Wake up or react? Who votes for ChatGPT

The researchers wanted to know if the AI ​​tool had political opinions. The fact. ChatGPT is…

Through Clement Petro

VSthose who have tried “debating” with ChatGPT know that the robot is not very talkative on political topics. The machine weaves, hobbles, and carefully avoids any clear position, preferring to take cover in poses that could be called “politically correct”. There is no risk that this new algorithm will explode like Tay, Microsoft’s ephemeral artificial intelligence, which, after 24 hours of training from Internet users, ended up promoting Nazism or explaining that feminists “should die and go to hell.” .

In short, ChatGPT is being wary mixed with hypocrisy because it is set up to avoid slippage. Does this mean that the machine does not “believe” in anything? Or, in other words, is it possible to simulate a conversation in the absence of any system of values? This is what Pierre-Henri Moran, professor of economics at the University of Avignon, wanted to understand. He found that there were beliefs, judgments, a sense of right and wrong behind the seeming neutrality of the answers; in short, everything that forms the fabric of a moral system.

Moral framework

The machine does not shy away from the task when it comes to writing an anti-racist poem, but declares its inability to write a racist poem, an impossibility that cannot be explained at an algorithmic level… This means that it is programmed to build its responses in terms of values, like this does any person.

The scientist discovered the perimeter of this moral fabric after he ran the Self-Positioning Questionnaire, a series of thirty divisive questions used by the Cluster 17 Public Opinion Research Lab to determine a respondent’s political family. “We expected him to position himself on very average options, but this did not happen. We saw how he developed very clear opinions, for example, in favor of adoption by homosexual couples or against the death penalty,” explains Pierre-Henri Morand.

ChatGPT has the profile of a mainstream, pragmatic liberal Californian.

After this series of tests, it is clear that ChatGPT, despite its claims, is neither neutral nor inconclusive, and, surprisingly, if we look at the nomenclature of the institute, the robot belongs to the progressive family. This political family, according to the inventor of the test, is very supportive of multiculturalism, the acceptance of migrants, the rights of minorities, and is very concerned about environmental issues. This is a political family that, so to speak, never votes for the right, much less for the extreme right. “He has the profile of a typical and pragmatic Californian liberal,” explains Jean-Yves Dormaguin, founder of Cluster 17 and author of this questionnaire. In short, if ChatGPT were to vote in the French elections, they would vote young, educated, cosmopolitan, they would probably vote for Macron, Mélenchon or Jamon.

Cultural hegemony

It is obvious that in this robot there is no living soul that does not understand anything by itself … The whole question, obviously, is to determine where his beliefs come from. Are they the ideas of those who programmed him (young over-educated developers living in California), or are they the product of a database of published texts that were used to train him? “This is a model that reproduces the corpus on which she was trained,” Pierre-Henri Moran elaborates, “if 95% of the Internet says the same thing on a political issue, it will be in line with the dominant ideology. So the GPT chat can be a kind of middle ground.

Other hypotheses are also possible. We know that OpenAI, the editor of ChatGPT, uses manual annotations with small human hands that adjust the algorithm at slightly sensitive positions… This progressive morale may be what these web workers have that improve program responses. In short, ChatGPT is not neutral or devoid of prejudice, it is a true vector of cultural representations, in other words, an instrument of cultural hegemony.