A place for majestic STEMLORD peacocking, as well as memes about the realities of working in a lab.
Rules
This is a science community. We use the Dawkins definition of meme.
I am not saying the two are equally comparable, but I wonder if the same "most rapid change in human written communication" could also have been said with the proliferation of computer-based word processors equipped with spelling and grammar checks.
All this really does is show areas where the writing requirements are already bullshit and should be fixed.
Like, consumer financial complaints. People feel they have to use LLMs because when they write in using plain language they feel they're ignored, and they're probably right. It suggests that these financial companies are under regulated and overly powerful. If they weren't, they wouldn't be able to ignore complaints when they're not written in lawyerly language.
Press releases: we already know they're bullshit. No surprise that now they're using LLMs to generate them. These shouldn't exist at all. If you have something to say, don't say it in a stilted press-release way. Don't invent quotes from the CEO. If something is genuinely good and exciting news, make a blog post about it by someone who actually understands it and can communicate their excitement.
Job postings. Another bullshit piece of writing. An honest job posting would probably be something like: "Our sysadmin needs help because he's overworked, he says some of the key skills he'd need in a helper are X, Y and Z. But, even if you don't have those skills, you might be useful in other ways. It's a stressful job, and it doesn't pay that well, but it's steady work. Please don't apply if you're fresh out of school and don't have any hands-on experience." Instead, job postings have evolved into some weird cargo-culted style of writing involving stupid phrases like "the ideal candidate will..." and lies about something being a "fast paced environment" rather than simply "disorganized and stressful". You already basically need a "secret decoder ring" to understand a job posting, so yeah, why not just feed a realistic job posting to an LLM and make it come up with some bullshit.
Llm detectors are always snake oil 100% of the time. Anyone claiming otherwise is lying for personal gain.
I just want to point out that there were text generators before ChatGPT, and they were ruining the internet for years.
Just like there are bots on social media, pushing a narrative, humans are being alienated from every aspect of modern society.
What is a society for, when you can't be a part of it?
I just want to point out that there were text generators before ChatGPT, and they were ruining the internet for years.
Hey now, King James Programming was pretty funny.
For those unfamiliar, King James Programming is a Markov chain trained on the King James Bible and the Structure and Interpretation of Computer Programs, with quotes posted at https://kingjamesprogramming.tumblr.com/
4:24 For the LORD will work for each type of data it is applied to.
In APL all data are represented as arrays, and there shall they see the Son of man, in whose sight I brought them out
3:23 And these three men, Noah, Daniel, and Job were in it, and all the abominations that be done in (log n) steps.
I was first introduced to it when I started reading UNSONG.
I'm the type to be in favor of new tech but this really is a downgrade after seeing it available for a few years. Midterms hit my classes this week and I'll be grading them next week. I'm already seeing people try to pass off GPT as their own, but the quality of answers has really dropped in the past year.
Just this last week, I was grading a quiz on persuasion and for fun, I have students pick an advertisement to analyze. You know, to personalize the experience, this was after the super bowl so we're swimming in examples. Can even be audio, like a podcast ad, or a fucking bus bench or literally anything else.
60% of them used the Nike Just Do It campaign, not even a specific commercial. I knew something was amiss, so I asked GPT what example it would probably use it asked. Sure enough, Nike Just Do It.
Why even cheat on that? The universe has a billion ad examples. You could even feed GPT one and have it analyze for you. It'd be wrong, cause you have to reference the book, but at least it'd not be at blatant.
I didn't unilaterally give them 0s but they usually got it wrong anyway so I didn't really have to. I did warn them that using that on the midterm in this way will likely get them in trouble though, as it is against the rules. I don't even care that much because again, it's usually worse quality anyway but I have to grade this stuff, I don't want suffer like a sci-fi magazine getting thousands of LLM submissions trying to win prizes.
The reason chatgpt would recommend Nike though is because of its human-based training data. This means that for most humans the Nike ad campaign would also be the first suggestion to come to mind.
I'm not saying LLMs aren't having an impact, or denying that said impact is negative, but the way people talk about them is infuriating because it just displays a lack of understanding or forethought on how these systems work.
People always talk about how they can tell something "sounds like chatgpt" or, as is the case here, is the default chatgpt answer, while ignoring the only reason it would be so is because of the real human patterns which it is mimicking.
Brief caveats: of course chatgpt is wildly fallible and when producing purely generative content it pulls from nowhere because it's just remixing unrelated sources, but for things within the normal course of discussion and output chatgpt's output is vastly more human-like than we want to pretend.
I would almost guarantee that Nike's "just so it" was the singularly most popular answer to this kind of assignment before chatgpt existed too.
As someone who has been a teenager. Cheating is easy, and class wasn't as fun as video games. Plus, what teenager understands the importance of an assignment? Of the skill it is supposed to make them practice?
That said, I unlearned to copy summaries when I heard I had to talk about the books I "read" as part of the final exams in high school. The examinor would ask very specific plot questions often not included in online summaries people posted... unless those summaries were too long to read. We had no other option but to take it seriously.
As long as there isn't something that GPT can't do the work for, they won't learn how to write/do the assignment.
Perhaps use GPT to fail assignments? If GPT comes up with the same subject and writing style/quality, subract points/give 0s.
Last November, I gave some volunteer drawing classes at a school. Since I had limited space, I had to pick and choose a small number of 9-10yo kids, and asked the students interested to do a drawing and answer "Why would you like to participate in the drawing classes?"
One of the kids used chatgpt or some other AI. One of the parts that gave it away was that, while everyone else wrote something like "I want because", he went on with "By participating, you can learn new things and make friends". I called him out in private and he tried to bullshit me, but it wasn't hard to make him contradict himself or admit to "using help". I then told him that it was blatantly obvious that he used AI to answer for him and what really annoyed me wasn't so much the fact he used it, but that he managed to write all of that without reading, and thought that I would be too dumb or lazy to bother reading or to notice any problems.
Did he get into the class after all that?
I have a similar background and no surprise, it's mostly a problem in my asynchronous class. The ones who have my in person lectures are much more engaged, since it is a fun topic and I don't enjoy teaching unless I'm also making them laugh. No dice with asynchronous.
And yeah, I'm also kinda doing that with my essay questions, requiring stuff you sorta can't just summarize. Important you critical thinking, even if you're not just trying to detect GPT.
I remember reading that GPT isn't really foolproof on verifying bad usage, and I am not willing to fail anyone over it unless I had to. False positives and all that. Hell, I just used GPT as a sounding board for a few new questions I'm writing, and it's advice wasn't bad. There's good ways to use it, just... you know, not so stupidly.
I was just commenting on how shit the Internet has become as a direct result of LLMs. Case in point - I wanted to look at how to set up a router table so I could do some woodworking. The first result started out halfway decent, but the second section switched abruptly to something about routers having wifi and Ethernet ports - confusing network routers with the power tool. Any human/editor would catch that mistake, but here it is.
I can only see this get worse.
The Internet was shit before LLMs
It had its fair share of shit and that gradually increased with time, but LLMs are like a whole new level of flooding everything with zero effort
I'd say it was weird, not shit. It was hard to find niche sites, but once you did they tended to be super deep into the hobby, sport, movies, or games.
SEO (search engine optimization) was probably the first step down this path, where people would put white text on a white background with hundreds of words that they hoped a search engine would index.
It's not just the internet.
Professionals (using the term loosely) are using LLMs to draft emails and reports, and then other professionals (?) are using LLMs to summarise those emails and reports.
I genuinely believe that the general effectiveness of written communication has regressed.
Yep. My work has pushed AI shit massively. Something like 53% of staff are using it. They're using it to write reports for them for clients, all sorts. It's honestly mad.
I've tried using an LLM for coding - specifically Copilot for vscode. About 4 out of 10 times it will accurately generate code - which means I spend more time troubleshooting, correcting, and validating what it generates instead of actually writing code.
I feel like it's not that bad if you use it for small things, like single lines instead of blocks of code, like a glorified auto complete.
Sometimes it's nice to not use it though because it can feel distracting.
truly who could have predicted that a glorified autocomplete program is best at performing autocompletion
seriously the world needs to stop calling it "AI", it IS just autocomplete!
I find it most useful as a means of getting answers for stuff that have poor documentation. A couple weeks ago chatgpt gave me an answer whose keyword had no matches on Google at all. No idea where it took that from (probably some private codebase), but it worked.
I'm glad you had some independent way to verify that it was correct. Because I've asked it stuff Google doesn't know, and it just invents plausible but wrong answers.
How did they estimate whether an LLM was used to write the text or not? Did they do it by hand, or using a detector?
Since detectors are notorious for picking up ESL writers, or professionally written text as AI-Generated.
They developed their own detector described in another paper. Basically, this reverse-engineers texts based on their vocabulary to provide an estimate on how much of them were ChatGPT.
They just asked a few people if they thought it was written by an LLM. /s
I mean, you can tell when something is written from ChatGPT, especially if the person isn't using it for editing, but is just asking it to write a complaint or request. It is likely they are only counting the most obvious, so the actual count is higher.