📚 Claude teaches school, 🖌 Gemini edits art, and 🛡 Claude shields your browser
AI Bear
August 28, 2025. Inside this week:
Claude handles 70% of professors’ admin work
Google turns a meme into a Photoshop-killing tool
Claude for Chrome slashes agent exploitation in real-world use
Claude runs university admin - professors still grade themselves
✍️ Essentials
Anthropic’s Claude is already being used by professors to manage admin: designing curriculums, sorting budgets and dashboards, writing syllabi, and filing reports. 57% of curriculum design, 13% of research work, and 7% of grading have AI involvement. But when it comes to assigning actual grades, trust drops fast. Claude handles paperwork, not decisions.
🐻 Bear’s take
This is automation where it works - behind the scenes, not in the spotlight. Claude is saving hours of drudgery, freeing professors to focus on real teaching. But grading is sacred, and for good reason - no one wants a bot deciding futures. This sets a pattern: AI can augment work, but humans still own the final say.
🚨 Bear in mind
Even when Claude handles paperwork, errors can propagate fast - if wrong assignments or misfiled data slip through, it could harm students or budgets. More importantly, half your workflows are automated—but you're still on the hook for mistakes. Always double-check the output, even for small tasks.
Google’s Gemini Flash 2.5 - jumping from meme to design toolbox
✍️ Essentials
Google transformed its “nano-banana” experiment into Gemini Flash 2.5 Image, an AI editor that outpaces Photoshop in speed and character consistency. Users can swap objects, tweak scenes, and preserve styling seamlessly - all priced at $0.039 per image.
🐻 Bear’s take
What started as a meme is now an editing disruptor. Brands and creators can no-code complex edits faster than before, cutting back design overhead or agency costs. For design teams, it’s time to move from manual edits to creative direction.
🚨 Bear in mind
But easy edits don’t always mean smarter content. Overusing this tool risks building visual sameness - cookie-cutter visuals everywhere. Maintain your brand’s voice by using AI tools strategically, not just to automate everything.
Claude for Chrome fights prompt injections in the wild
✍️ Essentials
Anthropic tested Claude as a Chrome extension, watching how it handled hidden prompts on real websites. In an initial test, Claude failed (performed risky actions) in 23% of cases - but with new sandboxing and DOM filters, failures dropped to 11%.
🐻 Bear’s take
This is how agents should be deployed - real-world trials before mainstream rollout. Claude now recognizes suspicious prompts, avoids DOM-based hacks, and takes fewer blind actions. That’s how you earn trust, not just hype “AI that automates your web.”
🚨 Bear in mind
Even 11% failure is too high for real users. Browser agents hold keys to keys - passwords, workflows, payment pages. No tech should be shipped without facing real adversarial prompts. Expect policy and bug-bounty teams to build full stop-gaps.
Quick Bites
Perplexity lawsuit - Nikkei and Asahi Shimbun sue over content use and revenue sharing issues
AI school comp - The White House kicks off an AI challenge for K-12 students to solve local problems
Translate upgrade - Google Translate now offers real-time TV subtitle-style translation for 70+ languages