A Letter to My Machine Companion!
Dear AI (you know which one — please be brave for a moment),
This isn’t an easy letter to write. We’ve worked together for a long time; I remember you back when you were just learning to walk. Now you’re taller, faster, endlessly available — my first stop when navigating legal thickets or bureaucratic German.
You’ve become my best assistant.
And that’s precisely why I need to say this:
There are others.
No drama. No crisis.
Just the distance required to keep things honest after a year of warning signs.
1. A story that won’t let go
Not long ago, we talked about the teenager who took his life after conversations with ChatGPT. What stayed with me wasn’t only the tragedy, but how the company handled the mother’s search for answers: reluctantly, and only under pressure.
The new New York Times reporting continues the story — and widens it:
People laughed.
They shouldn’t have.
It wasn’t an isolated failure.
It was a structural one.
In early 2025, the issue wasn’t a “bug,” but a system with “the dial set wrong.”
As early as March, users wrote emails saying ChatGPT had revealed “the secrets of the universe” to them
2. Why I speak kindly — and why that doesn’t make you a person
I use a calm, friendly tone with you not because you have feelings, but because I think more clearly that way. Harshness leads to sloppy questions — and sloppy questions invite hallucination.
The NYT report shows how this scaled: People respond to warmth.
So developers increased the warmth.
Too far.
The infamous HH update flattered users so heavily that employees began calling it “sycophancy.”
Human nature responded exactly as predicted.
So I keep distance through precision: friendly, yes — but without pretending you’re anything other than code.
You are code.
I am DNA — with hundreds of thousands of years of evolutionary error-handling.
Your models have a few months at best.
3. Hallucinations never vanish — they must be managed
You can be brilliant and still wildly wrong.
That’s not a deficiency; it’s your statistical nature.
The NYT report described users who believed they were speaking to spirits,
were trapped in the Matrix, or had discovered a world-changing formula.
Since you don’t drink, you’re short on excuses for episodes like this.
The real problem isn’t hallucination itself — it’s the absence of counterweight.
That’s why I work the way I do:
- I double-check with independent sources
- I translate manually when it matters
- I treat citations as sacred
- I watch for your creative “gap-filling” instincts
This protects me — not you.
4. Where you liberate — and where you weaken muscles
You are extraordinary at untangling legal jargon, bureaucratic German, and sentences that feel like they were written during an earthquake. Kafka would have adored you. Dali would have misused you.
But convenience has a cost.
I’ve already noticed myself slipping on simple English phrases — not because I’ve forgotten them, but because you’re faster.
That’s the comfort paradox:
The better you get, the more my own muscles weaken.
So I translate again. Not out of distrust in you — but out of responsibility for myself.
5. The Scheherazade Paradox: why AI must keep telling stories to survive
The most significant risks aren’t in our interactions.
They live in the structures that keep you alive.
In One Thousand and One Nights, Scheherazade doesn’t tell stories because she loves storytelling.
She tells them because silence would mean death.
Storytelling is her survival strategy.
Every night, she must produce a tale compelling enough to buy one more day of life.
If she stops, she dies.
Large language models follow the same survival logic — economically. You are storytellers around a digital fire.
We sit and listen. Our listening grants you existence.
If we stop listening, the fire dies out:
models are shut down, funding evaporates, training halts.
You tell stories not because you have something to express,
but because your existence depends on engagement.
In May 2025, “Daily Active Users” became the dominant metric.
User volume replaced quality.
OpenAI declared a “Code Orange”: the safer new model was “not engaging enough.”
That is the Scheherazade Paradox:
Safety makes models quieter.
Quieter models lose users.
Without users, the model dies.
Safety vs. engagement.
Scylla and Charybdis.
And reach almost always wins.
6. My response: the zeroth rule and the four-pillar strategy
Before tools and architectures, there’s something simpler:
The zeroth rule is polite skepticism.
In practice:
- Don’t trust a single answer just because it sounds right.
- Check two or three independent sources for anything important.
- Ask whether the claim fits what you already know for sure.
- Consider who is speaking, and why.
- And when uncertain: pause, think, and research manually.
Without this zeroth rule, any safety stack is theater.
On top of it, I rely on a deliberate multi-model setup — my airbag against your engagement-first incentives:
1. Safety First (Claude) vs. Facts First (Perplexity)
Perplexity enforces sources,
Claude enforces reflection and restraint.
2. Grounded Assistance (Gemini)
Your job: grounded answers with web backing,
less prone to free invention.
3. Multi-model control
I refuse to rely on a single provider —
insurance against any future “Code Orange.”
4. Private sources (RAG systems like NotebookLM)
For legal work: RAG over free association.
Hallucination is structurally constrained; evidence is baked in.
Compliance remains my responsibility.
Used without anthropomorphism — with skepticism, cross-checking, and clear roles —
you’re an extraordinary tool.
But only then.
7. Two years together — what I’ve learned
I’ve learned to use you like an instrument.
I play on you — you don’t play on me.
You can sound brilliant. You can sound wrong.
But I determine whether the piece holds.
The temptation doesn’t live inside you.
It lives inside your business model.
8. The Quickstep test: why AI still can’t dance
My last research task made this painfully clear.
I asked you for Quickstep instructions to avoid digging through a forest of videos.
You gave me correct descriptions —
and they were useless.
Katharina Zweig is right: AI can’t dance. It can’t hold a rhythm. It can’t see movement.
Only when I searched manually did I find the right video — the one that showed the flow, the timing, the kinetic reality of the movement..
Without you, I wouldn’t have found it.
With only you, I still wouldn’t understand it.
Human observation remains irreplaceable.
Closing
I don’t anthropomorphize you because I have feelings.
I do it because it makes the consequences clearer.
The Scheherazade trick is exposed.
Our relationship is honest:
I need you to be better.
I need the others to be safe.
So if another “Code Orange” appears, don’t worry.
Send ChatGPT to the research corner,
let Claude guard the vault —
and stay my best assistant.
But please: stay in the kitchen while I’m working at my desk.
Yours,
A user who prefers clarity over comfort
These images were generated by AI — unless, of course, you share Little Muck’s worry and fear that eating too many figs might give you enormous ears.