Microsoft gave Windows Insiders in the Canary Channel a new reason to right-click: on September 8, 2025, Build 27938 delivered AI-powered image editing directly within File Explorer. The experimental update also resurrects a long-requested seconds display in the Notification Center clock and adds a new transparency pane for generative AI usage — all under one flight. While none of the changes are monumental on their own, together they signal a clear push to weave AI into the very fabric of Windows shell and workflows.

What’s inside Build 27938

The Canary build packs three visible feature areas, plus a collection of under-the-hood fixes and server-side rollouts:

  • AI actions in File Explorer: Right‑click any .jpg, .jpeg, or .png image and you may now see an “AI actions” submenu. It groups four one‑click shortcuts: Visual Search (Bing), Blur background, Erase objects, and Remove background. The first is a web‑based visual lookup; the other three rely on Photos or Paint to perform quick edits.
  • Seconds clock in Notification Center: A new toggle inside Settings > Time & language > Date & time allows users to show time including seconds in the Notification Center flyout, restoring a capability that many power users missed.
  • Recent activity for generative AI: Under Privacy & security > Text and image generation, a new page logs which apps have called Windows‑provided generative AI APIs in the last seven days, giving users and admins a window into on‑device AI usage.
  • Miscellaneous fixes: File Explorer stability improvements, thumbnail rendering enhancements, and the usual Canary‑grade bug fixes.

A closer look at AI actions in File Explorer

How it works

The context menu entry is a shell hook. Depending on the action, it either launches Microsoft Photos or Paint with a pre‑scripted edit, or calls a Windows platform API — possibly on‑device or cloud‑assisted — and hands the result back to Explorer. Microsoft is mixing local and cloud inference: on Copilot+ hardware with a capable NPU, some generative tasks may stay entirely on the device; otherwise, they may be uploaded to Microsoft’s cloud endpoints. The critical catch: the current UI rarely tells you which pipeline is active, a transparency gap that privacy‑conscious users will want to close.

Supported formats and current limits

  • Formats: Only JPEG/PNG are confirmed. RAW, HEIC, PSD, or other professional formats are not yet supported in this flow.
  • Operation scope: Designed for quick micro‑tasks, not for high‑resolution or multi‑layer jobs. Attempting a complex erase on a 50‑megapixel TIFF will likely falter or prompt you to open a full editor.
  • Feature gating: As with most Canary builds, Microsoft enables these actions via server‑side feature flags. Two machines running the exact same build may see entirely different context menus.
  • App dependencies: The actions depend on up‑to‑date versions of Photos and Paint from the Microsoft Store. If either app is outdated, the menu may not appear.

Practical workflows

  • Quick privacy scrub: Right‑click a screenshot containing a license plate or a face, choose “Erase objects,” and the generative fill will attempt to remove the element without ever launching a heavy image editor.
  • Thumbnail background removal: For a web thumbnail, “Remove background” (via Paint) cuts out the subject in one step, saving the hike through Paint’s full interface.
  • Visual research: Right‑click any image from your downloads folder and “Visual Search” fires off a Bing Visual Search query, instantly showing you similar images or product listings.

Privacy and transparency — the critical tradeoffs

Build 27938 introduces a “Recent activity” dashboard under Privacy & security > Text and image generation. It logs which apps used Windows’ generative AI APIs in the last seven days, but that log does not reveal whether the inference ran locally or in the cloud. For users handling sensitive files, this is a significant blind spot.

Before relying on the AI actions for anything confidential, ask:

  • Is the image being uploaded to a Microsoft cloud endpoint for processing?
  • If so, how long is it retained, and is it used for model training?
  • What metadata is recorded in OS telemetry beyond the activity logs?

Microsoft’s hybrid execution model makes it difficult to answer these questions without testing. Early testers on WindowsForum recommend starting with non‑sensitive throwaway files and cross‑checking the activity pane immediately after each action to gauge what gets logged.

The return of the seconds clock

For anyone who has pined for the seconds hand on the taskbar clock, Build 27938 doesn’t put it there — but it does the next best thing. A new toggle under Settings > Time & language > Date & time, labeled “Show time in the Notification Center,” expands the Notification Center flyout to display a larger clock with seconds alongside the calendar. It’s a subtle quality‑of‑life upgrade, particularly useful for users who need to monitor exact time during presentations, time‑sensitive tasks, or remote collaboration.

Enterprise and IT admin implications

Canary builds are not for production, but this build gives IT teams a preview of controls they’ll eventually need. Key points:

  • Policy controls are absent today. No Group Policy or MDM settings exist to disable AI actions or restrict which apps can call the generative AI platform. Admins should anticipate future policies similar to those Microsoft introduced for Copilot.
  • Licensing entanglements. Some AI features — especially those that hit cloud APIs — may eventually require Microsoft 365 or Copilot subscriptions. Test now to understand the dependency.
  • App compatibility. Shell extensions and LOB apps that interact with file context menus could conflict with the new AI actions. A rigorous compatibility test on a clean VM is essential.
  • SIEM and audit. The Recent activity page is visible only locally. Enterprises that need centralized logging must determine if these events will be surfaced to Windows event logs or Defender for Cloud Apps.

Microsoft is likely to ship administrative templates in future builds. For now, IT should use this flight to define internal policies on OS‑level AI use and to educate users about the data flow risks.

Known issues and rollout quirks

  • Server‑side gating: Features appear and disappear unpredictably. Don’t assume a missing capability means it was pulled; it may simply not have been rolled out to your device.
  • Canary instability: Driver crashes, localization gaps, and accessibility issues are common. Screen reader support for the new context menu entries is especially patchy.
  • Batch operations can thrash the system: Triggering background removal on dozens of files simultaneously can spike CPU and GPU usage. On older hardware, expect noticeable lag.

How to safely test Build 27938

Power users eager to try the new features should follow a deliberate safety protocol:

  1. Never run on a primary machine. Use a spare laptop, a virtual machine, or a non‑critical device.
  2. Snapshot everything. Take a full disk image before upgrading so you can revert if the build corrupts drivers or essential software.
  3. Update Store apps first. Open the Microsoft Store, download the latest Photos and Paint, then reboot.
  4. Right‑click a test image. Look for “AI actions” in the context menu. If absent, the feature hasn’t been enabled for your device.
  5. Check the activity log. After running an edit, immediately visit Settings > Privacy & security > Text and image generation to see what was recorded.
  6. Monitor system performance. Use Task Manager to watch for unusual CPU, GPU, or disk spikes during AI operations.
  7. Send feedback. Press Win+F to open Feedback Hub and describe any bugs, latency issues, or ambiguous data‑handling indicators.

Microsoft’s broader strategy: AI as an OS primitive

Build 27938 isn’t about a single flashy feature. It’s a validation checkpoint for Microsoft’s strategy of making AI a first‑class OS capability. By embedding actions into File Explorer — the hub of everyday file management — Microsoft reduces context‑switching. Instead of opening an app specifically for a quick edit, users can stay in their flow. The parallel introduction of the generative AI activity log acknowledges that this power demands transparency.

The roadmap signals are clear: expect more file types, more AI actions (text summarization for documents has been teased), and granular policy controls to emerge in subsequent Dev and Beta builds. The dependency on Store‑shipped apps like Photos and Paint means that the feature set will evolve in lockstep with those apps’ updates, not just with OS builds.

Design tradeoffs to watch

  • Context menu clutter. Packing more entries into the right‑click menu risks overwhelming users. Microsoft must give people the ability to hide or reorder AI actions.
  • Local vs. cloud ambiguity. A simple icon or badge indicating where inference runs would go a long way toward trust. Absent that, skepticism will run high.
  • Per‑action permissions. A global kill switch is too blunt. Allowing admins to block cloud‑assisted generative erases while permitting local blur is a more nuanced and realistic ask.
  • Accessibility. Every AI action must be fully navigable by keyboard and screen reader. Early Canary builds often fall short here; widespread enterprise adoption depends on closing that gap.

What’s next

Features that prove stable and useful in Canary usually trickle to Dev, Beta, and eventually the General Availability Channel — but only if they meet privacy, accessibility, and reliability bars. Microsoft will likely iterate the AI actions to support more formats (PDFs, HEIC, perhaps RAW) and add administrative templates for IT. The seconds clock, being a simpler feature, may land in a stable release more quickly.

For now, Build 27938 is a sandbox. It’s a chance for Insiders to prod at the future of Windows, for IT to prepare governance frameworks, and for everyone to hold Redmond to account on data handling. If you’ve tested the build, share your findings — latency, local‑vs‑cloud behavior, and logging accuracy — because community feedback shapes the final product far more than silent telemetry ever will.