Why an AI assistant should not read your whole screen
An assistant that continuously records your screen has to keep a searchable copy of everything you looked at, including material that was never yours to keep: a colleague's salary in a shared spreadsheet, a customer's record, a password manager mid-unlock. Scoped capture gets you most of the same usefulness for none of that: the assistant sees the text you point at, at the moment you point at it. The practical test is not what a vendor promises to do with your data, but what they are technically able to do with it.
There is a genuinely good argument for an assistant that records everything. Context is what makes AI useful, gathering context is tedious, and the tidiest way to never gather it again is to capture continuously and search later. Every product built that way is solving a real problem, and the demos are convincing because the feature works.
The argument against it is not that these companies are careless. It is that continuous capture creates an asset that did not exist before, on your machine, and that asset has properties nobody chose: it is complete, it is searchable, it is durable, and its contents were decided by whatever happened to be on your screen rather than by anyone's intent.
The material you did not agree to keep
Think about a normal Tuesday. You open a shared spreadsheet and the tab next to the one you needed is compensation. A colleague screen-shares a customer record while you take notes. You unlock your password manager and the list of entries is visible for four seconds. Someone pastes a production credential into a channel and deletes it a minute later. A recruiter's email is open behind your editor.
None of that is yours to retain. Some of it is not legally yours to retain. In a continuous-capture model all of it is now in a local index that will still be there in eight months, and the deletion a colleague performed in the source system does not reach your copy. That is the part that is hard to undo: not a breach, just a record that outlives the thing it recorded.
"It stays on your device" is a real answer to a different question
Local storage genuinely does defeat the biggest threat, which is a vendor with a server full of everyone's screens. It is worth taking seriously and it is not marketing. But it answers the question "can the vendor read it", and the questions people actually run into are different ones.
- Your laptop gets stolen, or a colleague borrows it while it is unlocked.
- Your employer runs endpoint tooling that can read local files, which is normal and usually contractual.
- You are asked, in a dispute or an audit, to produce what is on your machine. A complete searchable index is discoverable in a way that your memory is not.
- Some future feature, yours or an integration's, is given read access to the index. The data was collected under one set of expectations and is used under another.
None of these require anyone to behave badly. They are the ordinary weather that any large, complete, durable record eventually meets. The way to survive them is not better promises. It is not having the record.
Scoped capture, and what it actually costs you
The alternative is to make the user the trigger. The assistant sees the text you selected, or the region you drew a box around, at the instant you asked for something. Nothing runs between requests. There is no index of your day, because no day was recorded.
It is worth being honest that this costs you something real. You cannot ask "what was that error I saw on Thursday", because Thursday was not kept. Retrieval across weeks of activity is the one thing continuous capture does that scoped capture cannot, and if that specific capability is what you need, you should buy the product that provides it with your eyes open.
What is surprising is how much of the value survives the trade. Most of what people actually want from an assistant is a transformation of something in front of them right now: rewrite this, summarize that thread, turn these notes into a ticket, read the text out of this screenshot. All of it works on a selection. The recall use case is a minority of the work and it absorbs the entire privacy cost.
Scoped memory is still memory
The middle ground is worth naming, because "record nothing" would be an overcorrection. What an assistant needs is not your history. It is what you are working on. A work session is a container you start deliberately and stop when the task is done: it collects what you ran and what you produced while it was open, and it collects nothing when it is not.
The difference from a recording is that its boundaries are chosen. You know when it started, you know what it covers, and the reason it does not contain a colleague's salary is not a filter that might fail: it is that you never ran an action on it. A record whose contents are decided by intent is a fundamentally different object from one whose contents are decided by what was on screen.
Questions worth asking before you install anything
These are the ones that produce different answers from different products, which is what makes them useful. Vague answers are themselves informative.
What permissions does it request, and why each one?
Screen Recording is the load-bearing one on macOS. An app that has it can read your screen whenever it is running. An app that does not have it cannot, whatever its settings say.
What runs when I am not asking for anything?
If the answer is a background capture, indexing, or sync loop, the product's real data footprint is that loop, not the feature you were shown.
What is retained, where, and for how long?
Ask specifically whether there is a searchable index of past activity, and whether deleting the source deletes the copy. Those are separate systems and they are often separate answers.
What leaves the device, and at what moment?
"Only what you send" is a good answer if you can point at the moment you sent it. "Encrypted in transit" is not an answer to this question.
Can I turn any of it off and keep the product useful?
If capture is the product, there is nothing to turn off. That is a legitimate design, and it tells you the trade you are making.
The honest summary
Continuous capture is the fastest way to make an assistant knowledgeable, and it is not a scandal. It is a trade, made once, whose cost arrives later and unevenly. Scoped capture is slower to feel magical and much easier to live with, because the material it never collected cannot leak, cannot be subpoenaed, and cannot be repurposed by a feature nobody has written yet.
If you want the detail of where a specific product draws that line, ours is written out on the security page, and the comparison with the recall model is on Khint vs Rewind.