4 ms·

>>Our findings reveal that these detectors consistently misclassify non-native English writing samples as AI- generated, whereas native writing samples are accu

>>Our findings reveal that these detectors consistently misclassify non-native English writing samples as AI- generated, whereas native writing samples are accurately identified. Furthermore, we demonstrate that simple prompting strategies can not only mitigate this bias but also effectively bypass GPT detectors, suggesting that GPT detectors may unintentionally penalize writers with constrained linguistic expressions. Essentially ChatGPT writes like a non native english speaker. It has to translate from it's computer language into english. I think it's more likely that humans will adapt to ai "style" then the other way round. Look at the elaborate and often exquisitely constructed language that books used 100 years ago. Does anyone think that the art of good writing has not degenerated?

Sharlin 3y ago
But that does not make sense; ChatGPT writes almost perfect, idiomatic copy, probably at the level of the >95th percentile of native speakers. I guess one could call it too good to be a native – but to get around that, you can just simply ask it to use vernacular and a more casual register. Whatever the difference from "average" English prose that an improperly trained ChatGPT detector has gradient descended upon, it is probably something entirely unintuitive and indescribable that also happens to manifest in non-native writing – a classic case of underfitting due to a lacking training corpus.
- simion314 3y ago
  >ChatGPT writes almost perfect, But I feel it responses look like a template where it fills the information in and many times fills it with bullshit. This could be the part where they trained the "chat" part on how to respond. Probably better data for training that is needed, and maybe data that would train it to response like a human and not like an LLM.
  - tjpnz 3y ago
    As a large language model created by OpenAI I cannot...
  - Sharlin 3y ago
    Yes, it does have ChatGPT-isms and a distinct style and "personality", but that’s entirely intentional and taught to it via RLHF. The model itself has no such bias, as is easily demonstrable by just asking it to change its style.
    - simion314 3y ago
      I never tried asking to change the style, I would love it to make it play with me some text adventure games but when I tried and some characters were drinking alcohol it bumped in with its concerns. It would be interesting to ask it to write as a non native english speaker and then explain itself what he did. Though it might bullshit us and invent crap sicne I do not thing the training data was also labeled with the author mother language.
      - Sharlin 3y ago
        Me: Please write a few paragraphs of text in English that demonstrate grammar mistakes that a native Russian English-as-a-second-language learner might make. ChatGPT+: Sure, here are a few paragraphs that demonstrate some common mistakes made by native Russian speakers when learning English: "Last weekend, I with my family went to the park. We liked very much the weather, it was very sunny and warm. We have picnic near to river and play the football. My brother, he is very good in playing football, make a lot of goals. After we tired, we come back home. In home, we cook a dinner together. My mother, she best cook I know, she make a soup and my father grill some meats. It was tasting so good. After the dinner, we watch a movie. I not remember name of movie, but it was about aliens who come to Earth. Movie was little scary, but I liked very much. I think, we must to go more often to the park, and make picnic. It bring joy and healthy for us. I like spend time with my family. It is always fun." In the text, the following common mistakes can be identified: incorrect word order, misuse of articles, incorrect prepositions, misuse of tense and incorrect verb forms, redundancy in personal pronouns, omission of auxiliary verbs, overuse of the verb 'make', and awkward phrasing due to direct translation from Russian. Me: Can you explain in more detail how these mistakes are specifically something that a Russian might make? ChatGPT+: Sure, I can explain in more detail: * Incorrect word order: In the sentence "Last weekend, I with my family went to the park.", the word order is influenced by the structure of Russian sentences. In Russian, it's quite common to structure a sentence like "I with someone did something" because the Russian language doesn't have a strict order for sentence elements. * Misuse of articles: Russian does not have definite (the) and indefinite (a, an) articles, so Russian speakers often either omit them or use them incorrectly in English. For instance, "We have picnic near to river and play the football." should be "We had a picnic near the river and played football." * Incorrect prepositions: Russian prepositions don't correspond directly to English prepositions, which often leads to mistakes like "near to river" instead of "near the river". Similarly, the phrase "in home" should be "at home" in English. * Incorrect verb forms and misuse of tenses: Russian verb tenses don't map exactly to English ones, which can lead to confusion. For example, the sentence "After we tired, we come back home." should be "After we got tired, we came back home." * Redundancy in personal pronouns: In the sentence "My brother, he is very good in playing football", the repetition of the subject (My brother/he) is due to the fact that in Russian, it's common to repeat the subject for emphasis, especially in spoken language. * Omission of auxiliary verbs: The sentence "I not remember name of movie" lacks the auxiliary verb 'do'. This is because Russian doesn't use auxiliary verbs in the same way as English. * Overuse of the verb 'make': Russian uses one verb ("делать") where English would differentiate between 'do' and 'make'. So, Russian speakers often use 'make' where 'do' would be more appropriate, like in "make a lot of goals" instead of "scored a lot of goals". * Awkward phrasing due to direct translation from Russian: For example, "It bring joy and healthy for us" is a direct translation from Russian, whereas in English we would say "It brings us joy and health."
- janalsncm 3y ago
  AI detectors aren’t trained by gradient descent, they’re trained by finding the probability distribution of next words. If the probability distribution matches that of the text in question, they mark it as “AI generated”. But this is an extremely flawed process. Consider the string “We hold these truths to be self _______”. Almost everyone knows the next word is “evident”. The probability distribution is basically 100% on that word. Does that mean if I quote the Declaration of Independence in a paper I was cheating? Of course not. But these AI detectors aren’t smart enough to know that. And there are plenty of teachers and professors who aren’t smart enough to understand the limitations of the tool.

Sharlin 3y ago

But that does not make sense; ChatGPT writes almost perfect, idiomatic copy, probably at the level of the >95th percentile of native speakers. I guess one could call it too good to be a native – but to get around that, you can just simply ask it to use vernacular and a more casual register. Whatever the difference from "average" English prose that an improperly trained ChatGPT detector has gradient descended upon, it is probably something entirely unintuitive and indescribable that also happens to manifest in non-native writing – a classic case of underfitting due to a lacking training corpus.

simion314 3y ago
>ChatGPT writes almost perfect, But I feel it responses look like a template where it fills the information in and many times fills it with bullshit. This could be the part where they trained the "chat" part on how to respond. Probably better data for training that is needed, and maybe data that would train it to response like a human and not like an LLM.
- tjpnz 3y ago
  As a large language model created by OpenAI I cannot...
- Sharlin 3y ago
  Yes, it does have ChatGPT-isms and a distinct style and "personality", but that’s entirely intentional and taught to it via RLHF. The model itself has no such bias, as is easily demonstrable by just asking it to change its style.
  - simion314 3y ago
    I never tried asking to change the style, I would love it to make it play with me some text adventure games but when I tried and some characters were drinking alcohol it bumped in with its concerns. It would be interesting to ask it to write as a non native english speaker and then explain itself what he did. Though it might bullshit us and invent crap sicne I do not thing the training data was also labeled with the author mother language.
    - Sharlin 3y ago
      Me: Please write a few paragraphs of text in English that demonstrate grammar mistakes that a native Russian English-as-a-second-language learner might make. ChatGPT+: Sure, here are a few paragraphs that demonstrate some common mistakes made by native Russian speakers when learning English: "Last weekend, I with my family went to the park. We liked very much the weather, it was very sunny and warm. We have picnic near to river and play the football. My brother, he is very good in playing football, make a lot of goals. After we tired, we come back home. In home, we cook a dinner together. My mother, she best cook I know, she make a soup and my father grill some meats. It was tasting so good. After the dinner, we watch a movie. I not remember name of movie, but it was about aliens who come to Earth. Movie was little scary, but I liked very much. I think, we must to go more often to the park, and make picnic. It bring joy and healthy for us. I like spend time with my family. It is always fun." In the text, the following common mistakes can be identified: incorrect word order, misuse of articles, incorrect prepositions, misuse of tense and incorrect verb forms, redundancy in personal pronouns, omission of auxiliary verbs, overuse of the verb 'make', and awkward phrasing due to direct translation from Russian. Me: Can you explain in more detail how these mistakes are specifically something that a Russian might make? ChatGPT+: Sure, I can explain in more detail: * Incorrect word order: In the sentence "Last weekend, I with my family went to the park.", the word order is influenced by the structure of Russian sentences. In Russian, it's quite common to structure a sentence like "I with someone did something" because the Russian language doesn't have a strict order for sentence elements. * Misuse of articles: Russian does not have definite (the) and indefinite (a, an) articles, so Russian speakers often either omit them or use them incorrectly in English. For instance, "We have picnic near to river and play the football." should be "We had a picnic near the river and played football." * Incorrect prepositions: Russian prepositions don't correspond directly to English prepositions, which often leads to mistakes like "near to river" instead of "near the river". Similarly, the phrase "in home" should be "at home" in English. * Incorrect verb forms and misuse of tenses: Russian verb tenses don't map exactly to English ones, which can lead to confusion. For example, the sentence "After we tired, we come back home." should be "After we got tired, we came back home." * Redundancy in personal pronouns: In the sentence "My brother, he is very good in playing football", the repetition of the subject (My brother/he) is due to the fact that in Russian, it's common to repeat the subject for emphasis, especially in spoken language. * Omission of auxiliary verbs: The sentence "I not remember name of movie" lacks the auxiliary verb 'do'. This is because Russian doesn't use auxiliary verbs in the same way as English. * Overuse of the verb 'make': Russian uses one verb ("делать") where English would differentiate between 'do' and 'make'. So, Russian speakers often use 'make' where 'do' would be more appropriate, like in "make a lot of goals" instead of "scored a lot of goals". * Awkward phrasing due to direct translation from Russian: For example, "It bring joy and healthy for us" is a direct translation from Russian, whereas in English we would say "It brings us joy and health."
janalsncm 3y ago
AI detectors aren’t trained by gradient descent, they’re trained by finding the probability distribution of next words. If the probability distribution matches that of the text in question, they mark it as “AI generated”. But this is an extremely flawed process. Consider the string “We hold these truths to be self _______”. Almost everyone knows the next word is “evident”. The probability distribution is basically 100% on that word. Does that mean if I quote the Declaration of Independence in a paper I was cheating? Of course not. But these AI detectors aren’t smart enough to know that. And there are plenty of teachers and professors who aren’t smart enough to understand the limitations of the tool.