Everyone writes now and everyone sounds the same

No Saffer has stood in a queue at Home Affairs for three hours, turned to the person behind him and said, "kindly revert soonest". Yet there it is in your inbox on Monday morning. Along with "trust this finds you well", which it seldom does.

We have somehow invented a second version of English that exists only in offices. I have a suspicion it was designed by a committee terrified somebody might think and sound human simultaneously.

One of my favourites is the recorded message. “We are currently experiencing higher than usual call volumes.” This has been true since 2007. At what point does “higher than usual” become usual? Nineteen years should probably do it. The phones are not unusually busy. The phones are always busy. Somewhere, a poor oke has spent nineteen years fiercely protecting the reputation of the farcical word “usual”.

Body corporates have their own language. "It was resolved that the matter be referred for further consideration."

Nothing was resolved. Dennis is supposed to phone the plumber, eventually, possibly soon, maybe asap (as slowly as possible). Unfortunately, the leak has settled in and now pays levies. And when Dennis finally phones, the plumber says he is around the corner. No, he is not around the corner. He has not yet left Kraaifontein.

Corporate South Africa was writing like a machine long before the machine existed. They made it themselves, slowly and by hand. Writing like this feels like comfort food. Nobody gets fired for "please advise accordingly". You cannot really disagree with it. You also cannot hold it responsible for much either. It sits in the inbox wearing a dull tie, looking busy and is easy to ignore.

We pretend this is a skills problem. It is not. Anybody who can order a fillet steak medium rare can write a clear, human sentence. Most people write this way out of unexamined habit. They saw somebody senior do it, somebody who appeared to be doing rather well and they copied it. For some it is the psychology of undervaluing one’s worth. They think their plain and clear sentences will show them up. Nope, what shows them up is the answer that is the same as everyone else’s. At least the sheep have the excuse of being sheep.

This is what interests me. Not the bad writing. Unexamined sidestepping. AI did not invent the habit. It made it easy.

Clear language tells people what happened, what must happen next and who is responsible for doing it. This creates an obvious difficulty. Somebody may later discover that you were wrong. Vague language solves that problem.

It also creates six new ones.

I work in insurance, where innocent little words have a habit of becoming expensive after a loss. "Temporary" is harmless until a vehicle has been kept at a different address for eight months. "Regularly" seems perfectly reasonable until somebody must decide whether twice a month counts. "Approximately" is an agreeable word right up to the moment a warehouse turns out to contain several million rand more stock than approximately suggested. People often soften language as they are trying to reduce risk. They are not reducing it. They are only moving it somewhere less visible, which, in my experience, is where it grows moss.

AI found heaps of this, read the lot and started churning it out by the bucketload. Which is hardly the machine’s fault. Feed something forty years of “my previous email refers” and one should not be shocked when it begins to sound like Procurement. Drivel in. Drivel out.

Before AI, people who could not write generally had the decency not to. Now they publish before breakfast, at volume, with glorious detachment. The machine did not make them lazy. It made the average sound like expertise. It gave them unsubstantiated credibility. The ultimate booby prize. Waffle with no syrup.

If you strip the mystique, a language model does one thing. Give it some words and it predicts what is likely to come next. "Once upon a" is followed by "time". Humans have written that sequence vastly more often than "once upon a time we paved paradise and put up a parking lot". The machine does not pause to consider whether this option might be a tad bit more interesting. Unlike Star Trek, it blindly goes where language has gone before.

And then, once upon a time Claude met Claudette.

She had read everything he had read, so they got on famously. Then somebody on your team put their name on it and sent it to you. It was well structured and it did not contain a single thing they thought. Lovely. But utterly useless.

I have a word for them. Claudettes.

Not the machine. The person who copies AI text, edits two words, proudly puts their name on it and presses send. I already have Claude. What I am short of is you.

The first time I saw this was on a letter to an insurer about a declined claim. Four paragraphs. Beautifully arranged. Almost no facts and the two it did contain were points the insurer already knew. The rest of it asked them to feel sorry for the client.

The problem was not that they were lazy. They did not see what I saw. That is fixable.

Do not fix it and you end up running a mediocre business where you ask a question, a machine answers it and Claudette delivers it to your desk. Or worse, to your clients desk.

If nobody on your team ever hands you something you disagree with, ask yourself why. Thinking produces disagreements. Consistent agreement usually means somebody worked out what you wanted to hear. That used to take effort. Now they paste the problem into a chatbot and you have it before the kettle boils. The other possibility is probably worse. They stopped trying as you shot down the last three.

The trouble starts when nobody checks. The machine has an uncanny ability to come up with sentences nobody thinks to question. Neat, tidy and plastic. And very sure of itself.

So is the Labrador arriving with the wrong ball.

There is another problem with polished writing. It feels true. Psychologists call it processing fluency. Fancy name, but nothing earth shattering. If something is easy to read, we assume it is more likely to be true. Rhyming sentences also feel more right than ones that do not rhyme. Apparently we will believe almost anything if it rhymes. Advertising was doing this long before psychology gave it a name.

AI has the ability to make almost anything sound sensible. Everything tastes like a Woolies ready meal. Perfectly balanced, nothing you would savour. You finish reading with the pleasant sensation that something happened.

Sometimes it has. Mostly you nod all the way through and remember nothing.

AI is not manufacturing knowledge. It is manufacturing the feeling that you understand.

Ask it a simple question and you invariably get a post grad memo back.

But this is not the part I distrust.

It is the agreeableness.

Some time ago I ended a paragraph with the phrase "digital zombies" and asked an AI tool what it thought. It said the phrase was tired and overused. I had never seen it used. So, mildly insulted that it was spurning a phrase I had spent 30 minutes constructing, I challenged it. Surprise, surprise, the machine immediately flipped. "Digital zombies" was now the best work I had done. Amazing, incredible and definitely worth keeping.

A marvellous recovery.

Truth be told, it had not really considered my question. Instead, it gave me the answer it thought I wanted. That is not judgement, that is flattery. I may well have been right. I thought I was. I always do. But if all I wanted was ego stroking agreement, I could have saved myself the AI subscription. Anyway, flattery is for the birds. When I am deciding whether an idea is any good, I need friction, a reason to change my mind. Most worthwhile ideas begin badly and you only find out which ones are worth keeping by being argued with until either the idea improves, or you admit it was nonsense. Both outcomes are handy. What is not so lekka is a second opinion that caves the moment you question it.

There is a longer term problem with taking the shortcut every time and it is not obvious at first. Do it often enough and you can no longer describe what you want. Eventually you no longer understand it either. Comfortably numb. Great song. But a terrible way to run a business.

Do not get me wrong. I use AI every day, we have a love-hate relationship. Giving it up would be downright inconvenient. It helps me research, question an argument. Occasionally, it even catches me talking total poppycock. Used properly it is a lekka tool. But it is a tool. It is not the bloke who has to phone the client back with bad news about his bakkie.

None of this is really a technology question anyway. People still want the same things they wanted in 1994. Electricity, water, cheap diesel and beating the All Blacks at Eden Park. Everything else is negotiable. They want to be told the truth, preferably by a human who really knows, in plain English. The tools have changed beyond recognition. The truth has not.

We keep asking whether a piece of writing is AI. I think it is much better to ask who wrote it. Not which machine. Which person.

AI can rattle off codswallop faster than politicians. It can also add a dash, break a sentence and throw in the odd anecdote about a fictional Oom or Aunty. That is not personality. Personality is what is left after a person has chosen what is important and what they stand for.

We have spent decades learning to write incognito. Mindless and proud of it. Then we made something Pontius Pilate would be proud of. We called it AI. And it turned out to be remarkably good at washing its hands.

Somebody, somewhere in the Republic of South Africa will read this headline, decide it sounds boxing clever, forward it to the team and ask them to kindly revert soonest.

Disclaimer: This article is provided for general information only and does not constitute insurance, financial or legal advice. Where relevant, it draws on publicly reported information. The scenario is fictional and is used to illustrate real insurance principles and common industry practices. It does not refer to any specific insurer, policyholder or claim.