AI Megathread
-
@Rathenhope alerted me to this:
https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing
Made me think of the intermittent discussions we’ve had about how to identify AI writing. Maybe this will be helpful or interesting for some of you!
she/they
-
@Tez I do find it interesting, for sure.
What I find a little “funny” (it’s not - I don’t have a better choice of words) is that you would think OpenAI and the other LLM providers would offer tools to detect their own LLM’s handiwork as a capitalistic venture.
This article made me think of the so-called “AI” detectors that I would contend venture towards snake oil, especially since they can generate false positives and negatives. The only people who could possibly make a “foolproof” detector are those who provide what you’re trying to detect. Even then, we’re getting into the whole schtick about random and pseudorandom number generation in computers, which is part of where LLMs get their “ideas” from.
-
@Tez said in AI Megathread:
Just as a note, many (if not most) of these “signs of AI writing” are in fact signs of professional writing as well.
The so-called “ChatGPT Dash” is just the em dash, widely used by pro authors and well-known in Emily Dickinson poetry. Rule of three, “has been described”, parallelism… most of these are common writing tools that many people just weren’t aware of before. ChatGPT is able to imitate those tools because it stole the published work of actual writers.
Now if your coworker who couldn’t string a coherent paragraph together suddenly starts using elegant triplets and juxtaposition, it’s probably a sign that they’re using AI writing. Otherwise, it doesn’t mean much. And that’s why, to @dvoraen’s point, there is no reliable tool for AI writing detection that doesn’t have a zillion false-positives with real writing.
-
Agreed with all @Faraday said above.
The only way you can truly tell if writing is LLM generated and not simply a style you’ve come to associate with LLM is to be comparative. It’s sort of like differentiating a student’s work from something their parent wrote, to use a reference from back in our day.
If you want to test someone that you can’t physically be with to monitor, the best way – which is not a foolproof way – is to get them to write something reflective, about a mutual experience if possible. You’ll more easily see the main flaw in LLM writing: When it makes shit up. An essay written by ChatGPT is going to look like any of the thousands of good essays written in the last hundred years. Because it’s copying them. It’ll probably even get most of the facts right. But a personal, reflective piece? Sure, the LLM can get the structure right, but it’ll just make shit up because there’s no googling for facts of someone personal experience.
-
@Pavel That only works when a student hands it in raw. Many I’ve encountered have been using AI to write the bulk of it and then editing and adding.
-
Students using AI is a real problem in public school, but what I find interesting is how it doesn’t seem to be as much of a problem in the homeschool community. When you take the pressure of grades off, and let kids write about things they’re passionate about, many (most?) of them don’t WANT to use AI.
-
@Faraday said in AI Megathread:
it doesn’t seem to be as much of a problem in the homeschool community
I can’t find a source for this, maybe I’m not searching with the right terms. Can you link me?
-
@Faraday has there been a wide-spread study done regarding AI use in public school vs homeschool? This sounds more anecdotal than factual.
-
@bear_necessities @Ashkuri Sorry, I should have clarified. This is not based on a specific study. It’s actually tough to find robust studies on anything homeschool-related, since it’s such a minority of families scattered across 50 states with very different homeschool regulations. But I would argue it’s more than just a limited anecdote. I participate in a lot of different homeschool communities, and it’s a sentiment I see shared a lot. Nevertheless, it is an opinion/observation not a conclusive fact.
-
it does line up with my own anecdotal experience to a degree, though I do need to qualify that with, it’s pretty split along political lines in my family and their homeschooling communities. because of course it is.
-
The only concrete data I can find is this HSLDA article (based on survey results of only a modest sample size) in which 70% of homeschool families reported that they do not use ChatGPT in their schooling, and those who do mostly use it for lesson planning (separate issue there, lol). I speculate that in the homeschool environment, the parent/teacher has both more opportunity to monitor the work and more familiarity with the writing style/capabilities of their kid, making it easier to mitigate unauthorized AI use.
Regardless, I shouldn’t have made such a broad unsubstantiated statement. What I actually meant to say was just that it’s interesting what can happen to students when you take away some of the common incentives to lean on GenAI (grades and BS assignments they don’t care about).
-
@Faraday said in AI Megathread:
@Tez said in AI Megathread:
Just as a note, many (if not most) of these “signs of AI writing” are in fact signs of professional writing as well.
The so-called “ChatGPT Dash” is just the em dash, widely used by pro authors and well-known in Emily Dickinson poetry. Rule of three, “has been described”, parallelism… most of these are common writing tools that many people just weren’t aware of before. ChatGPT is able to imitate those tools because it stole the published work of actual writers.
This. So very much this. I write professionally not as an author of novels or biographies or anything like that, but in the world of corporate communications. There are literally hundreds and hundreds of both externally-facing articles online and internally-facing intranet articles that I’ve either written or edited over the last five years in this job.
As I scan this article about signs of AI writing, I can see no less than five things that appear in my writing either because they’re a personal writing habit (m-dashes, rule of three) or because they’re in my company’s internal style guide (disclaimer-like language as a result of regulatory rules we’re adhering to, items being treated as proper nouns, excessive use of bold face when referring specific form fields or product titles).
Anything I write personally, I don’t use AI for at all and generally refuse to touch. Knowing my writing style, though, I’m just waiting for the day someone sees me using the MS Word autoformat for an m-dash (–) out of pure habit and accuses me of using ChatGPT to RP.
Anything I write professionally would almost certainly be pegged as written by AI, despite the fact that I have access to three different LLM products at work and largely refuse to use them–unless I’ve gotten a last minute request from a level of leadership I can’t delay. Even then, I use it to spit out a very rough draft at most, which I edit significantly.
On our team of seven people, one of my teammates consistently receives praise for her adoption and advocacy of AI tools. She’s also recognized by our boss’s boss as the team’s worst writer. I largely refuse to use them and am generally regarded as the team’s Luddite, and also get consistently praised as our best writer. The irony isn’t lost on me, but it does seem to be lost on management.
