AI See - or Rather, I don't by Neil McGowan

 

It’s been a bit of a grumpy month for me. I’ve had the dubious pleasures of AI foisted on me at work, and spent far too long looking for an off switch for it.

To be perfectly honest, I’ve never really bought into the hype about AI. I’ve always thought it’s a bit like 3D TV – quite niche, and ultimately not that useful. Because, fundamentally, it’s not really intelligence – it’s just a cleverly programmed algorithm with a vast trove of data to fall back on for pattern matching.

I tried to be fair when evaluating it, but gave up after a day when it confidently informed me that the word ‘strawberry’ contained the letter ‘R’ four times, and showed me a picture of a raspberry. It also insisted there were twenty-five years in half a century. Sigh.

Despite this, I’ve had a stream of people gushing over it. The most ironic comment was that it would ‘write my emails for me’ – all I had to do was proof read them and correct them…

Suggesting it would be as quick to just write them myself seemed to make no sense to the person concerned. Perhaps I was inadvertently speaking in Swahili, or something.

Jokes aside, though, I’ll admit the spelling question was an example I’d heard before and wanted to try it out myself. Sure enough, AI cannot spell. But it got me looking a bit deeper into how it works, and it turns out AI is really bad at knowing what a word is; all they ‘know’ is that certain strings match a token they hold in their training database. As strawberry consists of two words, it treats them as two tokens, and reports on the last part (berry) only. It doesn’t ‘know’ what the word or words are, only what token they match. Trying to get it to spell, or even count the letters in a word, is extremely difficult, and is made even harder if you start adding in support for other languages.

I can see uses for AI in the more accurately named ‘machine learning’, which uses the processing power and algorithms to spot patterns and oddities in vast troves of data. It’s remarkably good, for example, at spotting potential problems on scans and radiology images that even experienced clinicians would sometimes miss. For me, that’s where the future of it lies. In my more cynical moments, I wonder if it’s merely because the term ‘machine learning’ is not sexy enough for the marketing departments who are looking for some way to make a return on the billions in investment they’ve made.

But AI? Not for me. As a writer, one of the reasons I write is a love for words, and the joy to be had playing with words and phrases, using nuance and the intricacies of language to create a story that will (hopefully) entertain and stimulate. Software that’s little more than a confused auto-correct is never going to be able to manage that level of fluency with language. (The funniest description I read was that AI was nothing more than ‘Clippy on steroids’.)

One of my colleagues asked if I could use it to generate ideas, until they realised that you have to give it a prompt to start with, which is pretty much what an idea is – the hard work comes in fleshing out the detail. Ideas are not something I struggle with, anyway – I have lots of them, some good, and some less so. And if the AI has been trained on other writers’ works, then anything it comes up with would surely contain an element of plagiarism.

Anyway, I’ve managed to (mostly) get rid of it on my work laptop (sadly, Microsoft won’t take no for an answer and refuses to let me remove its offering, Copilot; I am of kindly nature and refer to it as coprolite). Thankfully my personal laptop runs Linux and is thus far free of infection. I’ve even found ways to stop the AI-generated answers appearing in search results. I may be fighting a losing battle, but for the moment, I’m blissfully free of it. Until next time :)