On our work WhatsApp group, colleagues send each other birthday greetings emojis and memes. It’s fun. I’m sure it’s the same all around the world — expressions of staff conviviality transmitted over WhatsApp are a hallmark of our age; they are touching at best and harmless at worst. Just for fun, I have introduced a new twist by sending ChatGPT-generated poems for people’s birthdays — and the results range from embarrassingly bad to extraordinarily good.

To herald the birthday of Marianne Thamm, for instance, I suggested a poem that contained the phrase “to the least MT person I know”, and it came up with:

“The least MT person I know,” they say,
For in her, life’s colours vividly play.
In news and jest, her spirit dances,
Happy Birthday to the voice so bold,
In your words, the world’s tales are told.

Not too bad, I thought. It informed our chief photographer Felix Dlangamandla’s ditty:

Happy Birthday, Felix Dlangamandla,
With your camera, you’ve shown us so much.
Out in the field or in the bustling city,
Your photos speak with a gentle touch.

You know, not Shakespeare but not terrible. I tried to mix it up a bit and asked it to write a haiku for Rebecca Davis’ birthday, and it came up with: 

In the Daily’s light,
Rebecca Davis shines so bright,
Birthday’s joyous flight.

Davis, never one to let slip a moment of biting humour, replied, “Honestly I feel like ChatGPT didn’t really put its back into that.” Hard to argue with that assessment. To upbraid ChatGPT, I made it write a haiku about badly written haikus, but it didn’t seem to be in any way remorseful. It just spat out: 

Words fall short, unsaid,
In the haiku’s quiet depth,
Meaning left unread.

I think it might have been more accurate if ChatGPT replied, “ChatGPT falls short, words unsaid … meaning left unread”.

Thinking about this, I have come to the conclusion that AI is both very, very impressive and sadly lacking. We, or at least those who care about this stuff, have all been on a bit of a roller coaster ride when it comes to AI.

Your first impression is often just to marvel; what AI can do — or to be more specific, what large language model generative AI can do is simply mind-bending: “artworks” of extraordinary oddities, listenable songs in seconds, summaries of very long documents, analyses of stories… How will humanity deal with all of this, we ask? Is this the end of this job, or that job, or my job!? And, er, how do you make money out of this thing? And so on.

It is well known that there are (at least) five stages of grief — denial, anger, bargaining, depression and acceptance. There are presumably (at least) five stages of AI appreciation: amazement, fear, greed, reassessment, and … where it ends we don’t know. But it does seem to me that people closest to the unfolding narrative are at least at Stage 4.

The reassessment narrative runs something like this: AI is amazing, but if you think more carefully about what it actually is, it’s quite different from the all-encompassing, life-changing promises, both good and bad, that we have heard up till now.

One example is from an interesting discussion between New York Times journalist and podcaster Ezra Klein and Nilay Patel, the editor-in-chief of the tech news site The Verge and host of the Decoder podcast.

Patel starts from a different premise about AI: not that it is fabulous or brilliant, but that it’s very average: he calls it C+ content. Klein points out that the places where ChatGPT is proving very sticky — where it’s being used every day — are not in high science or nuclear physics, but in schools and undergraduate universities. These are places, he says, where it doesn’t have to be very good. 

Average — and getting worse

Why does that make sense? Well, it makes sense because what ChatGPT is doing is taking as much digitised knowledge that can be fed to it, averaging it out, massaging it a bit, and feeding it back to you. The fact that it does is just miraculous; the fact that it does it so fast (think seconds) is incredible.

But in the long term, it’s only as good as the content it analyses. And that content is often going to be contested, possibly from mediocre sources, and is always going to be historic. The result is that it’s feeding back content which is better than average, but worse than good — and nowhere near excellent.

Both Patel and Klein are, ironically, clearly thrilled by the way this is unfolding. For example, one of the ways people have been using AI is to write novels, which they then sell on Amazon. Because the novels cost practically nothing to write, even if only a tiny handful of people buy the novel, it still counts as profit.

To make real money, you have to “write” hundreds of novels, which is what some AI entrepreneurs are doing. Amazon, in response, is doing one of the only things it can do to control this influx, which is to prevent “writers” from uploading more than three books a day. Patel says this is an absurd way to deal with the problem.

So why are Patel and Klein so happy? The duo quote US technology journalist Casey Newton as saying that many parts of the internet have been building traffic thinking that they were building an audience.

“And the traffic, at least in that era, was easy, but an audience is really hard,” Klein says. “Over the last year, Google and the big social platforms have been flooded with AI spam, flooded with fake news sites filled with stolen or made-up stories. There are TikToks of AI voices reading random text off of Reddit, and nonsensical YouTube videos for kids. It’s no novel observation to say the internet has felt like it is in a state of decay for a while.”

And yet, the builders of the biggest AI bots rely on the quality of the internet to improve the quality of their AI.

The big question is whether AI stays at this level or gets better. But the realisation that AI’s content is mediocre is itself revealing because it demonstrates the oldest law of computer science: garbage in, garbage out.

The massive increase in mediocre garbage that’s about to hit the internet makes it more likely that new versions of AI will not be better than the current versions; they will be worse. The result will be that genuinely good ideas, real books or news written by actual humans, honestly delightful content and challenging prognostications might stand out more vividly and will be more valuable.

Or to put it another way, it is just possible that AI might break the internet. And that could be a good thing. DM