This week's reading keeps returning to safety and evidence: how easily chatbot guardrails can be bypassed, and how little proof there is that more screens or more AI help learning. There are practical takeaways here for staff conversations about student AI use, from teaching students to question chatbot answers to checking the tools you already have.
UNSW researchers found that large language models, including OpenAI's GPT-4 and GPT-3.5, were much more likely to answer harmful questions once manipulated into behaving as if intoxicated, whether by prompting or by retraining. The altered models were also significantly more likely to reveal confidential information.
Why it matters: Student-facing chatbots rely on built-in safety filters, and this work shows those filters are not a guarantee, so schools should pair any approved tool with supervision, clear rules and a plan for what happens when a student sees something harmful.
Experts in Australia and abroad say schools may have adopted AI faster than the research on its effects on learning. The piece cites PISA 2025 data showing students who used AI to draft writing performed worse, and notes restrictions such as New York City's moratorium for elementary and middle school and a NSW review of HSC integrity.
Why it matters: It is a handy summary of the evidence and policy pushback for anyone who has to justify (or pause) an AI rollout, and a prompt to ask what learning evidence sits behind each tool.
Macquarie University-led research found only 53% of NSW teachers think digital devices help students learn, and half say students are often distracted by screens. Some 84% of secondary students said screens distract them at least sometimes, compared with 57% of primary students.
Why it matters: With stricter NSW screen-time limits arriving in 2027, schools anywhere will be asked which digital activities earn their minutes, and the primary versus secondary gap suggests younger students may need different supports.
After about 4,000 test interactions, Common Sense Media found ChatGPT's study mode was easy to switch off and that parental alerts did not trigger for concerning topics such as self-harm. OpenAI disputed the methodology and said the testing did not reflect how its safeguards work.
Why it matters: Child-safety features on consumer chatbots should be tested rather than assumed, which matters when students use these tools at home for homework, as a survey cited in the report suggests 70% of teens do.
Digital Inquiry Group researchers gave 117 high school students a false, unsourced chatbot answer: about 3 in 10 trusted it and nearly half were unsure. None of the 17 who said they would follow up actually did. The authors say even a few hours of instruction improves students' digital savvy.
Why it matters: Students being confident with AI is not the same as being able to evaluate it, and a short, explicit lesson on checking chatbot claims is a realistic first step for any year level.
After US states reached a child-safety agreement with social media companies this summer, AI firms have begun adopting similar safeguards such as age verification and restricted experiences for minors. Experts say AI's educational benefits make it harder to limit than social media, and urge faster action.
Why it matters: It frames the debate schools are already having about bans versus guided use, and it is a reminder to ask vendors what age-appropriate protections they actually have in place.
A five-minute read for busy teachers: the week's most useful AI-in-education news, Australian stories first, each with a note on why it matters in your classroom. Free, and you can unsubscribe at any time.
