Sorry guys, a post in English to make it more possible that guys like Sam Altman can read it. What is pretty unlikely. But hey, let's try :-) If you can, forward it, please ;-)
I finished reading an article at Times.com about #OpenAI trying to make Ai more democratic. That's a RED FLAG for me.
You all know the problem of what happens when you use todays LLMs. You get an answer as if it were a fact. But we do know in the meanwhile: We can't say if it is. The model just tries to be as close as possible with automated statistics to what the web and people would say. And by chance it's not wrong.
#LLMs are getting better and better in that. And if you ask a LLM today you get an answer which is actually useful most of the time.
At the same time, we know: There are biases, a lot of them!
And here comes the problem: In a society which is also political and public opinion is important for elections, which political bias should the LLM reflect?
To solve that problem, you could try to do the same as you did before: You feed the LLM with what most people in a democratic society agree upon. So you have a societal consensus.
Thats what OpenAI tried to do in different ways, which is layed out very well in this article. And I fall short here to explain how, but it's important to say that they failed, although is was important for learning. The reason why they failed is very clear: There is no such thing as A democratic opinion you could reflect. Because a society and what it thinks is in flow. Always. An AI can not be of any kind without biases and be democratic. As long as it behaves in a manner of "As if".
A solution might be, and for me as a journalist this sounds like a truism: An AI must behave as a librarian or an archivist. Its first job is to sort information, explain where you can find it, where it comes from, what quality the information has and how to understand it best. As soon as it starts prioritizing information as if it would be the only answer, have it's own opinion and so on, it starts manipulating opinion. Even in its weakest form, this is very dangerous. We have learned that from disasters like the Trump-election in the US, where Facebook (congrats for that achievement, too, Mark Zuckerberg), were heavily used to manipulate electors.
Whatever OpenAI tries to do to "democratize" it's tech, it's a red flag. It should actually do everything to not be democratic but keep its hands out of opinions. Else it is manipulating what people think. AI should be like a librarian, not like a thought leader.
And I even think that should be written in our regulations. Thierry Breton
Just imagine on what scale an AI by OpenAI could manipulate the opinion, even if it were not intended to do!
This important question gets asked in the article: "Would it ever truly be possible to democratize AI?"
No. Its purpose is not to have any direct influence on democracy at all but help people to be better democrats.
That counts for all tech, btw.
