chatbots always seem to be like this. Yesterday it was a specific website (for a paid service) with help offered by a chat bot built in to the UI. It was telling me where to find settings with specific instructions that just didn’t match the UI.
One of it’s answers was longwinded instructions on how to build an app that would change the settings via their API. There were a few sentences at the end suggesting a “low code” option where I could change the setting in the web ui.
I think what I was doing is a fairly common use case, a guide explaining how to achieve what I wanted would have been much quicker.
I have had this happen pretty often. I think even the wording used in the “prompt” can also influence whether the chatbot hallucinates a solution.
I remember recently asking a third party’s mandatory chatbot support system (if it can’t help, it then forwards you to a human) whether I was able to change the mapping of some fields in an upload to them without the need for a support agent. It walked me through the process but it didn’t work. Eventually, after three failed chatbot modifications to the instructions, I was forwarded to a human that confirmed I was not, in fact, able to do it myself. Such a fucking waste of time.
What gets me is that despite how often LLMs hallucinate a completely incorrect answer like this, for some reason people still blindly trust them and regurgitate their output without telling you where it came from.
You would think after the first time of encountering an output that is verifiably false people would at least have some hesitancy in the output, but they still keep blindly trusting the language guessing machine.
chatbots always seem to be like this. Yesterday it was a specific website (for a paid service) with help offered by a chat bot built in to the UI. It was telling me where to find settings with specific instructions that just didn’t match the UI.
One of it’s answers was longwinded instructions on how to build an app that would change the settings via their API. There were a few sentences at the end suggesting a “low code” option where I could change the setting in the web ui.
I think what I was doing is a fairly common use case, a guide explaining how to achieve what I wanted would have been much quicker.
I have had this happen pretty often. I think even the wording used in the “prompt” can also influence whether the chatbot hallucinates a solution.
I remember recently asking a third party’s mandatory chatbot support system (if it can’t help, it then forwards you to a human) whether I was able to change the mapping of some fields in an upload to them without the need for a support agent. It walked me through the process but it didn’t work. Eventually, after three failed chatbot modifications to the instructions, I was forwarded to a human that confirmed I was not, in fact, able to do it myself. Such a fucking waste of time.
What gets me is that despite how often LLMs hallucinate a completely incorrect answer like this, for some reason people still blindly trust them and regurgitate their output without telling you where it came from.
You would think after the first time of encountering an output that is verifiably false people would at least have some hesitancy in the output, but they still keep blindly trusting the language guessing machine.