Instead of a nefarious conspiracy, maybe it's just bad software

My guess is it is pretty hard to make AI that is both reliable and sensitive.

If a system is designed to give weight to users' reports of content offensiveness, it will bias toward the preferences of whichever users are most likely to report offensive behavior.

Within a Victimhood Culture, where one is encouraged to appeal to authority for justice, people are especially likely to flag content they find offensive. Within an Honor Culture, where people tend to directly confront offenders and where appealing to authority for justice is deemed a last (rather than first) option, people are not only reluctant to flag content but they are likely to respond to offense in a confrontational way that gets themselves flagged.

I think this dynamic probably explains the majority of the left-leaning bias in social media; it's far less a nefarious plot and far more a consequence of a fundamentally untenable design constraint.

Sam Altman @sama

we are working to improve the default settings to be more neutral, and also to empower users to get our systems to behave in accordance with their individual preferences within broad bounds. this is harder than it sounds and will take us some time to get right.

When Altman says, being “more neutral…is harder than it sounds", I think that's an understatement. In a platform where content reach is impacted by user behaviors like flagging, stated goals of neutrality and accuracy will be incompatible with the unstated goal of sensitivity. Further, when endorsements from academic institutions or public health authorities impact AI's assessment of information credibility, the same sort of bias will emerge—not because humans told the AI to be biased, but because humans told the AI to accept signals of credibility from sources that operate within a Victimhood Culture. As a result, ChatGPT might be reluctant to write a nice poem about someone the victimhood crowd tends to hate but happy to write one about someone they tend to admire.

🐺 @LeighWolf

The damage done to the credibility of AI by ChatGPT engineers building in political bias is irreparable.

This applies to real groups, as well, where shared values are the "software". Hence the hard-left/hard-right tilt in organizations that have moved toward Victimhood/Honor Culture and away from Dignity Culture.

Within a Dignity Culture (as people over the age of 40 might remember) one generally acquires status by showing restraint in the face of offense, rather than immediately challenging the offender to a fight or calling for the authorities to intervene.