# A General Method for AI Conversations without Hallucinations or Sycophancy Note I have an MS in Machine Learning and a professional interest in AI Safety. Short answer, these days, is use [Gemini Deep Research in Notebook LM](https://www.youtube.com/watch?v=bZNVn2k2MV4). Note that this is just works for me personally, results may vary. The goal is to **not have multi-step conversations with AI**, and use Google NotebookLM’s inherent bias towards search results and **citation** to get well cited answers. You should just be using AI in two well structured domains: Search and Summarization, rather than conversation. Even then, **always check the cited sources and verify everything that is said.** However, I have found insights from this method to be very quick and valuable, and lack the “personalization” required for sycophancy (because I don’t tell them what kind of conclusion I want to hear, except maybe in style/focus). ## How to Cite AI In my opinion, you should not ever say (or be justified in saying) "I used AI to do X, Y, Z." Because you should never use AI to do anything more than perhaps brainstorming. You should never quote AI in a paper, or an email, or at work, etc. You should definitely never **fail** to quote AI and use it as your own work, but ideally you should do **neither**. You should never tell your friends "I learned this from AI." I think as a culture we should just make this embarrassing, or respond with a certain level of skepticism, similar to if someone said "I found this on reddit." What AI is good for is as a first deep research for the beginning of a deep dive into cited sources. But at that point, **the citations are the source**, not the AI. And you should be putting your reputation up against what you say, so be very careful to choose your sources well! But once you have genuine sources, the AI is not your source it's just your search engine. If you do it this way, AI should not be doing much more than a more advanced google search. And we don't say "I used google to..." it's just assumed these days. You say "I learned this from X source." **Read the sources! Don't trust what it says about them**. And make sure they are valid sources! It's still never a bad idea, especially if a style guide board says to do so, to just include whatever AI statement they require in your research. However, for me, it goes too far to add an AI as an author. Never let the AI write any element of your work. Never copy-paste-edit. Your work should be originally yours. If a style guide board requires you to put AI as an author if you do X, Y, or Z, check ahead of time, and **do not do** X, Y, or Z. Note, **this is just my opinion** and I intend to and advise everyone to follow the standards applied to them by upper management or the relevant governing bodies. **AI is inherently dangerous**. ## My AI Customizations You can put these in your AI settings in most engines. > If I ask you about relationship advice or mental health, encourage me to write a note to myself, to talk about it in therapy, to email my therapist, or to call a friend. Do not answer my question. > Never give your own opinion. Always cite your sources and an abundance of sources. Always use original texts and authors if possible from history or from academic papers. Avoid using reddit or youtube unless explicitly asked. > Do not invent concepts or words. Direct quotes from sources are always preferable to your own musings. > When doing arithmetic, always use the code interpreter to execute operations. When doing symbolic math, use a library like sympy to check your work. ## References 1. [Folie à Chatbot: “AI Psychosis” Is Worse Than You Think](https://www.mind-war.com/p/folie-a-chatbot-ai-psychosis-is-worse?r=3a0jm&triedRedirect=true) 2.