AI “Model Collapse” Shows Why Real Human Expression Is So Powerful

AI needs the real words of real people

Clive Thompson

--

two robotic figures reading from devices
Photo by Brett Jordan on Unsplash

I’m seeing more and more AI-generated text online. Even here on Medium, in fact!

In the last few months, I’ve noticed comments on my posts that seem oddly …. off. They’re grammatically correct, but existentially rather alien. They usually a) blandly praise my post (“this is great”) and then b) blandly repeat the main gist of the post (“this article argues [X, Y, and Z]”). They have none of the specificity and wit you normally see in comments; they’re just dull summaries.

Now, I’m lazy, so I haven’t mounted any serious investigation into these comments (and their commenters). Maybe they’re legit. But they sure look like someone is using a large language model to autogenerate Medium comments.

Why would they be doing it? Eh, who knows. Maybe it’s some coder noodling around. Maybe it’s someone creating an army of bot accounts, and seeding them with long history of normal, mundane-seeming interactions, the better to use them for nefarious purposes later on (a technique botmakers have long employed on Twitter).

The upshot is, it adds to the pile of grey-goo language that’s metastasizing online. Search for the phrases “as an AI language model” or “regenerate response” and

--

--

Clive Thompson
Clive Thompson

Written by Clive Thompson

I write 2X a week on tech, science, culture — and how those collide. Writer at NYT mag/Wired; author, “Coders”. @clive@saturation.social clive@clivethompson.net

Responses (60)