Passing AI detection - Are we chasing our own tails?

Just a few passing thoughts on the subject.
It could go one of two ways.
A. Google (or whoever) will rapidly create more sophisticated detection software based on the millions of GPT outputs now being published (in one form or another) meaning anything we “think” is AI detectable now, will become “detected” pretty fast - rendering the pursuit of undetectable content pretty pointless.
B. Google get on board with AI content, especially with Bard in their pipeline, and find an alternative way to rank content with AI Gen in mind - rendering the pursuit of undetectable content pretty pointless.

I would love to get some more experienced hot takes and predictions on the situ

Welcome and a perfect intro in the Random Shower Thoughts category!

The objective for content creators is not to satisfy one of those AI detectors out there.

The objective is to satisfy clients and users.

When you read thru my review
13 AI Content Detection Tools tested and AI Watermarks
you also start to think something like

“What’s the point of these detectors, anyway”.

There’s a lot of agencies trying to use them for quality control tough, so that’s one reason many content writers work with them.

Google couldn’t care less about these 3rd party tools, they have the PalM model, which blows GPT3/3.5 out of the water anyways.

You’re so right. Eyes on the prize! Also really appreciated the Learnings from 100’s of Prompts video. Super Helpful - Thank you!

Google have long been absolutely accepting of machine generated or algorithmically generated content that provides value for users. Google itself is one of the most valued (by humans) websites on the entire web, and has that value completely on the way it crawls, indexes, processes, and formats data in a way that provides a valuable service for consumers.

Sites that do price comparisons via algorithms date back to the late 90s, just as Google does, and Google have been happy to index them, help people to find them, and of course, eventually replicate them and cut them out of the middle.

Sites that track sports data, events, and all sorts of other scrape-able or accessible data that a machine can take, process, and process into some new added-value format are everywhere. If people love them, Google loves them and helps those people, that Google want as customers, to find them.

Yet at the same time, approached from the other side of the value equation, people have used automation to create low-value but cheap content since before Google existed. There was software to generate thin-content doorway pages, some of which were so badly written, of such poor utility to humans (designed only to rank) that they had to be cloaked, or have redirects on them to stop people seeing how bad the pages were.

ChatGPT may be a completely new concept to many, but to those who were using the most advanced scraping and spinning software of ten years ago, the evolution is a lot less than you’d probably love to imagine. It is kind of like how computerization and super-advanced electronics may really produce things not that much in advance of really sophisticated purely mechanical solutions, if the mechanisms are complex enough. Remember, for all the sophistication of a human being, you are ultimately a collection of microscopic cells acting like tiny biochemical machines where each cell is incredibly simple, only the way they are networked together making all the ‘magic’ happen.

Detecting AI written content is always going to be not just possible, but relatively simple. My old friend Jennifer Slegg shared her tip at PubCon and since repeated it for all her Twitter followers, and it is almost laughably simple. You search for “regenerate response” and your keyword(s), and you’ll quickly find all the people so incredibly lazy they didn’t even proofread or edit the output. But then you go one step further. Look for a couple of sentences in the returned articles and search for other articles using those sentences… Because AIs use pattern-matching from the data they were trained on, and repeat those patterns in new outputs, you’ll find they often use and reuse the same sentences when given similar prompts - but this time you find all the users who did at least remove the “regenerate response”, thanks to getting the pattern from those who didn’t.

Try it for yourself. No need for fancy AI solutions to detect AI, just use the pattern-using of the AI itself to catch it. It really is that simple. :wink:

2 Likes