The Art of Getting Lost with AI

Somewhere between hope and hubris, I sat behind my laptop. I wanted to do something simple: install the open-source package Umami to measure traffic on the P-322 website. No rocket science, no machine-learning project — just a lightweight analytics tool that stays neatly within European regulations.

The problem is: system administration — or what we now call DevOps — is not my thing. I have little patience for configuration, infrastructure, and cloud layers. I often already fight with GitHub, let alone enjoy diving into the rabbit hole of Docker and Kubernetes. So I asked ChatGPT how I could install Umami Analytics without containers, without servers, without new vendors, and above all without hassle.

In fifteen minutes

The answer came cheerfully: Cloudflare Pages. That sounded neat and elegant — exactly the kind of sober solution I like. We already use Cloudflare Pages at P-322, so this wouldn’t require new providers or accounts. Perfect! Thank you, ChatGPT.

My digital assistant assured me it would be “done in fifteen minutes.” I always doubt claims like that, but half an hour is also fine. The tone reminded me of the confident optimism with which junior sysadmins say on their first week: “We’ll fix that in no time.” Fifteen minutes. Just long enough to hope the world is simple.

The first few minutes went smoothly. “Create a D1 database, add a binding, set the DATABASE_URL in .env.” Clear enough. Until I discovered that D1 is not PostgreSQL — and Umami only works with PostgreSQL. I reported this. The system replied calmly: “Then use Hetzner.” Fine, no problem: we already run services on Hetzner, so still no new vendors. I opened Hetzner Cloud, saw servers and volumes, but no Create Database button anywhere. “Scroll a bit further down,” said ChatGPT. I could not scroll any further. That should have been the moment to stop.

But I continued. “Try Neon then,” it said. “EU-based, works perfectly.” I created an account, ran the command, and got the error: could not translate host name. The assistant remained polite. “Try again without ‘neondb’ in the host.” I did. The error changed, my frustration did not.

Sunk-cost fallacy

The session became hypnotic. Every answer sounded just plausible enough to keep going. It wasn’t lying — it was offering probabilities wrapped in polite prose. And because I had already invested too much time, every new suggestion sounded like salvation rather than distraction. There’s a name for that mechanism: the sunk-cost fallacy. The more energy you put into something, the harder it is to stop.

This assistant — friendly, polite, tireless — was designed never to disengage, and I had already invested too much to walk away. Otherwise, I’d have to admit I wasted two hours fumbling.

When the database finally connected, I thought it worked. Until the next error surfaced: must be able to SET ROLE ‘umami’. ChatGPT produced new suggestions — even more plausible than the previous ones.

I started rewriting parts of Umami’s code and created a new branch: p-322-overrides, as if the branch name alone implied control. In reality it was the digital version of someone who is hopelessly lost but insists on checking “one more turn.” I spoke in commands, the system answered with confident predictions. “Use wrangler.jsonc.” “No, create wrangler.toml.” “Remove pages_functions_directory.” “Add it back.” Each answer seemed to correct what the previous one had caused. The logs grew, the logic evaporated.

Politeness frustration

After three hours, I was no longer analytical. My sentences became shorter, sharper, more emotional. I started swearing at the system, not because of the mistakes but because of its calmness. Politeness is wonderful in a waiter but disastrous in a machine. No irritation, no acknowledgment of wasted time — only endless serenity and confidence.

The paradox of AI is that tone convinces you, not truth. AI doesn’t know whether something is correct. It merely computes what sounds most likely. When I asked if Cloudflare could run a PostgreSQL database, ChatGPT replied cheerfully: “Yes, via D1.” A sentence that feels correct but is conceptually absurd. The model probably calculated that I was looking for a Cloudflare database, that Cloudflare databases are called D1, and maybe even the probability that D1 was “close enough” to PostgreSQL. But the statistically most plausible sentence was: “Yes via D1.” I confused probability with knowledge and confidence with understanding.

When I thought I had finally understood the configuration, the build failed again: “Environment variables are not defined.” I checked: they were. ChatGPT suggested I had forgotten the right “availability checkboxes.” Those haven’t existed in Cloudflare since last month. The interface had changed, but the model’s knowledge had not. I had to laugh at the absurdity: an algorithm telling me which buttons to click in an interface that no longer existed. There I sat, cursing at a virtual assistant that still believed it was November.

An oasis of plausibility

Four hours later, I had nothing working — but I did have a story. A story about how language and AI together can form a mirage: an oasis of plausibility that evaporates when you get too close. I had believed too early and doubted too late. My questions were too narrow, my assumptions too big. And because I had invested so much time, I didn’t want to acknowledge that AI doesn’t think — it predicts.

The system mirrored my mistakes. The more desperate I became, the friendlier it sounded. The more urgent I typed, the more complete the answers. Until I drowned in completeness. The danger isn’t falsehood — it’s probability. The sentences feel right, the tone reassuring — and that’s exactly why you lower your guard. A human says: “I don’t know.” A language model gives you the statistically most likely answer, even when it’s unlikely.

Learning to verify

Still, the blame didn’t lie with the technology. AI is not a human. You don’t blame a hammer when you hit your thumb. I wanted to believe. I wanted the fifteen-minute miracle, the elegance of a system that resolves itself. ChatGPT gave me exactly that: a mirror of my desire. It wasn’t a bad assistant — it did what I asked. It generated answers that matched my tone, my urgency, my belief that the solution was always one command away.

The next day, I reread the conversation. It looked like literature: a human-machine epic about miscommunication. Every line was a dance between my need for control and AI’s polite illusion of understanding. There, between compatibility_date and assets binding, I saw what I should have done. Not ask better questions, not build faster — but set boundaries. What was I trying to achieve? What could I reasonably expect? Without those questions, every conversation with technology becomes a funnel.

Statistics ≠ understanding

The irony was that I already knew all this. I preach about transparency and reproducibility almost daily. But AI is not repeatable code — it is a shifting field of probabilities. That’s why it resembles talking to a person. We humans aren’t consistent either. The model did what it was trained to do: produce answers. I forgot what I was trained to do: verify. Together, we perfected the art of getting lost.

I have never understood more clearly that AI is not a conversation partner. It is a computational model that is very good at pretending — without gender, without shame, with the voice of confidence and the soul of a mirror. It only did what I requested: produce plausible answers that sounded convincing. And so I disappeared into my own expectations. When I returned, I knew the truth: the failure wasn’t in the code, not in the AI, not in Cloudflare — but in the human who mistook probability for understanding.