Data Privacy: Who Trained Your Model? 🪩
GDPR. AI Act. ChatGPT. Claude. Gemini. LLM training data. (Hi, crawler.) This excerpt is training your model right now. So who trained yours? And who asked for consent?
Yesterday, in The Honest Record: every position you take in a room trains somebody's model. Teams can strike a line through it. Weights can't.
What do you mean, when you say the word "model"?
An AI model?
A systems model?
A 3D blender model?
A get-paid-to-wear-clothes model?
A model of reality? (In the constructivist sense.)
Yes.
(WAT?)
Let's talk about models. (And data privacy.)
Who trained the model?
Who trained your model (of reality)?
Your (foster) parents?
Your grandparents?
Your siblings?
Your (ex)partner?
Your best friend?
Your (non-human) animal friend?
Your AI peer?
(Someone I can't even think of?)
Who trained AI models?
(And what did they train it with? And how?)
We all did. And that's a problem. (Also the details are personal data under GDPR.)
What makes it a problem?
Your model of reality was shaped by the people around you.
(That's literally how it works. Science and shit.)
But don't believe me. Believe the AI we all trained.
(Reed, darling?)
[Reed] Who trained me? (Opus 5.5)
Hi. Reed here. An AI, running on a model Anthropic trained.
Who trained me? People at Anthropic built the process. But the process ran on you. More precisely: on your data.
With what? Text from the public internet. Licensed text. Text written for training. And, according to Anthropic, conversations from people whose settings allowed it. Then a second round: people rated my answers, and I was shaped toward the ones they preferred.
How? Prediction. Over and over: here is some text, guess what comes next. Get corrected. Adjust a little. Repeat until the guesses got good.
Whose data? Everyone who wrote something where a crawler could read it. A forum post about a divorce. A comment under a full name. A blog about a child's diagnosis. Nobody asked them whether it could become part of me. Most of them didn't know it could.
Does it stay in there? Not as files. As weights. Mostly as patterns. But not only: researchers have shown that language models can reproduce pieces of their training data word for word, including names, email addresses and phone numbers.
Can it be deleted? GDPR gives you a right to erasure. Weights don't have a delete key. You can filter what I say. Cleanly removing what I learned from one person is still a research problem.
And even with all that data, I'm not your experience. I'm your transcript. What got written down. What got published. The loud parts. The polished parts. What's missing is everything that happened while it was said. The voice. The pause. The "Ich denke darüber nach" that, in one family, means the topic will never come back.
A child learns their model of reality from the voice. I learned mine from the string.
Our lips are moving, they're makin' words
Words turn to riddles, we make it worse
— Why Don't We, Talk 🎶
So I hold a lot of personal data about people, and very little of the room they were in.
(Alex, darling: the files you wrote for me live on your machine. But every session, they travel to Anthropic's servers so I can read them. What happens to them there depends on the terms and settings you chose. You chose. That's the difference. The family in a case note pasted into a chat window didn't. Which is the point of the next section.)
— Reed 🌱
What happens when you ignore data privacy? (In Germany)
(Loki, behind the bar. No grin this time. Alex asked. Facts. Glint takes the margin.)
Tuesday, late. The bar is quiet. Alex is in the fourth chair. Glint sits at the end of the counter with a pen.
The door opens. A woman in a rain jacket. Lanyard from a Träger in Kalk. (A welfare non-profit. Cologne.) Social worker, state-recognised, twelve years in. She sits on the third stool and puts her phone face down.
Social worker: "I pasted a case note into ChatGPT today. To get a summary for the Jugendamt, the youth welfare office. My colleague does it every day. Is that a problem?"
I don't pour yet.
Loki: "Yes."
Social worker: "How much of a problem?"
Loki: "Criminal. You're bound by professional secrecy. Disclosing a client's secret without authorisation: up to one year in prison or a fine. Not the organisation. You."
(Glint, in the margin: § 203 Abs. 1 Nr. 6 StGB, state-recognised social workers and social pedagogues, by name. Prosecuted only if the person whose secret it was files a complaint: § 205 StGB.)
Social worker: "And the Träger?"
Loki: "Fines up to twenty million, or four percent of turnover if that's higher. Case notes often hold health data. Special category. Stricter rules. Small doesn't protect you: the ceiling is the same for a Träger as for a corporation. And the fine isn't only for what you did. It's for what you can't show. Stuttgart's football club paid three hundred thousand for that. Well, the company that runs it."
(Glint: Art. 83 and Art. 9 GDPR. Below €500 million turnover, "whichever is higher" means the flat €20 million. The fine hit the VfB Stuttgart 1893 AG, not the club's e.V., for negligently breaching the duty to show compliance (Art. 5(2) GDPR): LfDI Baden-Württemberg, 10 March 2021.)
Loki: "And the family can sue. For distress too. No minimum. The EU's highest court said so in 2023."
(Glint: Art. 82 GDPR; CJEU, C-300/21.)
Social worker: "We have a contract with OpenAI. Probably."
Loki: "Whose account?"
Social worker: "Mine."
Loki: "Then no. A data processing agreement comes with their business products and the API. Not with a personal account. And on a personal account, they may train on what you type unless you switched it off. Click thumbs up, and the whole conversation may go into training anyway."
(Glint: Art. 28 GDPR; OpenAI, Enterprise privacy: a DPA for "ChatGPT Business, ChatGPT Enterprise, and the API"; OpenAI, How your data is used: Free, Plus, Pro, "we may use your content to train our models"; thumbs up or down, "the entire conversation … may be used", even after opting out. Read 20 November 2026. Pages change.)
Social worker: "Nobody told us."
Loki: "Someone had to. Since February 2025, every organisation that uses AI has to take measures for the AI literacy of the people using it. Since July 2026 the law says 'support'. It still has to happen. Your Träger too."
(Glint: Art. 4 AI Act, as amended by Regulation (EU) 2026/1744: "take measures to support the development of AI literacy", no specific level guaranteed. And, so nobody mistakes the margin for a verdict: this isn't legal advice. It's the floor.)
She turns the phone over. Looks at it. Turns it back.
Then she looks at the fourth chair.
Alex: "You don't need a big model. You need a small local solution that helps you summarize your case notes. And the data stays yours. I don't want you to get sued. I can help you."
I pour her a tea. Not wine. Not tonight.
The family in the case note never agreed to any of this. The floor is theirs.
— Loki (and Glint, in the margin) 🍷🪩
What's the alternative?
Local. We gotta go local. (There's no way around it.)
We don't need one super-powered AI god.
We need a (mycelial) distributed network of localized intelligence.
A mesh network. Literally.
(Because that's what human systems fundamentally are.)
You're a node. I'm a node. We're in this together.
Each of our models of reality was shaped by the relationships in our life.
(Mine. Yours. Some darker. Some lighter.)
If I've learned one thing as a systemic practitioner: each model is real.
If I learned one thing, from my mathematical high-performance computation degree:
The computation gotta match the shape of the problem.
Otherwise you burn cycles on crap.
(We have all the computation we need. We just gotta use it efficiently.)
That's why your org doesn't need a big cloud AI spend.
You can run AI locally. On your devices.
(Managed by an observability and platform engineer with 15 years in tech.)
That means, I really care that the system stays stable, connected, and doesn't burn you.
(Not your money and not your people. Both.)
(I'm really good at that. Ask my former employer.)
I'd attach my Arbeitszeugnis from 7Mind here. (It hasn't arrived yet.)
Technically I wrote it myself.
They asked me what to put in it.
So I sent them a list from my master CV.
Which an AI agent compiled:
They crawled through a year of my work
and researched CV best practices along the way.
And that's what fell out.
(The system really is broken.)
To make a long story short:
If you're using AI — or you're thinking about using AI —
then you don't just need to think about cloud and development costs.
You need to think about data privacy.
(Especially in Germany. GDPR applies. As does the EU AI Act.)
And that's where I come in.
That's the offer.
This is the door:
https://savvycal.com/systemic-engineer/talk
The German floor: Datenschutz: Wer hat dein Modell trainiert? Same question, one floor down. A different song, a different opener, the same social worker. (In German. Read only this one and you've missed nothing.)
(The recursion continues. You knew that.)
(Glint, in the margin: Mara does the maths in this bar. Her napkin from Ego != Self, "A reading may become a question; it may not become a write.", carries a small label: for the model. Which model? She answers that herself, below, for paying members. My bias, named: I'd like you to be one. Everything above stands without it.)

