The Paste Test: Would You Put This in an AI Chat?

Look at the last thing you pasted into an AI tool at work.

A draft email? Fine. A customer complaint with an account number attached? Hmm. A patient note, a contract, a spreadsheet of member balances? Now you’re doing mental math about who might eventually see it.

That hesitation is the most honest privacy signal your team has. We call it the paste test, and it’s where every good AI privacy conversation should start.

What “Private” Actually Means

When someone asks us whether an AI tool is private, they usually mean one of four things:

• Will you learn from my data? The “no training” promise. Important, but it’s a statement of intent, which is the weakest kind of privacy.
• Can you see it? Even well-meaning providers have staff, contractors, hosts, and logs. Some approaches process data inside sealed, hardware-protected environments where the operator genuinely can’t look. That’s the difference between a locked door and a door with no keyhole on your side.
• Where does it go afterward? Prompts get logged, cached, and backed up. “Processed securely” and “kept for two years” can both be true at once.
• Can you prove it? In regulated work, this is the one that matters most. Pilots rarely stall because something went wrong. They stall because nobody can show that nothing did.

Promise, Control, Proof

Think of privacy as a three-rung ladder.

1. Promise: terms of service, contracts, “we don’t train on your data.” Easy to offer, impossible to verify.
2. Control: things you enforce yourself, such as redaction before a prompt leaves your network, access limits, short retention, or keeping the model inside your own environment.
3. Proof: the system demonstrates its own behavior through audit trails and cryptographic evidence of what ran.

Most of the market sits on rung one. Most regulated buyers eventually need rung three.

Don’t Build a Fortress Around a Shed

The strongest protection usually means higher cost, smaller models, and more engineering. Apply it to everything and you’ll spend a year securing data that didn’t need it.

Let the data set the rung:

If the data is…	Start with
Public or low-stakes	A provider promise and common sense
Internal, moderately sensitive	Controls: redaction, minimal context, short retention
Confidential or competitive	Provider-blind processing, or keep it in-house
Regulated or privileged	Proof you can hand to an auditor

Notice that “send less” shows up in every row. The cheapest privacy feature in AI is still not sending the thing in the first place.

The Paste Test at System Scale

The same test works when you’re building, not just chatting.

When we built an AI document assistant for a large U.S. non-profit trauma hospital, the first design decision wasn’t about the model. It was about the boundary: which documents the assistant would read, and which it would never touch. We limited it to verified, official policy documents and tied every answer to its source page.

That’s the paste test applied to a whole system. Before anyone asked “is it private?”, we had already answered the more useful question: what is this tool allowed to see? A tool that never touches the sensitive material has a much smaller privacy problem than one that does and promises to be careful.

We’re now building small proofs of concept around different privacy approaches, to learn which ones hold up in real workflows rather than demos. The early lesson so far: the teams that get furthest aren’t the ones with the fanciest technology. They’re the ones who decided, in plain language, what each kind of data is allowed to do.

Run the Test This Week

Have your team list three things they’d happily paste into AI, three they’d hesitate over, and three they would never paste.

That last list is the valuable one. It’s a map of the work AI could be doing for you if privacy were solved. Bring it to whoever owns your AI strategy, and if you’d like a second pair of eyes on it, we’re easy to reach.