Opinion “The pleasure is fleeting, the position ridiculous, and the expense damnable.” Famously about sex but more probably about golf, this quote is now most accurately ascribable to AI-enhanced search engines. 

Google proved this last week when a wrong answer from its AI competitor to Microsoft’s ChatGPT cost Alphabet a cool $100 billion on the markets, thus also proving another old saying, “to err is human, to to really muck up takes a computer.” 

Widely and doubtless accurately seen as the big G getting its knickers in a twist over a very flash demo of Microsoft’s OpenAI-boosted Bing, it may in fact be the bargain of the year for Alphabet. The share price will recover – a single very good demo is no more proof of a very good product than a very bad one is of catastrophe – and if there’s one thing that needs to be shown off to everyone it’s that mixing AI and search is a bad idea. It’s bad if it doesn’t work to drive traffic, and far worse if it does. If Google realizes this, it may yet dodge one hell of a bullet – as may we.

Search is as polluted by SEO as Chernobyl is by misplaced spicy isotopes. If the search results we see are what a company, any company, is feeding its machine learning model as a corpus of ranked facts, no amount of subsequent linguistic legerdemain will render it less rancid. Quite the opposite, as we argued recently. If anything, ChatGPT’s charming way with words makes it more dangerous to us humans. 

If the search companies have got some super new way to de-SEO-ify their output, we’d like to have that for ourselves, please. But they haven’t, so we can’t, and neither can their AIs. If it’s working the other way around and the AI models really are pulling roses out of the dung heap, then that would be the first thing the search companies would trumpet. Google is at pains to say that its search output  contains 50 percent less irrelevant content than eight years ago, but that’s just not the experience of people who use it. And if it did have a new and artificially intelligent way to make better, rather than prettier, search results – that would be the first thing it would say.

There is something the Bing and Google staffers can do to improve their products. It’s a trick many of us have learned, just append “Reddit” to the search term. If you’re feeling particularly intent, site:redddit.com really drives the hint home. It also works for Wikipedia, and to a lesser extent StackOverflow and, if you must, Quora.

This is no perceptual bias. A recent paper by Ohio State University researchers looked at some 10 million Reddit posts made over a couple of years, and found that, contrary to cynical dismissal of all of the internet being a fact-deforming hellscape, Reddit posts that had been fact-checked got a lot more engagement by ways of more comments and longer conversations. Moreover, those that were fact-checked and found good got the biggest boosts of all.

It turns out that algorithmic boosting of content based on engagement, while capable of generating ad revenue, can run with the wild abandon of a nuclear core run with its cooling pumps off. We knew that, the search engine giants knew that, but radioactive revenue looks so good to shareholders on the balance sheet.

Adding the sugar coating of plausible AI glibness is like replacing the broken core cooling system with a series of ornamental fountains. It is a very bad idea. A very good idea would be to use all those supercomputer clusters that are currently lying about what the James Webb Space Telescope can see, or what the cheapest 65-inch TV is – yes, that was one of the new Bing breakthroughs that so excited the breathless commentariat. Aim them at recreating the user-friendly ontology and ability to fact-check that makes Reddit actually useful. What? Too hard? Then don’t pretend. It destroys trust, and without trust you are nothing. 

There is little doubt that machine learning and what is still incorrectly called AI can and will bring new dimensions of knowledge discovery and use. Build it on the environmentally devastating wreckage of SEO’d search, and those dimensions will be dark. 

All we can tell from last week’s Bing v Google hypefest – sorry, algorithm: NEW WAR OF THE TECH GIANTS – is that Microsoft is better at marketing than Google, but that’s an astonishingly low bar.

The test for whether these developments are good or bad for us will be the transparency with which they are described. There’s plenty of small print around OpenAI saying it’s not as smart as you might think, but by this stage the Open bit of that name is as illusory as the AI. Let’s see testable metrics.

We know, boy do we know, that trusting any industry to report on its sins doesn’t work; it took the biggest of guns to crack open tobacco, and we’re not there yet with petrochemicals. We can’t afford to carry on polluting the data ecosystem, and AI-infused search shows no sign of clearing up. Prove it or lose it; until it earns its place in our digital lives, we’re not buying it. ® 

Source: https://go.theregister.com/feed/www.theregister.com/2023/02/13/column_ai_in_search/