good at fluent, confident, generic prose. It does not check facts, it can contradict itself over long passages, and it can repeat itself. The research paper that introduced GPT-3 lists these weaknesses openly.

The same paper contains a finding teachers should know. Researchers showed people short news articles of about 200 words, some written by journalists and some by GPT-3, and asked which were machine-written. For the largest model, people were right only about 52% of the time, barely better than a coin toss. A short student response is not a news article, but the lesson carries over: reading closely is not a reliable way to detect machine text.

The authors also list "fraudulent academic essay writing" among the possible misuses of models like theirs, next to spam and misinformation. So the concern is not imagined; the people who built the model raised it themselves.

An honest look at misuse

There are real limits on misuse today. Access to GPT-3 goes through OpenAI's API, which has a waitlist, and OpenAI says it reviews applications before they go into production and requires safeguards such as rate limits and user verification. Still, once an app is approved, anyone who signs up for it can use it, and text completion is what every one of these apps is built on.

The weak spots of these models suggest where assignments are most and least exposed:

None of this requires software to "catch" anyone. It mostly means asking for the thinking behind the writing, not just the finished page.

A worked example: critiquing an AI draft

One of the most useful classroom activities turns the tool into the object of study. The teacher generates a draft in advance, with a prompt like this, and brings it to class:

Write a 200-word persuasive paragraph for a high-school audience arguing that cities should ban cars from downtown streets. Include two statistics and one quotation from an expert.

Students then work in pairs with the printed paragraph:

1. Highlight every factual claim, especially the statistics and the quotation.
2. Try to verify each one using the library or reliable websites. A model that predicts likely words can produce numbers and quotes that look plausible but do not exist, and finding one is memorable.
3. Mark the reasoning: where does the paragraph assert something without support?
4. Rewrite it with real evidence, and write two sentences on what they changed and why.

The students practise exactly the skills a generated draft lacks: sourcing, checking and judgment.

How can a classroom use them openly?

Start with a clear, written policy for each assignment: allowed, allowed with disclosure, or not allowed. Students who know the rule are far less likely to drift into misuse, and a disclosure rule ("paste the prompt you used and mark the parts you kept") turns the tool into something you can discuss rather than hide.

Then use the tools where they help thinking rather than replace it. A generated paragraph can be a counter-argument to rebut, a first draft to critique, or a set of questions to answer in your own words. Comparing a model's summary of a reading with the reading itself shows students what gets lost or distorted. Asking for three different openings to the same essay and choosing the best one, with reasons, teaches revision.

Keep the parts that matter most in plain view: brainstorming in class, drafts with feedback, short conferences where students explain their choices. These were good teaching practices before language models, and they are also what makes a classroom resilient to them. Finally, talk with students about why the tools get facts wrong. Understanding that a model predicts likely words, rather than looking things up, is itself a useful lesson in how to read anything online.