Musings of a Restless Mind

I’m a SF writer and have participated in multiple SF conventions this summer. If you’re not familiar with the world of ‘Cons’, they are a ubiquitous practice where those formerly known as nerds and geeks gather and celebrate the culture of comic books, science fiction, fantasy, science, potential futures, etc., exchanging ideas in the hopes of grooming more folks like Jeff Bezos and Elon Musk who, unlike NASA, actually want to see humanity go to space. Almost every week there is an SF Con going on in some city across this nation. Even if you don’t know about the world of Cons, you might have noticed that the geeks have inherited the Earth. When was the last time you heard about a big movie not involving superheroes or at least SF and comic book themes?

At each of these gatherings I’ve participated in this year, AI (machine learning) has featured as a huge topic of debate. One of the things I’ve emphasized in my talks on AI is that the tools open to the general public were trained on data that has been separated from its source. As another writer put it, “On the internet everything seems equal,” but, as a historian, I learned that is far from true. It’s not even true that you can trust one validated source to be equally true in everything it says. Einstein might be an important source on Physics, but it’s foolish to take his views on footwear with equal gravity.

What brought this to mind today is that I was reading in one of my (carefully curated) news sources about a publisher of scientific journals, retracting dozens, and potentially, hundreds of published ‘scientific’ articles because of insider dealing in the peer review process or data manipulation (what we ordinary people would call lying) by the authors. The thought occurred to me, “Does AI ever unlearn what it’s been trained on?” Even if you kept track of the sources of the documents your AI scanned, wouldn’t it be prohibitively expensive to monitor each source for retractions and then somehow tell your AI tool to “forget” that document? As I noted in a short earlier article, a lot of the meaning that human readers absorb from a news story is conveyed by adjectives and adverbs. Do AI platforms internalize such modifiers as ‘facts’ rather than opinions? Do they even differentiate between facts and opinions? I had to solve this problem for the Navy once as I chronicled in my work biography, but then I’m just a voice crying in the wilderness. No one ever listens to Zathras.