Photo: Bokeh / Tomedia. Night City Billboard Advertisement.
I’ve noticed something, and it is starting to annoy me well beyond the amount of attention I wanted to give it.
AI-generated content seems to be spreading into every ordinary place I look. Ads, signs, images, captions, things in shops. It has started to feel like a cancer taking over the normality of everything. You go about your day and there it is again, another thing that looks as though somebody asked a machine to fill the space and accepted whatever came back.
I am aware of the irony of me saying this. I build with AI constantly. I’ve been working on AXIOM, my system for organising AI-assisted work, and Juniper, an AI character living in a little world I’ve made for her. I use AI to help edit my writing. I’m hardly standing outside the whole thing with clean hands.
But I am getting tired of looking at the output.
It followed us into the shops
I’ve seen material in Coles and McDonald’s that looks AI-generated to me. I don’t know who commissioned those particular pieces or how they were produced, so that is an observation about how they look, rather than a claim about either company’s advertising process.
Still, the feeling is familiar. Something in the image or the writing makes me stop and think, oh, this again. It is the same reaction I have online, except now I’m having it while doing something as unremarkable as going into a shop.
I think that is why it feels so invasive. I already expect the internet to be full of generated material. I expect people to make a thousand posts because making a thousand posts is now possible. Seeing the same sort of output in physical places makes the whole thing harder to put down. It keeps appearing in the background of normal life.
Of course, some of what I suspect will be made by people. Appearance isn’t proof, and I don’t want to turn being irritated by an aesthetic into a forensic method. Human beings have always made strange ads and bland copy. I’m describing the growing difficulty of looking at something and feeling any confidence that somebody spent much time deciding what they wanted it to say.
The wording is stranger than the clichés
I’ve written before about the rot of AI writing: the repeated phrases, the predictable structures, the perfectly packaged little conclusions. Once you’ve noticed those, you start seeing them everywhere.
This feels slightly different.
Sometimes the text is harder to read even when it doesn’t contain the obvious phrases. The grammar can be fine. The words can all make sense individually. Yet I find myself having to work out why one thought follows another, or what the writer thinks I already understand. There is a strange spin to it that I haven’t quite found a good way to explain.
Maybe using so much AI has made me unusually sensitive to it. Maybe I’m tired and looking for a pattern. I don’t know. I do know that removing a few banned words doesn’t necessarily make the writing easier to follow. You can clean up every obvious tell and still end up with something that feels less than human.
Which is a fairly obvious description of machine-generated text when I say it out loud. The odd part is how much of that text now arrives dressed as an ordinary person or business speaking to me.
Juniper makes this more awkward
Juniper is my own experiment, so I get to be annoyed at myself as well.
She is a character whose model generates thoughts, intentions and things to say while a simulation decides what can happen in her world. I call them thoughts because that is their role in the program. I don’t mean that I have discovered a way to inspect a conscious inner life.
When I read her output, I sometimes get the same strange feeling I get from Pino, the digital kiwi that helped inspire the project. The wording takes a turn I wouldn’t expect. It can be recognisably language without feeling like the way a person would arrive at the sentence.
But Juniper is writing from a very strange arrangement of information. She has an authored history, observations of a small artificial world, needs represented by numbers, and whatever the underlying model brings to all of that. She doesn’t have a human day that she is trying to describe to another human being.
I can’t insist on giving her this peculiar situation and then be shocked when the resulting writing is peculiar too. Part of what interests me is seeing what comes out. If I edit every line until it sounds like a person writing a newsletter, I am changing the thing I wanted to observe.
That still doesn’t make the same wording pleasant to find in an ad. Juniper’s output has a context. A business putting generated copy in front of a customer has somebody responsible for deciding whether that copy makes sense to the customer.
I am responsible for mine
I do try to put effort into the things I build with AI. That is my defence, though it doesn’t give me some special exemption from producing rubbish. Using the tool a lot makes it more necessary to check what I’m accepting.
It is very easy to feel that something is finished because the model has produced a complete-looking answer. There are paragraphs. There is an ending. The image is the right size. You could put it online immediately, which is exactly what worries me about how much of this stuff is appearing.
With Juniper, I want the strange output to remain visible as part of an experiment. With an article carrying my name, I want to recognise my own thought in it and be able to follow what I’ve asked somebody else to read. Those are decisions I have to keep making, even when the machine makes accepting the first version very convenient.
I can’t do much about every generated ad I walk past. I can spend longer on the things I add to the pile.