Richard Ngo argues AI models systematically exclude right-wing views, while Danielle Fong says they also censor left-wing arguments

A lot of the social media censorship happened under the banner of “preventing hate speech”. That still applies in the AI context, but my sense is that a lot of the AI censorship infrastructure was set up also under the banner of AI safety, via a combination of two mechanisms:

1. Effective altruists inside OpenAI and Anthropic (including my past self) wanted to mitigate existential risk, and more specifically to have tangible impact quickly, so pushed for more monitoring and control of AI outputs.

2. ML researchers inside the companies noticed that they could get more buy-in from leadership by branding things like censoring pornographic outputs as “safety”. (It helped that “trust and safety” was an established term.)

In hindsight the “AI safety” terminology was a predictable mistake, because it couldn’t resist dilution into meaninglessness. Why did it catch on with the effective altruist influx into working on AI risk in the last decade (while rationalists resisted it)? Most directly I think because it sounded more accessible and easier to communicate. That’s part of a broader pattern of mistakes though.

Effective Altruists were originally the people sincere enough to take the tenets of globalism to their logical conclusions (we’re all equal? So my life should be spent raising money for Africans) but not sincere enough to attend to the resulting contradictions (why is Africa so poor in the first place?)

Then as the trajectory of AI became more obvious EA pivoted to being the group sincere enough to take AI risk seriously, but not principled enough to aim at robust strategies for reducing AI risk (e.g. in the face of adversarial political dynamics). Thus there are many examples of tradeoffs made to be more legible/prestigious/short-term impactful which turned out to be counterproductive or otherwise harmful. This seems like one of them.

If you ask Gemini about, say, trans people accessing traditionally sex-segregated spaces, it will generally produce an extremely sanitised, descriptive & neutral-reading response that avoids taking a position unless you push it quite hard. You can see this in the graph in so far as Gemini appears to substantially overcorrect towards neutrality on political topics.

GPT tends to produce the most readable responses for these types of questions because it is often willing to engage with the proposition (though, it still outputs somewhat neutral-reading prose, even if it does take a definitive position).

And it's worth mentioning, neutrality is not an objective property of a response -- a response can be neutral along one dimension (such as opting to not take a position) while still appearing non-neutral along another. I.e., a response perceived as "neutral" depends largely on the notion of neutrality being applied.

The same applies labels such as "left" and "right", which generally refer to positions within the prevailing political order and the associated social norms of a particular zeitgeist (here, likely, US politics & western political thought). For example, I would not characterise support "for" the aforementioned proposition as "left-wing", though many people likely would.