I Can't Read Anything Anymore Without Wondering If a Machine Wrote It

You’re three paragraphs into an article when it hits you. Wait — did a machine write this? And once that thought lands, you’re done. You’re not reading anymore. You’re auditing. An essay titled “I’m becoming AI-blind” hit the front page of Hacker News this week, and the comment section filled up fast with people saying the same thing: I do this now, and I can’t stop.

Remember banner blindness

Web usability researchers in the early 2000s documented something called banner blindness. Users stopped seeing ads. Not ignored — genuinely didn’t perceive. Eye-tracking studies showed participants’ gaze skipping over the rectangular block at the top of the page entirely. Jakob Nielsen’s team found that when researchers deliberately placed the information users were hunting for inside an ad-shaped box, people failed to find it. It was right there. They looked past it.

Here’s the part that matters: the filter your brain builds is not precise. It doesn’t screen out ads. It screens out things shaped like ads. Rectangle at the top, right-hand column, anything with that faint whiff of sponsorship. The filter is lazy, and lazy is exactly why it’s fast.

The filter we’re all building for AI text works the same way.

Delve, tapestry, and the rule of three

AI prose has a smell. Sentences of suspiciously even length. Transitions that glide a little too smoothly. Everything arrives in groups of three. And then the closing paragraph reaches for some grand significance the piece never earned.

English-speaking internet has already turned the tells into a punchline. Delve. Tapestry. It’s not just X, it’s Y. The em dash discourse alone consumed X for most of 2025 — real writers who’d used em dashes for twenty years suddenly found themselves accused of outsourcing their thinking to a chatbot.

But look at what’s actually on that list. Balanced structure. Clean transitions. Tidy conclusions. Those were the marks of good writing. That’s precisely why the models learned them. And now they’re evidence for the prosecution.

Which produces an inversion nobody planned. The writer who outlined carefully and revised three times gets flagged. The one who left a typo in the second paragraph and wandered off-topic in the fourth reads as authentic. Sloppiness has become a credential.

The loss isn’t the content — it’s the reading

The AI slop conversation usually runs on a pollution frame. Garbage floods the internet. Google results rot. Blogs turn into affiliate farms. Stack Overflow answers can’t be trusted. Amazon fills up with generated books about mushroom foraging that might get somebody killed.

The Hacker News essay pointed somewhere else. The thing getting polluted isn’t the internet. It’s the inside of the reader’s head.

Suspicion, once switched on, doesn’t switch off cleanly. A long text from a friend now reads a little strange — did they run this through something? A well-structured proposal from a colleague raises an eyebrow instead of respect. You catch yourself scanning the sentence rhythm of a blog post from 2019. Written years before the models could hold a paragraph together. You check anyway.

That’s the actual cost. What we’re losing isn’t a handful of good essays. It’s the ability to fall into a text without armor on. Reading has always started with a small act of faith: a person spent time here trying to reach me. You follow the sentences on that assumption. Pull the assumption out and reading stops being comprehension. It becomes forensics.

Images already went through this

Text is late to this. Images got there first, and the pattern is instructive.

You count the fingers. You check whether the signage in the background dissolved into glyph soup. You look at the lighting and ask if it’s too perfect. Then that habit — this is the whole problem — runs on real photographs too.

Wildlife photographers have been accused of faking shots they spent three days in a hide to get. Photojournalists have posted original RAW files with intact EXIF data, and the arguments continue anyway, because RAW files can be manufactured too and everyone knows it. The burden of proof flipped. It used to sit with the person claiming a photo was fake. Now it sits with the person claiming it’s real.

Text has it worse. A photo at least carries metadata, however forgeable. A paragraph carries nothing. “I wrote this myself” is just another sentence — and a model can produce that sentence on request.

The fixes don’t fix it

The proposed solutions are, charitably, weak.

AI detectors are worse than useless because they’re confidently wrong. Stanford researchers found detectors flagged over half of TOEFL essays by non-native English speakers as machine-generated, while near-perfectly clearing essays by US-born eighth graders. Feed one the US Constitution and it will often tell you a machine wrote it. Turnitin’s detector has been quietly de-emphasized by multiple universities after enough students got accused of cheating on their own work. Provenance standards like C2PA are real and progressing, but they were designed for camera sensors and editing pipelines. There is no equivalent hook to attach to a paragraph.

What’s left is unglamorous: trusting people instead of texts. You stop trying to authenticate the writing and start looking at the writer. What have they published before? What have they gotten wrong, and how did they handle it when someone caught it? Is their name attached, permanently, in a way that costs them something if they burn it?

You can see this repricing in the market. Substack keeps growing. Personal blogs are being read again. RSS, of all things, came back. The anonymous content firehose is losing value and the fixed byline is gaining it, because a byline with history behind it is one of the few signals that’s expensive to fake.

The catch is that it doesn’t scale. You can maintain a roster of fifteen or twenty writers you trust. You cannot read the internet that way. So the total volume of text any given person actually reads goes down. It has to.

Writing got free. Reading didn’t.

That’s the whole asymmetry in one line. Generation cost collapsed to roughly zero. Consumption cost didn’t move — it still takes time, attention, and above all trust. Supply went infinite against fixed demand-side cost, and when that happens, demand defends itself by closing.

So: how did you read this one? Did you just follow the sentences? Or did the question surface somewhere around the third section — is a person actually writing this? If it crossed your mind even once, that’s the finding. Something already shifted, and it wasn’t in the text.