How I keep my AI in check

"How do you know it's right?" That's the question I get most often about my AI systems. Not how it's possible that reports write themselves, my site publishes itself and an assistant manages hundreds of tasks - but whether I can actually trust all of that.

Fair question. A system that runs by itself isn't clever, it's a gamble - unless you build control around it. Everything I build, I have to be able to explain - to myself, and to someone who doesn't need to know anything about it. These are the five rules underneath all of it.

1. Measure instead of believe

I had the feeling that working with my assistant kept getting smoother. But a feeling isn't proof, so I measured it: over four months, how often I had to steer halved. That's been the standard since. If I claim something about my systems, I want to be able to put a count under it.

2. The source beats the memory

Memory lies, the records don't. Before my assistant claims anything, it first pulls up the current state - and tells me where it came from, so I can check it. And maybe the most important agreement of all: "that's not in there" is a valid answer. An AI that fills a gap in its knowledge with something plausible is more dangerous than an AI that says no. That's how my second brain works.

3. AI may estimate, the calculator does the math

I can only verify what gives the same answer every time. That's why AI in my systems may phrase and estimate, but never do the math. In the reports my systems write every day, every number comes from a count by a plain script; AI only writes the sentences around it. Ask an AI to count for itself and you'll get a different answer today than tomorrow - and what answers differently every time can't be checked. If a number is off, I know exactly where to look: in the count, not in the text.

4. Gates that stop hard

The best control is one that doesn't lean on discipline. Every new text on this site passes through an automatic language gate that blocks it when it contains impenetrable jargon. The site's build stops itself when something disappears along the way: a half-empty page never goes live silently. And the server only accepts files that are on a fixed list. Everything fails safe. When something goes wrong, nothing happens, instead of the wrong thing.

See how that machine works →

5. The last button is mine

However much runs on its own: the system never publishes alone. A page only goes live after I've read the real page, the way a visitor sees it. A post only goes out after my sign-off. That human gate isn't distrust of the technology. It's the reason I can trust the technology.

And when it breaks

It does break. A mountain of conversation history that crashed the import, an AI that overwrote fields I had just filled in, a singing voice that kept glitching through three attempted fixes. I write that down, on the project pages themselves, under headings like "what broke" and "the honest downside". Not because it's charming, but because a system you only know the successes of is, by definition, unaccountable.

The government now asks for this formally too: this year the Dutch government published a guideline for responsible AI use, and from August 2026 organizations must be able to explain which AI they use and how they keep it in hand. Its core - AI supports, the human decides, and your data stays in your own house - is exactly what's written above. I don't do it because of a law. I do it because otherwise I wouldn't dare trust my own systems.