OpenAI just admitted it can’t identify AI-generated text. That’s bad for the internet and it could be really bad for AI models.::In January, OpenAI launched a system for identifying AI-generated text. This month, the company scrapped it.

  • @diffuselight
    link
    English
    2
    edit-2
    1 year ago

    The entropy in text is not good enough to provide enough space for watermarking. No it does not get better in longer text because you have control over i lot/chunking. You have control over top-k and temperature and prompt which creates infinite output space. Open text-generation-webui, go to the parameter page and count the number of parameters you can adjust to guide outcome. In the future you can add wasm encoded grammar to that list too.

    Server side hashing / watermarking can be trivially defeated via transformations / emoji injection Latent space positional watermarking breaks easily with post processing. It would also kill any company trying to sell it (Apple be like … you want all your chats at openAI or in the privacy of your phone?) and ultimately be massively dystopian.

    Unlike plagiarism checks you can’t compare to a ground truth.

    Prompt guidance can box in the output space to a point you could not possibly tell it’s not human. The technology has moved from central servers to the edge, even id you could build something for one LLM, another one not in your control, like a local LLAMA which is open source (see how quickly Stable Diffusion 2 Vae watermarking was removed after release)

    In a year your iphone will have a built in LLM. Everything will have LLMs, some highly purpose bound with only a few M parameters. Finetuning like LoRa is accessible to a large number of people with consumer GPUs today and will be commoditized in a year. Since it can shape the output, it again increases the possibility space of outputs and will scramble patterns.

    Finally, the bar is not “better than a flip of a coin. If you are going to accuse people or ruin their academic career, you need triple nine accuracy or you’ll wrongfully accuse hundreds of essays a semester.

    The most likely detection would be if someone finds a remarkable stable signature that magically works for all the models out there (100s by now), doesn’t break with updates (lol - see chatgpt presumably getting worse), survives quantisation and somehow can be kept secret from everyone including AI which can trivially spot patterns in massive data sets. Not Going To Happen.

    Even if it was possible to detect, it would be model or technology specific and lagging technology - we are moving at 2000miles and hour and in a year it may mot be transformers. They’ll be GAN or RNN elements fused into it or something completely new.

    The entire point of the technology is to approximate humanity - plus we are moving at it from the other direction - more and more conventional tools embed AI (from your camera not being able to take non AI touched pictures anymore to Photoshop infill to word autocomplete to new spellchecking and grammar models).

    People latch onto the idea that you can detect it because it provides an escapism fantasy and copium so they don’t have to face the change that is happening. If you can detect it you can keep it out. You can’t. Not against anyone who has even the slightest idea of how to use this stuff.

    It’s like gunpowder was invented and Samurai would throw themselves into the machine guns because it rendered decades of training and perfection, of knowledge about fortification, war and survival moot.

    On video detection will remain viable for a long time due to the available entropy. Text. It’s always been snakeoil and everyone peddling it should be shot.