What I Really Use AI For

Work with AI a lot and you get asked a particular set of questions: about code, about copy, about jobs. All fair — this site is proof of the first two. But when I recently wanted to know what I actually use AI for, I didn’t ask my memory. I asked the machine itself: it got the data export of my chats — 291 conversations from four months, alongside some 160 coding sessions — and the job of sorting them.

What came out is an inventory of the uses nobody much talks about. And the most inconspicuous finding is the biggest: there is no such thing as an “AI topic”. The same machine that tinkers with this website in the morning identifies an animal in the garden at noon and dissects a contract offer in the evening.

Show, don’t describe

The most underrated interface is the camera. “Any idea what might be digging here?” — a photo of a hole in the lawn, an assessment plus follow-up questions in return. Another animal I didn’t just want identified; I wanted to know right away whether it has to be reported in North Rhine-Westphalia. A handful of stones, with a request for educated guesses. And once, the photo became evidence: the utility company claimed my water meter reading was “implausible” — the picture of the meter said otherwise.

Describing is tedious and error-prone. Showing is faster — and the AI’s follow-up questions are often better than my description would have been.

The patient co-operator

A few conversations in the inventory stand out because they are no longer questions but working days. Upgrading my two domain controllers from Windows Server 2022 to 2025: more than 250 exchanges in a single conversation, from the ISO image to the final reboot. Getting started with Home Assistant and my Zigbee devices: 458 messages in one stretch, without a single branch — you feel your way forward together, error message by error message. Smaller but more dramatic: the moment a complete Signal chat history vanished on the PC. The phone went into flight mode before it could pick up the deletion — and then the two of us sat over the question of how to rescue the history from there.

I have also probed the limits of this kind of help: I seriously asked about an affordable robot that could operate a keyboard. Slow typing would do; Ctrl+Alt+Del is a must. Because the most capable software ends at the computer that no longer boots. (What happens when hardware acts up in earnest is a story of its own.)

Backbone in small matters

A use I hadn’t reckoned with: AI as backbone in dealings with institutions. The recycling center refused to take a small UPS battery — so I had the applicable rules looked up; arguing works differently with a citation at hand. The utility’s reply to my meter objection was unsatisfying — so we drafted the next step together. A mobile-phone offer, run against the existing contract; notice periods, looked up. None of this is spectacular. But it shifts a balance of power: those who can look things up cave in less often.

For others

The most striking item in the inventory: how often the answers weren’t for me at all. “Some motorway interchanges look like works of art from a bird’s-eye view” turned into a bobbin-lace pricking — a pattern from which lace can be made at home. For a friend’s 80th birthday: a year of AI access as a present, including the research into how to gift such a thing in the first place. After a week of hiking in the Azores with a friend: the search for a Portuguese restaurant for his birthday present. For a family we are friends with: the house hunt in the Ruhr area. For a friend with a new computer: Outlook contacts dug out of an old Windows 7 backup. For my granddaughter: more battery life for the iPhone — and a Sunday plan “for a fifteen-year-old and her grandpa”.

If this inventory yields one thesis, it is this: the benefit of AI rarely stays with the person who opens the chat. It gets passed on.

Verify, don’t trust

And then the kind of use closest to this site’s heart. An ad promises that fixed-maturity ETFs are the better, more lucrative alternative to saving — a sentence like that now goes straight into research. I haven’t handed my investments to the AI, but one question I couldn’t resist: how much AI bubble is actually sitting in my own ETFs. The answer was more nuanced than the headlines — the deciding still happens without it. An email allegedly from PayPal I suspected to be phishing and had the suspicion confirmed. And it works the other way round, too: when an AI statement about a dietary supplement didn’t add up for me, I pushed back with the source. Why this distrust is a method is explained here.

The inventory itself

Which leaves the method — itself an item for this list: for this article, the AI combed through 291 of its own chats and condensed them. The sediment belongs to the truth: why long drinks are called long drinks, where the Blue Danube got its name, when the school holidays start — plus some sixty untitled two-question chats and the occasional health question. Everyday life.

And right in the middle, the nicest twist. When the AI presented the longest conversations to me — 514 messages for the server upgrade — I asked back: “Are you accounting for forks in that count?” It wasn’t. Rephrase a question and you leave a discarded branch behind in the chat history; the machine had counted branches. After the correction the number shrank only a little — but it shrank. In an earlier report, a count like this corrected my memory. This time I corrected the count. Verify, don’t trust works in both directions — and thus belongs right at the top of the list of what I really use AI for.