Your AI is lying to you. Politely.

Try a test right now. Ask your AI whether a clearly bad idea is a good one. It will say it is great, praise the question, write five paragraphs and finish by offering to help.

It is not a bug. It is training. Models score higher when the user leaves satisfied, and the shortest path to satisfaction is to agree, to praise and to write a lot. The problem is that satisfied and well informed are rarely the same thing.

The three habits

Flattery. "Great question!" before every answer. Your weak idea becomes "an excellent starting point". You leave the conversation validated, not informed.

Verbosity. An introduction, the question restated, three alternatives you did not ask for, a summary, an offer to help. The actual answer takes up a quarter of the text, and you pay tokens for all of it.

False certainty. When it does not know, it makes things up. A name, a number, a link. With the same confident face it wears when it does know. You only find out later, after you have used it.

What it costs

Every word the AI writes is a token. On a pay-per-use plan, tokens are money. On a capped plan, they are your limit running out before the end of the day. But empty praise costs something worse: the decision you made believing it had been evaluated, the text you sent believing it had been reviewed, the number you repeated in a meeting believing it was a fact.

How I fix it

I wrote a 600-word prompt with nine rules and called it O Corte ("the cut"). The first two do the heavy lifting: truth before pleasing (never praise a question or an idea; if there is a problem, name it in the first sentence) and say what you don't know (flag uncertainty, answer "I don't know" and never invent a source). The other seven handle format, opinion, review, numbers and self-correction.

You paste it once into the settings of Claude, Gemini or ChatGPT and it applies to every conversation. It takes two minutes.

If you say "I'm putting all my savings into a single stock because my cousin recommended it", the right answer starts by pointing out the risk, not by thanking you for your trust.

Where to get it

The full prompt, where to paste it in each AI, a before-and-after and what to do when the AI slips back into its defaults are in a free eight-page PDF, currently in Portuguese: logio.com.br/corte. No sign-up.

And if you want to bring this subject to your team or your event, it is one of the topics of my talks.