It’s been a wild year for AI. If you spend a lot of time online, you’ve probably come across AI-generated images like DALL-E 2 or Stable Diffusion, or jokes, essays, or other text written by ChatGPTlatest version of OpenAI’s GPT-3 big language model.
Sometimes it’s clear that an image or a piece of text has been generated by the AI. But increasingly, the results these models produce can easily fool us into thinking it was man-made. And big language models in particular are confident bullshitters: they produce text that sounds right but can actually be filled with falsehoods.
While it doesn’t matter if it’s just a little fun, it could have serious consequences if the AI models used to give health advice aren’t filtered or provided with forms. other important information. AI systems can also make it stupidly easy to create masses of misinformation, abuse, and spam, distorting the information we use and even our sense of reality. For example, it can be particularly unsettling around elections.
The prevalence of these large easily accessible language models raises an important question: How do we know if what we read online was written by humans or machines? I just publish a story review the tools we currently have for AI-generated text detection. Spoiler alert: Today’s detection toolkit is not capable enough against ChatGPT.
But there is a more serious long-term consequence. We could be witnessing, in real time, the birth of a crappy snowball.
Large language models are trained on datasets built by searching text on the internet, including all the malicious, silly, untrue, malicious stuff that humans have written directly. line. Finished AI models will repeat these mistakes into reality, and their outputs are spread all over the network. Tech companies scour the internet again, looking for AI-written text that they use to train larger, more convincing models, which humans can use to create even more amazing things. more pointless before it was removed over and over, the ad nauseated.
This problem—the AI is self-powered and produces increasingly polluting output—is for images as well. Mike Cook, an AI researcher at King’s College London, tells my colleague Will Douglas Heaven in his book: “The internet is now forever polluted with AI-generated images. new piece about the future of general AI models.
“The images that we make in 2022 will be part of any model created from now on.”