Microsoft unveils Copilot+ PCs, sparking Recall privacy debate

On May 20, 2024, Microsoft introduced Copilot+ PCs, a category of Windows computers with dedicated AI hardware and features such as Recall. Microsoft’s announcement promised to retrieve material previously seen on the device through locally processed snapshots; that convenience also turned screen history into an asset that required protection.

The move matters because Microsoft wants to shift some AI functions from data centers to the PC. That could reduce reliance on an internet connection, speed up certain tasks and limit the amount of information sent to external services. But it also turns the computer into a device that observes and organizes more data about its user.

AI as a hardware requirement

Copilot+ PCs must include an NPU capable of at least 40 TOPS. An NPU, or neural processing unit, is a chip specialized in the calculations used by AI models; it is designed to handle them with lower power consumption than a conventional CPU or graphics card. Primary source

Microsoft is also setting minimum requirements of 16 GB of RAM and 256 GB of storage. The first models will come with Qualcomm’s Snapdragon X Elite and Snapdragon X Plus processors, an Arm-based architecture designed to combine battery life and performance. The company has announced devices from Acer, ASUS, Dell, HP, Lenovo and Samsung, along with its own Surface Laptop and Surface Pro. Primary source

The devices will go on sale on June 18, starting at $999 in the United States. The Copilot+ PC label therefore does not apply to just any computer compatible with Windows 11: it identifies a specific generation of hardware built to run Microsoft’s new AI experiences. Primary source

Recall turns activity into a searchable history

The most striking feature is Recall. Windows periodically takes snapshots of what appears on the screen and uses AI models so users can later search for what they saw or did. Microsoft’s promise is straightforward: retrieve a conversation, webpage, document or image even if the user does not remember the file name or the app they used.

In the Build demonstration, Recall lets users scroll through a timeline of their activity and submit queries in natural language. If someone remembers seeing a chart about a particular topic or a conversation with a specific person, the system aims to find it using the visual and textual content of the captures.

It is not a video recording of the screen, but a file of snapshots that the system processes and indexes. Microsoft says the data is stored locally on the device, is not sent to its servers and is not used to train its models. Recall requires signing in with Windows Hello and allows users to pause collection, delete specific moments or exclude apps and websites.

Privacy depends on more than keeping data on the PC

Local processing avoids one of the main concerns raised by cloud-connected assistants: a company centralizing millions of people’s usage histories. However, it does not eliminate the underlying problem. A file containing screenshots of emails, work documents, messages, purchases or visible credentials could be extremely sensitive if someone else manages to access the computer.

There is also a practical issue. The filters for excluding apps or pages will require users to know the feature exists, understand what information they do not want to retain and configure the exceptions before they need them. On a shared computer, in a workplace or when handling confidential data, that configuration burden could be significant.

Microsoft presented Recall as a tool users can pause and configure, although it was enabled by default in the initial announcement. Its usefulness depends precisely on recording a great deal of context. That tension will be difficult to resolve through configuration controls alone: the more complete the history, the better it can find information; the more complete it is, the greater the potential impact of unauthorized access.

A race to make AI useful in the operating system

Alongside Recall, Copilot+ PCs will include improvements for image generation and editing, live translation through Live Captions and locally processed camera effects. Microsoft says some of these features will work without an internet connection, a significant shift from the first waves of assistants that relied almost entirely on the cloud.

The initiative puts Windows on the same path as Apple and Google: making AI a basic device capability rather than a standalone app. For users, the test will be whether these features genuinely save time. For Microsoft, the immediate challenge will be proving that a computer that remembers more about its owner can also protect that information better.

“Local” is a location, not a guarantee

Running a feature on the device reduces some transfers to the cloud, but it does not remove risk. The useful questions are what is captured, where it remains, which account can retrieve it and what happens if another application or person gains access to the computer. A screen history aggregates material from services that never agreed to become a searchable database: disappearing messages, video calls, internal systems or a client’s data.

The assessment should draw the entire path. Start at capture, then follow image and text recognition, the search index, storage and the interface returning results. Record encryption, retention, exclusions, access logs and deletion at each stage. “It is not sent to Microsoft” answers only one question; it says nothing about a secondary user, privileged malware or a backup reading it.

How to test an operating system memory

Before enabling it in a sensitive environment, use a test profile. Display a temporary password, confidential document, private window and excluded application; then search for every item and inspect local files. Revoke access to a document and check whether its snapshot remains retrievable. Repeat the exercise using the real accounts and privileges of the intended computer.

The safest control is to collect nothing until the person understands the scope and enables it. Pause, filtering and deletion must be visible, but settings do not replace data-minimising architecture. An organisation needs policy controls that disable capture by application and verify device state, rather than trusting every employee to remember to exclude a window.

The transferable skill is to translate “on-device AI” into a threat model. Locate capture, index, keys, permissions, retention and deletion, then test with planted data. The same procedure applies to Recall, photo search or any assistant that promises to remember personal activity.

An update can change the risk

An assessment belongs to a specific version. A change in default activation, authentication, encryption or filtering can alter exposure without changing the Recall name. Before a major update, repeat the tests, check that exclusions remain in force and verify that deletion removes the index, snapshots and associated copies. Useful memory also needs a verifiable way to forget.

The test should end in a concrete decision: enable, disable or restrict the feature by profile. Record version, policy and date so it can be revisited. Without that outcome, a privacy report describes risks but does not change the computer’s exposure.

Configuration is part of the product being evaluated.