Spotting it

You cannot decide about something you have not noticed. This lesson is about noticing.

The vocabulary

Vendors frequently avoid the word AI — sometimes because the feature predates the current wave, sometimes because “AI” now makes people wary.

Words that usually mean it:

Smart — smart reply, smart compose, smart sort. Suggested — suggested replies, suggested edits. Auto — autocomplete, auto-summary, auto-enhance. Intelligent, assist, assistant, copilot. Magic — magic eraser, magic write. Enhance, generate, rewrite, summarise. Insights, highlights, for you.

None of these are lies. They are just not the word that would prompt you to think about it.

The visual tells

A sparkle icon. Four-pointed stars, in some arrangement, have become the near-universal glyph for a generative feature. If you see one, that is what it is.

Text appearing before you type it. Greyed-out completions ahead of the cursor.

A summary above the actual content — search results, a document, a thread.

Something offering to write for you anywhere at all.

The question that matters

Once you have spotted it, one question ranks above all others:

Does my content leave this device, and if so, where does it go?

That determines whether E4·L8’s paste test applies. A keyboard prediction running entirely on your phone is a different proposition from a summariser uploading the whole thread to a server.

Three broad answers:

On-device. Processing happens locally. Nothing is transmitted. Common for keyboard prediction and increasingly for phone features, and vendors advertise it when it is true — because it is a selling point.

Cloud, provider only. Sent to the vendor’s servers, processed, returned. Usual for anything more capable.

Cloud, third party. Sent onward to another company’s model. Common when a smaller product bolts AI on, and the one people least expect.

Where the answer is

Rarely in the interface. Usually in:

The privacy policy, under processing or third parties. A feature-specific help page — often clearer than the policy. The settings screen for that feature, where a data toggle sometimes lives.

For work software, your IT or data protection team will have asked this question already. They may be the fastest route to an answer.

If you genuinely cannot find out, treat it as cloud and third party. That is the assumption that keeps you safe if you are wrong.

Ask the product itself

Increasingly viable and worth trying: ask the assistant embedded in the product.

Is this feature processed on my device or on a server? Is my content used for training? Where is that documented?

Treat the answer as a pointer rather than an authority — E7’s rules apply to a model describing its own product as much as anything else. But it frequently names the document you need, which is most of the work.

Try it now (5 minutes)

Open your phone’s keyboard settings and your email settings. Find every feature matching the vocabulary or the icons above.

Pick the one handling the most sensitive content and find out where that content goes. If you cannot find out in five minutes, that itself is the finding.

Check your understanding

1. Which word most reliably signals an AI feature without saying 'AI'?

Advanced Smart Premium

2. The most important question about an embedded feature:

Which model does it use? Does my content leave the device, and where does it go? How much does it cost?

3. If you cannot determine where a feature sends your data, you should:

Assume on-device, since that is now common Assume cloud and third party Assume it is not AI at all

Check answers

Recap

Learn the vocabulary — smart, suggested, auto, magic, enhance — and the sparkle icon, because vendors often avoid the word AI. Once spotted, the question that matters is whether your content leaves the device and where it goes. The answer lives in the privacy policy or a feature help page rather than the interface, and if you cannot find it, assume cloud and third party.