AI Haluzinations...

[Mario]

The number of support emails where users have asked an AI first how to do something in IMatch is rising. Just got another one today.

The AI produced absolute nonsense, telling him to use non-existing commands in non-existing features. Even told him to look for files in a non-existing category.

This of course confused the user even more, and I've got a longer email with a two paragraph complaint about how hard it is to find the non-existing commands in IMatch 

After reading through the rand I found the actual question and answered it in one sentence - and a link to the IMatch help system where it is explained in great detail.

One would think that AI should be capable of reading the IMatch help system and don't invent stuff. Else it generates eveb more support load for me.

I use AI in software development quite often, to avoid reading 16,000 pages of documentation or to safe time writing tests or refactoring. My impressions are generally positive. It's useful when you know the limits of what it can do.

Did you ask AI for how something works in IMatch? If so, which, and did you get useful results?

[axel.hennig]

Did you ask AI for how something works in IMatch?

No, but just because the help is so great.

[Mario]

No, but just because the help is so great.

Thanks.

I've made a test just now, asking Microsoft Copilot a rather complex question.
And it got it almost right. But the error in the response is one of the things that sends users one a goose chase and cause frustration. I have highlighted the error (no such command in the context menu) below.
The rest of the answer is kinda right.

Problem is, if you don't know something, you can usually not tell that the AI is wrong either...

[Menace]

Unfortunately I have to admit that I use perplexity quite often for step by step instructions (Home Assistant, Blender, Photogrammetry and sometimes IMatch). I usually let the AI search more precisely, but it often happens that he takes old versions of the programs, some reddit comments or just fabulates, even with IMatch (when he suggested a panel for the autotagger that does not exist or I have not found so far).
I think if you work with AI you have to know about this weakness. I always link to the help page as a source for the AI and check the links first. I also often give him back tasks when I have doubts. That's how I get a lot of things solved that I couldn't manage before. But you should always doubt the AI first, not the programs or apps.
For biological species identification with automatic generation of categories and directories I use perplexity and it is a timesafer.

[sybersitizen]

Did you ask AI for how something works in IMatch? If so, which, and did you get useful results?

No, but I recently tried asking ChatGPT how to accomplish something in a popular photo editing app. It did not produce the results it said it would, it suggested using non-existent commands, and when it offered to provide a download of a file that it claimed would help me, the file could not be accessed for download.

AI in my experience often provides answers that are useless or obviously wrong.