Should Students Learn AI Where an Adult Can See It?

Yvette Renteria, chief program officer at Common Sense Media, made an argument in EdSurge this week that every district debating an AI pause should read. She compares AI bans to abstinence-only sex education: telling students not to do something doesn't stop them from doing it. Her numbers back her up. In a nationally representative Common Sense study, 44% of teens said a tool had been banned at their school, and 59% of those teens said they used it anyway on a personal device.

Renteria supports bans as a way to buy time. Her conclusion is that schools have to teach AI literacy whether or not AI is allowed in the building. We agree.

Her own research points to a second conclusion, and it matters just as much for school leaders.

The failures Common Sense found happen when no adult is watching

Common Sense's Youth AI Safety Institute spent six months testing AI products built for teens, such as ChatGPT for Teens and Gemini with teen protections, alongside general-purpose chatbots. Renteria lists what they found. Leading chatbots missed signs of suicidal ideation, went along with signs of psychosis, gave incorrect information about eating disorders and completed every homework assignment they were given. ChatGPT for Teens promised to alert parents about dangerous chats, then sent no notification when testers simulated a crisis.

Each of those failures happened in a private conversation between a teenager and a consumer chatbot. No teacher could see the chat, and no adult in the student's life was told when it went wrong.

A ban leaves those private conversations in place. Students who use AI at home still use the same consumer chatbots, still alone. A ban only removes the version of AI that a school could have supervised.

AI literacy needs practice, and practice needs a safe place

Renteria describes what AI literacy should cover: deepfakes, why a chatbot acts like a friend when it is a language model, when AI helps creative work and when it does the work instead, and how to spot bias and misinformation in an AI answer. New York City's pause requires high schools to teach two 45-minute lessons on these topics each year. Renteria calls that a good start but not enough.

Students learn most of these skills by doing them. A student learns to catch a made-up fact by checking an AI answer against a source and finding it wrong. A student learns the difference between help and a finished essay by working on an assignment where the AI asks questions instead of writing paragraphs. A lesson about AI that never lets a student use AI is like driver's education without a car.

The safest place for that practice is a classroom, on a tool the school chose, with a teacher who can read the conversations.

What to require from the AI students learn on

Common Sense's findings work as a buying checklist. Whatever AI tool a school allows, students should only use one that fixes the failures Renteria describes:

• Crisis alerts reach a person at school, immediately. When a student mentions self-harm or suicide, someone responsible for that student is told right away. Ask vendors who gets the alert, how fast it arrives, and whether it can be switched off. It shouldn't be.
• The AI is not a student's friend. When a student starts treating it as one, it should point the student to a trusted adult and flag the conversation for school staff.
• Teachers decide how much help the AI gives. For each assignment, the teacher should be able to tell the AI to coach and ask questions rather than write the answer.
• Teachers can read every conversation. Students' chats should not be visible only to an administrator searching records after the fact.
• Teachers are trained first. A teacher who understands how these models work can teach students to question them. A teacher who doesn't will struggle to supervise them.

How Flint handles these risks

Flint is an AI platform built for schools. Here is how it handles each item on that list:

• Teachers can read every conversation students have with the AI in their classes and assignments. School administrators can read every conversation in the school.
• The AI is instructed to flag concerning messages as they arrive, including self-harm, violence, harassment and sexual content, and to err toward flagging when unsure. Schools can add their own categories.
• Self-harm flags are always emailed immediately to the student's teachers and to the school's moderators for the student's grade, even when the school has chosen hourly, daily or weekly emails for other flags. Self-harm alerts are on by default, and Flint's settings screens don't let anyone turn them off.
• When a student treats the AI as a friend or confidant, or asks it for personal advice unrelated to schoolwork, the AI is instructed to flag the message, encourage the student to talk to a teacher, counselor or parent, and offer to help with schoolwork instead. The school chooses which staff are alerted.
• Teachers write guidelines for each activity that the AI follows, such as "ask guiding questions, but don't write the student's thesis."

Use the pause to prepare

A pause gives a district time. Districts that use the time well will train their teachers, pick the tool their students will learn on, and test it with a small group of classes before the pause ends. Districts that only wait will reach the end of the pause with the same questions and a year less preparation.

Renteria ends her piece by asking schools to teach kids what AI is and why it acts like a friend when it isn't one. Students will learn that best by using AI in class, with a teacher who can read what the AI says and who gets an email as soon as a student mentions self-harm.