Product news: ChatGPT's connected Drive Library and Claude Code 2.1.235 make source and permission review visible

This product-news report connects three current approaches to source-grounded AI work. OpenAI's August 13 ChatGPT update brings connected Google Drive content into Library. Anthropic's August 18 Claude Code 2.1.235 release tightens prompt, permission and notebook review behavior. Google positions Gemini for Education across teaching, learning and institutional work. The shared educational issue is whether users can identify the evidence an AI used, the authority it received and the review required before its output becomes course or research work.

OpenAI says users with the Google Drive plugin connected can browse My Drive and files or folders shared directly with them from ChatGPT's Library. A file can be added from the composer or with an @ mention, and supported Google Docs, Sheets and Slides can remain open beside the conversation while ChatGPT summarizes, compares or creates material. Where supported and authorized, ChatGPT can update the source file. The rollout covers Plus, Pro, Enterprise, Edu, Healthcare and Business users on the web, but it initially excludes Shared Drives and some collaboration features.

For education, convenience should not erase provenance. A cited Drive file may itself be outdated, shared under the wrong permissions or detached from a course's canonical version. When an AI can write back, the difference between reading, drafting and changing a source also becomes consequential. A defensible workflow should record the file owner, version, sharing scope, requested operation, generated change and person who accepted it.

Claude Code 2.1.235 addresses related controls in an agentic coding environment. The release fixes full prompt-cache invalidation when a language server reconnects, reducing unintended loss of working context. It corrects a keyboard path that could approve an edit and grant session-wide permission when a user meant to close a comment field. Notebook approval dialogs now explain when existing cell content cannot be read instead of silently omitting it. Permission dialogs more closely match the scope of the grant, and large cross-session messages fail explicitly rather than disappearing. Optional prompt spellcheck and lower background cloud-session resource use are smaller usability additions.

These changes do not prove that an agent's code or analysis is correct. They make the review boundary more legible. Students and researchers should still inspect the exact diff, run an independent test and separate a session's cached assumptions from evidence in the repository or dataset.

Gemini for Education supplies the institutional comparison. Google describes educators using it to plan lessons, differentiate materials and create assessments; learners using it for explanations, practice and writing feedback; and staff using it for communications, administration and research. Google also describes administrator controls and education data that is not human reviewed or used to train AI models. Those are vendor-described product and data boundaries, not independent evidence of learning effectiveness.

For Hong Kong schools and universities, the practical response is a source-and-permission ledger. Record each connected source, its owner and version, what the AI was allowed to read or change, which model or workspace acted, and the human check that followed. Connected tools can reduce friction, but educational trust depends on keeping evidence, authority and acceptance visible.