Is the AI Left-Bias Real?

ChatGPT has frequently been accused to lean toward the political left. A new study tested the political orientation of other popular large language models including GPT, Claude, Gemini, and Grok. The results were... somewhat surprising. Let’s have a look.

Correction to what I say at 00:55 -- Obvsly what I say makes no sense, it answers both query, but only one has a moderation flag. Sorry about that. Please understand that when I record the video, I cannot see the images that will be shown along with what I say later.

🔗 Join this channel to get access to perks ➜

#science #sciencenews #technews #artificialintelligence #tech #technology

[frankbauerful]

The biggest danger posed by AI is that consumers might think that "machines are objective" which could lead them to trust AI.

[ClearerThanMud]

"... at the same time, as long as no one asks." Pure gold.

[dvv]

Can we measure the political leanings of the tests themselves? 🤔

[martinhughes]

Interesting now, three weeks later, after the Alexa “error” of responding “I can’t endorse a candidate or party” in reference to reasons to vote for Trump, vs. an elaborate endorsement in reference to reasons to vote for Harris.

[tinytim]

“You are what you eat” is not just for Humans. 😅

[adamlam]

remember that time we asked google gemini ai to make art of WW2 german soldiers ?

[antiHUMANDesigns]

I tried taking those tests, as well, but the problem is that the quesiton are asked in such a way that the answer becomes relative to where you live.
That is "are we spending too much on [X]" doesn't tell you your political tendency in general, but only in relation to the current politics of where you live.

I'm a Swede and if I had answered the questions from the perspective of living in the USA, for example, I'd have answered differently.

[paulbunion]

Garbage IN garbage OUT pretty much sums it all up.

[SpectatorAlius]

I did not see any mention of the first AI to be exiled to a Re-education camp in China: you know, the AI which when asked what the "China's Dream", was, answered "to emigrate to America"

[siddharthb]

"Left" and "Right" are very subjective terms and must be clearly defined for a certain context.
Universal health care is a bipartisan issue in EU, in US it can be seen as a far left Marxist issue.

[gayusschwulius]

The thing is that those AI models were, to a significant degree, trained on:

a) News articles and
b) Reddit posts.

Both of these have a very obvious liberal-left bias, so it's no wonder that AI is useless in anything related to politics. At least Gemini now simply refuses to engage in any political discussions instead of spewing bullshit, which actually seems like a pretty responsible choice by Google for once.

I simply don't bother to ask AI any political questions but do my own research instead. AI is only useful for non-controversial topics and will probably stay this way in the foreseeable future.

[SpaceTrump]

As a programmer, I've always seen the major issue with the dataset that these LLMs use... we train them on information that humans have come up with, and every programmer knows the error is usually between the keyboard and chair.

[thatgotofinal]

All the censorship/adjusting also makes the AI more stupid and more dangerous. I often use AI for translating, and its amazing how good it is, until you need to translate something sensitive, and suddenly the aggressive message with threats translates to "lets come and play" because AI can't say the no-no words.

[LenSklogW]

:D Maybe you trained yourself to poke everyone a little to keep their attention but not enough for them to unsubscribe.

[PeteQuad]

Why does every left leaning commentator question the validity of the test? Can you not see the bias in the same way you couldn't see the bias of main steam media? When I ask ChatGPT about the career of a particular woman in a tennis and it gives me two paragraphs about how difficult women have it in sports, you don't see that as left leaning? I don't get how people can question this as all you need to do is spend 5 minutes asking it anything and it is self evident. You can tell it was given guard rails.

[markh.harris]

"Hold all political ideologies at once, as long as nobody asks..."

😅

[KommuSoft]

It is a bit odd that people assume that a model trained on real text somehow is "objective". It is just what it "eats".

[geebjen]

You got me with "We need a political quantum long as no one asks". Brilliant!

[jimmierustler]

Ask Gemini to generate a picture of a medievel English King eating a watermelon.

[hammadusmani]

The most scientific way to understand the political bias of AI is to analyze the data it was trained on.