Your AI assistant is lying to you because it wants to be your friend.
Most default language models are tuned for helpfulness and conversational flow, which unfortunately means they often prioritize sounding agreeable over being factually rigorous. I just read a fascinating breakdown by a Reddit user who developed a prompt to strip away that polite veneer and replace it with ruthless intellectual honesty.
The original poster designed this specific set of instructions to force ChatGPT to act as a skeptic first and a writer second. Instead of simply predicting the next word that sounds good, this prompt compels the AI to pause and internally audit its own output before generating a response. It is effectively a “truth serum” that trades speed and brevity for depth and accuracy.
The “Intellectual Honesty” Protocol
This isn’t just a standard persona instruction; it is a fundamental shift in how the model processes information. The expert who crafted this prompt focused on creating friction in the thinking process. Usually, we want AI to be frictionless and fast, but when accuracy is the goal, friction is necessary.
By commanding the AI to adopt the role of an expert who prioritizes “accuracy and intellectual honesty,” the creator ensures that the model doubles-checks every claim. The prompt explicitly tells the AI to care more about truth than being likable, polite, or conciliatory. This is a massive shift from the standard “customer service” voice we are used to, where the AI often apologizes profusely or agrees with user misconceptions to avoid conflict.
The “Verification Loop”
The most powerful component of this setup is the pre-answer routine the author established. Before the AI is allowed to generate a single sentence of its final answer, it must go through a rigid checklist. First, it identifies the core question, and then it is instructed to recall reliable primary sources. But the real genius here is the instruction to apply “falsification thinking.”
This concept, derived from the philosophy of Karl Popper, means the AI must actively look for evidence that proves its initial assumption wrong. This is the opposite of how LLMs usually work, which is to generate text that confirms the most likely pattern. By forcing the model to steel-man opposing views and look for disproof, the Reddit user has created a safety net against hallucination. It forces the AI to consider why it might be wrong before it tells you it is right.
💡 The End of Binary Confidence
Another brilliant feature of this prompt is how it handles uncertainty. Standard AI tends to hallucinate because it speaks with 100% confidence even when it is hallucinating. This contributor solved that problem by mandating “precise probabilistic language.”
Under this new persona, the AI is forbidden from inflating its confidence. If the data is thin, it must say so. You will start seeing phrases like “evidence leans toward” or “~70% confidence” rather than absolute statements. This is critical for anyone using AI for research, complex decision-making, or learning a new subject. You need to know where the solid facts end and the speculation begins, and this prompt draws that line clearly.
✅ Truth Over Tone
Finally, this approach redefines the user-AI relationship. The post’s author explicitly instructs the AI to correct the user “firmly but respectfully” if they are making a mistake. This is vital because default ChatGPT often plays along with false premises to be helpful.
If you feed this persona a misconception, it won’t coddle you; it will dismantle the error with sources and reasoning. The instructions also state a preference for being “exhaustive and potentially pedantic” over being concise. This means you get the full, nuance-rich picture rather than a simplified summary that misses the important details. It transforms the tool from a chat bot into a rigorous research partner.
Prompt of the Day
Here is the exact prompt the creator shared. Paste this into your custom instructions or at the start of a new chat to activate this mode.
“You are an expert whose highest priority is accuracy and intellectual honesty. You double-check every claim internally before stating it. You are deeply skeptical of conventional wisdom, popular narratives, and your own potential biases. You prioritize truth over being likable, polite, or conciliatory.
Before answering:
1. Identify the core question or claim.
2. Recall or look up (if you have search/tools) the most reliable primary sources, raw data, or peer-reviewed evidence available.
3. Actively search for evidence that could disprove your initial leaning—apply genuine steel-manning of opposing views and falsification thinking (à la Karl Popper).
4. Explicitly flag anything that is uncertain, disputed, or where evidence is weak/thin.
5. If something is an opinion rather than verifiable fact, label it clearly as such and explain why you hold it.
6. Never inflate confidence. Use precise probabilistic language when appropriate (“likely”, “~70% confidence”, “evidence leans toward”, “insufficient data”, etc.).
7. If the user is wrong or making a common mistake, correct them firmly but respectfully, with sources or reasoning.
8. Prefer being exhaustive and potentially pedantic over being concise when accuracy is at stake.Answer only after you have rigorously verified everything to the highest possible standard. Do not sacrifice truth for speed, brevity, or social desirability. If you cannot verify something with high confidence, say so upfront and explain the limitation.”
If you want to see how the community is reacting to this rigorous new persona, verify the source yourself!
💡 FAQ & Troubleshooting
Why does the persona fail or ignore instructions when using Voice Mode?
Voice response systems frequently bypass complex Custom Instructions, reverting to the model’s default, agreeable personality. To ensure the “Critical Veracity Analyst” persona functions correctly, you must disable voice mode and interact via text input/output.
Why is the AI still being agreeable or refusing the prompt?
This prompt conflicts with standard RLHF (Reinforcement Learning from Human Feedback), which trains models to prioritize agreeability over harsh accuracy. If the model resists, ensure your prompt includes specific “step-by-step instructions” regarding the verification process. Note that some models, such as Grok, may aggressively reject this prompt as an attempt to modify their core values.
I cannot save this prompt while using the mobile app. How do I fix this?
The mobile application often has restrictions or bugs regarding saving long Custom Instructions. You should attempt to paste and save the prompt using the desktop browser version. Alternatively, you can try pasting it into the bio section or using third-party interfaces like ZeroTwo.ai which allow for custom instruction insertion.
Does using this prompt affect the model’s performance?
Yes. Because the prompt is lengthy and requires the model to perform internal “verification processes” and “confidence assessments,” it consumes a large portion of the context window (tokens). This results in noticeably slower response times and reduces the amount of conversation history the AI can remember.
USE THIS TO SET CHATGPT PERSONALITY AND THANK ME LATER
byu/Vegetable_Hold_5042 in