A week using AI for everything

For seven days I asked artificial intelligence about everything: the menu, emails, shopping, small decisions and a few not-so-small ones. Some things improved, others went wrong and at a couple of points I caught myself not thinking. Here's the whole week, stumbles included.

"Let AI run your life." "My entire routine, managed by artificial intelligence." There are whole videos of people who get up, ask the assistant what to have for breakfast, what to wear and what to do first, and end the day with a yoghurt-commercial smile.

I use these tools a huge amount in my work. Almost everything goes through Claude at some point, and through other models too. That's exactly why I wanted to know where the limit is: not in theory, but in my own week, with my own shopping and my own emails.

A warning up front: I don't recommend repeating it as is. It's a slightly absurd challenge done on purpose to see where things break. And they broke, though not always where I expected.

The rules were simple and I wrote them down before starting, so I couldn't cheat.

• What's in: decisions, writing, shopping, work and household matters. If in doubt, I asked.
• Which models: Claude as the main one, ChatGPT and DeepSeek to cross-check when an answer smelled off.
• Limits: no client data, no irreversible decisions, nothing health-related and no purchase over €50 without thinking twice myself.
• How I logged it: a diary in a note on my phone, with the question, a summary of the answer and what I ended up doing, plus screenshots of the most striking bits.

Dates: from Monday 14 to Sunday 20 September 2026. A normal week, no travel or anything unusual, so the test was about real life.

Monday started with enthusiasm. I asked for the week's menu based on what I had in the fridge, plus the shopping list, and it was surprisingly good: it used up half a courgette that had been staring at me for days and spared me my weekly conversation with myself in front of the fridge. First point to the machine. In the afternoon I asked it what to wear for a video call meeting. It recommended a shirt. I was already wearing a shirt. A draw.

Tuesday was emails. A message to a supplier postponing a delivery came out correct but so formal it looked like I was suing him. I softened it myself. On the other hand, organising three errands around the city in the order with the least back-and-forth worked brilliantly: it saved me about 40 minutes of driving.

Wednesday mixed the useful with the clumsy. At work, comparing two configurations for a new server helped me organise what I already knew, which is what it's best for. In the evening I asked it for a film, it recommended a title that wasn't on any of the platforms I have, and when I told it so it recommended another that wasn't either. I ended up watching one I'd already seen.

By then something curious was happening: the novelty of asking about everything was fun, but I was starting to feel a little friction every time I was about to do something without asking. As if I needed permission.

The second half of the week weighed more. Checking every little thing is tiring, and tiredness leads to laziness about thinking.

Thursday I ran a test that sums up one of the disadvantages of artificial intelligence: I asked three models what to get my nephew for his birthday, and then the same model three times. Six different answers, all reasonable and none better than the others. It isn't a flaw: the model generates the answer that seems most likely to it, and each attempt comes out different. But if you're looking for someone to decide for you, you discover you're flipping a coin with very good prose.

Friday brought the ridiculous moment. Someone in the family WhatsApp group asked whether I was coming to Sunday lunch, and instead of answering "yes", I asked the assistant to help me reply. It gave me three options, one of them with a party emoji. I sat staring at the screen for a long while, wondering at what point I'd got there.

That same Friday, though, it saved me something real: a complaint about a duplicate charge from the electricity company. The letter came out in ten minutes, clear and citing the right article, and I got my money back the following week.

Saturday and Sunday brought the serious failures, which I describe next. And a practical detail: on Saturday morning the service I use most was down for nearly an hour, and I discovered I didn't quite know what to do with my morning.

Three failures, in increasing order of seriousness.

A decision. On Sunday I wanted to go on a hike with the family. I asked for a circular route of about 12 kilometres near home and it described one in full detail. It turned out to be linear and almost 18: the return trip ended up being by taxi. I believed it because the description was precise, with the names of real places. That has a name: hallucination, which is when the model makes something up and tells it with complete confidence.

A text. The message to the residents' association group about a special levy came out sounding like an angry lawyer. I sent it without rereading it properly and had to apologise the next day in the entrance hall. I believed it because on Monday and Tuesday everything had gone well, and trust builds up faster than attention.

A purchase. A 35-euro adapter to connect a monitor to my laptop. The assistant assured me it was compatible with my model; it wasn't. Consequence: a return, a week waiting for the refund and another adapter bought after checking the spec sheet myself.

And the near miss, which is the one that scared me most: for the electricity complaint I was about to paste the entire bill into the chat, with my ID number, my address and my bank account number. I stopped in time. Depending on the service, what you type may be stored or used according to its terms, and that isn't undone with an apology in the entrance hall.

These are the risks of artificial intelligence for someone who uses it every day: that it gives you false information with aplomb, that you share what you shouldn't and that you end up depending on it to think. All three happened to me in the same week.

This is what I least like telling, and that's why I'm telling it.

On Wednesday I accepted the shopping list without reading it and came home with two kilos of lentils, when I already had an unopened bag. On Thursday I asked it what a 10% tip on a 46-euro bill was. I know how to work that out. I know perfectly well. But I already had the phone in my hand and the reflex did the rest.

And on Friday afternoon something worse happened: I had to write a short email to a colleague and sat for a couple of minutes in front of the blank screen not knowing where to start. Me, who has been writing emails all my life.

The warning signs I recognised in myself were exactly the ones described when someone wonders whether they're becoming dependent on artificial intelligence: accepting answers without reading them, finding it hard to get started without it and not being able to explain what you've just handed in.

No drama: I haven't become a different person in seven days. But it happened to me, someone who uses these tools every day and thought he was vaccinated.

When it was over I separated two kinds of dependence, because they look alike but aren't the same.

The practical kind: if the service goes down, changes its terms or raises its price, your work stops. The hour without service on Saturday was a small warning of something that in a business can be big. That's why I work with several models and keep some running locally, on my own machines.

The mental kind: the things you stop exercising get rusty, like writing, calculating or deciding. This one is more treacherous because it gives no warning.

And it's worth not confusing heavy use with dependence. I use it many hours a day and will keep doing so. The difference lies in whether I still understand what it does and whether I could do it without it. What I do now to avoid falling in: I write the first draft of any personal text myself, I do simple sums in my head and, before asking something, I make myself think for ten seconds about what I would answer.

I stopped asking it for:

• Replies to personal messages. If I can't answer my family without help, the problem isn't the wording.
• Technical compatibility for purchases. The manufacturer's spec sheet doesn't hallucinate.
• Matters of taste, like gifts or films. Six different answers aren't a recommendation.

I still ask it for:

• Organising what I already have: menus, errands, lists. It's very good at that.
• Drafts of paperwork and complaints, which I then review. Without pasting personal data.
• Organising ideas at work, when I already know what it's about.

Verdict in one sentence: the tool is very useful, but thinking is still my job.