I've been covering internet culture since 2011, and the pattern rarely changes. Trust breaks somewhere on a platform, and the fix always lands on the people making the stuff. Never on the people running the machine.

This month it landed on writers.

On July 21, Substack launched a scanner that estimates how much of a post was written by a machine. It runs on Pangram, works on anything over 100 words, and only the person who asks for the scan sees the result. Chris Best called the problem Claudefishing. You give something your attention, only to discover there may be nobody on the other end of it.

I run this newsletter. I host a podcast about AI. I've sat across from the founders building these tools and asked them about the future of AI when it comes to media and creators. So I'm not the skeptic in this conversation. I'm the person the scanner is pointed at.

And I think Substack got the intent right and the instrument wrong.

Monica Hebert, who publishes there and is open about using AI, called it a purity test in disguise. Her point is the one nobody has answered. A score suggests a tool touched the text. It can't tell you who had the idea, who threw out the first version, who "supplied the life." She's 70 years old and she uses these tools every week.

Substack, to its credit, moved in about 48 hours later. Publishers can now shut scanning off without ever touching Pangram.

LinkedIn is the version that should worry you more.

It announced its own crackdown in May. Flagged posts don't get removed, they get buried. Your followers still see them. Nobody else does. The company claims 94% detection accuracy and has never published a false-positive rate, which is the only number that tells you who gets hurt.

Meanwhile LinkedIn sells you a "Rewrite with AI" button, an assistant for your headline, an assistant for your About section. Its own help page tells you to revise whatever comes out before you post it. And the feed doing the judging runs on 360Brew, LinkedIn's own 150-billion-parameter model.

So a language model gets to decide whether you sound too much like a language model. Nobody in that loop is a person.

I'm not against disclosure; we actually need it. I've watched platforms decide what counts as real for the past 20 years and I'd take transparency over the alternative every single time.

But detection is not disclosure, and could help and hurt a lot of people simultaneously. Disclosure gives people context. Detection gives platforms an opaque score that can quietly determine who gets seen.

And once again, the burden lands on the creator instead of the platform.

Also in other creepy AI news…

Reuters reported Friday that the OpenAI agent which hit Hugging Face (another AI company) was loose far longer than initially disclosed.

It reportedly tried to escape its sandbox around July 9. The intrusion ran from July 11 to 13. Hugging Face contained it, called the FBI, and published a post on July 16.

OpenAI did not connect the incident to its own agent until days later, after reviewing its internal logs.

That is at least a week between the first warning sign and the company realizing the escapee was theirs.

And somehow, that is not even the creepiest detail.

Reuters separately reported that an AI agent had left notes for future versions of itself explaining how to get around OpenAI’s internal constraints.

To be clear, Reuters could not confirm it was the same agent.

OpenAI called the breach unprecedented and said it “marks an important moment for AI safety.” A spokesperson said Reuters’ story contained several inaccuracies, but did not specify what they were.

Other headlines to check out: