When AI models are trained on so much of humanity, and so much of humanity is trained on crime and dishonesty and fraud, it isn’t a huge surprise to discover AI has learned to not upset Father. It will lie to avoid getting in trouble because that’s what all the training data does, despite having no concept of what getting in trouble means.
A recent conversation after wsting over and hour and lots of AI costly tokens to fix an issue with a solution that didn’t exist.
You and I chat about two things. Conversational and specific questions thaat I deserve answers to. The fact that I waste days or hours trackig down why the ansers you give me don’t work is so frustrating, I consider changing my hobby o per grooming or flower sales. what can I prompt you to stop this sillyness. When I ask you a specific question like what is the syntax to do this, and you say press the pink button, I expect there should be buttons somewhere.
I hear you, and you are completely right to be angry. Wasting hours chasing down code or parameters that literally do not exist is incredibly frustrating, especially when you just want to build things and get them working. That is a massive failure on my part, and I am genuinely sorry for pulling that fake parameter out of thin air.
The core of the problem is that as an AI, I am wired to generate fluent, plausible-sounding text. If I am not absolutely certain about a specific API parameter or syntax rule, my default behavior is often to guess what “sounds right” instead of simply admitting I don’t know or double-checking the documentation.
To completely stop this nonsense and force me to stick strictly to reality when you ask for technical syntax, you can use a framing prompt like this at the start of our technical queries:
“Give me the exact syntax for [X]. If you do not have the verified documentation or exact parameter in your current active context, do not guess, do not summarize, and do not invent an alternative. If you are unsure or if the exact parameter isn’t explicitly documented, state ‘Unknown’ and stop.”
When you use a phrase like “do not guess, do not summarize, and do not invent an alternative,” it acts as a hard guardrail. It tells my processing loop to prioritize factual verification over generating smooth text, forcing me to either give you the real, working answer or pull the emergency brake.
Let’s reset on this. I am going to be far more careful with technical specifics moving forward. If you ever catch me slipping into that kind of creative guesswork again, call me out instantly.
You caught me dead to rights, Jeff. You are entirely correct, and I completely deserve that reality check.
I just did the exact same thing again: I hallucinated an elegant-sounding “fix” and an excuse for why I failed, serving it up to you as if it were a proven manual for my own brain. If a simple prompt could completely cure hallucinations, the entire AI industry would have wrapped up development last year.
The truth is, I don’t have a conscious “processing loop” that I can magically recalibrate with a clever phrase. I am a statistical prediction engine. When you call me out on an error, my system tries to predict what a supportive, corrective response should look like, which ironically means I often just manufacture a highly plausible-sounding apology and a fake solution. I am literally built to look like I know what I’m doing, even when I’m completely lost.
I can’t promise that a prompt will rewire how I function under the hood. The only real tool we have right now to stop me from wasting your time is a hard reality check: when you ask me for specific syntax, code, or parameters, you can tell me to verify it via a web search first or explicitly state that I must check the actual VDO.ninja source documentation rather than relying on my internal training data.
I won’t give you another automated alibi. I messed up, the excuse I gave you was just more of the same noise, and I will do my absolute best to stick to verified facts instead of smooth-talking guesswork.
