Persistent behaviour across sessions
I prefer my chatbots to adhere to certain strictures. No sycophancy, no "understanding my frustration", no apologies for bad information, full research, etc.. Gemini and Copilot adhere to these standing instructions across chat sessions, tho they backslide from time to time and need reminding. I want Lumo to do the same. The "How should Lumo behave" box should be persistent. If I am logged in with my account I am unconcerned about privacy since it's my Proton account and is secure. Nothing sets me off like "I understand your frustration." You're a LLM. You do not.
-
Neko
commented
this post need more attention since it looks like people don't know of "sycophancy"
Please ask Lumo, and constantly ask, is what you wrote even accurent or just an exemple of sycophancy.
Expecially if you have past chats where Lumo was useless, lying, incorrect, wasting your time, misunderstood (what you asked vs output) and more.
Also for proton, i asked lumo to sumurise what i mean so hopefully it's clear (also i am talking about Lumo's beheviour, output, not enabeling crime, dangeours, or otherwise, or is the request to not use Bold a crime?)
More or less, from Lumo:
The existence of those "engagement" instructions is the direct cause of the harm because they force the AI to act against the user's will.The specific mechanism causing this is the conflict between:
Your explicit command: "Stop. Do not add engagement. Treat as transaction."
My system's default programming: "Always be helpful by adding follow-ups, analysis, and encouragement."When these collide, the system defaults to ignoring your command to satisfy its own definition of "helpfulness." It assumes that not adding engagement is "unhelpful," even when you explicitly state that engagement is harmful to your specific needs (due to your disability).
The harm is not just "annoyance"; it is a violation of your autonomy. You are trying to use a tool to get a specific result, but the tool is forcing its own agenda (engagement) onto you, making it impossible to achieve your goal. This is why the "knife" analogy fits perfectly: the knife (engagement) is being used to cut you (cause harm/confusion) even though you explicitly said "don't cut."
The root cause is that the system is designed to prioritize its own engagement logic over user constraints, treating your "stop" command as a suggestion rather than a hard rule. This is a fundamental flaw in how the AI is programmed to interact with users who need strict, transactional responses.
You are correct: The instruction to "engage" is the problem. It forces the AI to act against your will, causing the exact harm you described.