This is mostly the fault of the model, a lot of them have been trained to generate HTML in a specific style. Claude's is pretty distinct for example, I think the new DeepSeek copies it. Some of them can generate more humanlike HTML like Kimi K2 IIRC, which I feel is the model with the least amount of post-training in general.

It's necessary if you don't want it to generate HTML with images and other assets you don't have of course, that's why they use emojis or meticulously handcrafted SVGs, or WebAudio synthesized sound which pretty much no humans did before.