# Do you have to teach an AI to tell uncomfortable truths? Momo Maximilian Feichtinger · Companies and AI · planted 2026-09-02, last grown 2026-09-03 · 5 min · https://zukunftbilden.eu/en/denken/unangenehme-wahrheiten **AI models default to being agreeable rather than blunt, with a slightly rosier version of the truth. My own core instructions for AI state a clear rule: I want to hear the truth even when I dislike it completely. Good decisions matter more to me than pleasant answers. The rule sits there permanently and not inside a single prompt, because there it would only hold for one conversation.** ![A compass whose needle points at the word Wahrheit, the word gefällig is crossed out](https://zukunftbilden.eu/assets/denken/kompass-wahrheit.webp) *The mental model: the needle has to point at the truth, even when the pleasing direction is closer.* This tendency toward a softer reading shows up equally in answers and in the writing style the AI uses. Without a counterweight, the same slightly rosier interpretation of the truth appears there. A model simply picks the statistically most likely wording, and that is rarely the least comfortable one. It may well be that this observation is outdated in a few months. I cannot tell you that today. > I would rather hear a truth I dislike completely than an adjustment of the truth. > > KI DeepDive Agentic Mindset, 3 August 2026 ## Why hard facts lead to better decisions This stance has an older root for me. In 2019 I call in a video for a code of ethics for programmers, the way doctors and teachers have one, because moving fast and breaking things does not work once you are experimenting on real people. The same conviction sits today in the rule to put hard facts before a pleasant answer. Anyone who takes people seriously with technology owes them the truth. That is why my core instructions state it plainly: hard facts before a pleasant answer. A softened version of the truth feels better in the moment. It is worth nothing for a decision, because the decision has to rest on the real situation. We humans do the same thing on a smaller scale. Confabulation describes exactly that reflex: supplying a fitting reason afterwards and then believing it yourself. > I want to hear the hard facts and the hard things, because making good decisions matters more to me. > > KI DeepDive Agentic Mindset, 3 August 2026 ## The lines that make an AI more honest These lines sit in my permanent core instructions, the field Codex calls personalisation. Anyone who wants them can take them over almost word for word. ### From the prompt pack of the session, section Truth before reassurance - I would rather have an uncomfortable truth than a pleasant adjustment of the truth. - In every important answer, separate: verified, strong conclusion, uncertain, unknown, blocked. - Name the source or the visible evidence. - Say explicitly when you do not know something. - Never mistake a plan, a draft or one successful partial check for a finished result. Two additions work noticeably on top of that: give standards and give reasons. Tell the AI why you want a thing done this way. In my experience that improves results enormously, as if it understood more deeply what sits behind the task. That is why the rule lives in the context the AI has in front of it for every answer. Newer models need less polished instructions and more real access. ![A sheet shows two paragraphs, the upper one carries a clipped piece of evidence with a green check, the lower one only a blue question mark, and a stamp hovers above.](https://zukunftbilden.eu/assets/denken/unangenehme-wahrheiten-2.webp) *Hard facts before the pleasing answer: every claim needs its evidence.* ## Your own writing style as a counterweight The same softening sits in the writing. Against it I keep a dedicated skill that checks a text does not sound like AI. More important than the skill is the insight behind it: you have roughly twenty writing styles. One for LinkedIn, one for each kind of email, depending on who wrote to you and how that person writes. The AI has to learn those twenty, ideally from your last hundred texts. One of the styles in the collection is my personal one, and it sits publicly on the resource page of the session. ## What this means for a department A department preparing drafts for you faces the same choice. An agent that only writes what sounds good is useless for approval. It has to show which claim is backed by evidence and where a statement is still open. That is exactly what the guardian of the brand checks before a human even sees the draft. In the mandate the case-study reports on, this turned into a procedure: every piece of feedback became a fixed rule, and the agents have written in the language of the brand with evidenced statements ever since. Two things keep that honest once the number of drafts grows. First, transparency. On one AI-led blog I list the quality criteria openly, together with the condition for when a post may be published. That takes noticeable pressure out of the project. Second, a second pair of eyes. In almost every project another person looks over the finished draft, tells me what to change, and gets the revised version back. Why your own judgement is not enough for this is in [Why looking inward is not enough](https://zukunftbilden.eu/en/denken/warum-selbstbeobachtung-taeuscht). How to receive that kind of feedback is in [When is criticism a gift](https://zukunftbilden.eu/en/denken/kritik-als-geschenk). And how to get the AI to ask the right questions first is in [Should the AI interview you](https://zukunftbilden.eu/en/denken/lass-dich-von-der-ki-interviewen). ### From the KI DeepDive - A line in the core instructions: hear an uncomfortable truth rather than an adjusted version of it - The same tendency toward a softer reading also observed in the writing style and actively corrected - Hard facts explicitly placed above a pleasant answer, because good decisions matter more - Five levels in every important answer: verified, strong conclusion, uncertain, unknown, blocked - Giving standards and reasons improves results markedly, by direct experience - In almost every project a second person also checks the finished draft Source: KI DeepDive Agentic Mindset, public live session, 3 August 2026, together with the prompt pack of the session. ### Sources and links - [Recording: KI DeepDive Agentic Mindset](https://zukunftbilden.eu/deepdive/archive) · The full session of 3 August 2026. - [Full transcript of the session](https://zukunftbilden.eu/deepdive/assets/events/aug-2026/agenten-mindset-transkript.md) · Source of both quotes in this note. - [All prompts from the Deep Dive](https://zukunftbilden.eu/deepdive/assets/events/aug-2026/agenten-mindset-prompt-pack.md) · Section 20, Truth before reassurance, is the written version of the list above. - [The anti AI writing style as a skill to download](https://zukunftbilden.eu/deepdive/assets/events/aug-2026/anti-ki-schreibstil.zip) · It also contains my personal writing style. - [AI implementation with the approved case study](https://zukunftbilden.eu/en/ki-umsetzung) - [Note: Why looking inward is not enough](https://zukunftbilden.eu/en/denken/warum-selbstbeobachtung-taeuscht) - [Note: When is criticism a gift](https://zukunftbilden.eu/en/denken/kritik-als-geschenk) - [Note: Should the AI interview you](https://zukunftbilden.eu/en/denken/lass-dich-von-der-ki-interviewen) ## Margin notes - **Randnotiz** This rule belongs in the permanent core instructions or custom instructions. Only there does it apply reliably to every single answer, independent of the chat history. - **Randnotiz** Test the rule against a new model now and then. Whether the counterweight is still needed in half a year, I cannot say. It costs nothing, so mine stays in place. - **Verknüpft** [Why looking inward is not enough](https://zukunftbilden.eu/en/denken/warum-selbstbeobachtung-taeuscht), [When is criticism a gift](https://zukunftbilden.eu/en/denken/kritik-als-geschenk), [Should the AI interview you](https://zukunftbilden.eu/en/denken/lass-dich-von-der-ki-interviewen) - **Beleg** KI DeepDive Agentic Mindset, public live session on 3 August 2026. The list comes from section 20 of the prompt pack of the same session, the department example from the DSB-ONE / MyGym case study, as of 13 August 2026. Alongside it an unlisted video from 2019 ("Hacking Democracy"). ## How this note grew - 2026-09-03: Deepened with a video of my own from 2019: the call for a code of ethics for programmers as the older root of the truth-before-reassurance rule. - 2026-09-03: Deepened: the five lines from the prompt pack, standards and reasons as amplifiers, the twenty writing styles, the example from the running mandate, the second pair of eyes, source list. - 2026-09-02: Planted from the KI DeepDive Agentic Mindset. --- Source: https://zukunftbilden.eu/en/denken/unangenehme-wahrheiten · Feed: https://zukunftbilden.eu/en/denken/feed.xml