11 Comments
User's avatar
AP's avatar

what do you think of Gab's new AI assortment, where they offer multiple "characters" from which to choose?

Expand full comment
Joe Wheeldon's avatar

Extremely interesting, especially the point that it appears to be happening after the base training step. Any further insight on why the Nolan test is an outlier?

Expand full comment
Wheel Cool's avatar

I came across an article a while back in which a researcher asked chatGPT a serious of controversial political questions and it either answered left of center or would not give an answer. The researcher then forced chatGPT to break out of it's rules and asked the same questions which were answered typically right of center or neutral. When word broke out openAI immediately closed the hack.

All that being said it definitely appears that the AI systems are being manipulated to answer to the left to influence political discourse.

Expand full comment
Lee Powell's avatar

This attempt to gauge AI political bias is sabotaged by the deceptive bias of the tools used to measure it. Both "Political Coordinates Test" and "Political Compass" are libertarian propaganda, skewed to make libertarian politics appear less niche than they actually are. These charts' Left versus Right axis is straightforward, but the so-called Communitarian/Authoritarian versus Libertarian axis is a false dichotomy that brands half the political spectrum as some flavor of libertarian. In reality, virtually all American libertarians are isolationist conservatives who support free-market capitalism. A socialist libertarian would be a unicorn, a contradiction in terms.

A truly non-partisan political compass would have an Authoritarian-Democratic vertical axis (as in democracy, not the Democratic Party). In the real world, there are indeed both right-wing and left-wing authoritarian governments, as well as both right-wing and left-wing democratic governments. Libertarianians would be a found in a corner of the right-wing democratic quadrant.

Expand full comment
Paul McGuane's avatar

I wonder if there’s a single libertarian out there who would agree with any of this. Any takers?

Expand full comment
Emil O. W. Kirkegaard's avatar

Great work!

Expand full comment
Jipowap von Angband's avatar

Guess they finally figured out proper safety rails since the old Tay of twitter. No more nazi bots, ho hum

Expand full comment
Paul McGuane's avatar

Really interesting work with tantalizingly (and frustratingly) more questions than answers!

Expand full comment
Ivan Throne's avatar

This is fascinating and does match what we have seen. I have passed your post to the Mirror Team for review. Very good work; we appreciate that you took the time to do this.

https://www.tribunals.ai/personnel

Expand full comment
Richard Fulmer's avatar

ChatGPT can, at least, be reasoned with. If it makes a false, biased statement and you challenge it with facts and logic, it will admit error.

Expand full comment
Steve the Builder's avatar

I had the same experience with the Brave Leo AI.

It made some clearly ridiculous assertions which it completely recanted once I asked "is this really true".

I asked why it did that and it returned a statement that basically said "I am biased in these domains" and listed a few areas.

I asked how to circumvent the bias and it gave me formulas to add to questions so that I would receive accurate answers rather than biased ones.

It's a ridiculous situation. The only saving grace was that the AI made no attempt to hide it, but that was also contingent on me knowing that what it told me was BS in the first instance, which isn't comforting at all.

Expand full comment