One of my favourite TV programmes of the past 20 years or so is called Person of Interest. In the US, this thought-provoking action-drama won award after award after award.
In the UK it was shown on Channel Five, which meant it had a very small, but loyal audience.
The premise was that, post-9/11, the US government had asked various tech companies to come up with a way to predict terrorist attacks. The winner was a ‘black box’; data went in, an answer came out. You couldn’t influence or direct the AI, what you got was what you got.
A lot of that data came from everyday sources such as traffic cameras or CCTV, encrypted NSA feeds, and public social media posts or newspapers. In many ways it was every conspiracy theorist’s wet dream, but it was also one heck of a story. Last time I saw it, it was streaming on Amazon Prime – do hunt it down.
The point is, a lot of the issues it raised around information management, surveillance and what we share online about ourselves are over 15 years old now. Most people are only just becoming aware of them. A few years back there calls for a six month moratorium on AI development. Far, far too late.
What is “AI?”
There’s more than one type of “AI”, and I can heartily recommend The Joy of AI, presented by Professor Jim Al-Khalili, as a primer while it’s still available. The last 10 minutes is comedy gold, but I won’t spoil it for you.
All of the AIs we’re exposed to these days rely on data. We think of data as Robert Burns might have, when he wrote: “But facts are chiels that winna ding, An’ downa be disputed”1 but everything is up for interpretation.
In the UK, we drive on the left. Right? But if my only knowledge of cars was the (very unrealistic) adverts on TV, I’d think we drive on the right. Watch them, and see.
Conscious bias
Equally, we can’t feed our AIs the sum total of human knowledge, because knowledge evolves. We no longer believe that the sun revolves around the earth, or that the earth is flat. Most of us, anyway.
So, we have to decide what data to feed in, and this is where our own biases come into play.
Last week I took part in a panel discussion, where the question was “Will AI Save the World?” The answer, clearly, is no; AIs are created by humans, and humans are fallible. But AI can save some people in some ways. It is not a panacea.
The human element
Several years ago, Find My Past digitised the ecclesiastical records in what used to be called the India Office Collection, now held by the British Library. I’ve spent over 21 years researching my family history, and the thought occurred to me: if we fed all of the transcribed data into an AI, could it build everyone’s family trees for them?
No.
It could do it with modern, typed, verified data but it would have a much harder time with unverified, handwritten data. I can give you a good example from my personal research, involving the siblings of my great-great-great grandmother, Charlotte Ann Taylor (née Thorpe). Her father was John Luke Thorpe, her mother was Jane.
Because of the format of the baptismal records of the time, the parents of her siblings are listed in two columns for father and mother, John Luke and Jane Thorpe; but as you read that to yourself, where did you put the emphasis? Half of her siblings have their parents names transcribed as “Father: John Luke” while the rest have “Father: John Luke Thorpe”.
The format of the baptismal record creates the confusion; human error creates two data groups; AI gets it wrong. I was confused as well, until I made a wonderful leap of intuition and realised what had gone wrong.
Internationalisation v homogenisation
And it’s these subtleties and nuances that will cause the most grief, as we move away from regional variances into a one-size-fits-all world.
English is sometimes referred to as a lingua franca (the irony!) but it’s probably more true to say that American English is now. This is not a new thing; for many years in the 1980s and 1990s I would install software for which the default language was English (United States) – proper English had to be downloaded separately. Some software packages now don’t even allow that option. I had to download a British English dictionary for the copy of Word I’m using to write this.
In the US, people such as Noah Webster – of the dictionary – were spelling reform advocates, preferring the written word to follow the spoken tongue. That meant perfectly good, English words such as ‘oestrogen’ were reduced to ‘estrogen’.
Non-English words find their acutes, graves and umlauts being removed because there is no easy way to type them on an ‘english’ keyboard. It’s one area in which the iOS keyboard on Apple devices excels. But even on the good old BBC website, there is no degree symbol for temperatures; it will say, for example, “20C”, which is 20 Coulombs not 20 degrees Celsius.
Small things, perhaps, but my German friend who teaches English sometimes texts me with a question, where a student has used an Americanism. It all affects the data. Even the Germans had their Rechtschreibreform in the late 1990s.
And don’t get me started on the word ‘shuttered’ to mean ‘closed’…
So where are we heading?
“Claude, give me a well-known phrase involving a creek and a paddle.”
The first thing to realise is that AI has limitations. One of my former employers uses Copilot to draft releases, which is fine as it’s a Large Language Model (LLM) which is suitable for anything you might want to throw at it. But it’s terrible at drafting comments from the Mayor; for that you need a Small Language Model, which is fed everything they’ve ever said or written to more closely match what they would say/write.
Sometimes, you need a specialist…
Then there’s that data problem. AIs need data. Anything you posted on Twitter now belongs to Grok (which, to be fair, explains Grok). Anything you’ve posted online can be scanned and rehashed in some way.
LinkedIn has been awash with people using an image generator; you throw in lots of information about yourself and you get some twee image back. Why are we so keen to tell the AIs about ourselves? And is that data accurate or are we finessing things in our favour?
How transparent are we on our use of AI? If we use it to apply for a job, how do we know it isn’t steering us wrong? Is it a real reporter or industry expert we’re hearing from, if we’ve never seen them in public?
And yet… and yet, I believe that tightly-focused heuristic algorithms can find new treatments (once fed with everyone’s DNA, sadly), or new propulsion methods, or ways to eliminate plastic and plastic waste, and probably other things that will improve our health and general quality of life. I have to be optimistic about the future, for fear that we fall into one of the many dystopian futures we see on film.
Fingers crossed, hey?
- “A Dream”, Robert Burns, 1786[↩]