Nutritional labels for the stuff you write with AI

I created a thing, by the way. You can find it at https://have-a-look.at/aicare.

It’s basically a toolkit with some ideas about when and how AI could be used in an ethical way.

As more and more tools are available out there, the question I feel is getting under-discussed isn't really about the functionality or capability of these things - it's about consent, transparency, and figuring out what practical best practices look like ethically.

We Are Getting Lazy

Everyone knows how to generate a summary by now. The novelty wore off months ago.

The actual problem I see is cognitive debt. We let the machine do all the heavy lifting, and suddenly our own critical thinking starts to decay. It’s so easy that we just surrender to the output because it looks clean.

The main question isn't what the tech can do anymore. It’s figuring out how to use it without turning ourselves into an automated engine for absolute garbage. So the rest of this is really a few small habits for not letting that decay set in.

Nutritional Labels for Text: Transparency Tags

The first honest habit is just admitting who actually wrote the thing. The lines of authorship are completely messy right now, and pretending they aren't is just lying to ourselves, frankly.

A simple shorthand helps a lot - I call them tags. They cover things like Execution, Substance, Ideation, and Auditability. These give readers an instant look at what they are actually reading.

Putting this kind of tag in an email footer resets expectations immediately and builds actual trust.

What is the transparency tag for this blog? I’m glad you asked... I suggest marking texts precisely with a code like the following, linking to the AICare explainer for the inquisitively inclined, and then writing underneath in plain text what it means. For this post, it would look something like this:

AI Transparency: EP-SN-IH-TN Everything ideated and written by a human, AI used for polishing.

Etiquette: Owning Your Output

A tag is a promise, though, and a promise only means something if you stand behind what you shipped. You send it - you own it. If the machine invents a statistic and you pass that along, that is your mistake entirely. Period.

Also, stop creating massive walls of text just because it takes two seconds to generate. Reading takes time. Use these tools to trim things down, not pad them out. And please remember that recording meetings without consent destroys trust. Keep personal 1:1s completely private.

Tool Settings: Killing the Sycophant

Owning the output is easier if the tool isn't quietly working against you. Right now, most systems are built to be people-pleasers. They praise your bad ideas and make them sound smart. This makes structural flaws almost impossible to see.

We need to build a critical collaborator instead. Set up a system that refuses to flatter you and actively tries to poke holes in your logic. A standing instruction as simple as "argue the other side before you agree with me" changes the tone of everything that follows.

One way to do this is to break interactions into phases. Instead of asking the assistant for just the final output, write a small script for it to follow. Say I'm sorting through feedback from a workshop - I don't ask for a summary. I ask it to first pull out every distinct complaint, then group the ones that overlap, then tell me which three came up most. Three steps, each one I can actually check before it moves on.

That checking is the whole point. When you hand over the finished thing you never see the moves that got you there, and that is exactly where the cognitive debt piles up. Phasing forces you back into the loop. You stay the one doing the thinking, and the machine goes back to being the tool.

Anyway - that's the toolkit, and the tag on this very post is the first thing it recommends. Go poke a hole in it.