3 ms·

And yet, this is ChatGPT's attempt at generating a college essay: https://acoup.blog/2023/02/17/collections-on-chatgpt/ https://acoup.blog/2023/02/17/collectio

by vkou 4y ago

And yet, this is ChatGPT's attempt at generating a college essay: https://acoup.blog/2023/02/17/collections-on-chatgpt/ https://acoup.blog/2023/02/17/collections-on-chatgpt/ Looking at the actual essay it produced, I don't need to know anything about Roman history to know that the essay sucks. Looking at the professor's markup of the essay, it becomes very clear that for someone who knows a lot about Roman history, the essay sucks - a lot. And it's not like it was prompted to write about an esoteric topic! According to the grader, the essay made 38 factual claims, of which 7 were correct, 7 were badly distorted, and 24 were outright bullshit. According to both myself, and the grader, way too much heavy lifting is done by vague, unsubstantiated, overly broad statements, that don't really get expanded on further in the composition. But yes, if we're looking to generate vapid, low-quality, low-value content spam, ChatGPT is great, it will produce billions of dollars of value for advertisers, and probably net negative value for the people reading that drivel.

photochemsyn 4y ago
What is you sequentially fed ChatGPT with samples of the course professor's own writing, and then asked it to write an essay on the subject of interest? As the article notes, optimization is possible: > "For example, high school and college students have been using ChatGPT to cheat on essay writing. Since current recognition of AI generated content by humans involve identifying ChatGPT’s signature overly-academic voice, it wouldn’t surprise me if some kids on TikTok figure out a system prompt that allow generation such that it doesn’t obviously sound like ChatGPT and also avoid plagiarism detectors." A decent student might go to the trouble of checking all the factual claims produced in the essay in other sources, thus essentially using ChatGPT to write a rough draft then spending the time saved on checking facts and personalizing the style. I don't even know if that would count as serious cheating, although the overall structure of such essays would probably be similar. Running 'regenerate response' a few times might help with that issue, maybe even, 'restructure the essay in a novel manner' or similar.
- vkou 4y ago
  I can't see why feeding it another 50, or even 500 papers is going to make it any better at making the categories of non-explicit inferences that its currently failing at, when the terabytes upon terabytes and hundreds of thousands of papers of training data did not. It could train it to speak in the mannerisms of a particular author, but that's the least interesting thing in this context, and it'll still be speaking in banalities and nonsense.
  - ShamelessC 4y ago
    Going from "zero-shot" inference to many-shot is known to improve accuracy. I think you are moving goalposts a bit far! in any case. Sure, it's suffering at the college level; but the type of student using this isn't hoping for an A, they are hoping not to fail completely. Which, as far as I can tell - ChatGPT will give you about the same chance of passing as the existing strategy of "do it all in an hour before handing it in because you procrastinated". Probably better in the case of grade school.

by vkou 4y ago

photochemsyn 4y ago

What is you sequentially fed ChatGPT with samples of the course professor's own writing, and then asked it to write an essay on the subject of interest? As the article notes, optimization is possible: > "For example, high school and college students have been using ChatGPT to cheat on essay writing. Since current recognition of AI generated content by humans involve identifying ChatGPT’s signature overly-academic voice, it wouldn’t surprise me if some kids on TikTok figure out a system prompt that allow generation such that it doesn’t obviously sound like ChatGPT and also avoid plagiarism detectors." A decent student might go to the trouble of checking all the factual claims produced in the essay in other sources, thus essentially using ChatGPT to write a rough draft then spending the time saved on checking facts and personalizing the style. I don't even know if that would count as serious cheating, although the overall structure of such essays would probably be similar. Running 'regenerate response' a few times might help with that issue, maybe even, 'restructure the essay in a novel manner' or similar.

vkou 4y ago
I can't see why feeding it another 50, or even 500 papers is going to make it any better at making the categories of non-explicit inferences that its currently failing at, when the terabytes upon terabytes and hundreds of thousands of papers of training data did not. It could train it to speak in the mannerisms of a particular author, but that's the least interesting thing in this context, and it'll still be speaking in banalities and nonsense.
- ShamelessC 4y ago
  Going from "zero-shot" inference to many-shot is known to improve accuracy. I think you are moving goalposts a bit far! in any case. Sure, it's suffering at the college level; but the type of student using this isn't hoping for an A, they are hoping not to fail completely. Which, as far as I can tell - ChatGPT will give you about the same chance of passing as the existing strategy of "do it all in an hour before handing it in because you procrastinated". Probably better in the case of grade school.