Hacker News Reader

Comment by Kim_Bruning | Hacker News Reader

What you want is a sort of AI Witches' Scales?

https://en.wikipedia.org/wiki/Oudewater#Buildings

From the linked article: " After the weighing, they received an official certificate proclaiming them not a witch... Certificates would state that 'the body weight is in proportion to its build'. The reasoning behind this is the old belief that a witch has no soul and therefore weighs significantly less than an ordinary person".

I love this piece of history so much, thank you for sharing it! Adding this one to my bookmarks :-)

Which is obviously misguided. You can't determine from somebody's BMI if they have a soul. A soul only weighs three quarters of an ounce [1], or about 1/10th a cup of coffee. Well within even the daily variation of a person's body weight. (/s)

1: https://en.wikipedia.org/wiki/21_grams_experiment

Does my soul look fat? Is it good or bad to have an obese or skinny soul? Do they float? Are they charged, can you trap one in a Faraday cage? Is that what the Ghostbusters were collecting?

Also:

“The Waag is still open as a tourist attraction, and official certificates are available.”

Maybe with modern science we can make someone much lighter than they ought to be. (Put helium in their stomach?) Then they would be forced to issue a certificate that said that they could not guarantee that the person is not a witch :D

Not quite. Oudewater offered fair scales. What this guy's talking about are still rigged scales, just oppositely rigged scales. The only reason nobody was ever found a witch there was because the underlying concepts were BS.

You're not wrong.

I'm trying to think if you can indeed do something with a fair test.

Isn't the underlying concept of AI detectors BS as well? You could say something like "We cannot guarantee beyond reasonable doubt that this content is AI generated". Which one could argue is also always true.

This is something that I suspect will get easier in the next couple years as AI research progresses. I could imagine some sort of reversal tool being made such that "given this output and these weights files, how likely is it that these weights produced this output with a reasonable input?"

I don't know, that might be incredibly difficult forever, or it might be something that someone has a breakthrough in and manages to do.

Perhaps it's only useful for particularly egregious examples, and wouldn't be able to detect an AI text that was subsequently edited. That would be fine. When I was a TA in college, we ran students' papers through a plagiarism detector, which would spit out some % number based on what it thought was the amount/likelihood of plagiarism. The number itself was bunk, but across a class of 30 people, if one or two papers had much higher numbers than the other 28, I could click and look at it's justification. Sometimes it was a meaningless false positive, other times it was "here are the paragraphs that were copied verbatim from Wikipedia".

There also seems to be some willingness to have the models watermark their content, and LLMs provide very subtle methods to do so. For example for sufficiently long texts you could bias the token selection to achieve a very specific frequency for certain words; and declare any text where the word "the" makes up between 5.1% and 5.7% of all words, and "which" about 1% of all words to likely be generated by that AI.

One major issue is the feedback loop: There is lots of entrophy to hide a signal in regular English (whether you follow the methods of Stylometry or Steganography), and even an unmodified AI model has a stylometric signature. But if a significant portion of text we consume ends up being written by a handful of AI models, over time people will adopt many of these features. This is obvious right now, with ChatGPT having a certain writing style and some people unconsciously copying that style, and it will remain true no matter how subtle these things become.

> that might be incredibly difficult forever, or it might be something that someone has a breakthrough in and manages to do.

well, why not train an AI on it?

Sure but...since no one actually is a witch (in the sense they were testing for), a fair scale is also one guaranteed to say that they're not a witch.

Presumably the scales have at some point vindicated an actual witch in error.

Reply on news.ycombinator.com