Copilot Pages Hands-On: Microsoft’s AI Notebook Edges Out Google, but Privacy Concerns Persist

Microsoft’s Copilot Pages has entered the AI note-taking arena, and after extensive hands-on testing, it’s clear that Google’s NotebookLM finally has a serious rival. But while Copilot Pages shines with its intuitive, flexible design and powerful language models, it also forces users to confront a critical trade-off: convenience versus data privacy.

Launched initially for Microsoft 365 Business subscribers and now available to anyone with a free Copilot account, Copilot Pages is a persistent, editable workspace where you can draft, research, and iterate with AI assistance. I put it through its paces with real-world projects—including compiling a comprehensive list of hummingbird species in Mexico—and compared it head-to-head with NotebookLM. The results reveal a tool that’s faster and more adaptable for everyday workflows, but one that also raises urgent questions about how our data trains the models that power these experiences.

How Copilot Pages Works: A Sidebar-Centric Approach

Copilot Pages lives inside the Copilot app (on Windows, macOS, iOS, and Android) or on the web. You can create a new page directly from the Copilot interface by switching from “New Conversation” to “New Page,” or by converting an existing AI chat into a page with the “Edit in a Page” option. The workspace is deceptively simple: a blank canvas in the center and a chat sidebar on the right.

From the moment you start typing, Copilot feels like a smart notebook. You can paste text, upload files, or simply begin drafting your own content. Highlight any text and a formatting toolbar appears—bold, headings, bullet points—letting you structure the page without breaking your flow. The real power, however, is the sidebar Ask box. Ask a question, and Copilot pulls answers from its broad knowledge base, displaying results in a modal. You can then apply, reject, or regenerate those suggestions directly onto the page.

This sidebar-first workflow is fundamentally different from NotebookLM’s three-panel layout. Google’s tool forces you to define sources upfront—documents, web links, or text snippets—and the AI chat only uses that bounded knowledge. Copilot Pages, by contrast, taps into the full might of Copilot’s generative models (backed by OpenAI’s GPT technology) for every query, offering a broader, more fluid research experience. For exploratory tasks, that freedom is a game-changer; for rigorous academic work, it can be a double-edged sword.

Hands-On: The Hummingbird Test and Real-World Use

During testing, the project that best illustrated the difference was my hummingbird species list. I asked both tools to produce a table of all hummingbird species found in Mexico, sorted by prevalence, with conservation status and region.

Copilot Pages returned a more complete list—over 50 species—and correctly organized them by region (e.g., Yucatán Peninsula, Pacific Slope). When I requested it to re-sort using eBird data, it seamlessly revised the table right on the page, requiring only a click on “Apply Revision.” The level of granularity and direct edit-in-place capability felt natural and efficient.

NotebookLM, in contrast, generated a table of just 17 species and substituted habitat (mountains, gardens) for region. Because NotebookLM restricts its answers to sources it initially discovers or you provide, the narrower output likely reflects the limitations of its source corpus. For a casual researcher, that incompleteness might be a dealbreaker. For a scholar who needs every claim traceable to a specific document, it’s a feature.

The editing experience also highlighted Copilot’s strengths. When I pasted a long article, Copilot stripped out ads and images (albeit slowly), then reorganized the text into bulleted lists on command. While that “organization” felt reductive for prose-heavy content, it demonstrated the AI’s ability to restructure information quickly—a boon for note-takers who want to distill meeting minutes or lecture notes into actionable outlines.

Feature Comparison: Where Each Tool Excels

To cut through the marketing, here’s how the two platforms stack up:

Copilot Pages’ core advantage is its flexibility: the AI doesn’t just answer questions, it becomes an active collaborator within the document itself. You can ask it to “rewrite this paragraph in a more academic tone” or “add a table of contents,” and it modifies the page content directly. NotebookLM keeps AI-generated outputs siloed in the chat panel, requiring manual copying.

NotebookLM fights back with structured learning features that students adore. The audio overview turns documents into a podcast-like conversation between two AI hosts—perfect for auditory learners. Mind maps and timelines provide visual aids that Copilot Pages lacks entirely. And NotebookLM’s source-lock gives teachers and compliance officers the assurance that no external data is slipping into student work.

The Privacy Divide: Training Models vs. Protecting Data

Perhaps the most consequential difference between the two services is what happens to your data after you close the page. NotebookLM explicitly states that it does not use user entries to train Google’s AI models. That promise is gold for businesses handling proprietary data, researchers with unpublished findings, or anyone wary of feeding the machine.

Copilot Pages, unfortunately, takes the opposite stance by default: Microsoft uses your contributions to improve its models. The setting is buried in the app’s privacy controls and can be disabled, but the opt-out nature means many users will unknowingly participate. During my testing, I toggled it off immediately for sensitive projects, but the responsibility falls on the user to know this exists.

For IT administrators, this default could be a compliance nightmare. If an employee pastes a confidential client document into Copilot Pages without adjusting settings, that data could theoretically influence model training. NotebookLM’s upfront refusal to train on user content simplifies legal review and third-party risk assessments. Until Microsoft moves to a “no training by default” policy for its consumer and enterprise tiers, this will remain a significant differentiator.

Accuracy, Hallucinations, and the Verification Problem

Both tools share the same generative DNA, meaning they can fabricate facts with unnerving confidence. In the hummingbird test, Copilot’s reliance on broader data (including eBird) produced more thorough results, but I still checked every entry against the official eBird taxonomy. A handful of subspecies were missing, and one species’ conservation status was outdated—reminders that AI outputs are drafts, not gospel.

NotebookLM’s constrained approach reduces the surface area for hallucination but introduces its own risk: if your source documents are incomplete or biased, the AI will faithfully reproduce those gaps. Neither tool provides sufficient source attribution in the final output. Copilot sometimes cites web links in the sidebar, but those can vanish after revisions. NotebookLM offers a clickable source chip within chat responses, yet those aren’t ported into the notes themselves.

Best practices are simple but essential:
- Always cross-reference AI lists against authoritative databases.
- Treat every Apply/Reject step as a mandatory editorial checkpoint.
- Maintain an audit trail by exporting the final version and noting which sources informed key claims.

Design Gaps: What Microsoft Must Fix

Copilot Pages is promising, but it ships with frustrating omissions. An “Add to Page” button in the sidebar would eliminate the clumsy Apply/Reject modal that disrupts flow. Image embedding—screenshots, diagrams, photos—should be table stakes in 2025, yet neither platform supports it. For visual subjects like biology or design, this is a critical blind spot.

Collaboration is another weak point. You can share a static snapshot of your page via a web link, but there’s no real-time co-editing, no comment threads, no version history. NotebookLM offers shared notebooks and public templates, making it far more suitable for group projects. If Microsoft wants Copilot Pages to be a serious knowledge workspace, it must add granular permissions and live editing.

Content cleanup also needs work. Pasting from the web leaves behind stray markup and takes too long to process. A dedicated “Paste as plain text” shortcut or automatic stripping would improve the experience immensely.

Who Should Pick Which? A Practical Guide

Choose Copilot Pages if:
- You need a fast, iterative drafting environment that feels like a smart notepad.
- You value direct editing where AI can rewrite your content in place.
- You prefer a broad knowledge graph over rigid source constraints for exploratory research.
- You’re comfortable managing privacy settings and verifying AI-generated facts.

Choose NotebookLM if:
- Source provenance and reproducibility are non-negotiable (e.g., academic papers, legal briefs).
- You want built-in study aids like audio overviews, mind maps, and timelines.
- You need collaborative notebooks with strict access controls.
- Privacy is paramount and you cannot risk accidental model training.

For most knowledge workers, a hybrid approach works best: use Copilot Pages for brainstorming and drafting, then switch to NotebookLM when assembling final, source-bound deliverables. This way you harness the speed of Microsoft’s models without sacrificing auditability.

Enterprise Implications: The AI Note-Taking Arms Race

Microsoft’s strategy is clear: weave Copilot into every corner of Microsoft 365, from Word to Teams to this new Pages canvas. The integration potential is staggering—imagine a future where your Copilot Page automatically syncs with a OneNote notebook or a SharePoint library. Google, on the other hand, is selling a walled garden of responsible AI, betting that educational institutions and compliance-heavy industries will prefer its transparent source model.

For IT departments, the immediate priority is data governance. Before rolling out Copilot Pages, administrators must:
- Disable model training tenant-wide via PowerShell or Microsoft 365 admin center.
- Configure DLP policies to block upload of sensitive file types.
- Educate users on the difference between the two tools and when to use which.

NotebookLM’s simpler compliance story may win over cautious CISOs, but Copilot’s superior language understanding could ultimately prove stickier for daily productivity. The winner will be whoever ships robust collaborative features and image support first.

Final Verdict: A Strong Contender with a Privacy Asterisk

Copilot Pages is the most intuitive AI-powered notebook I’ve used. It strips away the rigidity of source-first tools and delivers a truly conversational editing experience. The hummingbird test, the article restructuring, the seamless revisions—all point to a tool that understands how people actually work.

Yet that ease comes with strings attached. The default model-training setting is a corporate overreach that puts the burden on users to protect their own data. Until that changes, every recommendation must include a big, bold privacy caveat.

Google should be worried. Copilot Pages outclasses NotebookLM in raw flexibility and usability. But Microsoft should be worried too—because in an age of increasing data consciousness, a tool that silently trains on your work is a liability waiting to explode. The AI notes race is just getting started, and the real prize will go to whoever builds a canvas that’s both powerful and trustworthy.

Try Copilot Pages today, but first visit those privacy settings. And for anything that must be airtight, keep NotebookLM in your toolkit. In the world of AI note-taking, the smartest move is to use both—with your eyes wide open.