Slop Detection

If you're reading this page, you probably already know what "slop" means in the context of artificial intelligence. But in case you don't, I'll take a shot at a definition here:

Slop is a derogatory term for any kind of content generated by an AI model.

Text can be slop. Images can be slop. Videos can be slop. Slop is becoming more pervasive, and people are building better tools to distinguish human-created content from slop. Today on the Hard Fork podcast they interviewed Max Spero, founder and CEO of a company called Pangram Labs that makes slop detection software. I hadn't heard of Pangram, but found the conversation insightful.

Max touched on the concept of perplexity, a metric historically used in natural language processing:

So part of the reason that we started Pangram was because a lot of these, all the existing AI detection systems were pretty flawed in different ways. But I think the main thing is that most of them were using this metric called perplexity, which was considered state of the art at the time. So AI text on average is less confusing to a language model. It's lower perplexity. And human written text has things that surprise a language model.

This approach has a lot of flaws. For example, any document that the AI model has memorized would also be low perplexity. For example, the Declaration of Independence or English language learners as well, who just write in more simple English. So we do something completely different. Instead, we are training our own classifier network. So for example, we might have an essay on Moby Dick written by a seventh grader. And then we'll ask like an LLM to also write an essay on Moby Dick in the style of a seventh grader. And so then our model is able to learn the differences between A and B and figure out like what AI text actually looks like. And so because it's a neural network, and not a metric, we're able to improve it with more data and more compute and make it a lot better.

I was curious to try out Pangram, so tonight I signed up for an account and did an audit of my own personal website to see how good it is at detecting slop. Since I created all of the content on my site, I know with certainty which posts are real and which ones I regrettably fired from a slop cannon in a moment of weakness.

Pangram's findings are accurate: it correctly identified the small handful of slop posts I published in the last few months, and even found the blobs of intermingled slop where I let an AI agent co-author parts of a post alongside me.

Of the 100,000-ish words on my site, 2.8% are slop.

Here are the results:

Building your own slop detector

I whipped up a little open-source repo that uses Pangram's API to detect slop in my content. It cost me $6 to audit my whole site. I could have brought that number down if I hadn't scanned every single page in its entirety, but I just decided to go all in. Pangram also has a Chrome extension that can analyze webpages and social media posts on the fly.

See github.com/zeke/slop-detector for the code and detailed results.

Keeping me honest

I'm really impressed with Pangram's accuracy, and how easy it is to use. In a time where I feel myself increasingly dependent on AI for work, I'm happy to have this new tool to help me keep myself honest as I create new content. Gotta keep it human. :)