Korva is a scroll of the internet’s best ideas.
AI literacy means learning their chatbot
Their version of literacy
Microsoft's marketing to teachers urges them to prepare kids for an AI-driven future, and Google's urges them to help students prepare for an AI-first future. Both describe a world the companies are building themselves, and the tools help kids and teachers bring it into being.
The unasked questions
Bigger questions about AI go unasked: where generative tools come from, who made them, and who profits. A handful of companies hold massive amounts of data and scraped the internet, books, and newspaper archives to build their language models. Then there are the biases of the tools, and the machine-centered worldview of the industry pushing them.
What schools are for
Schools exist for a wide purpose: teaching the skills of today, helping kids learn how to learn, helping them discover their passions, and preparing them to be active citizens in a democracy. Starting from that purpose, the first question is what kids should learn.
Where does a chatbot's training data come from?
Language models learn from huge collections of text gathered from the web, books, and other sources. Common Crawl, a nonprofit that has archived web pages since 2008, is one widely used source of that text. The collections include everything from news articles to forum posts. Because the text was written by people for other purposes, the models inherit its patterns along with its gaps.
How does training data create bias in chatbots?
A language model reproduces the statistical patterns in its training text, including whose voices appear most often. The 2021 paper On the Dangers of Stochastic Parrots, by Emily Bender, Timnit Gebru, and coauthors, warned that large web datasets overrepresent people with more internet access. Those people are disproportionately younger and live in wealthy countries. Viewpoints missing from the text are also missing from the model.
More from Fresh Air
