The Honest Middle
There are two easy stories about artificial intelligence. One says it will save us. The other says it will end us.
I don't live in either.
I live in the honest middle — where a tool this powerful is neither saviour nor threat, but a question, asked quietly of each of us.
That question is the whole of my stance: does this make me more myself, or less? Everything below follows from it. Not a list of things to fear — a map of where to pay attention, and the commitments I hold so the tools carry the load without taking the wheel.
You will find pieces that count the dangers of AI. Forty-seven of them, or some other number built to sound precise. I understand the instinct; the risks are real, and naming them is honest work. But a number invites you to fear on schedule, and it ages the moment the world moves. So I have not counted. I have mapped.
Five places a thoughtful person actually touches when they work with AI. A map, not a scoreboard. It expects to change — and it will.
Privacy, client material, intellectual property, and what a tool quietly keeps. The convenience of pasting everything in is real. So is the cost of not knowing where it goes.
Confident wrong answers, inherited bias, someone else's defaults dressed as neutral, fluent text that only sounds true. It is most persuasive at exactly the moment it should not be trusted.
The slow flattening of your own voice. Judgment outsourced one small decision at a time. Attention scattered, purpose drifting while the days get easier. The quiet risks are the ones without an alarm.
Tools built without a problem worth solving. Assistants with no off-switch. AI used on people without telling them. No plan for the day it fails.
Power concentrating in a few hands, the environmental cost, surveillance made cheap, the human quietly displaced. The part no single person controls — and no honest map leaves out.
Each of these is a version of the same thing — a way AI can make you a little less yourself. Which is why I hold a shorter list. The commitments that keep me more.
Seven, because that is the count my own practice produces — not a number chosen for effect. Each one I can defend, and each one I actually keep.
I name what I'm solving, and who it serves, before I reach for AI. If I can't, I'm not ready to build it.
AI drafts, researches, accelerates. It never decides what I believe or what I stand for.
I know where my inputs come from and what a tool retains. Client and private material is not raw fuel.
What you tell my tools lives with you — local-first, exportable, never sold, never used to train. The Mirror keeps nothing once you leave. The Companion lives in your own browser.
Anything I publish or send is verified first — facts, numbers, names. AI sounds right even when it is wrong.
My lived experience is the one thing AI cannot manufacture. It sharpens my drafts. It never stands in for the part only I can say.
Disclosure, plainly — and a way to reach me. It is now the baseline the law asks for in Europe. I was already doing it.
The test under all sevenWould I accept this being used on me — or on someone with less power, less information, fewer options than I have? If not, I change it before it ships.
The serious voices — the people actually building this — say much the same thing at the scale of nations. Govern it. Be honest about the risk. Don't pretend. The most credible of them, writing about what an AI-reshaped world will owe people, marks where policy runs out:
“The need for people to find meaning, purpose and agency… is something to be collectively worked out by society as a whole, not something policy can directly address.”
— Dario Amodei, Anthropic, June 2026
Policy can buy us the time. It cannot hand anyone their own life back. That last part — staying yourself as the tools grow — is the work I do, at the scale of one person. The honest middle, small enough to actually practise.
Increasingly, the law says it must tell you. Since August 2026, Europe's AI Act requires AI to disclose itself when you're speaking with it, and — with a grace period into December — to mark what it generates in a way machines can read. The tools for this exist: content credentials that travel with a file, invisible watermarks that survive edits. Some of the tools I use already mark their output; many still don't.
But most tools still don't. So the absence of a mark is not proof a thing is real. Checking well still means provenance, watermark, source — and the part no tool replaces, your own judgment.
We disclose. We don't fake. And we don't ask a watermark to do the work that discernment must.
No. The Mirror keeps nothing once you leave, and no one — not even me — ever reads what you write. It was built so I can't. Private by design, not by promise.
No. What you tell my tools lives with you. The Companion runs in your own browser; nothing is sold, and nothing is used to train a model.
Several. I work across commercial models and open source, and I choose the tool for the task rather than pledge loyalty to one lab. What stays constant is not the vendor but the rules: I know what each tool retains, I keep client and private material out of what I can't control, and I judge a model by how it behaves, not by how it markets itself. No single company's promise is load-bearing here. That is deliberate.
Yes. It is reviewed every month or two, as regulation, tools, and my own practice move. Each change is dated, and I approve every one by hand. The judgment stays human, here too.
The clearest proof of all this is the thing itself. Five minutes with an AI that reflects rather than answers, and keeps nothing when you leave.
Meet the Mirror →