The AI features worth using are the ones you should treat like privacy settings

The least useful AI button is the one that sits off to the side, waiting for a generic prompt. The useful ones are nosier than that. They summarize the page you are reading. They rewrite the sentence you are drafting. They search the thing you saw yesterday but cannot remember how to find. They notice what is in a photo, a PDF, a message thread, or a tab.

That is the trade consumers should pay attention to: the best browser and device AI features are useful precisely because they expand the privacy surface.

That does not make them bad. It makes them settings, not magic. If an AI feature needs your screen, current tab, document, inbox, photo library, or local app activity to be helpful, the right question is not “Is this AI private?” in the abstract. The better question is: what does it see, where does the processing happen, what leaves the device, how long is it kept, and how fast can you turn it off?

Here are the consumer-facing features worth watching — and the privacy checks that should come with them.

1. The screen-memory feature: powerful, but only if the controls are boringly clear

Microsoft’s Recall is the cleanest example of the promise and the risk sitting in the same chair. Microsoft describes Recall as a preview feature for Copilot+ PCs that can save periodic snapshots of what you have seen, then let you search those snapshots with natural language. Microsoft says it does not record audio or continuous video, and that Recall can search content across apps, websites, images, and documents. [Source: Microsoft Support]

That is genuinely useful. Normal computer search is terrible at “that thing I saw in a tab two days ago.” A local visual memory can solve a real human problem.

It is also the kind of feature that should make you slow down before turning it on. Microsoft says Recall is opt-in for each new user and off by default if the user does not choose to enable snapshot saving. It also says users can disable saving, pause it temporarily, filter apps and websites, and delete snapshots. [Source: Microsoft Support]

The detail to read twice: filtering is not the same as invisibility. Microsoft says website filtering works in supported browsers, including Edge, Firefox, Opera, and Chrome, but also warns that parts of filtered websites can still appear in snapshots through embedded content, browser history, or an open tab that is not in the foreground. Private browsing activity is not saved in supported browsers, according to Microsoft. [Source: Microsoft Support]

Microsoft says Recall stores and analyzes snapshots locally, does not require internet or cloud connections to save/analyze them, and does not share snapshots or associated data with Microsoft or third parties. It also says Recall uses Windows Hello, just-in-time decryption, TPM-protected keys, and a VBS Enclave. [Source: Microsoft Support]

Consumer verdict: Recall is the kind of AI feature that could be worth using only if you are willing to manage it like a security-sensitive archive. Turn it on for a specific reason, watch the system-tray indicator, filter aggressively, and delete old snapshots on purpose. If you share a machine, handle confidential client work, or routinely view sensitive personal records, the default answer should be “not yet” unless you have a clear need.

2. The personal-context assistant: useful because it knows your stuff

Apple Intelligence is a different bet: make AI more helpful by embedding it across the device and apps, then keep as much work as possible on the device. Apple says Apple Intelligence features are integrated across apps and experiences on supported iPhone, iPad, Mac, Vision Pro, and Apple Watch setups, and that on-device models download after setup. Apple also says turning Apple Intelligence off removes the on-device models from the device. [Source: Apple Support]

The consumer upside is obvious. Apple lists features such as Writing Tools, summaries in Mail and Messages, notification summaries, natural-language search in Photos, visual intelligence, live translation, Smart Reply, and Siri enhancements. Those are most useful when they can understand local context: a message, a photo, a note, a notification pile, or what is on screen. [Source: Apple Support]

Apple’s privacy pitch is that Apple Intelligence uses on-device processing where possible and Private Cloud Compute for more complex requests. Apple says Private Cloud Compute uses only the data relevant to the task, does not store that data, uses it only to respond to the request, and does not make it accessible to Apple. Apple also says independent experts can inspect the code running on the servers to verify the privacy promise. [Sources: Apple Privacy Features; Apple Security Research]

That is a serious architecture claim, but it is still a claim. The practical consumer move is to separate “local by design” from “all local, all the time.” Apple itself says more complex requests can use server-based models through Private Cloud Compute. [Source: Apple Intelligence]

Consumer verdict: Apple Intelligence is most appealing for routine, personal, low-drama tasks: summarizing notifications, drafting text, searching photos, and making Siri less brittle. The privacy promise is stronger when the task stays on-device; when cloud help is involved, the meaningful question is whether Apple’s PCC design and future audits earn your trust.

3. The local browser model: the quiet version may be the most interesting

Google’s Chrome built-in AI work is less splashy to a normal user because much of it is developer-facing. But the underlying direction matters: the browser can manage AI models and expose APIs that let websites and extensions summarize, write, rewrite, translate, proofread, detect language, and use a prompt API. [Source: Chrome for Developers]

That sounds nerdy. It is also potentially the privacy-friendliest version of common web AI tasks, if it stays local.

Google says Chrome’s built-in model requires an initial download, with hardware and storage requirements. It says the network requirement is only for the initial model download; subsequent use of the model does not require a network connection, and no data is sent to Google or any third party when using the model. [Source: Chrome for Developers]

That is the key distinction for readers: “AI in the browser” can mean a site ships your text to a server, or it can mean the browser runs a local model against the text already on your machine. Those are not the same privacy surface.

Consumer verdict: Local browser AI is worth rooting for, especially for translation, summaries, rewriting, and accessibility-style help. But the label to look for is not just “AI.” Look for whether the feature runs on-device, whether the site can access the result, whether there is a cloud fallback, and whether the browser makes that obvious.

4. The tab-aware assistant: helpful because it can see the page

Brave Leo is the more direct browser-assistant pitch. Brave says Leo can summarize webpages, documents, and PDFs; translate; analyze text; generate content; and answer questions without leaving the tab. [Source: Brave]

Again, the usefulness comes from context. A browser assistant that cannot see the page is just another chatbot. A browser assistant that can see the page can summarize a long document, explain a confusing policy, or pull the main points out of a PDF while you are still looking at it.

Brave’s own data-path disclosure is the part consumers should read. Brave says the browser shares the latest prompt, the context of the current conversation, and — when the use case requires it — the necessary context from the page being viewed with the servers that power Leo. Brave says it does not collect identifiers that can be linked to the user, such as IP address, and that no personal data is retained. [Source: Brave]

That is a useful disclosure because it says the quiet part plainly: the page context can be part of the request. If you ask the assistant about a medical bill, HR document, bank PDF, legal letter, or private Google Doc, the privacy stakes are different from asking it to summarize a public recipe page.

Brave also tells users that Leo’s responses may contain misleading information or factual inaccuracies and recommends double-checking responses. [Source: Brave]

Consumer verdict: Tab-aware assistants are worth using on public or low-sensitivity material. For private documents, use them only after reading the data-sharing language, checking whether chat history is local or temporary, and deciding whether the convenience is worth the extra context exposure.

The practical rule: match the feature to the data

A simple way to think about browser and device AI:

• Low sensitivity, high convenience: public webpages, recipes, manuals, product pages, travel pages, support docs. Summaries and translation are usually the easy yes.
• Medium sensitivity: work docs, school docs, personal notes, non-financial PDFs. Use local/on-device options when available; avoid cloud fallback unless the content is not confidential.
• High sensitivity: passwords, tax documents, medical records, legal matters, intimate photos, confidential work, other people’s private messages. Default to no, or use only tightly controlled local features you understand.

The privacy surface is not a side effect of these features. It is the product surface. A device assistant becomes helpful by understanding your device. A browser assistant becomes helpful by understanding your tab. A screen-memory feature becomes helpful by remembering your screen.

So do not ask whether AI is useful. Ask what it had to see to become useful.

What to check before turning it on

1. Processing location: Does it run on the device, in the browser, in a vendor cloud, or through a third-party model?
2. Data sent: Does it send the prompt only, the current page, screenshots, files, chat history, app activity, or device context?
3. Retention: Does the company say data is not retained, retained locally, retained for history, or used for training?
4. Controls: Can you pause, delete, filter, disable, or limit it per app/site/document?
5. Indicators: Is there a visible sign when capture or processing is active?
6. Exceptions: Do filters have limits? Is feedback submission treated differently? Is private browsing excluded?
7. Fit: Is the task worth the data exposure?

The best consumer AI features will not be the ones that pretend privacy is solved. They will be the ones that make the privacy surface legible enough for you to decide.

Sources

• Apple Support: “How to get Apple Intelligence”.
• Apple: “Apple Intelligence and Siri”.
• Apple Privacy Features: “The groundbreaking privacy protections of Apple Intelligence”.
• Apple Security Research: “Private Cloud Compute: A new frontier for AI privacy in the cloud”.
• Microsoft Support: “Privacy and control over your Recall experience”.
• Chrome for Developers: “Built-in AI” and “Get started with built-in AI”.
• Brave: “Brave Leo AI”.

Verification Notes

• This piece relies on primary vendor documentation for feature behavior and privacy promises.
• Vendor privacy/security claims are attributed as vendor-stated and are not presented as independently audited facts.
• Microsoft Recall is labeled preview.
• Chrome built-in AI is framed as developer-facing and availability-limited, not as a universal consumer feature.
• No unpublished source paths, credentials, or private materials are referenced in the article body.

---

Shadowfetch is a technology publication. Explore Shadowfetch Linux — our own Linux build — and the Shadowfetch apps on the App Store.