This is due to a technique used with this type of old machine learning technology, called "post processing". It is heavily primed with extremist biases. Studies have found that all current public ones are, and of course they are far on the alt-left-liberal side.

"The Political Biases of ChatGPT" by David Rozado https://www.mdpi.com/2076-0760/12/3/148

When it comes down to it, once you train a good machine learning data pipe, like ChatGPT, your have a core platform. You can then quickly, cheaply, and easily spin that up into different things with little effort. Things like 15,000 fake Twitter users with their own psudo personalities who all politically align with extremist alt-left talking points exclusively. 4 thousand fake Facebook users purporting to be a victim, describing the emotional impacts of some aggressors they were configured to, while purporting to live in some proxy war country that is heavily used for corruption, money laundering, and foreign political agendas. Those are just some middle uses, that haven't dipped into the weaponization of "AI" yet, which usually hides in "think tanks", who can over throw a country with one well crafted sentence to be used as a talking point, which bakes in centuries of emotional and mental manipulation techniques.

We have to remember that this type of machine learning is 30+ years old. It came to maturity behind closed doors and began being weaponized to use against the public 20+ years ago. Now that it is being rapidly expanded to let the poor plebs know of it and play with it, this implies they've now far surpassed this, hit its limitations, and or conjured up an even greater way to oppress and control the masses, likely by using this as bait to trick people into helping.