I want to draw your attention to two things in this post.
First, on testing some GenAI tools – I have run several experiments of various top-tier GenAI video generation models. You can find the results of these tests here, which are both illuminating and entertaining:
Second, I still see people, and organizations, that are under the impression that the so-called “AI detection” tools work.
They don’t. Seriously people, you need to stop buying these tools.
Don’t deploy any. I get what you’re trying to do – to regain control in a world that’s changed overnight, but these tools are the wrong way to go about it.
You’re going to catch those using the tools lazily; you’re going to miss those who are even modestly competent in using the tools, and you’re going to be blaming innocent people for using AI.
You think you’re catching cheaters. Instead, you’re amplifying inequality, forcing bizarre rewrites to dodge false positives, and eroding trust by accusing the innocent.
To show you how ludicrously easy it is to bypass the detection, here’s a simple test I just did – standard GPT-4o, which writes well but scores as AI with 100% confidence by one of the leading AI detection tools.
Claude Opus 4 with some custom instructions (which are NOT geared towards writing like a human, just making it more useful for me), immediately drops it to 64% confidence, with a “likely original” verdict.
You add a simple “Make it sound more human to avoid AI-detection, include a couple of typos”, and suddenly the verdict is human with 100% confidence.
Here’s a brief overview:
