You Ask, I Answer: Detection of AI Content?

In today's episode, Zhao Lee asks, how can people determine whether some content is created by AI or human? So this is a very popular question, understandably so. And there are some tools out there that can uh detect some of the language patterns of generic models. So, for example, if you were to type into you know AI content detector into Google, you'll get a whole bunch of different web services that all say, like we can detect uh AI content and plagiarism, all this, that, and other stuff. And these tools do that to a greater or lesser degree of success.

The reason they can do that is because there are predictable statistical distributions in the way that uh large language models like the GPT family that uh chat GPT uses that are detectable that uh that are uh you can find in the generic models. So the generic model means someone using just off-the-shelf chat GPT with no customizations, no plugins, no anything. It's just the the stock-based model, and the prompt they're putting it is so generic that the um the model is is essentially writing and doing most of the heavy lifting. Um it's funny, these tools, these detection tools, they're iffy in their quality. I took um a blog post that uh Katie and I had written for um the Trust Insights newsletter, and I I fed it in.

And there are sections that were clearly marked, this is section that's been generated by AI, and the section is not. And one of the tools missed everything. Completely missed it. Um one of the tools marked everything as AI, even the parts that we know were human written. And then two of the tools kinda halfway winged it, right?

Some one of the tools more or less got eh, okay. Um, but none of them got it right. None of them were got it perfectly right. None of them said, yep, this is the AI part, this is the non-AI part. And that's a problem, right?

So these tools do exist. Their quality to write now is hit or miss. And here's the part that's that's tricky. They are tuned for the open AI family of models. So GPC 3.5, GPT 4.

With the release of uh Facebook's Llama large language model set into open source and the proliferation of dozens, if not hundreds, of variations, these tools can't do that anymore. These tools are incapable of detecting language created by different models that have different model weights, different parameters, essentially all the different settings that these other tools use that will make their text have statistically significant distributions, but different distributions than open AIs. And so there really isn't a way to ironclad detect uh the use of AI. The other way that these tools will fall down depends on the process. So if you wrote an outline as your prompt and you had chat GPT write out the post, there's a decent chance that at least some of these tools would correctly identify it.

If you did it in reverse, you said open AI, you write the outline, because I can't think of what to write, and once I have my prompts as a writer, I'll do the writing. These tools will not detect that usage of AI. Even though AI was involved in the content creation process, the final step was done by human, and those statistical distributions will not exist nearly as much or as strongly as a machine generated version. So I would say the only surefire way to know whether content was created by machine or human is to ask. And if it's your content, hopefully you would know.

But if it's content created by your company, having policies in place as to the situations in which AI is permissible to use or not permissible to use is is critically important because these models will continue to evolve. Just the open source models alone are evolving so fast and getting such specific capabilities that the plagiarism detection, well, it's not plagiarism, the the AI content generation detection algorithms are going to get less and less useful. And here's why. If you take an open source model and you tune it towards a very specific task, like just writing blog posts or just writing emails, what's going to happen is those tools will have very different language distributions. And so something looking for the generic model is not going to see that.

And again, the fact that we see so much innovation happening on the open source side means you're going to have dozens, if not hundreds of models to try and keep up with. And uh you're gonna, as a if you were marketing, you know, AI content detection software, you're gonna have a real hard time doing that. So that's the answer. The answer is there are tools, they're unreliable, and they will continue to get to be unreliable. They'll actually get w less reliable over time as models proliferate.

So good question, though. Thanks for asking. Talk to you next time. If you like this video, go ahead and hit that subscribe button.