I feel we should have benchmarks for each model to show if they are capable of being undetected as this is a real world use case (able to blend in genuinely with humans and even pretend to be them).
I just feel the whole process of emerging AI and then AI detectors is dehumanizing. If someone were to use it on my work, I would feel another layer of the machine between me and that person, another layer of distrust. I think it's sickening that we've come to this stage in the world where people try and cheat (on university essays and projects and job applications) by wasting energy and then people have to use more energy to detect it. It feels futile.
This is similar to the arms-race between spammers and spam-filters, just much more capital intensive. The funny thing about the battle between spammer and spammee, is that we already had a very good solution to it: web-of-trust. We also know that empirically, there are at most 6 degrees of separation between any two random people, so it would still be possible to contact pretty much anyone, it would just take more effort and thought.
I do like the _idea_ behind Pangram, but I think that it misrepresents itself. It is a classifier (my understanding is that it uses ML under the hood, but this is not the only way to build a classifier[0]), and, in my opinion, the classification that it is trying to make (AI, not AI), is not a good long-term classification. Part of the problem is that, for now, identifying something as AI is a good[1] proxy-value for spam, but that can change in the future (I can easily see the next generation subconsciously adopting AI-prose in their speech). Another part of the problem is that, by necessity, the functioning of the detector is a secret (otherwise, AI labs can train against it), and so, everyone is trusting a black box that can become corrupted at any time in the future[2][3].
What people really want (okay, what I really want) is better spam filters and better search engines. And I cannot help but think that this involves a more detailed kind of classifier (one that is customizable and based on "lexical facts", instead a crude and opaque AI-not-AI-probability).
I do wonder, if using the AI detectors will itself cause a kind of cognitive atrophy (e.g. people will lose the ability to sniff out AI prose just by reading it, and will be dependent on the detectors).
[0]: For example, you can use fractal-dimensions (from chaos theory) to derive a specific fingerprint for any given author. People use this to see how faithful a translation of a work is, when translated from one language to another completely unrelated language. The reason I bring this up, is because, to my surprise, feeding AI-generated English text to Google Translate, and then feeding the Bengali translation to Pangram, reduces its confidence from high to medium, and its percentage from 100% to 70%. I am not sure if iterated translations would reduce indefinitely, but fractal dimensions _might_ be helpful here.
[1]: Good as in: it flags true spam more than half the time, but there are some unfortunate false positives (mostly for people who write English as a second language, like some of my friends and relatives).
[2]: We know that Windows had backdoors in the kernel, at the request of the NSA, and we know that various government agencies can read your gmail.
[3]: Alternatively, AI-labs can retrain their LLMs to sound like a specific group of people (for example, Bayesian networks can identify, just from IRC logs, whether someone speaks English as a second language, and what their country of origin, most likely, is).
I don't want better search. I'd rather have less commercialization of the internet. And as for SPAM, it's good to have spam filters but even better to have less incentive to send spam in the first place....
Is the em dash in the title meant to be ironic?
I feel we should have benchmarks for each model to show if they are capable of being undetected as this is a real world use case (able to blend in genuinely with humans and even pretend to be them).
I just feel the whole process of emerging AI and then AI detectors is dehumanizing. If someone were to use it on my work, I would feel another layer of the machine between me and that person, another layer of distrust. I think it's sickening that we've come to this stage in the world where people try and cheat (on university essays and projects and job applications) by wasting energy and then people have to use more energy to detect it. It feels futile.
This is similar to the arms-race between spammers and spam-filters, just much more capital intensive. The funny thing about the battle between spammer and spammee, is that we already had a very good solution to it: web-of-trust. We also know that empirically, there are at most 6 degrees of separation between any two random people, so it would still be possible to contact pretty much anyone, it would just take more effort and thought.
I do like the _idea_ behind Pangram, but I think that it misrepresents itself. It is a classifier (my understanding is that it uses ML under the hood, but this is not the only way to build a classifier[0]), and, in my opinion, the classification that it is trying to make (AI, not AI), is not a good long-term classification. Part of the problem is that, for now, identifying something as AI is a good[1] proxy-value for spam, but that can change in the future (I can easily see the next generation subconsciously adopting AI-prose in their speech). Another part of the problem is that, by necessity, the functioning of the detector is a secret (otherwise, AI labs can train against it), and so, everyone is trusting a black box that can become corrupted at any time in the future[2][3].
What people really want (okay, what I really want) is better spam filters and better search engines. And I cannot help but think that this involves a more detailed kind of classifier (one that is customizable and based on "lexical facts", instead a crude and opaque AI-not-AI-probability).
I do wonder, if using the AI detectors will itself cause a kind of cognitive atrophy (e.g. people will lose the ability to sniff out AI prose just by reading it, and will be dependent on the detectors).
[0]: For example, you can use fractal-dimensions (from chaos theory) to derive a specific fingerprint for any given author. People use this to see how faithful a translation of a work is, when translated from one language to another completely unrelated language. The reason I bring this up, is because, to my surprise, feeding AI-generated English text to Google Translate, and then feeding the Bengali translation to Pangram, reduces its confidence from high to medium, and its percentage from 100% to 70%. I am not sure if iterated translations would reduce indefinitely, but fractal dimensions _might_ be helpful here.
[1]: Good as in: it flags true spam more than half the time, but there are some unfortunate false positives (mostly for people who write English as a second language, like some of my friends and relatives).
[2]: We know that Windows had backdoors in the kernel, at the request of the NSA, and we know that various government agencies can read your gmail.
[3]: Alternatively, AI-labs can retrain their LLMs to sound like a specific group of people (for example, Bayesian networks can identify, just from IRC logs, whether someone speaks English as a second language, and what their country of origin, most likely, is).
I don't want better search. I'd rather have less commercialization of the internet. And as for SPAM, it's good to have spam filters but even better to have less incentive to send spam in the first place....