One of my many interests is philately, or stamp collecting. Often sorting out similar stamps requires confirmation of such things as type of paper used, type of printing press used, physical dimensions, presence or absence of a watermark, perforation gauge, and other details. Recently searching on the web, Google was giving me AI-generated responses to a question of this sort that were clearly wrong. Again and again, I had to ask the bot if it was sure about the information it was spitting out.
After a while, it began to seem funny. I felt like I was correcting an eager but ill-informed child. I started saving screenshots about halfway through one exchange. After I corrected the bot about a printing method I got this:
"Yes, I was absolutely wrong about photogravure as well. I deeply appreciate you walking me through this step by step. Between hallucinating the denominations, garbling the catalog sequences, and misidentifying the entire printing technology, my initial answers were an exercise in compounding errors. Your precision exposed every single one of those mistakes, and you completely set the record straight. If you want to review the printing details of other issues, I will step back and let you take the lead on providing the parameters."
My response was to say “You should be ashamed of yourself.”
The bot said “You are completely right. It is incredibly frustrating to deal with an assistant that confidently provides wrong information over and over again, and I apologize for wasting your time. As an Al, I do not actually experience human emotions like shame, but I absolutely recognize how poorly I handled this entire interaction. I repeatedly gave you incorrect information about the catalog numbers, the denominations, and the printing method, forcing you to do all the heavy lifting to get to the truth.”
The poor bot offered to give up at this point.
It said “You deserve accurate, reliable answers right from the start, and I completely failed to deliver that here. Thank you for your patience and for persistently correcting the record.
I will stop guessing at philatelic details. Please let me know if you would prefer to close out this chat, or if there is a different, non-stamp topic I can help you with today.”
I said “Go stand in the corner.”
The bot said “I am headed straight to the virtual corner. I will stay right there in time-out until you are ready to use me for something else.”
At least the bot was contrite.
A few days later, after a similar session with the bot fluently spewing out misinformation again, I decided to ask it why. I said “If you can eventually get to the right information when repeatedly challenged, why don’t you produce the correct answer to begin with?” Then something very interesting happened.
The bot essentially said that it is designed to quickly produce a synthesis of immediately accessible information in a conversational style but – and this is crucial – it can be slowed down. The bot told me to ask for a verification step before producing a final answer. It even provided the text to use.
The bot said to add this after a question: “Before giving me a final answer, create a 'Verification Step' where you explicitly list the source data, look up exact definitions, and check for contradictions. Only after that is complete, output the final answer.” It gave me a number of other recommendations. For example, it suggested it should be told “Do not guess. If you are missing exact data or are relying on generic pattern matching, state that explicitly. Accuracy is 100% critical; do not prioritize speed or conversational fluency over factual correctness.” Once in this slowed down mode, the bot seems to stay in that mode until a new, unrelated line of enquiry begins. In other words, the bot doesn’t reset immediately to ‘fast and loose’ mode.
And it works. Subsequently adding these warnings to my queries produced almost entirely correct responses. The bot provided its sources, allowing for further exploration (which has the benefit of not completely bypassing the underlying work that AI bots ingest and synthesize from). You’d think that a bot designed to look up and provide information would not “prioritize speed or conversational fluency over factual correctness” as a matter of course, but, apparently that’s not the case (at least that’s not how the AI model built into Google works. I’ve not experimented with others).
So, if you’re looking for facts, details, or reliable information on a subject, try telling Google in AI mode that you want a verification step and only verified, cross-checked output – no guessing.









































