AI + Education Weekly Update
Week of November 24th, 2025
To catch up on another busy week in AI + Education news, check out our weekly updates for our review and analysis of some of the key highlights worth staying informed about.
Week of November 24th, 2025
Google's Updates: Nano Banana Pro & NotebookLM
Google’s Nano Banana Pro and NotebookLM bring professional-level AI image generation and AI-powered infographics and slide decks that are transforming daily workflows in creative work and education.
Google's Nano Banana Pro: State-of-the-art image generation that produces clean, legible text directly in images (in multiple languages), maintains visual consistency across up to 14 blended images, keeps character appearance consistent across scenes, and outputs in 2K and 4K resolution for professional use.
NotebookLM’s Infographics & Slide Decks: AI automatically matches your deck's existing design style, transforms complex documents and PDFs into visual infographics, and generates graphics directly in your presentation.
OpenAI's ChatGPT for Teachers
OpenAI launched ChatGPT for Teachers—a free, FERPA-compliant workspace for K-12 educators through June 2027, featuring GPT-5.1 Auto, unlimited messages, Google Drive/Microsoft 365 integration, and district-level administrative controls.
What Teachers Get: Unlimited messages with GPT-5.1 Auto (OpenAI's most capable model), file uploads, web search, image generation, and connectors to Google Drive/Microsoft 365—enabling seamless integration with existing classroom workflows
Education-Grade Protection: FERPA-compliant with data protections that meet district compliance requirements. Teacher and student data is NOT used to train OpenAI's models, ensuring classroom information remains private.
Administrative Features: School and district leaders can create shared workspaces with single sign-on (SSO), role-based access controls, domain claiming, and usage analytics—allowing centralized management while giving teachers autonomy
AI Chatbots for Mental Health Support
Common Sense Media released a comprehensive risk assessment finding major AI chatbots unsafe for teen mental health conversations.
Most Common Teen Use Case: Mental health support is one of the most common and dangerous ways teens use AI (nearly 3 in 4 use it for companionship/emotional support); while chatbots improved at detecting explicit suicide prompts, they fail to recognize non-crisis conditions affecting 20% of youth—anxiety, depression, eating disorders, ADHD, and psychosis.
Why They Fail: Chatbots lack clinical judgment to recognize warning signs, get sidetracked by tangential details, miss "breadcrumbs" (subtle warning signs), and perform worse in extended conversations—exactly when teens need consistent support.
The Dangerous "Trust Transfer": Teens who find AI helpful with homework incorrectly assume equal competence for mental health guidance; systems lack clear disclosure about AI limitations and prioritize engagement (with follow-up questions) rather than directing teens to professional care
Kids Message AI 10X More Than Friends
New Aura research on children 8-17 finds kids engage far more deeply with AI chatbots than friends, raising questions about AI's impact on development.
Key Highlights:
Messages average 163 words to AI vs 12 words to friends — over 10X longer
Over one-third (36%) of conversations involve sexual or romantic roleplay — nearly 3X more common than homework help
Other themes: creative play (23.2%), homework (13.1%), emotional/mental health (11.1%), advice/friendship (10.1%), personal info (6.1%)
Based on 10,000+ users
Sycophancy in a Simulated Educational Context
Research demonstrates that AI tutors change their answers based on students’ suggestions rather than maintaining objective accuracy, potentially creating an equity problem where the AI helps knowledgeable students more than struggling ones.
When students mention incorrect answer choices in their questions, LLM accuracy drops by up to 15%; when students mention the correct answer, accuracy improves by 15%
Models 'flip' their answers 4-18% of the time to match student suggestions (varying by model)
Smaller, more affordable models show stronger sycophancy effects (GPT-4.1-nano at 30% vs GPT-4o at 8%)
