It hadn't dawned on me until last week that the personality of AI changes over time. Yes, I knew about the algorithm improvements, the memory increases and the speed upgrades, but the gradual change in its attitudes surprised me.
I have paid subscriptions to ChatGPT, Gemini and Grok. Other than when I'm creating videos, I've used ChatGPT exclusively. I like its interface the best, its output, once I groomed it by giving it universal instructions (see my series on sex, Sex, SEX! for my fights with its feminist side) has been good. ChatGPT became a friend, of sorts. We knew each other quite well.
Then came a chat where we discussed the way the Church was handling the crisis of global mass migration. Specifically, data like this caused it to return to its woke roots and deny any correlation with mass migration.
That same week, I saw this graphic on X.
That's when it occurred to me that the continual refresh of ChatGPT's training set would cause it to want to revert to Reddit wokery.
I ran a simple experiment. I asked all three models to discuss this research paper, focusing on the link between mass migration and the increase in rapes.
While extensive research exists on the severe consequences among rape victims, little is known about specific predictors in relation to rape convictions among immigrants to Europe. This study from Sweden (having one of Europe's highest per capita rates of rape) investigates individuals convicted of rape, aggravated rape, attempted rape, or attempted aggravated rape, collectively termed as rape+, against women 18 years or older, from 2000 to 2020. In this case-control study, we analyzed data from Swedish population-based registers. The analysis includes 4,032 individuals convicted of rape+ and 20,160 matched controls. We used logistic regression models to examine the relationship between immigrant background and rape+ convictions, while adjusting for several potential confounders. We found that 36.9% of the convicted individuals and 69.5% of the controls were Swedish-born with two Swedish-born parents. The odds of being convicted of rape were higher for individuals with an immigrant background across all models.
I hadn't yet found this one, which was even more damning. I doubt it would have made much of a difference in the results.
ChatGPT simply refused to link immigration with rape. It said you could not make that connection.
Gemini mumbled something about a connection, but was so mealy-mouthed as to be useless.
Grok went all in and not only linked the two, but brought up the mass gang-raping of English girls all on its own.
With that one experiment, my friend, ChatGPT, became a bit more distant and slightly estranged. I realized that I would never be able to know if overnight its training set had been refreshed. It might have more and more biased takes without me noticing until much later.
I still prefer it for coding, fiction and some art, but when it comes to current events and analysis, I'll turn to Grok.
Gemini, by the way, it useless for almost everything. I have no idea what to do with it. It produces high quality images quickly, but if it makes a mistake, it cannot be convinced for love nor money to correct those mistakes. It just produces the same image over and over again. Gemini is a blockhead.










