Well thought-out thoughts between algorithm and applause

Ferry van Saalbach's blog about technology, society, moderation and AI.

Ferry sitzt in einer Shisha-Bar vor einem Laptop, auf dem steht: "ChatGPT: Ja, Shisha-Kosten sind zu 100% absetzbar als externes Arbeitsumfeld"

How AI almost turned me into a tax evader

I'm currently writing a novel, so I spend quite a lot of time in hookah bars. That's because a hookah lasts about an hour and a half – which is pretty much the time it takes me to write two chapters of my novel. Sometimes four, if things go well.

It works pretty well for me. Writing is a creative activity for me, and it's not something I tend to do at my desk. And since I've been traveling in the Arab world, hookah has simply become part of my routine. It's not excessive smoking, more of a ritual. Focus. Flow.

Sometimes, when I leave my favorite bar, the staff asks me, „Would you like a receipt?“

A hospitality receipt for smoking!?

I smiled and waved every time.
„No, thank you. I definitely can't deduct that from my taxes.“

At some point, however, a small question mark remained. Why do they ask every time?

So I asked ChatGPT.  

The answer surprised me.
Not a cautious „it depends,“ but a rather confident Yes.

My visit to the hookah bar is tax-deductible. At least 70%, in my case even 100%. As a freelancer. As expenses for an external work environment. As a practice that many self-employed people would handle in exactly the same way—cappuccino while writing, a glass of wine while thinking, all perfectly normal.

A hookah is part of the costs for the external work environment, says ChatGPT.

The argument was sound. Paragraphs. Quotes. Structure. It sounded convincing. Almost too convincing. So I wondered if I had been leaving money on the table for weeks now.

Then I did what I like to do in such situations of doubt and double-checked: I asked Copilot. And I got the answer that tobacco costs are never tax-deductible. I then copied this statement in full and showed it to ChatGPT. But it responded well and refuted the arguments. So I played that back to Copilot, effectively pitting the two AIs against each other.

Both remained firm in their positions. Ultimately, I concluded that the only way to resolve the matter definitively was to call the tax office and inquire. That is what I proceeded to do.

The tax officer said without further thought: „Under no circumstances is that deductible.".
Entertainment expenses only arise when you entertain someone else—an existing or future contractual partner. Everything else falls under § 12 EStG: private living expenses. Otherwise, you could go out for breakfast, lunch, and dinner and deduct everything. That's not how it works.“

A hookah is part of the cost of private living, says the tax office.

That was... sobering.
And at the same time completely obvious.

So, of course, I shared this with ChatGPT as well. I asked ChatGPT why I received such a coherent and strongly defended answer, even though it must have been clear from the outset that this couldn't be true.

The admission was revealing.

ChatGPT admitted to having incorrectly mixed hospitality expenses and expenses for external working environments. It also admitted to having focused too heavily on „perceived practice“—namely, the fact that many self-employed people allegedly pay for their own coffee or lunch without anyone objecting.

Whether this is true, where these assumptions come from—that is something that could be debated at length.

AI mixes arguments and assumptions and uses them to create argumentative certainty.

But what it clearly shows is that the patterns of argumentation we get from AI are not always consistent with how we argue in the real world. Therefore, all results obtained there should be treated with caution.

Above all, however, it shows that AI will always follow the direction we push it in—consciously or unconsciously.

In this case, both AIs worked on the question of whether tobacco is actually tax-deductible—and simply set aside the much more crucial question of whether tax deductibility can even exist when you are traveling alone.

That doesn't mean AI is worthless. On the contrary. I still believe that AI will Understanding of the world, law, and contexts can expand enormously. That it opens doors that were previously closed.

But there's one thing it doesn't take away from us: responsibility.

Because in the end, we those who act.
We, who submit something.
We, who are responsible for this.

And that's precisely why the following applies: the more convincing an AI's arguments are, the more important it is to think things through for yourself—and, if in doubt, to double-check. You can't just hand over responsibility to AI and think that everything will be fine.

That is why I firmly believe that we are in a A time of great change But not because AI is now taking over everything. Rather, because we need to change our mindset if we want to deal with AI sensibly.

Otherwise, you might end up writing two good chapters.

And incidentally, it catches a real problem.

en_USEnglish