I'll moderate that statement, it isn't completely incorrect. It's a perceived behavior, which I believe is largely attributed to prompt quality and style. While the root issue is user knowledge base, and being able to leverage the tools properly, the different tools do respond differently to prompts and prompts need to be structures differently to work best. If the user does not drove deep enough to understand them methods and limitations of the tools, the observed behavior is simply that it isn't consistent.

(FYI, you can wipe all your stored data and start with a fresh model any time, so if you've botched the customization you can just start fresh when you want to. You can also detect which AI model version you apply in case you don't like how a particular one responses, though I would say in general the newest is likely to be the best quality unless you have a specific reason to revert.)

I am by no means an expert, I am very much still exploiting and learning AI tools. But I started out asking each tool what it does best, what its known weaknesses are, how to best verify accurate and thorough responses from it, and how best to structure a prompt to receive accurate and truthful responses and avoid hallucination and other errors. Then I asked each tool the same about the other tools. This built a basis of understanding that I then experimented with on familiar topics so I could easily verify or validate the output. I found some models work better with the way I prompt, and the information I tend to use AI for, than others, and so I have gravitated to those tools more than others. Currently I find Claude and Gemini to be my most used AI platforms, Claude for informational research and background, like an interactive encyclopedia, and Gemini is better with more recent inputs, news, forums, aggregating opinion pieces and summarizing a broad view of a topic. Particularly with opinion pieces, it's easy to aggregate a majority opinion and several dissenting opinions with direct references for each example so you can look for yourself. Done platforms need the target question first followed by relevant data and boundaries, some need a pile of data and valid sources first and the request and clarification at the end. Some perform better with more complex tasks with more data, and some are better at keeping it simple. Dialing in your prompt and the type of request to the platform you use is a big part of being successful with the current AI tools.