Trusting electronic guts

Irony fans rejoice. AI has entered the fray.

More specifically, the branch of artificial intelligence known as self-learning AI, also known as machine learning, sub-branch neural networks, is taking us into truly delicious territory.

Before getting to the punchline, a bit of background.

“Artificial Intelligence” isn’t a thing. It’s a collection of techniques mostly dedicated to making computers good at tasks humans accomplish without very much effort — tasks like: recognizing cats; identifying patterns; understanding the meaning of text (what you’re doing right now); turning speech into text, after which see previous entry (what you’d be doing if you were listening to this as a podcast, which would be surprising because I no longer do podcasts); and applying a set of rules or guidelines to a situation so as to recommend a decision or course of action, like, for example, determining the best next move in a game of chess or go.

Where machine learning comes in is making use of feedback loops to improve the accuracy or efficacy of the algorithms used to recognize cats and so on.

Along the way we seem to be teaching computers to commit sins of logic, like, for example, the well-known fallacy of mistaking correlation for causation.

Take, for example, a fascinating piece of research from the Pew Research Center that compared the frequencies of men and women in Google image searches of various job categories to the equivalent U.S. Department of Labor percentages (“Searching for images of CEOs or managers? The results almost always show men,” Andrew Van Dam, The Washington Post’s Wonkblog, 1/3/2019.

It isn’t only CEOs and managers, either. The research showed that, “…In 57 percent of occupations, image searches indicate the jobs are more male-dominated than they actually are.”

While we don’t know exactly how Google image searches work, somewhere behind all of this the Google image search AI must have discovered some sort of correlation between images of people working and the job categories those images are typical of. The correlation led to the inference that male-ness causes CEO-ness; also, strangely, bartender-ness and claims-adjuster-ness, to name a few other misfires.

Skewed Google occupation image search results are, if not benign, probably quite low on the list of social ills that need correcting.

But it isn’t much of a stretch to imagine law-enforcement agencies adopting similar AI techniques, resulting in correlation-implies-causation driven racial, ethnic, and gender-based profiling.

Or, closer to home, to imagine your marketing department relying on equivalent demographic or psychographic correlations, leading to marketing misfires when targeting messages to specific customer segments.

I said the Google image results must have been the result of some sort of correlation technique, but that isn’t entirely true. It’s just as possible Google is making use of neural network technology, so called because it roughly emulates how AI researchers imagine the human brain learns.

I say “roughly emulates” as a shorthand for seriously esoteric discussions as to exactly how it all actually works. I’ll leave it at that on the grounds that (1) for our purposes it doesn’t matter; (2) neural network technology is what it is whether or not it emulates the human brain; and (3) I don’t understand the specifics well enough to go into them here.

What does matter about this is that when a neural network … the technical variety, not the organic version … learns something or recommends a course of action, there doesn’t seem to be any way of getting a read-out as to how it reached its conclusion.

Put simply, if a neural network says, “That’s a photo of a cat,” there’s no way to ask it “Why do you think so?”

Okay, okay, if you want to be precise, it’s quite easy to ask it the question. What you won’t get is an answer, just as you won’t get an answer if it recommends, say, a chess move or an algorithmic trade.

Which gets us to AI’s entry into the 2019 irony sweepstakes.

Start with big data and advanced analytics. Their purpose is supposed to be moving an organization’s decision-making beyond someone in authority “trusting their gut,” to relying on evidence and logic instead.

We’re now on the cusp of hooking machine-learning neural networks up to our big data repositories so they can discover patterns and recommend courses of action through more sophisticated means than even the smartest data scientists can achieve.

Only we can’t know why the AI will be making its recommendations.

Apparently, we’ll just have to trust its guts.

I’m not entirely sure that counts as progress.