Right Tool, Wrong Job

There’s a joke about a drunk hunting around the ground below a bright street light, and when asked what he’s looking for, says that he dropped his keys somewhere up the road.  So why is he looking here? “The light is better here,” he says.

OK, so you didn’t come here for the humor, but for the sharp analogies that jokes can prompt (self-referential reference here). Although this one isn’t my analogy: this is cribbed from AI expert Gary Marcus, a professor at NYU, on a smart podcast discussing the promise and limitations of AI with Ezra Klein, about our obsession with ChatGPT and our attempts to have it solve any multitude of problems – not because it’s good at them, but because the light is so much better under it. 

There have been thousands of words written about ChatGPT and its miraculous capabilities, massive shortcomings or apocalyptic dangers. This isn’t one of those pieces.  It’s more about what it doesn’t do well, what what it does do well could do for journalism and why we should be looking elsewhere to fill the gaps it can’t.

To be sure, ChatGPT is very good at some things.  It’s an astounding language model, which means it can produce human-sounding text in multiple styles at scale, and will doubtless upturn any profession industry that requires writing as an output – business executives and copywriters, for example.  That doesn’t mean it’ll put people out of work – although it certainly could; it’s more that people who aren’t great at expressing themselves might get a tool to help them on that front, just as a calculator helped people who weren’t great at doing math with pencil and paper. Given the right prompts, it can turn mediocre ideas into acceptable prose – a low bar, perhaps, but then again lots of writing ain’t Shakespeare. (There are a whole set of other questions about equity and the new “AI divide” between those who have access to such tools and those who don’t, but that’s a topic for another day.)

And, as I noted earlier in an earlier post, it’s uncannily good at “understanding” questions, and if combined with a good backend search engine, could well revolutionize search as well.  (As Microsoft is doubtless thinking with its investment into OpenAI and integration of ChatGPT into Bing.)

What are its weaknesses?  It can’t do math, for one, as CNet’s ill-fated experiment in ChatGPT-generated stories demonstrated.  And that’s pretty basic skill for journalism.  Nor is it great at discerning fact from fiction, as any number of people have shown. And while it can create new content in multiple styles, all that is ultimately based in some broad way on words that have been written before.  It isn’t original in that sense. And, per Gary Marcus:

What it’s bad at is abstraction. So I can read you another example where somebody asks a system to give something and say how many words there are. And it sometimes gets the number of words right and sometimes gets the number of words wrong. So the basic notion of counting a number of words is an abstraction that the system just doesn’t get. There are lots of abstractions that deep learning, in fact all abstractions of a certain narrow technical sense, these systems just don’t get at all.

So it’s not all that helpful to criticize it for not being original, for not understanding concepts, or for not performing great (or even mediocre) journalism; that’s not what it’s built to do. After all, you don’t complain that Excel does a bad job of writing stories; it’s not supposed to. At heart, ChatGPT is a language model that does an astoundingly good job at putting words one after another that cohere.  It doesn’t “understand” any of them; it doesn’t analyze them, or facts, per se. It’s taking large amounts of data and predicting, based on the words it has ingested, how to create something new.

And when those words are largely accurate, it can give a pretty good answer. But when those words are riddled with inaccuracies, not so much. But journalism is often about new words: new facts, new analysis, and new ideas. It’s math, at one level.  It’s analysis.  It’s inferences and weighing of conflicting statements.  It’s verification of new facts.  Which leaves a system like ChatGPT without a lot of training data to work from.

For that – at least in my limited understanding of AI and technology – you need something more like symbolic logic.  You need a system that can take data, analyze it, look for patterns that are “interesting” or “insightful” and surface them, whether in text or some other format.  That’s what we were building when I was at Reuters with Lynx Insight.  Language generation was the least interesting part of it; what we wanted was smart pattern recognition.  Does this financial data suggest some major shift in company strategy?  Are corporate insiders bailing out?  And so on.

Pages: 1 2