How do machines understand what we say and write? How do Alexa, Siri and Google Home respond to our requests? How does WhatsApp know which words we want to type while sending a message? In short: how do artificial intelligences understand us? Reading what we write, especially what we write online.
These software they devour the Internet, leafing through pages and pages and pages of articles, documents, posts on social networks, comments and texts and learn. They are great at doing it and this thing works very well, getting better and better as time goes on. But there is a problem. The problem is that these texts (the ones that machines read to learn) they are a mirror of who we are as people: if we are racists, male chauvinists, misogynists, conspirators or deniers, so will the AI. They will be because we taught them to.
The trouble has long been known within the scientific community that deals with artificial intelligence, natural language processing and machine learning and many researchers are working to solve it. Even inside a Meta, who however decided to tackle the question from an unusual side: not at the mouth but at the source, not at the output stage but at the input stage. Trying to improve what machines read to learn.
Artificial intelligences between racism and the ethical question
by Emanuele Capone
Using an AI to improve Wikipedia
Zuckerberg’s company, which controls (among others) Facebook, Instagram and WhatsApp, has decided to intervene on Wikipedia. Because? A little bit because it is among the sites most consulted by AI to study (Gpt-3, the NLP software developed by OpenAI and financed by Microsoft, has read it all), but above all because it is an excellent source also for us humans: it is among the 10 most visited sites in the world and often constitutes the first resource for those looking for information on historical and important figures. Which, however they don’t all have the same space: according to the Wikimedia Foundation itself, only 20% of the biographies on the English site concern female characters and the percentage drops even more if they are women belonging to minority groups, such as female scientists or African or Asian women. Yes: by its own admission, Wikipedia has a gender problem. And the AIs that learn from there have (will have) too.
Here Meta enters the field: Angela Fan, researcher in the Artificial Intelligence division of the Menlo Park company, has designed an “open and reproducible scientific method” that “will contribute to increasing the cultural representation of women and minorities on the Web”. And which among other things is based on artificial intelligence.
According to what was told, for Fan it was even a personal problem: in the third grade she was asked to write an essay on a historical figure to whom a book was dedicated in the school library. She wanted to do it about Eleanor Roosevelt, but since there were no books about her, it was somehow forced to talk about her husband Theodore, 26th president of the United States. If it happened today, the students would consult Wikipedia. And they would probably face the same problem.
The case
The artificial intelligence that writes stories together with human beings
by Francesco Marino
Not just women: how the Meta project works
The AI model developed by Fan in conjunction with Claire Gardent, narrator of the University of Lorrainein France, he should be able to do research online and then write biographical texts, precisely in the style of Wikipedia.
It works like this: AI searches sites for relevant information about characters and writes a draft (complete with citations) using as a sample the data used in 1527 biographies of women belonging to marginalized groups. The intention is that these drafts “may constitute a start point for people who write content for Wikipedia and fact-checkers “, speeding up the work and therefore “helping to increase the publication of biographies dedicated to underrepresented groups”.
From a technical standpoint, the process of creating a bio begins with the use of a RAG architecture (the acronym stands for Retrieval-Augmented Generation) based on large-scale pre-training, which teaches AI to identify only relevant information, such as the place of birth or where the person attended school. Then we move on to drafting of the text and finally to the bibliography, with links to the sources consulted. The idea is to arrive at a Wikipedia entry that collects all the necessary elements and includes the first years of the character’s life, his school background and career.
From what has been written so far, it is easy to understand that this complex work is even more complex in the case of marginalized or underrepresented groups online, for which less information is available.
Despite this, Meta’s idea is to devote even more attention to the project, identifying “other under-represented groups on Wikipedia, in addition to women, such as transgender or non-binary people, for which there are forms of prejudice that also affect information sites “. In the end, the result will be twofold: we will have a richer Wikipedia and consequently the AIs that learn from there will be more educated, informed and inclusive. And we can also write a scholastic research on Eleanor Roosevelt (who still has a page on Wikipedia).