As schools grapple with how to spot AI in students' work, the FTC cracks down on an AI content detector that promised 98% accuracy but was only right 53% of the time.
Text AI detection is such a crapshoot. When you consider what it’s doing, it’s just reproducing plain language. It’s not going like, “He walked through the rubble [beep], his boots crunching on the remnants of what was [boop].” The closest you tend to get is tells like a no-proofreading “As an AI language model” slipped into an academic paper. Or AI-isms associated with certain models and datasets, like a high tendency to talk about “shivers down the spine” and such things. But even AI-isms are not themselves evidence of AI use. They become model tropes because of how common they are in human writing.
Image AI detection can be more reliable, depending on the tool, because (if I understand right) diffusion models produce “noise” in a very particular way, that an automated tool could be trained to detect. And it’s in a way that a human is highly unlikely to accidentally produce.
But that’s the crux of it, as far as I can tell. For it to be reliably detectable, it needs to have distinctly non-human characteristics embedded in the output. If it doesn’t, good luck.
Yeah for sure, although you do see some common patterns in AI generated text because it tends to reuse same structure a lot. Like you often notice stuff like “A is not only B, but it’s actually C”, etc. Another tell comes from the small context size, so you end up with a bunch of independent statements that don’t necessarily connect to each other. With human writing you often have an idea introduced and then developed gradually towards some conclusion. But as a whole, I agree that these are just common tropes from regular human writing, and it’s pretty much impossible to definitively say something was written by a human or not.
Amusingly, image detection is kind of turning into arms race now with people figuring out techniques like adding noise perturbation that throws off the detectors.
Text AI detection is such a crapshoot. When you consider what it’s doing, it’s just reproducing plain language. It’s not going like, “He walked through the rubble [beep], his boots crunching on the remnants of what was [boop].” The closest you tend to get is tells like a no-proofreading “As an AI language model” slipped into an academic paper. Or AI-isms associated with certain models and datasets, like a high tendency to talk about “shivers down the spine” and such things. But even AI-isms are not themselves evidence of AI use. They become model tropes because of how common they are in human writing.
Image AI detection can be more reliable, depending on the tool, because (if I understand right) diffusion models produce “noise” in a very particular way, that an automated tool could be trained to detect. And it’s in a way that a human is highly unlikely to accidentally produce.
But that’s the crux of it, as far as I can tell. For it to be reliably detectable, it needs to have distinctly non-human characteristics embedded in the output. If it doesn’t, good luck.
Yeah for sure, although you do see some common patterns in AI generated text because it tends to reuse same structure a lot. Like you often notice stuff like “A is not only B, but it’s actually C”, etc. Another tell comes from the small context size, so you end up with a bunch of independent statements that don’t necessarily connect to each other. With human writing you often have an idea introduced and then developed gradually towards some conclusion. But as a whole, I agree that these are just common tropes from regular human writing, and it’s pretty much impossible to definitively say something was written by a human or not.
Amusingly, image detection is kind of turning into arms race now with people figuring out techniques like adding noise perturbation that throws off the detectors.