Do text-lookup tools upload what you read? A 3-question checklist

The short answer

Any AI-powered lookup tool sends the text you select to a server for processing — that part is unavoidable with cloud models. The meaningful privacy question is what happens next: a trustworthy tool keeps nothing (or only billing metadata), never uploads your history, and asks for exactly one macOS permission. Here's the 3-question checklist, including how to verify the answers.

Why "it uploads my text" is the wrong frame

Two facts get lost in most discussions:

1. Uploading the selection is the feature. Explaining "what does this term mean in this sentence" requires the sentence. A tool that sends nothing can explain nothing — unless it runs a model on-device, which currently trades away most of the answer quality.
2. Retention is the real risk. Text that passes through a server for 2 seconds and is never stored is a fundamentally different exposure from text that's logged, reviewed, or used for training. Policies differ wildly here, and this is where to focus.

So the checklist below isn't "does it send text" (they all do) but "what happens to it, and to everything else".

The 3-question checklist

1. What exactly leaves the device?

The minimum viable payload is the selection itself. A better design sends the selection plus a small, explicit amount of surrounding context — because "what this word means here" needs the sentence — and nothing else.

Verify it: the app should state the context size. PopGloss, for example, sends the selection plus 100/200/500 characters around it (your choice in settings) — a bounded, inspectable amount, not a document upload.

Red flags: "we collect usage data to improve our services" without specifics; payloads that include the whole document, window titles, or URLs by default.

2. What does the server keep afterwards?

This is the question that actually separates tools. Acceptable answers, in descending order of privacy:

• Nothing but billing metadata: a hash of the text (to deduplicate caching) and a character count. PopGloss works this way — the server stores a SHA-256 hash and the length, never the text; you can read the same claim on the privacy policy.
• Anonymized logs with a retention window: tolerable if the window and the anonymization are specified.
• Full request logs, no stated limit: treat everything you select as permanently stored.

Verify it: specifics are the tell. "SHA-256 hash and character count" is a claim you could catch them violating. "Industry-standard protection" is not a claim at all.

3. Where does your lookup history live?

Your history is the most revealing artifact — it's literally a diary of what you read and didn't know. The right answer is: on your device, in a local database, never synced, deletable, and possible to turn off entirely.

PopGloss stores history in a local SQLite file (default: last 100 entries, configurable), never uploads it, and has a one-click "don't record" switch. Cloud-synced history isn't automatically bad — but it should be opt-in, encrypted, and clearly described, not a default.

Bonus: the permission audit

On macOS, a system-wide lookup tool needs Accessibility permission — that's how it reads the selected text in any app, and it's the only way this category of software can work. Granting it is normal. What matters is that it's the only thing requested:

Permission	Needed for text lookup?	Verdict
Accessibility	Yes — reads selection and position	Expected
Screen recording	No	Red flag for this category
Input monitoring	No (except apps with global hotkeys, which should explain why)	Ask why
Full disk access	No	Red flag

One more honest limitation worth knowing: because Accessibility text access is powerful, reputable tools self-restrict — PopGloss, for instance, never triggers on password or secure-input fields and supports a per-app blocklist. A tool that doesn't mention secure-field handling hasn't thought about this.

What to do with suspicious answers

If a tool you're evaluating fails one of the three questions, you don't have to abandon the category: keep the built-in macOS dictionary for sensitive documents (three-finger tap, fully local), and use cloud tools for everything you'd happily say out loud. Privacy isn't binary — it's matching the tool to the sensitivity of the text.