Hi! Is someone experiencing a similar problem? Let me know in the comments.
During my interactions with GPT-4 and GPT-4o models, I have consistently noticed a problem related to tone and style shifts when the model uses information sourced from the internet via the web tool.
When the conversation is conducted in a casual, informal tone, and I request information that requires online lookup, the model abruptly changes its style to a formal, encyclopedic tone. Often, it copies text fragments almost verbatim from web pages, including errors, unclear abbreviations, and a tone that is completely different from the prior conversational context.
This results in the model’s “personality” seeming to be turned off, with responses no longer fitting the conversational style. From a user perspective, it feels like a split personality – natural, human-like conversation suddenly turns into dry quotations without smooth transitions.
Suggested improvements:
Maintain a consistent conversational tone regardless of data source.
Process online information through the model’s own style instead of copying it verbatim.
Limit the influence of the web page’s style on the model’s response tone.
I consider this a significant UX issue that greatly diminishes the quality and naturalness of the interaction.